Hand Book of Applied Economic Statistics

HANDBOOK OF APPLIED
ECONOMIC STATISTICS
edited by
Aman Ullah
University of California
Riverside, California
David E. A. Giles
University of Victoria
Victoria, British Columbia, Canada
M A R C E I
MARCELDEKKER,
INC. NEWYORK BASEL HONGKONG
D E K K E R
Library of Congress Cataloging-in-Publication Data
Handbook of applied economic statistics/edited by Aman Ullah, David E. A.

Giles .
p. cm.-(Statistics, textbooks, and monographs; v. 155)
Includes bibliographical references and index.
ISBN 0-8247-0129-1
1. Economics-Statistical methods. I. Ullah, Aman. 11. Giles, David E.
A. 111. Series.
HB137.H36 1998
330 ' .O 1'5 195-dc2 1 97-47379
CIP
This book is printed on acid-free paper.
Headquarters
Marcel Dekker, Inc.
270 Madison Avenue, New York, NY 10016
tel: 212-696-9000; fax: 212-685-4540
Eastern Hemisphere Distribution

Marcel Dekker AG
Hutgasse 4, Postfach 812, CH-4001 Basel, Switzerland
tel: 44-61-8482; f a : 44-61-261-8896
World Wide Web

http://www .dekker.com
The publisher offers discounts on this book when ordered in bulk quantities. For
more information, write to Special Sdes/Professional Marketing at the address
below.
Copyright 0 1998 by MARCEL DEKKER, INC. All Rights Reserved.
Neither this book nor any part may be reproduced or transmitted in any form or
by any means, electronic or mechanical, including photocopying, microfilming,
and recording, or by any information storage and retrieval system, without
permission in writing from the publisher.
Current printing (last digit):

1 0 9 8 7 6 5 4 3 2 1
PRINTED IN THE UNITED STATES OF AMERICA

Preface
Many applied subjects, including economic statistics, deal with the collection of
data, measurement of variables, and the statistical analysis of key relationships and
hypotheses. The attempts to analyze economic data go back to the late eighteenth
century, when the first examinations of the wages of the poor were done in the United
Kingdom, followed by the the mid-nineteenth century research by Engle on food
expenditure and income (or total expenditure). These investigations led to the early
twentieth-century growth of empirical studies on demand, production, and cost func-
tions, price determination, and macroeconomic models. During this period the sta-
tistical theory was developed through the seminal works of Legendre, Gauss, and
Pearson. Finally, the works of Fisher and Neyman and Pearson laid the foundations
of modern statistical inference in the form of classical estimation theory and hypoth-
esis testing. These developments in statistical theory, along with the growth of data
collections and economic theory, generated a demand for more rigorous research in
the metholodogy of economic data analysis and the establishment of the International
Statistical Institute and the Econometric Society.
The post-World War I1 period saw significant advances in statistical science,
and the transformation of economic statistics into a broader subject: econometrics,
which is the application of mathematical and statistical methods to the analysis of
economic data. During the last four decades, significant works have appeared on
econometric techniques of estimation and hypothesis testing, leading to the appli-
cation of econometrics not only in economics but also in sociology, psychology, his-
tory, political science, and medicine, among others. We also witnessed major de-
velopments in the literature associated with the research at the interface between
econometrics and statistics, especially in the areas of censored models, panel (lon-
gitudinal) data models, the analysis of nonstationary time series, cointegration and
volatility, and finite sample and asymptotic theories, among others. These common
grounds are of considerable importance for researchers, practitioners, and students
of both of these disciplines and are of direct interest to those working in other areas
of applied statistics.
...
111
iv PREFACE
The most important objective of this volume is to cover the developments in

both applied economics statistics and the econometric techniques of estimation and
hypothesis testing. It is in this respect that our book differs from other publications
in which the emphasis is on econometric methodology. With the above purpose in
view, we deal with the material that is of direct interest to researchers, practitioners,
and graduate students in many applied fields, especially economics and statistics. It
covers reasonably comprehensive and up-to-date reviews of developments in various
aspects of economic statistics and econometrics, and also contains papers with new
results and scopes for future research. The objective behind all this was to produce a
handbook that could be used by professionals in economics, sociology, econometrics,
and statistics, and by teachers of graduate courses.
The Handbook consists of eighteen chapters that can be broadly classified into
the following three groups:
1. Applied Economic Statistics
2. Econometric Methodology and Data Issues
3. Model Specification and Simulation
Chapters 1-5 belong to Part 1 and they are applied papers dealing with impor-
tant statistical issues in development economics and microeconomics. The chapter
by Davies, Green, and Paarsch reviews the literature on using economics statistics,
such as income inequality and other aggregate poverty indices, and they make a
strong case for the use of disaggregated dominance criteria to make social welfare
comparisons. They also discuss some statistical issues related to parametric and non-
parametric inference concerning Lorenz Curves, with reference to stochastic dom-
inance. In contrast, Kramer’s chapter develops two ways of looking at inequality
measurement: the first, a preordering based on majorization defined over income
vectors, and the second, an axiomatic-based approach in which axioms are defined
over inequality measurements. The chapter also includes the empirical application
of inequality measurement primarily focused on dealing with the fact that data is mu-
ally grouped by quantile. Ravallion’s chapter addresses an important issue of persis-
tence in the geography of poverty. It proposes a methodology for empirically testing
the validity of two competing explanations of poverty: an individualistic model and
a geographic model. His proposed approach contributes to our understanding of the
determinants of poverty and provides information for policymakers regarding which
policy interventions are likely to be most effective for its alleviation. The chapter by
Deolalikar explores another dimension of the poverty issue in developing countries,
that is, whether decreased spending on government health programs will reduce the
demand for public health services by the poor and hence will adversely affect the
health status of the poor. This question is analyzed using data from research con-
ducted in Indonesia. The chapter also attempts to address the shortcomings of the
existing literature. Finally, the chapter on mobility by Maasoumi reviews two differ-
ent, but related, approaches to testing for income mobility and shows that the two
PREFACE v
ways converge to the same ordering of states. The relationship between this ordering
and the partial ordering given by Lorenz dominance is shown.
Chapters 6-10 and 15-17 deal with econometric methodologies related to dif-
ferent kinds of data used in empirical research. Chapter 6 by Russell, Breunig and
Chiu is perhaps the first comprehensive treatment of the problem of aggregation as
it relates to empirical estimation of aggregate relationships. It is well known that the
analysis of individual behavior based on aggregate data is justified if the estimated
aggregate relationships can be consistently disaggregated to the individual relation-
ships and vice versa. Most empirical studies have ignored this problem; those that
have not are reviewed in this chapter. Anselin and Bera’s chapter details another
data problem ignored in the econometric analysis of regression model: the problem
of spatial autocorrelation and the correlation in cross-sectional data. This chapter
reviews the methodological issues related to the treatment of spatial dependence in
linear models. Another data issue often ignored in empirical development economics
and labor economics is related to the fact that most of the survey data is based on
complex sampling from a finite population, such as stratified, cluster, and systematic
sampling. However, the econometric analysis is carried out under the assumption of
random sampling from an infinite population. The chapter by Ullah and Breunig
reviews the literature on complex sampling and indicates that the effect of misspec-
ifying or ignoring true sampling schemes on the econometric inference can be quite
serious.
Panel data is the multiple time series observations on the same set of cross-
sectional survey units (e.g., households). Baltagi’s chapter reviews the extensive ex-
isting literature on econometric inference in linear and nonlinear parametric panel
data models. In a related chapter, Ullah and Roy develop the nonparametric kernel
estimation of panel data models without assuming their functional forms. The chapter
by Golan, Judge, and Miller proposes a maximum-entropy approach to the estimation
of simultaneous equations models when the economic data is partially incomplete.
In Chapter 15 Terasvirta looks into the modeling of time series data that exhibit non-
linear relationships due to discrete or smooth transitions and to regimes’ switching.
He proposes and develops a smooth transition regression analysis for such situations.
Finally, the chapter by Franses surveys econometric issues concerning seasonality in
economic time series data due to weather or other institutional factors. He discusses
the statistical models that can describe forecasts of economic time series with sea-
sonal variations encountered in macroeconomics, marketing and finance.
Chapters 12 and 18 are related to the simulation procedures and 11, 13, and
14 to the model and selection procedures in econometrics. The chapter by DeBene-
dictis and Giles surveys the diagnostic tests for the model misspecifications that can
have serious consequences on the sampling properties of both estimators and tests.
In a related chapter, Hadi and Son look into diagnostic procedures for revealing
outliers (influential observations) in the data which, if present, could also affect the
estimators and tests. They also propose a methodology of estimating linear models
vi PREFACE
with outliers, which is an alternative to computer-intensive quantile estimation tech-

niques used in practice. Next, the chapter by Dufour and Torrks systematically de-
velops the general theory of union-intersection and sample split methods in various
specification testing problems in econometrics. They apply their results for testing
problems in the SURE model and a model with MA(1) errors. In contrast to the an-
alytical approaches of specification testing, the chapter by Veal1 provides a survey
of bootstrap simulation procedures that is especially useful in small samples. The
book concludes with the chapter by Pagan, in which he debates about the calibra-
tion methodology of estimation and specification analysis. Several thought-provoking
questions are raised and discussed.
In summary, this volume brings together survey material and new methodolog-
ical results which are vitally important to modern developments in applied economic
statistics and econometrics. The emphasis is on data problems, methodological is-
sues, and inferential techniques that arise in practice in a wide range of situations
that are frequently encountered by researchers in many related disciplines. Accord-
ingly, the contents of the book should have wide appeal and application. We are
very pleased with the end product and would like to thank all the authors for their
contributions, and for their cooperation during the preparation of this volume. We
are also most grateful to Benicia Chatman, University of California at Riverside, for
the efficient assistance that she has provided, and to the editorial and production
staff at Marcel Dekker, especially Maria Allegra and Lia Pelosi, for their patience,
guidance, and expertise.
Aman Ullah
David E. A. Giles
Contents
...
Preface LLL
Contributors LX
Part I Applied Economic Statistics
I. Economic Statistics and Social Welfare Comparisons: A Review I

James B. Davies, David A. Green, and Harry J. Paarsch
2. Measurement of Inequality 39
Walter Krarner
3. Poor Areas 63
Martin Ravallion
4. The Demand for Health Services in a Developing Country:
The Role of Prices, Service Quality, and Reporting of Illnesses 93
Anil B. Deolalikar
5. On Mobility I19
Esfandiar Maasoumi
Part 2 Econometric Methodology and Data Issues
6. Aggregation and Econometric Analysis of Demand and Supply I77

R. Robert Russell, Robert I!Breunig, and Chia-Hui Chiu
7. Spatial Dependence in Linear Regression Models
with an Introduction to Spatial Econometrics 23 7
Luc Anselin and Anil K. Bera
8. Panel Data Methods 29 I
Badi H. Baltagi
vii
viii CONTENTS
9. Econometric Analysis in Complex Surveys 325

Aman Ullah and Robert I!Breunig
10. Information Recovery in Simultaneous-Equations’
Statistical Models 365
Amos Golan, George Judge, and Douglas Miller
I I. Diagnostic Testing in Econometrics:Variable Addition, RESET,
and Fourier Approximations 383
Linda DeBenedictis and David E. A. Giles
12. Applications of the Bootstrap in Econometrics and
Economic Statistics 419
Michael R. Veal1
13. Detection of Unusual Observations in Regression and
Multivariate Data 44 I
Ali S. Hadi and Mun S. Son
14. Union-Intersection and Sample-Split Methods in Econometrics
with Applications t o MA and SURE Models 465
Jean-Marie Dufour and Olivier Torrb
Part 3 Model Specification and Simulation
I5. Modeling Economic Relationships with Smooth Transition

Regressions 507
Timo Terasvirta
16. Modeling Seasonality in Economic Time Series 553
Philip H a m Frames
I7. Nonparametric and Semiparametric Econometrics of Panel Data 579
Aman Ullah and Nilanjana Roy
18. On Calibration 605
Adrian Rodney Pagan
Index 619
Contributors
Luc Anselin, Ph.D. Research Professor, Regional Research Institute and Depart-
ment of Economics, West Virginia University, Morgantown, West Virginia
Badi H. Baltagi, Ph.D. Professor, Department of Economics, Texas A&M University,

College Station, Texas
Anil K. Bera, Ph.D. Professor, Department of Economics, University of Illinois,

Champaign, Illinois
Robert V. Breunig Graduate Student, Department of Economics, University of Cal-

ifornia at Riverside, Riverside, California
Chia-Hui Chiu Graduate Student, Department of Economics, University of Cali-

fornia at Riverside, Riverside, California
James B. Davies, Ph.D. Professor and Chair, Department of Economics, University

of Western Ontario, London, Ontario, Canada
Linda F. DeBenedictis, M.A. Senior Policy Analyst, Policy and Research Division,
Ministry of Human Resources, Victoria, British Columbia, Canada
Anil B. Deolalikar, Ph.D. Professor, Department of Economics, University of Wash-

ington, Seattle, Washington
Jean-MarieDufour, Ph.D. Professor, C.R.D.E. and Department of Economic Sci-

ences, University of Montreal, Montreal, Quebec, Canada
Philip Hans Franses, Ph.D. Associate Professor, Department of Econometrics, Eras-

mus University Rotterdam, Rotterdam, The Netherlands
ix
x CONTRIBUTORS
David E. A. Giles, Ph.D. Professor, Department of Economics, University of Victo-

ria, Victoria, British Columbia, Canada
Amos Golan, Ph.D. Visiting Associate Professor, Department of Agricultural and

Resource Economics, University of California at Berkeley, Berkeley, California, and
Department of Economics, American University, Washington, D.C.
David A. Green, Ph.D. Department of Economics, University of British Columbia,

Vancouver, British Columbia, Canada
Ali S. Hadi, Ph.D. Professor, Department of Statistics, Cornell University, Ithaca,

New York
George Judge, Ph.D. Professor, University of California at Berkeley, Berkeley, Cal-

ifornia
Walter Kramer, Dr. rer. pol. Professor, Department of Statistics, University of

Dortmund, Dortmund, Germany
Esfandiar Maasoumi, Ph.D. (FRS) Professor, Department of Economics, Southern

Methodist University, Dallas, Texas
Douglas Miller, Ph.D. Assistant Professor, Department of Economics, Iowa State

University, Ames, Iowa
Harry J. Paarsch, Ph.D. Associate Professor, Department of Economics, University

of Iowa, Iowa City, Iowa
Adrian Rodney Pagan, Ph.D. Professor, Economics Program, The Australian Na-
tional University, Canberra, Australia
Martin Ravallion, Ph.D. Lead Economist, Development Research Group, World

Bank, Washington, D.C.
Nilanjana Roy, Ph.D. Assistant Professor, Department of Economics, University of

California at Riverside, Riverside, California
R. Robert Russell, Ph.D. Professor, Department of Economics, University of Cali-

forni a at Riverside, Riverside, Ca1i forni a
Mun S. Son, Ph.D. Associate Professor, Department of Mathematics and Statistics,

University of Vermont, Burlington, Vermont
CONTRIBUTORSxi
Timo Teriisvirta, Ph.D. Professor, Department of Economic Statistics, Stockholm

School of Economics, Stockholm, Sweden
Olivier Torres, Ph.D. Maitre de Conferences, U.F.R. MathCmatiques, Sciences

kconomiques et Sociales, Universitk de Lille, Villeneuve d’Ascq, France
Aman Ullah, Ph.D. Professor, Department of Economics, University of California

at Riverside, Riverside, California
Michael R. Veall, Ph.D. Professor, Department of Economics, McMaster University,

Hamilton, Ontario, Canada
This page intentionally left blank
HANDBOOK OF APPLIED
ECONOMIC STATISTICS
Economic Statistics and
Social Welfare Comparisons
A Review
lames B. Davies
University of Western Ontario, London, Ontario, Canada
David A. Green
University of British Columbia, Vancouver; 5rItish Columbia, Canada
Harry J. Paarsch
University oflowa, Iowa Ci& Iowa
1. INTRODUCTION
We present a selective survey of the literature concerned with using economic statis-
tics to make social welfare comparisons. First, we define a number of different sum-
mary income inequality measures and social welfare indices as well as functional
summary measures associated with disaggregated dominance criteria. Next, we de-
scribe the theoretical basis for the use of such measures. Finally, we outline how data
from conventional sample surveys can be used to estimate the functionals that can
be interpreted in terms of social welfare and compared in decision-theoretic terms
with other functionals.
While we define and discuss some of the properties of popular summary in-
equality indices, our main focus is on functional summary measures associated with
disaggregated dominance criteria. In particular, we examine the estimation and com-
parison of Lorenz and generalized Lorenz curves as well as indicators of third-degree
stochastic dominance.* In line with a growing body of opinion, we believe that the
*The estimation of summary indices is discussed, for example, by Cowell (1989); Cowell and Mehta
(1982); and Cowell and Victoria-Feser (1996).
I
2 DAVIES
ET AL.
partial-ordering approach using disaggregated dominance criteria is more attrac-

tive than the complete-ordering approach using summary indices, since the former
requires only widely appealing restrictions on social preferences, whereas the lat-
ter requires the (explicit or implicit) choice of a specific form for the social welfare
function.
The issues in making applied welfare comparisons are well reviewed in sev-
eral places; see, for example, Atkinson (1975) and Cowell (1979).One of the most
important of these issues concerns the definition of income. Applied researchers
face choices between money income versus broader definitions including imputa-
tions; before- versus after-tax income; measuring over a short versus a long horizon;
and even the choice between income and alternative measures of welfare, such as
consumption. Other important conceptual issues concern the choice of unit (individ-
ual? family? household?) and whether one should examine total income, income per
capita, or perhaps income per adult equivalent. In what follows, we assume that an
appropriate definition of income has already been chosen.
Practical difficulties in making welfare comparisons center around measure-
ment and related problems. Official data, for example from tax records, omit income
components, are contaminated by avoidance and evasion, and sometimes do not al-
low the researcher to use the desired family unit. Survey data are affected by dif-
ferential response according to income and other characteristics, and misreporting.
Also, both of these sources tend to neglect in-kind income. These problems should be
borne in mind by applied researchers working with the techniques discussed here.
In Section I1 we define the bulk of the notation used and describe different
summary inequality indices as well as functional summary measures which are often
used to relate economic statistics to social welfare comparisons. Section I11 provides
an axiomatic foundation for many of the measures listed in Section 11. We show how
data from conventional sample surveys can be used to estimate the functionals of in-
terest in Section IV, and in Section V we discuss how observed covariates can be in-
troduced into this framework. We summarize and present our conclusions in Section
VI. Section VII, an appendix, describes how to access several programs designed
to carry out the analysis described here. These programs reside in the Economet-
rics Laboratory Software Archive (ELSA) at the University of California, Berkeley;
this archive can be accessed easily by using a number of different browsers (e.g.,
Netscape Navigator) via the Internet.
II. DIFFERENT ECONOMIC MEASURES OF

SOCIAL WELFARE
In this section we set out notation and define several scalar summary indices of eco-
nomic inequality and social welfare as well as two functional summary measures
associated with disaggregated dominance criteria. We confine ourselves to the posi-
tive evaluation of the behavior of the various measures. The normative properties
ECONOMIC STATISTICS AND WELFARE COMPARISONS 3
of the measures are discussed in Section 111. Nonetheless, it is useful from the out-
set to note some of the motivation for the various indices. In this section we pro-
vide a heuristic discussion, an approach that mimics how the field has developed
historically.
Economists have always agreed that increases in everyone’s income raise wel-
fare, so there is natural interest in measures of central location (the mean, the me-
dian), which would reflect such changes. But there has also always been a view that
increases in relative inequality make society worse off. The latter view has some-
times been based on, but is logically separate from, utilitarianism. Given the interest
in inequality, it was natural that economists would like to measure it. Historically,
economists have proposed a number of essentially ad hoc methods of measuring in-
equality, and found that the proposed indices were not always consistent in their
rankings. This gave rise to an interest in and the systematic study of the normative
foundations of inequality measurement. Some of the results of that study are surveyed
in Section 111.
One central concern in empirical studies of inequality has been the ability to
allocate overall inequality for a population to inequality between and within spe-
cific subpopulations. Thus, for example, one might like to know how much of the
overall income inequality in a country is due to inequality for females, how much is
due to inequality for males, and how much is due to inequality between males and fe-
males. One reason for the popularity of particular scalar measures of inequality (e.g.,
the variance of the logarithm of income) is that they are additively decomposable
into these various between- and within-group effects. The decompositions are often
created by dividing the population into subpopulations and applying the inequality
measure to each of the subpopulations (to get within-group inequality measures) and
to the “sample” consisting of the means of all the subpopulations (to get a between-
group measure). The more subpopulations one wants to examine, the more unwieldy
this becomes. Furthermore, in some instances, one is interested in answering the
question, how would inequality change if the proportion of the population who are
unionized increases, holding constant all other worker characteristics? The decom-
positions allowed in the standard inequality measures do not provide a clear answer
to this question. This point will be raised again in Section IV. In the following sec-
tion, the reader should keep in mind that indices like the generalized entropy family
of indices possess the decomposability property. For a more detailed discussion of
the decompositions of various indices, see Shorrocks (1980, 1982, 1984).
We consider a population each member of which has nonnegative income Y
distributed according to the probability density (or mass) function f ( y ) with corre-
sponding cumulative distribution function F ( y ) .
A. Location Measures
Under some conditions, economists and others attempt comparisons of economic
welfare which neglect how income is distributed. It has been argued, for example
4 DAVIES
ET AL.
by Harberger (1971), that gains and losses should generally be summed in an un-
weighted fashion in applied welfare economics. This procedure may identify poten-
tial Pareto improvements; i.e., situations where gainers could hypothetically com-
pensate losers. Furthermore, it is possible to conceive of changes that would affect
all individuals' incomes uniformly, so that distributional changes would be absent.
For these reasons, measures of central location are a natural starting point in any
discussion of economic measurement of social welfare.
I. Per Capita Income

Perhaps the most common measure of welfare is aggregate or per capital income.
Focusing on the latter, we have mean income
El. = WI = jo rf(r)dY
in the continuous case or
in the discrete case.* Such a measure is easy to calculate and has considerable intu-
itive appeal, but it is sensitive to outliers in the tails of the distributions. For example,
an allocation in which 99 people each have an income of $1 per annum, while one
person has an income of $999,901 would be considered to be equivalent in welfare
terms to an allocation in which each person has $10,000 per annum. Researchers
are frequently attracted to alternative measures that are relatively insensitive to be-
havior in the tails of the distribution. One measure of location that is robust to tail
behavior is the median.
2. Median Income
The median is defined as that point at which half of the population is above and the
other half is below. In terms of the probability density and cumulative distribution
functions, the median solves the following:
Alternatively,
t(0.5) = F-' (0.5)
*Hereafter, without loss of generality, we shall focus almost exclusively on the continuous case.
where F - ' ( - ) is the inverse function of the cumulative distribution function. Clearly,
other quantiles could also be entertained: for example, the lower and upper quartiles,
which solve
c(0.25) = F-'(0.25) and c(0.75) = F ' ( 0 . 7 5 )

respectively.
Using the example considered above, the median of the first allocation would
be $1, while that of the second would be $10,000. Clearly, tail behavior (or dis-
persion) is important in ranking allocations. Accordingly, a natural progression is
to consider measures of the scaZe of the distributions being considered, such as the
variance. However, alternative income distributions differ in more than just scale.
They also differ in shape, and both scale and shape can affect the degree of inequality
which observers perceive in a distribution. This leads to an array of different possi-
ble inequality measures, each member of which is an acceptable measure of scale,
but each of which reacts differently to shape.
B. Scale Measures and Inequality Indices

Consider two alternative income distributions which are related in the following way:
If a # 0 then these distributions differ in location, and if b # 1 they differ in scale.

Any index which rises monotonically in b, but which is invariant to a , is a measure
of the scale of a distribution. Standard examples include the variance and standard
deviation, but a wide variety of other scale measures is possible.
How do measures of scale relate to measures of inequality? Some scale mea-
sures, such as the standard deviation, make sensible inequality measures under
some circumstances. Other scale measures, such as the interquartile range discussed
here, are woefully inadequate inequality measures. The reason is that real-world
distributions differ in ways other than simple differences in location and scale. An
adequate inequality measure must, at a minimum, increase when income is trans-
ferred from any poorer person to any richer person. Not all scale measures satisfy this
criterion. Moreover, there is considerable interest in relative inequality measures,
which are defined on incomes normalized by the mean, and are therefore indepen-
dent of scale.
In this subsection we first look at three types of measures of scale: quantile-
based measures, the variance, and the standard deviation. We then go on to a few
popular relative inequality indices: the coefficient of variation, the variance of log-
arithms, and the Gini coefficient. Holding mean income constant, these indices are
all inverse measures of social welfare. Finally, we examine Atkinson's index and the
generalized entropy family of indices.
6 DAVIES
ET AL.
I. Quantile-Based Measures of Income
Quantile differences or quantile ratios have been employed to provide rough-and-

ready descriptions of the degree of income dispersion. Attention is often paid, for
example, to the interquartile range ((0.75)-((0.25) or the 90-10 percentile ratio
c(0.90)/c(0.10). While such differences or ratios respond to some changes in scale,
they may remain invariant in the face of major changes in the distribution of income.
Each is completely insensitive, for example, to redistributions of income occurring
exclusively in a wide, middle range of incomes.
2. Variance of Income
A standard measure of scale is the variance of Y , defined by
An important characteristic of the variance, which carries through to the related

measures like the standard deviation and the coefficient of variation, is that it is
highly sensitive to the tails of the distribution. Given that distributions of income
and related variables are typically skewed positively, in practice the sensitivity of
these measures is generally greatest to the length of the upper tail. Another impor-
tant property of the variance is that it is additively decomposable both in terms of
income components and population subgroups; see Shorrocks (1980,1982,1984).
The decomposability of the variance also carries through to convenient decomposi-
tions of the standard deviation and the coefficient of variation.
3. Standard Deviation of Income
A drawback of the variance is that it is in the units squared of income. Another

measure of scale, which has the same units as the mean, is the standard deviation,
which is defined as
a = G
4. Coefficient of Variation of Income

We now turn to some measures of relative inequality-i.e., to indices which are de-
fined over income normalized by the mean. These measures are invariant to a partic-
ular kind of change in scale, one where all incomes change equiproportionally. Such
measures are, of course, also insensitive to the choice of units of measurement (e.g.,
dollars versus thousands of dollars).
The first measure of relative inequality we define is the coefficient of variation
t which is simply the standard deviation divided by the mean:
0
t = -
w
Like the variance and the standard deviation, in comparison to other popular mea-
sures, t is especially sensitive to changes in the tails of the distribution.
5. Variance ofthe Logarithm of Income

An apparently attractive measure of income inequality, which is often employed, is
the variance of the logarithm of Y . One reason for the frequent use of this index
may lie in the popularity of “log-earnings” or “log-income” regressions. R2 gives
an immediate measure of the proportion of inequality explained by the regressors
if the variance of the logarithm of Y is accepted as an appropriate inequality mea-
sure, and inequality can be decomposed into components contributed by the various
factors.
In the case of the variance, and its related measures, we have seen that the
use of a linear scale gives great weight to the right tail of the distribution. Apply-
ing a logarithmic transformation reduces this effect.* Thus, introducing the transfor-
mation
2 = log Y
we define the variance of the logarithm of Y by

00
where
is the average value of log Y.? Note that the variance of the logarithm of Y is inde-
pendent of scale, as is the standard deviation of the logarithm of income ox,with
which it may be used interchangeably.$
*In some cases, the transformation can actually go too far as is discussed in Section 111.
?Here, we have used the fact that the prohatiilitv drnsity function of Z is
d
.fz(.) = .fw; = /(exp(z)) exp(2)
since
y = exp(t) and dy
- = exp(2)
d2
$The coefficient of variation of the logarithm ofincwmr TZ is not used since 5 8 depends on scale and would
fall, for example, with an equiproportionate increase in all incomes. Thus, changing from measurement
in dollars to cents would cause TZ to fall.
8 DAVIES
ET AL.
6. Gini Coefficient
Perhaps the most popular summary inequality measure is the Gini coefficient K .
Several quite intuitive alternative interpretations of this index exist. We highlight
two of them. The Gini coefficient has a well-known geometric interpretation related
to the functional summary measure of relative inequality, the Lorenz curve, which is
discussed in the next subsection. Here, we note another interpretation of the index,
which may be defined as
K = L
-
!
21-L
1”I”
0
(U - v (f (U)f (U) du dv
In words, the Gini coefficient is one half the expected difference between the in-
comes of two individuals drawn independently from the distribution, divided by the
mean p.
One virtue of writing K this way is that it draws attention to the contrasting
weights that are placed on income differences in different portions of the distribution.
The weight placed on the difference lu - 211 is relatively small in the tails of the
distribution, where f ( u ) f ( v )is small, but relatively large near the mode. This means
that, in practice, K is dramatically more sensitive to changes in the middle of the
income distribution than it is to changes in the tails. This contrasts sharply with the
behavior of many other popular inequality indices which are most sensitive to either
or both tails of the distribution.
7. Atkinson’s Index
Atkinson (1970)defined a useful and popular inequality index which is based on the
family of additive social welfare functions (SWFs)of the form
The parameter E governs the concavity, and therefore the degree of inequality aver-
sion, shown by the function. Note also that for E equal to zero the function merely
aggregates all incomes, and therefore ranks the same as the mean, given a constant
population.
Atkinson’s index @ is defined as
where ~ D isE“equally distributed equivalent income”-the income such that if all

individuals had income equal to then W would have the same value as the
actual income distribution. With the SWF given by (l),this yields
The parameter E plays a dual role. As it rises, inequality aversion increases, but,
in addition, the degree of sensitivity to inequality at lower income levels also rises
with E . In the limit, as E goes to infinity, the index is overwhelmingly concerned with
inequality at the bottom of the distribution. While the sensitivity of this index to
inequality at different levels can be varied by changing E , it is always more sensitive
to inequality that occurs lower in the distribution.
8. Generalized Entropy Family of Indices

An important class of inequality measures is the generalized entropy family. These
measures are defined by
p2 is defined as one half of the square of the coefficient of variation (t2/2).

The first
and second entropy measures of Theil (1967)are p1 and PO, respectively. Letting c
equal 1 - E , one obtains
+
@E =
{ 1 - [c(c - l)p,:
1 - exp(-pcL
1]’/C, c < 1,
c=o
c #0
Hence, for the particular form of Atkinson’s index given by some value of E , there is
always a corresponding, ordinally equivalent, generalized entropy measure.
All members of the generalized entropy family of inequality indices are based
on some notion of the average distance between relative incomes. These indices do
not take into account rank in the income distribution in performing this averaging,
which makes them fundamentally different from the Gini coefficient, for which rank
is very important.
The attraction of the generalized entropy family is enhanced by the fact that
it comprises all of the scale-independent inequality indices satisfying anonymity
10 DAVIES
ET AL.
and the strong principle of transfers which are also additively decomposable into in-
equality within and between population subgroups; for more on this, see Shorrocks
(1980). Furthermore, all indices which are decomposable (i.e., not necessarily addi-
tively) must be some positive transformation of a member of the generalized entropy
family; for more on this, see Shorrocks (1984).
C. Disaggregated Summary Measures

Each of the above measures summarizes the information contained in F ( y ) in a sin-
gle number and, in the process, discards a great deal of information and can mask
important features of the distribution. As an alternative, researchers have sought to
implement disaggregated summary measures which provide more information con-
cerning the shape of the distribution, but which are still convenient to use. We shall
examine two.*
I. Lorenz Curve
The Lorenz curve (LC) is the plot of the cumulative distribution function q on the
abscissa (x-axis) versus the proportion of aggregate income held by the quantile c ( q )
and below on the ordinate (y-axis). The qth ordinate of the LC is defined as
Note that C(0) is zero, while C(1) is one. A graph of a representative LC for income
is provided in Figure 1. The 45" line denotes the LC of perfect income equality. The
further is the LC bowed from this 45" line, the more unequal is the distribution of
income. The Gini coefficient K can also be defined as twice the area between the 4.5"
line and the LC L(q).Thus,
2. Generalized Lorenz Curve

Like the coefficient of variation, the variance of logarithms, and the Gini coefficient,
the LC is invariant to the mean. Thus, while it is the indicator of relative inequality
par excellence, it does not provide a complete basis for making social welfare com-
parisons. Shorrocks (1983)has shown, however, that a closely related indicator is a
*Howes (forthcoming) provides an up-to-date and more technical discussion of other measures related to
these.
ECONOMIC STATISTICS AND WELFARE COMPARISONS II
LC
0 1
Q
Figure I Example of Lorenz curve.
valid social welfare measure. This indicator is the generalized Lorenz curve (GLC)
which has ordinate
and abscissa q as another functional summary measure. Note that G ( q ) equals L ( q )

times p. Thus, G(0) is zero, while G(1) equals the mean E[Y]. A graph of a repre-
sentative GLC for income is provided in Figure 2.
As discussed in the next section, when the LC for one distribution lies above
that for another (a situation of LC dominance), then the distribution with the higher
LC has unambiguously less relative inequality. A distribution with a dominating
(higher) GLC, on the other hand, provides greater welfare according to all social
welfare functions defined on and increasing in individual incomes and having ap-
propriate concavity.
D. Stochastic Dominance
In some of the discussion we shall use stochastic dominance concepts. These were
first defined in the risk-measurement literature, but it was soon found that they par-
alleled concepts in inequality and social welfare measurement. We introduce here
12 DAVlES ET AL.
GLC
mu
I
I
I
I
I
I
I
II
Figure 2 Example of generalized Lorenz curve.
the notions of first-, second-, and third-degree stochastic dominance for two random
variables Y1 and Y2, each having the respective cumulative distribution functions
F1 (y) and Fz(y). First-degree stochastic dominance holds in situations where one
distribution provides a Pareto improvement compared to another. As discussed in
the next section, second-degree stochastic dominance corresponds to GLC domi-
nance. Finally, third-degree stochastic dominance may be important in the ranking
of distributions whose LCs or GLCs intersect.
I. First-Degree Stochastic Dominance

The random variable Yl is said to dominate the random variable Y2 stochastically in
the first-degree sense (FSD) if
and
In words, strict FSD means that the cumulative distribution function of Yl is every-
where to the right of that for Yz.
2. Second-Degree Stochastic Dominance

The random variable Y1 is said to dominate the random variable Yz stochastically in
the second-degree sense (SSD) if
and
l y [ F 2 ( u )- F l ( u ) ] du > 0 for some y
Note that FSD implies SSD. If the means of Y1 and Y2 are equal (i.e., &[Y1] = &[Y2]=
p), then Y1 SSD K2 implies that Y1 is more concentrated about p than is Y2.
3. Third-Degree Stochastic Dominance

The random variable Yl is said to dominate the random variable Y2 stochastically in
the third-degree sense (TSD) if
and
with the following endpoint condition:*
111. SOCIAL WELFARE AND INCOME INEQUALITY
It is possible that everyone may be better off in one distribution than in another, For
example, this may happen as a result of rapid economic growth raising all incomes.
In this case, there is an actual Pareto improvement. As a result, there will be fewer
individuals with incomes less than any given real income level Y as time goes on.
___
*Corresponding endpoint conditions can he stated for FSD and SSD, hut they are satisfied trivially.
14 DAVIES
ET AL.
In other words, we will have a situation of first-degree stochastic dominance. More

generally, we may have cases where some individuals become worse off, but the
fraction of the population below any income cutoff again declines. Under a suitable
anonymity axiom, the change would be equivalent in welfare terms to a Pareto im-
provement. Thus, it appears that first-degree stochastic dominance may sometimes
be useful in making social welfare comparisons. This is confirmed by Beach et al.
(1994).
In situations where first-degree stochastic dominance does not hold, we need
to consult some of the measures of inequality and welfare defined in the previous sec-
tion. But on which of the wide variety of popular indicators of inequality or welfare
should we focus? To answer this question, Atkinson (1970) argued that we must rec-
ognize that each indicator either maps into a specific social welfare function (SWF)
or restricts the class of admissible SWFs. (This argument was spelled out precisely
by Blackorby and Donaldson 1978.) In order to decide which inequality or welfare
indicators are of greatest interest, we must examine the preferences embodied in the
associated SW Fs.
A. Principle of Transfers, Second-Degree Stochastic Dominance,

and Lorenz Curves
Atkinson (1970) provided the first investigation of the SWF approach, using the ad-
ditive class of SWFs:
where U’(y) is positive and U”(y) is nonpositive. (If U ( y ) is thought of as a utility

function, then this is a utilitarian SWF. However, U ( y ) may be regarded, alterna-
tively, as a “social evaluation function,” not necessarily corresponding to an indi-
vidual utility function.) Because U(y) is concave, W embodies the property of in-
equality aversion. Of course, when U ( 0 ) equals zero and U’(y) is constant for all y,
W ( Y ) = € [ Y ]= f!L (3)
This SWF corresponds to the use of per capita income to evaluate income distribu-
tion. Note that by using the mean one shows indifference to income inequality.
Atkinson (1970) pointed out that strictly concave SWFs obeyed what has come
to be known as the Pigou-Daltonprinciple of transfers. This principle states that in-
come transfers from poorer to richer individuals (i.e., regressive transfers) reduce so-
cial welfare. Atkinson showed that this principle corresponds formally to that of risk
aversion and that a regressive transfer is the analogue of a mean-preserving spread
(MPS) introduced into the risk-measurement literature by Rothschild and Stiglitz
(1970).
Atkinson also noted that a distribution F1 (y) is preferred to another &(y) ac-
cording to all additive utilitarian SWFs if and only if the criterion for second-degree
stochastic dominance (SSD) is satisfied. Finally, he showed that a distribution Fl (y)
would have V ( Y 1 ) which weakly exceeds W(Y2) for all U ( y ) with U’(y) positive
and U”(y) nonpositive, if and only if L I( q ) , the LC of F1 (y), lies weakly above
&(q), the LC of F2(y), for all q . In summary, Atkinson (1970) showed the following
theorem.
Theorem 3.1. The following conditions are equivalent (where Fl (y) and F2(y)
have the same mean):
(i) W(Y1) 3 W(Y2) for all U ( y ) with U’(y) > 0 and U”(y) 5 0.
(ii) F2(y) can be obtained from Fl (y) by a series of MPSs, regressive trans-
fers.
(iii) F1 (y) dominates F2(y) by SSD.
(iv) L1( q ) 2 .c2(q) for all q .
Dasgupta, Sen, and Starrett (1973) generalized this result. First, they showed
that it is unnecessary for W ( Y )to be additive. A parallel result is true for any concave
W ( Y ) . Dasgupta, Sen, and Starrett also showed that the concavity of W could be
weakened to Schur or “S” concavity. These generalizations are important since they
indicate that Lorenz dominance is equivalent to unanimous ranking by a very broad
class of inequality measures.
To a large extent, in the remainder of this chapter we shall be concerned with
stochastic dominance relations and their empirical implementation. The standard
definitions of stochastic dominance, which were set out in the last section, are stated
with reference to an additive objective function, reflecting their origin in the risk-
measurement literature. Therefore, for the sake of exposition, it is convenient to con-
tinue to refer to the additive class of SWFs in the treatment that follows. This does not
involve any loss of generality since we are studying dominance relations rather than
the properties of individual inequality measures.* Dominance requires the agree-
ment of aZZ SWFs in a particular class. As the results of Dasgupta, Sen, and Starrett
show for the case of SSD, in situations where all additive SWFs agree on an inequal-
ity ranking (i.e., there is dominance), all members of a much broader class of SWFs
also may agree on the ranking.
*In the study of individual inequality indices, it would be a serious restriction to confine ones attention
to those which are associated with additive SWFs. This would eliminate the coefficient of variation, the
Gini coefficient, and many members of the generalized entropy family from consideration. Nonadditive
SWFs may be thought of as allowing interdependence of social preferences toward individual incomes.
For additional discussion of these topics, see Maasoumi (1997)and the references cited therein.
16 DAVIES
ET AL.
B. Some Welfare Properties of Summary Inequality and

Welfare Indices
As implied above, the mean ranks distributions in the same way as any additive SWF
when U”(y) equals zero. In contrast, the median is generally inconsistent with the
SWF framework. This is because the median concerns itself only with the welfare
of the median individual. Another measure with a similar property is the Rawlsian
SWF; in that case, the income of the worst-off member of society is treated as a
sufficient statistic for welfare.
Note that the mean violates the strong form of the principle of transfers: it does
not change in response to any regressive transfer. Note, however, that the mean obeys
the weak form of this principle since it never increases as a result of such a transfer.
In contrast, the median may well rise when a regressive transfer occurs. This would
be the case, for example, if a small amount of income is transferred from people in
the bottom half of the income distribution to persons around the median, without
altering the ordering of individual incomes.
The variance of the logarithm of income 0; may also violate the weak form of
the principle of transfers. This is because 0; is not convex at high levels of income;
see Sen (1973, p. 29) for more on this. While this flaw calls into question the uncriti-
cal use of o:, its use is justified where a suitable parametric form of F ( y ) is assumed.
For example, when F ( y ) is lognormal, 0; is a sufficient statistic for inequality.
As noted in Section 11, the variance and standard deviation are dependent on
scale. Therefore, they are questionable as inequality measures. The coefficient of
variation t does not suffer from this property. It also always obeys the strong form of
the principle of transfers.
It is evident from the definition of t that there is no additive SWF of the form
given in (2)to which it corresponds. For more on this, see Blackorby and Donaldson
(1978). This is also true for the Gini coefficient K . Like a number of other popular
summary measures (e.g., Theil’s index, see Sen 1973, p. 35), however, these mea-
sures correspond to SWFs within the S-concave family.
C. Second-Degree Stochastic Dominance and Generalized

Lorenz Curves
It is traditional in much of the inequality measurement literature to focus attention
on relative inequality, i.e., to examine the distribution of income normalized by its
mean. This reflects the early development of the LC as a central tool of inequality
measurement, the fact that most popular inequality indices are relative, and perhaps
also the influence of Atkinson (1970).
Inequality comparisons are, of course, only a part of welfare comparisons.
Therefore, Atkinson (1970) argued for supplementing LCs by comparisons of means.
He noted the following result.
Theorem 3.2. If €[Yl] 3 €[Y2] and C l ( q ) 2 &(q) for all q, then F l ( y ) is

preferred to FL (y) according to second-degree stochastic dominance.
Shorrocks (1983) went beyond this sufficient condition and established the equiva-
lence of GLC dominance and SSD. Using his results, we can state
Theorem 3.3. The following conditions are equivalent (where Fl (y) and F2(y)
may have different means):
(i) W(Y1) 2 W(Y2) for all U(y) with U’(y) > 0 and U”(y) 5 0.
(ii) F l ( y ) dominates F 2 ( y ) by SSD.
(iii) G ( q ) 1 G2(q) for all q .
Theorem 3.3 is of great practical importance in making welfare comparisons be-
cause the means of real-world distributions being compared are seldom equal. Also,
Shorrocks (1983) and others have found that, in many cases when LCs cross, GLCs
do not. Thus, using GLCs greatly increases the number of cases where unambiguous
welfare comparisons can be made in practice.
D. Aversion to Downside Inequality and Third-Degree

Stochastic Dominance
Another approach to extending the range of cases where unambiguous welfare com-
parisons can be made, beyond those which could be managed with the techniques
of Atkinson (1970), has been to continue to restrict attention to relative inequality,
but to adopt a normative axiom stronger than the Pigou-Dalton principle of transfers.
This axiom has been given at least two different names. Shorrocks and Foster (1987)
referred to it as transfer sensitivity. Here, we follow Davies and Hoy (1994, 1995)
who referred to it as aversion to downside inequality (ADI).
AD1 implies that a regressive transfer should be considered to reduce welfare
more if it occurs lower in the income distribution. In the context of additive SWFs,
this clearly restricts U ( y ) to have U”’(y), which is nonnegative in addition to U’(y)’s
being positive and Il”(y)’s being nonpositive.
The ordering induced by the requirement that W(Y1) be greater than or equal
to W(Y2) for all U ( y ) such that U ’ ( y ) is positive, U”(y) is nonpositive, and U”’(y)
is nonnegative corresponds to the notion of third-degree stochastic dominance (TSD)
introduced in the risk-measurement literature by Whitmore (1970). TSD initially
received far less attention in inequality and welfare measurement than SSD, despite
its embodiment of the attractive AD1 axiom. This was, in part, due to the lack of
a readily available indicator of when TSD held in practice. Shorrocks and Foster
(1987) worked to fill this gap.
When LCs do not intersect, unambiguous rankings of relative inequality can
be made under SSD, provided the means are equal. When GLCs do not intersect,
18 DAVIES
ET AL.
unambiguous welfare rankings can also be made under SSD. When either LCs or
GLCs intersect difficulties arise.
Suppose that €[Yl] equals E[Y2]and that the LCs C l ( q ) and & ( q ) intersect
once. A necessary condition for Fl(y) to dominate F2(y) by TSD is that C l ( q ) is
greater than or equal to &(q) at lower incomes and that C1 ( q ) is less than or equal
to & ( q ) at higher incomes. In other words, given that the LCs cross once, the only
possible candidate for the “more equal” label is the distribution which is better for
the poor. This is because the strength of aversion to downside inequality may be so
high that no amount of greater equality at high income levels can repair the damage
done by greater inequality at low incomes.
Shorrocks and Foster (1987) added the sufficient condition in the case of singly
intersecting LCs, proving the equivalent of the following theorem for discrete distri-
butions:
Theorem 3.4. If €[Yl] = €[I‘ll and the LCs for Fl(y) and F2(y) have a single
intersection, then F1 (y) is preferred to F2(y) by TSD if and only if
(i) The LC Ll(q)cuts the LC &(q) from above.

(ii) V[YI] I V[Y21.
This theorem provides a new and important role for the variance in inequality
measurement. It has been extended by Davies and Hoy (1995)to the case where the
two LCs in question intersect any (finite) number of times n. Lambert (1989) has also
extended the analysis to the case where &[Y1]does not equal E[Y2]by investigating
rankings of distributions when GLCs intersect. Davies and Hoy (1995)proved:
Theorem 3.5. If € [ Y l ] = €[I‘ll, F l ( y ) dominates F2(y) by TSD if and only if,
+
for all Lorenz crossover points i = 1 , 2 , . . . , ( n l), h:(qi) 5 h$(qi),with A?(qi)
denoting the cumulative variance for incomes up to the ith crossover point { ( q ; ) for
distribution j .
Since multiple intersections of LCs are far from rare in applied work, this result
has considerable practical value. Its implementation has been studied by Beach,
Davidson, and Slotsve (1994) and is discussed in the next section.
IV. FROM POPULATIONS TO SAMPLES:

UNIVARIATE CASE
Typically, it is far too expensive and impractical to sample the entire population to
construct F ( y ) . Thus, researchers usually take random samples from the population,
and then attempt to estimate F ( y ) as well as functionals that can be derived from
F ( y ) , such as LCs and GLCs. Often, researchers are interested in comparing LCs
(or GLCs) across countries or, for a given country, across time. To carry out this sort
of analysis, one needs estimators of LCs and GLCs as well as a distribution theory
for these estimators.
In this section, we show how to estimate LCs and GLCs using the kinds of
microdata that are typically available from sample surveys. We consider the case
where the researcher has a sample { Y l , Y2, . . . , Yiy} of N observations, each of
which represents an independent and identical random draw from the distribu-
tion F(y).*
A. Parametric Methods
Because parametric methods of estimation and inference are the most well known to
researchers, we begin with them.t Using the parametric approach, the researcher as-
sumes that F ( y ) comes from a particular family of distributions (exponential, Pareto,
lognormal, etc.) which is known up to some unknown parameter or vector of param-
eters. For example, the researcher may assume that
F(y; p ) = 1 - exp (-i) ,
when Y is from the exponential family with p unknown, or
when Y is from the Pareto family with 00 and 01 unknown, or
-00 < p < 00, 0

' > 0,y > 0
when Y is from the lognormal family with p and o2 unknown.
Given a random sample of size N drawn from F ( y ) , a number of estimation
strategies exist for recovering estimates of the unknown parameters. The most effi-
cient is Fisher's method of maximum likelihood. The maximum likelihood estimator
M of p in the exponential case is
*Many large, cross-sectional surveys have weighted observations. For expositional reasons, we avoid this
complication, but direct the interested reader to, among others, the work of Beach and Kaliski (1986)
for extensions developed to handle this sort of complication. We also avoid complications introduced
by dependence in the data, hut direct the interested reader to, among others, the work of Davidson and
Duclos (1995).
TA useful reference for the epistemology of distribution functions is Chipman (1985).
20 DAVIES
ET AL.
while the maximum likelihood estimators of 60 and 61 in the Pareto case are
TO= min[Yl, Y2, . . . , Yiy]
and those of p and 0

' in the lognormal case are
N
Ci=llog Y,
M=
N
In each of these cases (except that of To), the maximum likelihood estimators
are consistent and distributed normally, asymptotically.* One can also derive the
pointwise asymptotic distribution of both the LC and the GLC. This is easiest to do
in the exponential case where the GLC is linear in the parameter to be estimated.
Note that
so
by a central limit theorem, where p0 is the true value of p, so one can show that for
each q , a pointwise characterization of the asymptotic distribution is
*The estimator To of 80 is superconsistent and converges at rate N rather than a; its asymptotic dis-
tribution is exponential. This is a property of extreme order statistics when the density is positive at the
boundary of the support; see Galambos (1987). Because To converges at a rate N , which is faster than
the rate JX for the other estimator T l , it can he treated as if it actually equals 00, and thus ignored in
an asymptotic expansion of T1.
ECONOMIC STATISTICS AND WELFARE
COMPARISONS2 I
Using A, a realization of the maximum likelihood estimator M , one can estimate

G ( q ; p ) at q by
and calculate its standard error S € [ g ( q ;A)] by substituting h for po to get
Note that in the exponential case, the LC is invariant to the parameter p since
Gm - [ q
G(q) - -
C ( q ;p ) = - + (1 - q ) log(1 - q ) ]
WI P
so no estimation is required.
In the other two examples, the LCs and GLCs are nonlinear functions of the
parameters. We find it useful to consider these examples further, since they illustrate
well the class of technical problems faced by researchers. In the Pareto case, where
Now
so the LC is
L(q;01) = [l - (1 - q)(@\-l)'@y
Because G ( q ; 80,01) and C ( q ;0 , ) are continuous functions of 00 and 01, the maxi-
mum likelihood estimators of them are
22 DAVIES
ET AL.
u q ; T l ) = [1 - (1 - 4 ) (T1- l)/Tl3
respectively. These estimators are also consistent.*
To find the asymptotic distributions of G ( q ; To, T1) and L ( q ; T l ) , one needs
first to know the asymptotic distribution of T l . Now
where one can treat To as if it were 80 because To is a superconsistent estimator of

80.The random variable
Zl = log (Z)
is distributed exponentially with unknown parameter 81, so its probability density
function is
f.&) = 81 exp(-Olz), O1 > 0, z > 0
Thus, Ti is a function of the sample mean Z N of N independently and identically
distributed exponential random variables { Z , , Z 2 , . . . , Z N }where, by a central limit
theorem (such as Lindeberg-Levy), ZN has the following asymptotic distribution:
with 8: being the true value of 81. Because Tl is a continuous and differentiable
function of the sample mean Z N ,we can use the delta method (see, for example, Rao
1965)to derive Tl’s asymptotic distribution. We proceed by expanding the function
T1 ( Z N )in a Taylor’s series expansion about r[zN],
which equals l/O:. Thus,
*When T1 is a consistent estimator of 81; the true value of 81, then
plim ~1 = 8:
N+CC
Also, if, for example, L(T1) is a continuous function of T1, then
plim L(T1) = L (plim Tl) = L(8:)

N+m N+CC
by Slutsky’s theorem; see, for example, Rao (1965).

Noting that Tl evaluated at € [ Z N ]equals Oy and that dT1 ( € [ z , ] ) / d Z ~equals (8:)'

and ignoring the remainder term R: since it will be negligible in a neighborhood of
€ [ z ~ which
], equals l/Oy, one obtains
Thus, to find the asymptotic distribution of L ( q ; Tl), for example, expand L ( q ; T1)
in a first-order Taylor's series about the point 07, with Rk being the remainder,
to get
With a minor amount of manipulation, one then obtains
Similar calculations can be performed for the estimator G ( q ; TO,Tl)of S(q; 80,81).*
In the lognormal case, the GLC solves
where e(q; p , a2)is implicitly defined by
) and G ( q ; p , a
In this case, both e(q; p , a
' '
) are only defined numerically. For
a specific q, conditional on some estimates r?i and s^2, one can solve the quantile
equation (5) numerically and then the GLC equation (4). To apply the delta method,
one would have to use Leibniz's rule to find the effect of changes in M and S2 on
the asymptotic distribution of G ( q ; p , 0') at q . A similar analysis could also be per-
formed to find the asymptotic distribution of C(q; p , a2).As one can see, except in a
few simple cases, the technical demands can increase when parametric methods are
used because the quantiles are often only defined implicitly and the LCs and GLCs
can typically only be calculated numerically, in the continuous case. Moreover, the
*Note that one could perform a similar large-sample analysis to find the asymptotic distribution for
estimators of such summary measures as the Gini coefficient, but this is beyond the focus of this
chapter.
24 DAVIES
ET AL.
calculations required to characterize the asymptotic distribution of the pointwise es-

timator of either the LC or the GLC can be long and tedious. This, of course, is not
the main drawback of this approach.
The main drawback of the parametric approach is that it requires the researcher
to impose considerable structure on the data by assuming F ( y ) comes from a particu-
lar family of distributions. Some researchers, such as McDonald (1984), have sought
richer parametric specifications and have provided specification tests for these mod-
els, while Harrison (1982) has demonstrated the value of a careful empirical inves-
tigation using this parametric approach. Because the parametric approach is not an
entirely satisfactory solution, other researchers have sought to relax assumptions
concerning structure and have pursued nonparametric methods.
B. Nonparametric Estimators of Lorenz Curves and Generalized

Lorenz Curves
The first researcher to propose a nonparametric estimator of the LC was Sendler

(1979).* Unfortunately, Sendler did not provide a complete characterization of the
asymptotic distribution of the LC. This task was carried out by Beach and Davidson
(1983), who, in the process, also derived the asymptotic distribution of the GLC.
Beach and Davidson were interested in conducting nonparametric estimation
and inference concerning the set of population ordinates { ,C(qi)I i = 1, . . . , J } cor-
responding to the abscissae {q;li = 1, . . . , J } . When J is nine and the qis are
{0.1,0.2, . . . ,0.9}, for example, Beach and Davidson would be interested in es-
timating the population LC vector of the deciles
L = (L(O.l), C(O.2), . . . , L(0.9))’
where
with Yi = Y ( q i ) . To carry out this sort of analysis, one must first order the sam-
ple { Y l , Y2, . . . , Y N }so that Ycl) 5 Y(z) 5 . . - 5 Y ( N ) .Beach and Davidson de-
fined E ( q ) , an estimator of the qth population quantile c(q), to be the rth-order
statistic Y(rl where r denotes the greatest integer less than or equal to qN.Thus,
*McFadden (1989)and Klecan, McFadden, and McFadden (1991) have developed nonparametric pro-
cedures for examining SSD, while Anderson (19%) has employed nonparametric procedures and FSD,
SSD, and TSD principles to income distributions. Xu, Fisher, and Wilson (1995)have also developed
similar work. Although these procedures are related to estimation and inference concerning LCs and
GLCs, for space reasons we do not discuss this research here.
E = ( E ( q l ) , E ( q 2 ) . . . . , E(qJ))' is an estimator of the vector of quantiles 6 =

(c(q1),( ( q 2 ) , . . . , ( ( Q J ) ) ' and has the asymptotic distribution
where todenotes the true value of the vector 6 and where
The asymptotic distribution of the LC estimator
depends on the joint asymptotic distribution of the estimator
of the parameter vector
where
with ri being the greatest integer less than or equal to q ; N . Note that G is, in fact,
the GLC and that G is an estimator of the GLC. Beach and Davidson showed that the
distribution of &V(G - 0')is asymptotically normal, centered about the ( J 1) +
zero vector, and has variance-covariance matrix Cl' where
with (Ay)'
being A2(q;), the variance of Y given that Y is less than c"(qi).
Now, L is a nonlinear function of G. By the delta method, Beach and Davidson
showed that it has the asymptotic distribution
26 DAVIES
ET AL.
where
where Vp denotes the gradient vector of the function to follow with respect to the
vector 8.
Beach and Davidson (1983) appear to have ignored the implications of their
results of GLCs. These notions were first applied by Bishop, Chakraborti, and Thistle
(1989).
C. Testing for (Generalized) Lorenz Curve Equivalence

and Dominance
Once a researcher has estimated either the LCs or the GLCs for two populations, it
is natural to think about testing the null hypothesis of LC (or GLC) equivalence or
that of LC (or GLC) dominance. Testing the null hypothesis of equivalence turns out
to be much easier than testing the null hypothesis of dominance, so we outline first
how tests of equivalence can be carried out.
Suppose that there are two independent samples, 1 and 2, of size N1 and N2
respectively, and one seeks to test whether Ll equals L2 (or Gl equals &) at a count-
able number of points J . (This null hypothesis can also be written as II = L1 - L 2 =
0.)Assume that consistent estimators of the variance-covariance matrices of L1 and
L2 (or G1 and 6 2 ) exist, and denote them V1 and Vp, respectively. A test of LC (or
GLC) equivalence is referred to by Engle (1984)as a Wald test. It involves calculat-
ing the quadratic form
(L1 - L2)’ [v1 + n;l

v2 - l
(L1 - L2) = P’V-lP
This statistic has an asymptotic distribution which is x2 with J degrees of freedom.

If LC (or GLC) equivalence is rejected, then one may want to know further
which particular ordinates differ from one another. One way of providing such infor-
mation is through multiple comparisons. Beach and Richmond (1985),building on
the work of Richmond (1982),have derived the formulae for multiple comparison
intervals (Savin 1984).
Alternatively, a researcher may wish to test the null hypothesis of LC (or GLC)
dominance. LC dominance implies that
with
at a countable number of points J , so the null hypothesis can also be written as I’I 2
0. Testing this hypothesis involves deciding whether the random variable L1 - Lz
(or G1 - 6 2 for GLC dominance) is in the nonnegative orthant of RJ (or R-’+’). A
number of authors have built on the work of Perlman (1969) to provide solutions
to this problem; see, for example, Kodde and Palm (1986)and Wolak (1987, 1989,
1991).To calculate the test statistic LR for realizations 1 1 and & of L1 and L2 and
estimates 91 and 92 of V1 and Vz, requires one to solve the following constrained,
quadratic programming problem:
which we rewrite as
min (p - p ) ’ V 1 ( p- p) subject to p 3 0
(P)
Letting
p = argmin(p - p)’v-’(p - p) subject to p >_ 0
the test statistic is
LR = (p - p)’V-’(p - p)
which has asymptotic distribution

.I
lim Pr[LR 2 c ] = Pr[X2(j) 3 c ] w ( J ,J - j , V)
N+KJ
;= 1
where Pr[x2(j) 2 c ] denotes the probability that a x2 random variable with j de-
grees of freedom exceeds some constant c , and w ( J , J - j , V) is a weighting function.
D. Nonparametric Analysis when Lorenz Curves Cross

Sometimes in empirical studies, LCs cross. Extending the work of Shorrocks and
Foster (1987), in which they investigated but one crossing, Davies and Hoy (1995)
have considered how to make welfare statements when there are n crossings. One
can summarize the results of Davies and Hoy (1995)in terms of the cumulative co-
efficients of variation (CCVs):
28 DAVIES
ET AL.
In particular, if T(q)for population 1 is always less than z(q) for population 2,

then population 1 dominates population 2 in the social welfare sense.
Beach, Davidson, and Slotsve (1994)derived an estimator of the vector
7=(T(ql),T(q2),. . . T(qj+l))’
as well as the asymptotic theory required to test for dominance of the cumulative
coefficients of variation. In particular, they derived the asymptotic distribution of
the estimator
with ri again being the largest integer less than or equal to qiN. As with the test of
LC (or GLC) dominance, a test of the null hypothesis of CCV dominance for two
populations involves deciding whether TI - T2 is in the nonnegative orthant of
RJ+~
V. FROM POPULATIONS TO SAMPLES WITH

OBSERVED COVARIATES
In many empirical applications, considerable observed covariate heterogeneity ex-

ists. This sort of heterogeneity is of interest for a number of reasons. One important
reason is that the sources of differences in welfare between populations arise from
this covariate heterogeneity. To the extent that some covariate heterogeneity can be
manipulated by policy, some scope exists for improving welfare.
Thus, consider the random variable Y which is observed conditional on a 1 x k
random covariate vector X.Denote the probability density function of Y conditional
on X equaling x by f(y1x). As above, one can define the conditional qth population
quantile of Y conditional on X equaling x by
Hqlx)
4 = W(qIx)lx) = f(YM dr
One can also define the conditional LC ordinate by
aqlx) =
pX)
rf ( Y M dr -
-
A y x ) rf(ylx) 4Y
I;rf(rlx) 4 1x1
A LC of Y conditional on X equaling x is the plot of (4, L ( q l x ) ) .The conditional
GLC is the plot of ( q , € [ Y I x ] C ( q l x ) ) , which we denote by ( 4 , G(qlx)).
A. Parametric Methods
As in Section 1V.A one could assume that f(ylx) comes from a particular parametric
family. For example, suppose
logY = x p + v
where V is distributed independently and identically normal having mean zero and
variance 02.One could then use the method of maximum likelihood to get estimates
of p and 02, and these estimates in turn could be used to estimate g ( q l x ) or L(qIx).
Since the maximum likelihood estimators of p and o2 are asymptotically normal
and since both the LC and the GLC are functionals of f(ylx), one can then use the
delta method to find the pointwise asymptotic distribution of G ( q l x ) and .C(qlx). Of
course, as in the univariate case, the major criticism of the parametric approach is
that it imposes too much structure on the data.
B. Adapting Nonparametric Methods

At the other end of the spectrum is nonparametric analysis. To estimate LCs (or
GLCs) nonparametrically in the presence of observed covariates, one would simply
break up the observed covariates into cells, and implement, for example, the meth-
ods of Beach and Davidson (1983)on each cell. This approach is feasible when the
number of covariates is small and if the covariates are discrete. When the number
of covariates is large or if covariates are continuous, then the methods described in
Beach and Davidson (1983)are impractical. A natural solution to this quandary is a
method that imposes some structure on f ( y ) x ) , but which is not wholly parametric
(i.e., a semiparametric method).
C. SemiparametricMethods
The idea behind any semiparametric method is to put enough structure on the distri-
bution of Y given X equals x so as to reduce the curse of dimensionality that arises in
nonparametric methods. In this section, we focus on one particularly useful approach
(that of Donald, Green, and Paarsch 1995)to estimating GLCs semiparametrically.
These methods can be adapted to recover estimates of LCs, but in the interest of
space we omit the explicit parallel development of methods for LCs.
Donald et al. (1995) translated techniques developed for estimating spell-
duration distributions (Kalbfleisch and Prentice 1980)to the estimation of income
distributions. The main building block in their approach is the hazard function. For
a nonnegative random variable Y with associated probability density function f ( y )
and cumulative distribution function F ( y ) , the hazard function h ( y ) is defined by
the conditional probability
30 DAVIES
ET AL.
where S(y) is the survivor function. From (6) one can see that the hazard function
is simply a transformation of the probability density (or mass) function. One key
result from the literature on spell-duration estimation is that the conditional nature
of h ( y ) makes it easy to introduce flexible functions of the covariates and to entertain
complex shapes for the hazard function.
Donald et al. introduced covariates using an extension of a proportional haz-
ard model in which the range of Y is partitioned into P subintervals Yp = [$, y:),
where Yp n Y4 = 0 for all p # q with U,‘=, Yp = [0,00)’ and they allow the
covariate effects to vary over these subintervals.* They referred to these subinter-
vals as covariate segments. In particular, following Gritz and MaCurdy (1992),they
replaced xip in Cox’s model with
where l[Yp] is an indicator function equaling one if Y is contained in the set Yp

and zero otherwise, x;(Yp) is a 1 x k vector of covariates defined on the set Yp,
I t
pp is a k x 1 vector of unknown parameters, and j? = (/I’ , j?’,. . . , /?‘’)I is a
K x 1 parameter vector with K = P x k. Within this specification, the covariates
can shift the hazard function up over some regions and down over others, providing
the possibility of quite different shapes for the hazard function for individuals with
different covariate vectors.
There are several options for specifying the baseline hazard ho(y). One could
assume a particular density function, such as the Weibull, and use the associated
functional form for the hazard. This may aid in tractability, but imposes unwarranted
shape restrictions on the density. Donald et al. approximated ho(y) using a step func-
tion. In particular, they created a set of dummy variables corresponding to each of
the segments [F, %.+I) for j = 1, . . . , J where J is finite and estimated a parameter
associated with each of these “baseline” segments. The properties of a proportional
*This class of models was first introduced by Cox (1972, 1975). Specifically, the hazard rate for person i
conditional on x;,a particular realization of the covariate vector X,is
h(y1xi) = exp(xiS)ho(y)
where ho(y) is the baseline hazard function common to all individuals and is a vector of unknown
parameters. An important shortcoming of this specification is the restriction that individuals with very
different covariate vectors have hazard functions with the same basic shape, and that any particular
covariate shifts the entire hazard function up or down relative to the baseline specification. It can be
shown that if a particular element of the vector is negative, then F(ylx) for a person with a positive value
for the associated covariate and zero values for all other covariates first-degree stochastically dominates
F(ylO), the cumulative distribution function for a person with zero values for all covariates. This is quite
a strong restriction.
ECONOMIC STATISTICS AND WELFARE COMPARISONS 3I
hazard model with this form of baseline specification are discussed in Meyer (1990).
The major advantage of this approach is that, with a sufficiently large value for J ,
it can capture complicated shapes for the hazard, while allowing for very straight-
forward transformations from hazard to density estimates and then to estimates of
GLC ordinates. This latter occurs because sums rather than integrals are involved.
The main disadvantage is that density estimates are very “spikey,” including spikes
such as those induced by focal-point income reporting, e.g., integer multiples of
$1000. The latter spikes may prove distracting when they are not the main focus of
the analysis.
For a sample of size N , the logarithm of the likelihood function specifica-
tion is
N
i= 1
where the dependent variable for individual i falls in the jl*th baseline segment, Y;
is the set of Y values corresponding to the j t h baseline segment, Yp(,) is the set of
Y values corresponding to the pth covariate segment which itself is associated with
the j t h baseline segment, and Y, is less than y j if Y, is not right-censored. The vector
a contains the J baseline parameters and the notation a; indicates the element of
a corresponding to the jth baseline segment. Note that
a, = log “;+l ho(u) du]
For consistent estimates, Donald et al. required that the covariate values not change
within the baseline segments; they may, however, vary across baseline segments.
For the jth baseline segment, an estimate of the hazard rate is
where a, is the estimate of the jth element of the baseline parameter vector and &p
is an estimate of B p . An estimate of the survivor function is
32 DAVIES
ET AL.
The discrete form of the baseline hazard makes estimating the survivor function very
simple. An estimate of the probability mass function is then
f<,Ix) = R Y j - %,j+1Ix>
IX)
Define 6, a ( K + J ) vector of unknown parameters with true value 60. Donald

et al. assumed that the maximum likelihood estimator D satisfies
and that a consistent estimate of exists.* For segment j , they defined
which is a ( K + J ) column vector with elements
and
and
where t3, = rj - rj+l.Thus, the standard errors of the density estimate at each
baseline segment are easy to construct given d and an estimate of E. The discrete
form of the baseline hazard is helpful in this construction since one requires only
summations rather than integration.
One can also recover estimates of quantiles conditional on the covariates. The
qth quantile of Y conditional on x is defined by
*In practice, Donald et al. use the inverse of the Hessian matrix of the logarithm of the likelihood function
(divided by the sample size) to estimate E.
In the discrete case, however, estimates of the quantiles are easily found by E(qlx) =
yj if *
S(3jlx) 2 (1 - 4 ) ’ %Yj+llX>
and E(qlx) = YJ if
j(YJlx) 2 (1 - q)
Donald et al. characterized the limiting distribution of E(qlx) for different values
of q .
The main advantage of the approach of Donald et al. is its combination of flex-
ibility and tractability. Convergence for their likelihood function can be obtained
relatively quickly and easily. The transformations from d to estimates of the hazard
and then to the density functions described above are straightforward. At the same
time, the specification imposes few restrictions on the shape of the density function
and permits quite different shapes for different covariate vectors. In addition, the ap-
proach admits the examination of both decomposability of subgroups and marginal
effects of covariates. Finally, this approach provides a consistent means of address-
ing top-coded data since top-coded values are just right-censored “spells.””
Once a consistent estimate of f(ylx) is obtained, one can then estimate the
GLC ordinates conditional on the covariates. When the Y s are discrete, the ordinate
of the GLC at q is
J
G(qlx) = l[Yj 5 Hqlx)lYjf(r/Ix)
j= I
where
and where c(q1x) is the conditional qth quantile of Y
*In sample surveys, individuals with high incomes can often be identified by answers to other questions
on the survey instrument. For example, the richest man in a particular geographical region might have
a wife and six children. Thus, an interested individual might infer the man’s income from information
concerning where he lived and his household characteristirs. To provide confidentiality, those who con-
duct surveys often list high incomes as greater than some value, say $100,000 and higher. This practice
is often referred to as “top-coding.”
34 DAVIES
ET AL.
to be
where
VI. SUMMARY AND CONCLUSIONS
We have presented a selective survey of the literature concerned with interpreting

economic statistics in terms of social welfare. We started by defining and review-
ing the properties of location measures and the most popular summary inequality
indices. Recently, use of summary indices has been augmented, and increasingly
replaced, in applied work by the use of disaggregated dominance criteria. A sum-
mary inequality index must correspond to a single, specific, social welfare function
or narrow family of such functions. In contrast, dominance criteria require only that
social preferences should satisfy certain widely appealing axioms of social choice.
The principal dominance criteria examined here have been LC and GLC dom-
inance. In both cases, the main distributional judgement required is that one should
object to a regressive income transfer; i.e., these criteria embody the Pigou-Dalton
principle of transfers. Both criteria are equivalent to SSD when the mean is invariant.
However, the LC departs from SSD in that it is unconcerned with changes that merely
raise or lower mean income without affecting individuals' income shares. Therefore,
strictly speaking, the Lorenz Criterion is an indicator of relative inequality rather than
of social welfare. A further criterion we have discussed briefly, and which is receiv-
ing increasing use in practice, is AD1 (equivalent to TSD when means are the same).
AD1 comes into play in the comparison of distributions whose LCs intersect. If the
cumulative coefficients of variation for one distribution are less than those of another,
cumulating up to each successive point of LC intersection, then there is dominance
from the viewpoint of all individuals who are more concerned with inequality lower
in the distribution, i.e., from the viewpoint of everyone averse to downside inequality.
We have surveyed results on the implementation of these dominance criteria,
beginning with the parametric and nonparametric estimation of LCs and GLCs. We
have also reviewed the results on tests for LC and GLC equivalence, which is rela-
tively easy, and dominance, which is more difficult. We have also reviewed results
on the nonparametric implementation of the AD1 criterion.
A recent development in the literature has been the estimation of GLCs which
depend on covariates. We have discussed limitations of parametric and nonparamet-
ric approaches, and how these limitations can be avoided through the use of tractable
semiparametric methods. The latter can be made sufficiently flexible to reflect com-
plex effects of covariates. A further advantage of the approach is that it provides a
consistent means of addressing top-coded data.
As the review conducted in this chapter shows, recent literature has provided
powerful theoretical and statistical tools which can be applied in social welfare com-
parisons. At this point there is an important challenge to workers on income distri-
bution topics to make greater use of these techniques in their analysis of the data.
VII. APPENDIX
In this appendix, we describe how to access several programs designed to carry out
the analysis described above. These programs reside in the Econometrics Laboratory
Software Archive (ELSA) at the University of California, Berkeley; this archive can
be accessed easily using a variety of different browsers via the Internet. To contact
the staff at ELSA, simply send e-mail to
elsa@econ.berkeley.edu
For those with access to an Internet browser like Netscape Navigator, simply click
on the Open icon and then enter
http://econ.berkeley.edu
This will put you into the HomePage of the Department of Economics at the Univer-
sity of California, Berkeley. A number of options exist. Click the option
[Research Facilities]
You will then be on the page for which the icon
eELSA (Econometrics Laboratory Software Archive)
exists. Click on either the icon denoted Code or the icon denoted ~ o and
c scroll
through until you find the entries you desire.
The following describes the entries:
a. A Nonparametric Estimator of Lorenz Curves
FORTRAN code for implementing Beach and Davidson (1983)
b. A Nonparametric Estimator of Generalized Lorenz
Curves
FORTRAN code for implementing Bishop, Chakraborti, and Thistle (1989)
c. A SemiParametric Estimator of Distribution
Functions with Covariates
Gauss code for implementing Donald, Green, and Paarsch (1995)
36 DAVIES
ET AL.
ACKNOWLEDGMENTS
Green and Paarsch wish to thank the SSHRC of Canada for research support and
their co-author, Stephen G. Donald, for allowing them t o borrow from their unpub-
lished work. Most of this paper was written while Paarsch was the Arch W. Shaw
National Fellow at the Hoover Institution, Stanford, California. Paarsch would like
to thank the Hoover Institution for its hospitality and support. The authors are also
grateful to Gordon Anderson and Charles M. Beach as well as two anonymous refer-
ees for useful comments and helpful suggestions.
REFERENCES
Anderson, G. (1996), Nonparametric Tests for Stochastic Dominance in Income Distributions,

Econometrica, 64, 1183-1193.
Atkinson, ,4.(1970), On the Measurement of Inequality, Journal ofEconomic Theory, 2,244-
263.
Atkinson, A. (1975), The Economics ofhequality, Oxford University Press, Oxford.
Beach, C., K. Chow, J. Formby, and G. Slotsve, (1994), Statistical Inference for Decile Means,
Economics Letters, 45, 161-167.
Beach, C. and R. Davidson (1983), Distribution-Free Statistical Inference with Lorenz Curves
and Income Shares, Review of Economic Studies, 50, 723-735.
Beach, C., R. Davidson, and G. Slotsve (1994), Distribution-Free Statistical Inference for
Inequality Dominance with Crossing Lorenz Curves, typescript, Department of Eco-
nomics, Queen’s University, Kingston, Canada.
Beach, C. and S. Kaliski (1986), Lorenz Curve Inference with Sample Weights: An Applica-
tion to the Distribution of Unemployment Experience, Journal ofthe Royal Statistical
Society, Series C, 35,38-45.
Beach, C. and J. Richmond (1985), Joint Confidence Intervals for Income Shares and Lorenz
Curves, International Economic Review, 26,439-450.
Bishop, J., S. Chakraborti, and P. Thistle (1989), Asymptotically Distribution-Free Statistical
Inference for Generalized Lorenz Curves, Review of Economics and Statistics, 71,725-
727.
Blackorby, C. and D. Donaldson (1978), Measures of Relative Equality and their Meaning in
Terms of Social Welfare, Journal of Economic Theory, 18,59-80.
Chipman, J. (1985), The Theory and Measurement of Income Distributions, in R. Basmann
and G. Rhodes, Jr. (eds.), Advances in Econometrics, vol. 4, JAI Press, Greenwich, CT.
Cowell, F. (1979),Measuring Znequality, Philip Allan, Oxford.
Cowell, F. (1989), Sampling Variance and Decomposable Inequality Measures, Journal of
Econometrics, 42,27-41.
Cowell, F. and F. Mehta (1982), The Estimation and Interpolation of Inequality Measures,
Review of Economic Studies, 49,273-290.
Cowell, F. and M. Victoria-Feser (1996),Robustness Properties of Inequality Measures, Econo-
metrica, 64, 77-101.
ECONOMIC STATISTICS AND WELFARE
COMPARISONS
37
Cox, D. (1972), Regression Models and Life-Tables, Journal ofthe Royal Statistical Society,
Series B, 34, 187-202.
Cox, D. (1975), Partial Likelihood, Biometrika, 62,269-276.
Dasgupta, P., A. Sen, and D. Starrett (1973), Notes on the Measurement of Inequality, Journal
of Economic Theory, 6, 180-187.
Davidson, R. and J.-Y. Duclos (1995), Statistical Inference for the Measurement of the Inci-
dence of Taxes and Transfers, typescript, Department of Economics, Queen’s Univer-
sity, Kingston, Canada.
Davies, J. and M. Hoy (1994), The Normative Significance of Using Third-Degree Stochastic
Dominance in Comparing Income Distributions, Journal of Economic Theory, 64,520-
530.
Davies, J. and M. Hoy (1995), Making Inequality Comparisons when Lorenz Curves Intersect,
American Economic Review, 85,980-986.
Donald, S., D. Green, and H. Paarsch (1995), Differences in Earnings and Wage Distributions
between Canada and the 1Jnited States: An Application of a Semi-Parametric Estima-
tor of Distribution Functions with Covariates, typescript, Department of Economics,
University of Western Ontario, London, Canada.
Engle, R. (1984), Wald, Likelihood Ratio, and Lagrange Multiplier Tests in Economics, in Z.
Griliches and M. Intriligator (eds.), Handbook of Ecunometrics, vol. 11, North-Holland,
Amsterdam.
Galambos, J. (1987), The Asymptotic Theory of Extreme Order Statistics, 2nd ed., Krieger,
Malabar, FL.
Gritz, M. and T. MaCurdy (1992), Unemployment Compensation and Episodes of Nonemploy-
ment, Empiricd Economics, 17, 183-204.
Harberger, A. (1971), Three Basic Postulates of Applied Welfare Economics, Journal of Eco-
nomic Literature, 9, 785-797.
Harrison, A. (1982), A Tail of Two Distributions, Review ofEconomic Studies, 48,621-631.
Howes, S. (forthcoming), Asymptotic Properties of Four Fundamental Curves of Distributional
Analysis, Journal ojEconometrics.
Kalbfleisch, J. and R. Prentice (1980), The Statistical Analysis of Failure Time Data, Wiley,
New York.
Klecan, L., R. McFadden, and D. McFadden (1991), A Robust Test for Stochastic Dominance,
typescript, Department of Economics, MIT, Cambridge, MA.
Kodde, D. and F. Palm (1986), Wald Criteria for Jointly Testing Equality and Inequality Re-
stric tions, Econometrica, 54, 1243- 1248.
Lambert, P. (1989), The Distribution and Redistribution of Income, A Mathematical Analysis,
Basil Blackwell, Cambridge, MA.
Maasoumi, E. (1997), Empirical Analysis of Inequality and Welfare, in M. Pesaran and P.
Schmi d t (eds .), Handbook of Appl ied Microeconometrics, Basil Blac kwell, London.
McDonald, J. (1984), Some Generalized Functions for the Size Distribution of Income, Econo-
metrica, 52,647-663.
McFadden, D. (1989), Testing for Stochastic Dominance, in T. Fomby and T. Seo (eds.), Stud-
ies in the Economics of Uncertainty: In Hunor of Josef Hadar, Springer-Verlag, New
York.
Meyer, B. (1WO), Unemployment Insurance and Unemployment Spells, Econometrica, 58,
757-782.
38 DAVIES
ET AL.
Perlman, M. (1969),One Sided Problems in Multivariate Analysis, Annals of Mathematical

Statistics, 40, 549-567.
Rao, C. (1965),Linear Statistical Inference and Its Applications, Wiley, New York.
Richmond, J. (1982),A General Method for Constructing Simultaneous Confidence Intervals,
Journal ofthe American Statistical Association, 77,455-460.
Rothschild, M. and J. Stiglitz (1970), Increasing Risk. I: A Definition, Journal ofEconomic
Theory, 2,225-243.
Savin, N. (1984),Multiple Hypothesis Testing, in Z. Griliches and M. Intriligator (eds.), Hand-
book of Econometrics, vol. IT, North-Holland, Amsterdam.
Sen, A. (1973), On Economic Inequality, Oxford University Press, London.
Sendler, W. (1979),On Statistical Inference in Concentration Measurement, Metrika, 26,109-
122.
Shorrocks, A. (1980),The Class of Additively Decomposable Inequality Measures, Economet-
rica, 48,613-625.
Shorrocks, A. (1982), Inequality Decomposition by Factor Components, Econometrica, 50,
193-2 1 1.
Shorrocks, A. (1983),Ranking Income Distributions, Economica, 50,3-17.
Shorrocks, A. (1984),Inequality Decomposition by Population Subgroups, Econometrica, 52,
1369-1 385.
Shorrocks, A. and J. Foster (1987),Transfer Sensitive Inequality Measures, Review of Eco-
nomic Studies, 54,485-497.
Theil, H. (1967),Economics and Information Theory, North-Holland, Amsterdam.
Whitmore, G. (1970),Third-Degree Stochastic Dominance, American Economic Review, 60,
457-459.
Wolak, F. (1987),An Exact Test for Multiple Inequality and Equality Constraints in the Linear
Regression Model, Journal of the Americnn Statistical Association, 82, 782-793.
Wolak, F. (1989),Local and Global Testing of Linear and Nonlinear Inequality Constraints in
Nonlinear Econometric Models, Econometric Theory, 5, 1-35.
Wolak, F. (1991), The Local Nature of Hypothesis Tests Involving Inequality Constraints in
Nonlinear Models, Econometrica, 59,981-995.
Xu, K., G. Fisher, and D. Willson (1995),New Distribution-Free Tests for Stochastic Domi-
nance, Department of Economics, Dalhousie University, Working Paper 9502.
Measurement of Inequality
WaIte r Kriime r
University of Dortmund, Dortmund, Germany
1. THE PROBLEM
This chapter surveys the statistical measurement of inequality, where “inequality”

is a property of the elements of the set
00
D:=UR:
n=2
where R; = {x = (xl,. . . , x,)(xi E R,xi 3 0, c x i > 0).

For concreteness, I will often argue in terms of income distributions, but most
of what follows also applies to variables other than income. The discussion is in
terms of vectors rather than distribution functions, as a reliance on the latter would
automatically restrict attention to measures which are symmetric and population in-
variant (see Section IV.A.2 and IV.A.8), and it is important for many problems to be
more permissive here.
Given some vectors x and y from D,the problem is to rank these vectors ac-
cording to inequality, whatever “inequality” may mean. In a sense this is similar
to comparing distributions with respect to location and dispersion (scale), only less
intuitive: while everybody “knows” what is meant be scale and location, and (to a
lesser extent) by skewness and kurtosis, the inequality dimension commands much
less intuitive appeal, and, in fact, allows for views which are outright contradictory.
Accordingly, this survey first discusses ordinal approaches to inequality; i.e.,
it considers some widely accepted concepts of “inequality” that allow at least some
distributions x and y to be ranked. Section I11 considers cardinal measures of in-
equality, Section IV discusses the axiomatic foundation of these measures, Section
V addresses various empirical issues that arise when data are grouped or otherwise
39
40 K~MER
not completely available, and Section VI concludes with some recent developments
in multidimensional inequality.
As the focus here is on mathematical aspects of the ordering of given vectors x
and y, it is only fair to point out that in most applications these problems are dwarfed
by the complications and ambiguities involved in the definition and measurement of
the basic vectors x and y themselves. This is perhaps most obvious with income in-
equality, where the outcome of any analysis depends much more on how income is
defined, on the unit (individuals versus households), and on the accounting period
than on the statistical procedures applied afterward. However, such problems of ba-
sic definition and measurement differ so much across applications that there is no
hope of an adequate treatment in a general survey such as this.
II. O R D I N A L COMPARISONS OF INEQUALITY
There is wide agreement across value systems and applications that if some vector
is subjected to a progressive transfer, then inequality should be reduced.* Formally,
let y = xT,where both y and x are of dimension n and where T is an n x n matrix
of the form
with 0 5 p 5 1. Then yi = xi and y differs from x only in two components, say

i and j , who are each replaced by their weighted arithmetic mean. In the context of
income inequality, such a transfer means that the richer of the persons z and j gives
some or all of the surplus to the poorer (but no more; i.e., the richer does not thereby
become poorer as the poorer has been before; both incomes stay the same if x; = xj).
It was first suggested as an unequivocal reduction in inequality by Pigou (1912)and
Dalton (1920) and is often called a Pigou-Dalton transfer, but some prefer Arnold’s
(1987) more evocative “Robin Hood transfer” instead.
Obviously, a sequence of successive Robin Hood transfers should lead to suc-
cessive reductions in inequality, which in turn leads naturally to the majorization or-
*Or at least: should not increase. In what follows, “reduction” excludes the limiting case of “no change.”
If “no change” is included, I say “weak reduction” (and similarly “increase” and “weak increase”).
MEASUREMENT
OF INEQUALITY 4I
dering in D.* Let x = ( x ( l ) , . . . , x ( , , ) ) and y = ( y ( ~ ).,. . , Y ( ~ ) ) , with Cqi) =

C y(i), be ordered such that X ( I ) Ix ( 2 ) i . L qn)and y(i) I
* y(z) 5 . - L Y ( ~ ) .
Then, by definition, x majorizes y (i.e., x is more unequal than y: x 2~ y) if
k k
i= 1 i= 1
and a central result in majorization theory, due to Hardy, Littlewood, and Polya
(1934), states that x majorizes y if and only if y = xT1 T2 . - . TL with finitely many
matrices of the form (2).
The obvious drawback of this (the only) universally accepted preordering is
that only very few pairs of elements x and y from D can be compared. Both vectors
must have the same dimension and the same sum of elements, which renders this
ordering almost useless in empirical applications.
One way out of this dilemma is to base a comparison on one of the scalar-valued
measures of inequality to be discussed in Section 111. As these measures all involve
value judgments of one sort or the other, one might compromise here and declare
one vector y more equal than some other vector x if Z(x) 2 Z(y) for all inequality
measures I ( . ) in some conveniently chosen class. For instance, almost by definition,
the majorization order on D is the one implied by the set of all Schur-convex functions
I : D + R ( a function I ( . ) is Schur convex if it respects the majorization ordering),
and by restricting this set of functions, one might hope to obtain an order which is
richer (i.e., allows more pairs of vectors to be ranked).
Quite surprisingly, however, if we restrict I ( - ) to be symmetric and convex in
+ +
the ordinary sense (i.e., Z(hx ( 1 - h ) y ) 2 hZ(x) (1 - h ) Z ( y )for all 0 5 h 5 1;
this implies Schur convexity), or even further to be of the form
with a convex function g : R + R,we still end up with the very restrictive majoriza-
tion ordering > M (Mosler 1994, Proposition 2.2).
A major breakthrough in this respect was the discovery by Atkinson (1970)
and Dasgupta, Sen, and Starrett (1973) that the preordering on D becomes much
richer, and in fact identical to the familiar Lorenz order, if we require in addition to
(4) that the measure be population invariant:
Z(x) = Z(x, x, . . . , x) (x repeated rn times) (5)
*Actually apreordering, to be precise. A reorder is a partiul order if x 5 y and y 5 x implies x = y. In

what follows, “order” is short for “preorder” and a preorder is an order if either x 5 y or y 5 x.
42 KRAMER
-1
Figure I A typical Lorenz curve: convex, monotone, passing through (0,O)and (1, 1).
and homogeneous of degree zero:
Then y is more equal than x with respect to all inequality indices in that class if and
only if the Lorenz curve of y lies everywhere above the Lorenz curve of x, where the
Lorenz curve of x (and similarly for y) is obtained in the usual way by joining the
points (0,O) and
by line segments in a two-dimensional diagram. This preordering, which I denote

by “>L,” allows comparisons of vectors with different means and dimensions. When
applied to vectors with identical means and dimensions, it reduces to “>M.” (Note
that “ 3 ~and” “ 3 ~implies
” in increase in inequality.)
A typical Lorenz curve is shown in Figure 1: it is convex, with slope x(i)/X in
the interval ((z - l ) / n ,i/n), passing through ( 0 , O ) and ( 1 , 1).
The proof that the Lorenz order in D is implied by inequality indices with
properties (4), (5) and (6) is an immediate consequence of the fact that with (5) and
(6),we can without loss of generality assume x and y to have identical means and
dimensions: if X = $ xi # 7 and/or dimension (x) = n # rn = dimension (y),
consider
MEASUREMENT
OF INEQUALITY 43
2 = (x,x, . . . ,x) x repeated rn times

-
7 = g(y, y, . . . , y) y repeated n times (8)

Y
Then 2 has the same Lorenz curve as x, and 7 has the same Lorenz curve as y,
and f ( x ) = f ( 2 ) and f ( y ) = f ( 7 ) for all functions with properties (5) and (6)
and the equivalence of the Lorenz order with the ordering implied by functions with
properties (4), (S), and (6) follows from the fact that the majorization ordering is
implied by functions with property (4) alone.
If one does not want to subscribe to (6), but requires instead that inequality
remains unaffected if the same positive amount is added to all components of x:
Z(x + h e ) = Z(x) (e = ( 1 , . . . , l ) , h > 0) (9)

one obtains a different ordering which turns out to be equivalent to absolute Lorenz
domination as introduced by Moyes (1987): With L , ( p ) the conventional Lorenz
curve of x, the absolute Lorenz curve is
and x is more unequal than y in the absolute Lorenz sense (in symbols: x y) if
L A , ( p ) 5 L A , ( p ) . This gives another preordering on D which reduces to majoriza-
tion whenever the latter applies, but it incorporates a more leftist ideal of equality, as
first pronounced by Kolm (1976), for pairs of income vectors which cannot be ranked
by majorization.
There is an infinity of additional preorderings on D which are induced by fam-
ilies of “compromise inequality indices” (Eichhorn 1988, Ebert 1988, Bossert and
Pfingsten 1989), i.e., indices with the property
Z(x + t(hx + (1 - h ) e ) ) = Z(x) (11)

where 0 5 h 5 1 and t is any scalar such that x + t ( h x + ( l - h ) e ) E R;. Obviously,
this requirement reduces to (9) for h = 0 and to (6) for h = 1, but it allows some
“intermediate” value judgments too: while (6) incorporates the “rightest” view that
an equiproportional increases in all incomes leaves inequality the same, whereas
an equal absolute increase reduces it (this is not immediate, but follows from (6) in
conjunction with some minor regularity conditions) and (9)incorporates the “leftist”
view that an equal absolute increase in income leaves inequality the same, whereas
an equiproportional increase increases it (which can likewise be shown to follow from
(9) plus some regularity conditions), the requirement (11) implies that inequality
is increased both by an equiproportional increase of all incomes and by an equal
absolute reduction, with the h parameter governing the closeness to the pure leftist
and rightist view, respectively: As h tends to zero, one tends to take the leftist point
of view, and the rightist point of view is taken as h tends to one.
44 WMER
Given A, the preordering “>A” induced by Schur-convex compromise indices

satisfying (11)amounts to
which in turn is equivalent to 3 LM 7, where % = (x - x e ) / ( E +

1 - A) and
7 = (y - ye)/(Ay + 1- A), with e = (1, . . . , 1) a vector of ones (Pfingsten 1988a).
Given the multitude of pairs of vectors x and y in D which still cannot be
ranked, the primary purpose of any such preordering is not the ranking as such (since
this is often a rather disappointing exercise), but rather to provide a benchmark test
for various procedures that affect inequality, like taxation in the case of income in-
equality. Given any such preordering, it is both sensible and relevant to ask whether
some tax functions will always (i.e., irrespective of the pretax income vector x) reduce
inequality.
For the Lorenz order, the answer is well known (see Eichhorn et al. 1984): A
tax function t ( t ) will produce an after-tax income vector y which is majorized by
any pretax income vector x if and only if t ( t ) is incentive preserving and progres-
sive (i.e., if both residual income t - t ( z ) and average tax rates t ( t ) / z are increasing
functions oft).* For the ordering implied by ( l l ) , the tax functions compatible with
that ordering are derived in Pfingsten (1988b). Lambert (1993)gives a detailed dis-
cussion of such issues, and for the effects on data transformations on Lorenz curves
in general see Fellman (1976).
111. CARDINAL COMPARISONSOF INEQUALITY
As Lorenz curves, whether ordinary, absolute or generalized,t do often cross, there

remain many pairs of vectors in D which cannot be ranked by any of the methods from
Section 11. In an empirical comparison involving 72 countries, Kakwani (1984) for
instance found more than 700 Lorenz crossings among the 2556 possible pairwise
comparisons, hence the desire to attach some cardinal measure of inequality to a
given distribution.
The best known of these cardinal, scalar-valued measures is the Gini index
G ( x ) , defined as twice the area between the Lorenz curve and the 45” line of equality
*Note that progressiveness by itself does not guarantee a reduction in inequality. It is easy to find examples
where income inequality in the Lorenz sense increases after a progressive income tax: take two incomes
of 5 and 10 and a progressive tax rate of 0%up to 6 and 100%afterwards.
TA generalized Lorenz curve is an ordinary Lorenz curve multiplied by x. It is most useful in welfare
rankings of income distributions (Shorrocks 1983, Thistle 1989).
MEASUREMENT
OF INEQUALITY 45
(the concentration area). It perfectly captures Lorenz’s original intuition that, as the
bow is bent, so inequality increases, taking a minimum value 0 when all x; are
equal (i.e., when the Lorenz curve is equal to the 45” line), and a maximum value of
(n - l)/n whenever all xi except one are zero. It depends on the underlying vectors
x only via the Lorenz curve of x, so it is obviously symmetric, homogeneous of degree
zero, and, by construction, compatible with the Lorenz ordering on D.The history of
this “mother of all indices of inequality” is nicely summarized in Giorgi (1990).
There are various equivalent expressions for the Gini coefficient, like Gini’s
own expression
where
i,j=l
is Gini’s mean difference. Rewriting the double sum in (14)as

n n n
i,j=l ;>;=I i= 1
one easily sees that
where the latter expression is most useful when discussing sensitivity issues (see
below), or for showing that A(x)/2x is indeed equal to twice the concentration area.
Yet another algebraic identity was unearthed by Lerman and Yitzhaki (1984),
and independently by Berrebi and Silber (1987),who show that twice the Gini index
+
equals the covariance between the “rank gaps” r; := i - ( n - i 1) and the “share
gaps” s; = (x(;)- x(n-;+l))/ x(;)of the data. This identity, for instance, implies
that the Gini index is never smaller than 1/2 for symmetric distributions and that
a necessary condition for G ( x ) < 1/2 is that x is skewed to the right. It can also
be used to facilitate the empirical computation of the Gini index from large sets of
individual data, where competing measures were often preferred to the Gini index
only because of computational convenience.
In the context of income equality, an interesting intuitive foundation of the
Gini index based on (13) which does not rely on the geometry of the Lorenz curve
was suggested by Pyatt (1976):Choose randomly some unit, say i, with income xi,
and choose independently and randomly another unit x;. Then, if i keeps his income
if x; 5 xi and receives the difference if x; > xi, the Gini index is the expected profit
from this game, divided by average income.
46 KRAMER
A similar idea has led Blackburn (1989)to a nice interpretation of changes in

the Gini index: Let x and y be income vectors in R; such that G ( x ) > G(y). Then
G ( x ) - G(y) equals the percentage of x that units in x below the median have to
transfer to units above the median to make G ( x ) equal to G(y).
There are various generalizations of the Gini index, such at Yitzhaki’s (1983)
generalized Gini coefficients or Mehran’s (1976) linear measures of income inequal-
ity, which like the Gini index can be expressed as functions of the Lorenz curve,
but which avoid certain shortcomings of the Gini index such as its lack of transfer
sensitivity (see below).
The Yitzhaki family of inequality indices is defined as
with one coefficient for each value of a parameter U > 1. For U = 2, this is the
standard Gini index G ( x ) , but as U increases, higher weights are attached to small
incomes; the limit as U + 00 of G,(x) is 1 - %(I)/%, so in the limit inequality
depends only on the lowest income (given x), expressing the familiar value judgment
introduced by Rawls, that social welfare (viewed as a function of inequality) depends
only on the poorest member of society.
Mehran’s (1976) linear measures of inequality are defined as
where W ( p ) is some increasing “score function” not depending on x, which allows

one to incorporate value judgments as to the magnitude of a decrease in inequality
that is engendered by a Robin Hood transfer involving individuals a given number
of ranks apart: As is shown in Mehran (1976), the decrease in inequality decreases
as the rank of the recipient increases (i.e., transfers among the rich are less impor-
tant than transfers among the poor) if and only if the score functions has a strictly
decreasing derivative. We will return to this issue of transfer sensitivity in Sec-
tion IV.
Again, the family defined by (18) includes the Gini index as a special case
(take W ( p ) = 2 p ) , but it covers various other indices as well, allowing for a unified
discussion of their properties. For instance, it is easily seen that an index defined by
(18)is always compatible with the majorization order, but strictly compatible (i.e.,
strict majorization always leads to a strict reduction in inequality) only if W ( p ) is a
strictly increasing function of p . Taking W ( p ) = 0 if p is less than the proportion of
the xi’s below X and W ( p ) = 1 if p is equal to or larger than the proportion of the xi’s
above X, we obtain the maximum vertical distance between the egalitarion 45” line
and the Lorenz curve, and the well known result that this distance is not affected by
transfers either below or above X.
Historically, this maximum vertical distance between the 45” line and the
Lorenz curve has come in various guises; it is also known as the “maximum equaliza-
MEASUREMENT
OF INEQUALITY 47
tion percentage” (the relative amount of income transfers necessary for total equal-
ization), and it is also algebraically identical to one half of the relative mean
deviation:
2x
This measure was suggested, more or less independently, by Bresciani-Turroni
(1910),Ricci (1916),von Bortkiewicz (1931),Schutz (1951), Kuznets (1959),and
Elteto and Frigyes (1968), to name a few of the authors who have given it their at-
tention. For simplicity, I will refer to it as the Pietra index P ( x ) , in honor of Pietra
(1914/ 15).*
In the context of income inequality, it has long been viewed as a major draw-
back of both the Gini and Pietra indices and of related indices derived from them
that they are not explicitly linked to some social welfare function W ( x ) . t This un-
easiness was first expressed in Dalton’s (1920) proposal to use the “ratio of the total
economic welfare attainable under an equal distribution to the total welfare attained
under the given distribution” as a measure of income inequality; it has led Atkinson
(1970) to define inequality as A ( x ) = 1 - y/Z, where the scalar y is the income
that, if possessed by everybody, would induce the same social welfare as the actual
income vector x.
Given an additively separable welfare function W ( x ) = U ( x i ) / n ,where
U ( . ) is utility, and a utility function
this approach leads to a family of inequality measures given by
A&(%)=
I 1-
geom. mean of x
arith. mean of x
(E
(E
2 0, E # 1)
= 1)
*See Kondor (1971) or Chipman (1985, pp. 142-143) for a hrief sketch of its rather long and winding
genesis.
?Of courhe, an implicit link can in most cases easily he constructed, as shown by Sheshinski (1972) for
the Gini index. In view of Newhery (1970), who shows that the Gini index is incompatible with additive
separable welfare functions, these implied welfare functions are sometimes rather odd.
48 KRAMER
where the parameter E measures aversion to inequality: As E + 00, A,(%) ap-

proaches 1-x(ll/Z,similar to Yitzhaki’s index G , ( x ) . Also, with E = 1the Atkinson
index is one minus the ratio of the geometric to the arithmetic mean, and with E = 2
it equals one minus the ratio of the harmonic to the arithmetic mean.
IV. AXIOMATIC APPROACHES TO THE MEASUREMENT

OF INEQUALITY
A different approach to avoid the ad hockery implicit in much of the measurement

of inequality, other than the welfare-based development above, is to specify a set of
axioms or minimal requirements that an index should obey. Ideally, this leaves one
with a single index formula, which is then characterized by these initial axioms.
The first and most obvious axiom is that, as a result of a Robin Hood transfer,
the inequality index decreases:
( A l ) Z ( x T ) 5 Z(x) for a matrix T of the form (2).

This minimal requirement is violated by at least one well-known measure of inequal-
ity, the empirical variance of the logs of the data: It is easy to construct examples
where income is shifted from rich to poor and inequality increases (Cowell 1988).*
Another seemingly obvious axiom is symmetry: For any permutation matrix P ,
we should have
(A2) I ( % )= Z ( x P ) .
Together, (Al) and (A2) amount to Schur convexity (and vice versa):
(As) x LM y * 5 I(y).
These axioms, though disputable, are widely accepted. The next two constitute a
watershed, separating leftist from rightist measures of inequality:
(A4) Z(hx) = Z(x) for all h > 0 (rightist).

+
(AS) Z(x he) = Z(x) for all h > 0 (leftist).
It is easily seen that the only functions I : D + R satisfying (Al)-(A5) are constants,
so if these are excluded there is no inequality measure that is both leftist and rightist
at the same time.
As a compromise, one might in this context also settle for the Eichhorn-Bossert-
Pfingsten axiom:
(A6) Z(x + t ( A x + (1 - A)e) = /(x) (see Eq. (11).
*As this can happen only for very high incomes, where the very rich give to the not so very rich, the
significance of this aberration is in practice much disputed, and this measure continues to he widely
used.
MEASUREMENT
OF INEQUALITY 49
Indices which satisfy (A6)remain constant when an income vector x is replaced by

a weighted mean of x and e = (1, . . . , l),so by adjusting weights one can fine-tune
one’s degree of social rightism or leftism as closely as desired.
Another, independent watershed is provided by the next two axioms, which
separate measures of industry concentration from measures of the inequality of in-
come and wealth:
(A7) Z(x, 0) = Z(x) (i.e., appending zeros does not affect inequality).
(A8) Z(x, x, . . . , x) = Z(x) (the “population invariance”: replicating a pop-
ulation rn times does not affect inequality).
A prominent index satisfying (A7) but not (A8) is the Herfindahl coefficient
It is often used in legal disputes about merger activities (see, e.g., Finkelstein and
Friedberg 1967),where (relatively) small firms do not matter for the degree of con-
centration in the market, and where the axiom (A8) would in fact constitute a liability.
Unlike (A4) and (A5), (A7) and (A8) are nontrivially compatible with each
other and (Al)-(A4): consider the well-known Theil index
where 0 ln(0) is taken to be 0. However, most other measures of inequality satisfy

either (A7) or (A8), but not both.
Another important set of axioms concerns decomposability: If we partition the
x-vector into x = (x(’),. . . , x ( ~ ) where
), x ( ~is) ni x 1 and Cni = n, it appears
desirable that we have
(A9) Z(x) is uniquely determined by Z(x(1)),n;, and Z(i).

Or in words: Knowing inequality in the subgroups and the subgroups’ size and mean
suffices to determine overall inequality.
Many well-known inequality measures such as the Gini index, the Pietra in-
dex, or the logarithmic variance fail this test, sometimes by wide margins. For the
Gini index, it can be shown that decomposability obtains if the subgroups do not
overlap, but other indices are not decomposable even then.* For the logarithmic
variance and the Pietra index, Cowell (1988)gives an example where, even when all
*At least not in the sense defined by (A9). However, as shown by Kakwani (1980, pp. 178-181), the Gini
index can be computed from information about factor incomes. This issue of decomposition by factors
versus decomposition by subgroups has engendered its own literature, which we do not have space to
cover here.
50 KRXMER
subgroups are strictly ordered by income, it is possible to find a change in income

such that (a) mean income in every group is constant, (b) inequality in every group
goes up, but still (c) overall inequality goes down.
A stronger version of (A9)is additive decomposability, defined as
where e(i) is an ni-vector of ones, w; >_ 0, and CE1wi = 1. This means that
overall inequality can be expressed as a weighted sum of within-group inequalities,
plus between-group inequality, defined as the overall inequality that would obtain if
there were no inequality within the groups.
Still stronger is what Foster (1983) calls Theil decomposability:
i.e., the weights are the income shares of the groups. Not surprisingly, the Theil
coefficient T ( x ) from (23) is decomposable that way.
A final set of axioms refers to “transfer sensitivity” (Shorrocks and Foster
1987):These principles strengthen the Pigou-Dalton axiom (A3)by requiring that the
reduction in inequality resulting from a Robin Hood transfer should ceteris paribus
be larger, the poorer the recipient. Relying solely on (A3),one could have “a sit-
uation in which a millionaire made a small (regressive) transfer to a more affluent
millionaire and a simultaneous large (progressive) transfer to the poorest person in
society” (Shorrocks and Foster 1987, p. 485), but where the combined effect of these
transfers is that inequality increases: while the regressive transfer increases inequal-
ity and the progressive transfer reduces inequality, the axiom (A3) by itself puts no
restraint on the relative magnitude of these effects, so one needs an additional axiom
to prohibit such eccentric behavior.
(A12) I (y’) 5 I (y) whenever y’ and y differ from x by a Robin Hood transfer
of the same size, with spender and recipient equal amounts apart, but
where the recipient in y’ is poorer.
This notion can also be formalized by defining the sensitivity of an inequality index
I ( - ) ,evaluated at the components x(;) and x(;) of some vector x, as
-k , i , ; )
Sl(x, i , j ) = lim
s+o 6
whenever this limit exists, where
MEASUREMENT
OF INEQUALITY 5I
For the Gini index, using (16), this is easily seen to be
2
S&, i, j ) = ____ ( j - (25)
i.e., the sensitivity depends only on the rank of the units involved in the transfer, not
upon their incomes. In particular, this implies that equal transfers of incomes have
most effect where the population is densest, which is usually in the center.
On the other hand, for the Atkinson family, sensitivity is easily seen to be
proportional to - l/xFjl, so sensitivity increases both with E and with x(i),
given Ix(,) - and with x(,) - x(;)given x(,): ceteris paribus the decrease in
inequality is larger, the poorer the recipient and the larger the income gap between
spender and recipient.
Given some set of axioms (often augmented by normalization restrictions on
the range, or requirements concerning continuity and differentiability), the following
questions arise: (1)Are the axioms consistent with each other? (2) Are all the axioms
really necessary? (3)What do indices which satisfy these axioms look like?
The first question is usually answered by exhibiting some specific measure
that satisfies all requirements. The second question is trickier. While some axioms
are easily seen to be implied by others ((A3) by ( A l ) and (AZ), (A9) by (AlO), (A8) by
(All)), others are not: As Russell (1985)demonstrates, at least two of the require-
ments that Cowell (1980)imposes to characterize the CES class of inequality indices
are already implied by the others, and this implication is anything but trivial to see.
Usually, such questions of minimality are settled by exhibiting, for every axiom, at
least one index that fails this test but satisfies the others (Eichhorn and Gehrig 1982).
The third question has generated a minor industry, producing results of the
type: Any function I : D + R with continuous first-order derivatives satisfying (AZ),
(A4), (A8) and (AlO), plus I ( e ) = 0, can be expressed as a positive scalar multiple
of some function
IAx) = n c ( c1- 1) k[(;)c

1= 1
- 13
with some c E R (Shorrocks 1980). This class of indices has become known as the
generalized entropy family; as special cases corresponding to c = 1 and c = 0
(defined as the limit of I,(x) as c -+ 1 or c -+ 0), it includes the Theil index T ( x )
defined in (23) and another index proposed by Theil (1967):
1= 1
In the same vein, Foster (1983)shows that an inequality index satisfies (A3), (A4),
and ( A l l ) if and only if it is a positive multiple of the Theil coefficient T ( x ) , and
52 KRAMER
Eichhorn (1988) shows that the only functions I : D -+ R satisfying (A3), (A6), and
I ( e ) = 0 are
Ax + ( 1 - A)e
AT+ (1 - A)
with some Schur-convex function f such that f ( e ) = 0.
V. EMPlRlCAL IMPLEMENTATION
The empirical application of measures of inequality raises various issues which set
this branch of statistics apart from others. A first and minor problem is the proper
inference from a sample to a larger population. As samples are typically large, or
nonrandom, or populations rather small as in the context of industrial concentration,
relatively little work has been done on this.*
Much more important, in particular in the context of income inequality, is the
incompleteness of the data: Typically, income figures are available only for certain
quantiles of the population, and the rather voluminous literature on inequality mea-
surement with incomplete data can be classified according to the amount of addi-
tional information available.
For the case where only selected points of the Lorenz curve are given (i.e., frac-
tions of total income received by fractions of the total (ordered) population), Mehran
(1975)gives the most extreme Lorenz curves that are compatible with these points,
in the sense that the resulting concentration areas are the smallest and the largest
possible.
Obviously, the upper bound is attained by joining the observed points by
straight lines, as in Figure 2. This is at the same time an upper bound for the true
underlying Lorenz curve. Likewise, an obvious lower bound to the true underlying
Lorenz curve is given by extending these lines, as again in Figure 2. However, as
these line segments do not form a Lorenz curve, Mehran (1975) proposes tangents
to the true Lorenz curve at the observed points such that the concentration area and
thus the Gini index are maximized, and he gives a recursive algorithm to compute
the tangents' slopes. Although this curve does not necessarily bound from below the
true Lorenz curve, it gives the Lorenz curve which represents (in the sense of the
Gini index) the most unequal distribution compatible with the data.
If, in addition to selected points of the Lorenz curve, we are given the interval
means di), the interval endpoints a;-] and a ; , and the fraction of the population in
*See, for instance, Sendler (1979)on the asymptotic distribution of the Gini index and the Lorenz curve,
or McDonald and Ransom (1981)for the effects of' sampling variability on the bounds below.
MEASUREMENT
OF INEQUALITY 53
Figure 2 Given 2 nontrivial points P1 and P2, the true underlying Lorenz curve must pass
through the shaded area.
the various nonoverlapping subgroups that compose the vector x = (%('I, . . . , x ( ~ ) ) ,

the bounds on the Gini index derived by Mehran can be sharpened as shown by Gast-
wirth (1972). The famous Gastwirth bounds* are based on the expression G ( x ) =
A(x)/2? and on the decomposition
i,j=1 i= 1
where Aci) is Gini's mean difference for group i. In conjunction with
this decomposition provides immediate bounds for the Gini index, with the lower
one being attained for A(i) = 0, (i = 1 , . . . , m ) , i.e., when there is no inequality
within the groups, and the upper bound being attained when the observations in a
given group are placed at both ends of the interval in a proportion such that the group
mean is z(~):
*In part already in Pizzetti (1955); see Giorgi and Pallini (1987).
54 KRXMER
i,j=l
Still sharper bounds can be obtained from information on symmetry or skewness of

the distribution within groups: for instance, if distances between observations are
increasing, as often happens in the higher brackets of income distributions, it can
be shown that
A useful rule of thumb when no such extra information is available is provided by

the rule, as proposed by Champernowne in an unpublished manuscript and as
formally derived in Cowell and Mehta (1982):Approximate the unknown true Gini
+
index by i G ~ ( x ) $ G u ( x ) . As shown by Cowell and Mehta, this rule gives the
exact Gini index when the observations within each group are split at di) and spread
evenly across the subintervals (ai-1, di)) and (di),ai) such that the group mean is
preserved (the “split-histogram technique”), but it gives a good approximation to the
true Gini index also for many other disturbances.
In the same vein, one obtains upper and lower bounds also for other measures
of inequality, which, like the Gini index, are decomposable by population subgroups
when the subgroups are nonoverlapping (basically the Gini index plus the gener-
alized entropy family from (26);see Shorrocks 1984):The lower bound is given by
a within-group distribution where everybody receives ?idi),and the upper bound is
given by placing the observations at the ends of the group interval as above. How-
ever, if the measure is not the Gini index, the+-: rule has to be adapted: the optimal
approximation-in some sense-to the true inequality index is now f lower bound
+; upper bound (Cowell and Mehta 1982).
Using similar techniques, Gastwirth (1975)extends the bounds above to arbi-
trary inequality measures that can be expressed as ratios of a measure of spread to X
(which in view of (13) and (19)comprises the Gini and Pietra indices a special cases,
but covers many other indices as well, possibly after suitable transformations).
As an alternative to bounds, it is often suggested to fit a theoretical Lorenz
curve to selected values L ; ( p ) either by interpolation, such that the fitted curve
passes through the points (pi,L ( p ; ) ) , or by some least-squares technique, where
the curve is chosen from some given family, and to compute the desired measure of
inequality from the fitted Lorenz curve.
When doing interpolation, Gastwirth and Glauberman (1976) propose cubic
polynomials for the respective intervals whose derivatives at the endpoints match
MEASUREMENT
OF INEQUALITY 55
those of the true underlying Lorenz curve* (Hermite interpolation). The resulting
function need not be convex, i.e., need not be a Lorenz curve itself, but it is easily
seen (see, e.g., Schrag and Kramer 1993)that the Hermite interpolator is convex and
monotone if in each interval
i.e., if the group means stay clear of the endpoints of the intervals.
As to fitting parametric Lorenz curves, various families have been proposed,
whose usefulness however seems much to be in doubt as the implied indices of in-
equality often violate the Gastwirth bounds. One such popular family, suggested by
Rasche et al. (1980)is
where 0 < a! 5 1 , 0 < B 5 1. The implied Gini coefficient is

2 1 1
G ( x ) = 1 - - B ( - , - - 1) (35)
a! Q B
where B(., .) is the familiar Beta function, so this approach amounts to first fitting
(34)to the observed points of a true Lorenz curve via some nonlinear least-squares
technique and approximating the true Gini coefficient by (35).
However, as shown by Schader and Schmid (1994), the approximations thus
obtained often fall outside the Gastwirth bounds (for curves of the form (34)and for
other families as well), so the consensus is that such approximations do not work (see
also Slottje 1990).
VI. MULTIDIMENSIONAL INEQUALITY AND POVERTY
There are various additional research areas which impinge on inequality, such as
the welfare ranking of income distributions (Shorrocks 1983, Thistle 1989),the mea-
surement of interdistributional inequality (Butler and McDonald 1987, Dagum 1987),
or the parametric modeling of income distributions and the inequality orderings that
are implied by the parameters (Chipman 1985, Wilfling and Kramer 1993), which
I do not cover here. Rather, I pick the issues of multidimensional inequality and
poverty to show how the concepts introduced above can be extended.
Take poverty. Although among certain sociologists, this is taken to be almost
synonymous to inequality (“relative deprivation”), the majority consensus is that
*Rememberthat the Lorenz curve has the slope x ( , ) / X in the interval ((i - l)/n, i / n ) , so its (left-hand)
slope in pi is sip.
56 KRAMER
poverty is something different, and needs different measures for quantification. Fol-
lowing Sen (1976), this is usually done by first identifying a poverty line z below
which an income unit is considered poor, and then combining the poverty character-
istics of different units into an overall measure of poverty. Two obvious candidates
are the “head-count ratio”
n*
H(x) = -
n
where n* is the number of units below the poverty line, and the “income gap ratio”
i.e., the normalized per unit percentage shortfall of the poor. Individually, both mea-
sures have serious shortcomings-the head-count ratio is completely insensitive to
the extent of poverty, and might even fall when income is transferred from a poor
person to somebody not so poor, who thereby moves above the poverty line, and the
income gap ratio takes no account of the numbers of the poor-but they can be com-
bined into a measure with various desirable properties (Sen 1976):
+
S ( x ) = H ( x ) [ l ( x ) (1 - Z ( x ) ) G ( x ) ] (38)
-
where G ( x ) = G ( x ( l ) ,. . . , x(,,*))is the Gini index of the incomes of the poor. This
measure reduces to H ( x ) . Z(x) if all the poor have the same income, and it increases
whenever an income of a poor person is reduced (while incomes above the poverty
line always receive uniform weights, entering only via the head-count ratio H ( x ) ) .
For large n*, the coefficient S ( x ) can be shown to be almost identical to
i.e., it is a weighted sum of the income gaps of the poor, with weights increasing
as income gaps increase, which at the same time points to a serious shortfall of this
measure of poverty; it is based solely on the ranks, not on the distances of the incomes
of the poor, so there has been an enormous literature in the wake of the seminal paper
by Sen (1976), surveyed in Seidl (1988), which goes on from here.
Comparatively little work has been done on the second important issue of mul-
tidimensional inequality. While it is widely recognized that a single attribute such
as income is often not sufficient to capture the phenomenon whose inquality is to be
determined, the statistician’s toolbox is almost empty here.
There is a short chapter on multivariate majorization in Marshal1 and Olkin
(1979), where majorization among n-vectors x and y is extended to majorization
among M x n matrices X and Y : By definition, X > M Y if and only if Y = X P
with some doubly stochastic matrix P. When m = 1, this is equivalent to Y =
OF INEQUALITY57
MEASUREMENT
XT1, T2, . . . , TL with finitely many matrices of the form (2), but when rn 2 2 and
n 2 3, the latter condition is more restrictive, as simple examples show (Marshal1
and Olkin 1979, p. 431).Therefore I denote the latter preordering by ">T."
As shown by Rinott (1973),a differentiable function f:R""" + R respects
the preordering > T if and only if f ( x ) = f ( x P ) for all permutation matrices P , and
if for all j , k = 1, . . . , n,
where f i k ( x ) = af/ax;k evaluated at X . Similar to the univariate case, this con-

dition could presumably be used to suggest new indices and screen old indices of
inequality. However, as the rationale for the >r preorder is much less compelling
for rn > 1 as it is for rn = 1 (with data matrices, even fewer pairs will in general be
ordered in a given application than if we had only vectors), the condition (40) does
not immediately suggest a convenient index of multivariate inequality.
Similar to the univariate case, these majorization orderings can be embedded
in a more general order, which in turn is based on a multivariate generalization of the
Lorenz curve (Koshevoy and Mosler 1996).For X E RY"", define the Lorenz zonoid
of X as the convex hull of the set
and define
For rn = 1, this boils down to the univariate Lorenz order, as L Z ( X ) is then the
subset of R2 enclosed by the Lorenz curve L x ( p ) and the "dual Lorenz curve"
E x ( p ) := 1 - L x ( p - 1) (the area of which is equal to the Gini index G ( x ) ) . For
rn > 1, the Lorenz zonoid is a convex subset of the unit cube in Rrn+'.
The multivariate Lorenz ordering defined by (42) allows matrices of different
means and row dimensions to be compared. It is further related to univariate Lorenz
dominance by the fact that X 21,Y if and only if the generalized Lorenz curve of
d'X is below the generalized Lorenz curve of d'Y for all coefficient vectors d such
that d ; 2 0 and x ( d ; ) = 1 (Koshevoy and Mosler 1996, Theorem 3.1).
Other generalizations from univariate to multivariate concepts of inequality
exploit the parallel between inequality and choice under uncertainty (Kolm 1977).
Atkinson and Bourgignon 1982),or the obvious relationship between multivariate in-
equality and the decomposition of inequality by factor components (Shorrocks 1988,
Maasoumi 1986, Rietveld 1990), but as this field has not yet reached the maturity
for a useful survey, I had better close my survey here.
58 KRAMER
ACKNOWLEDGMENTS
I am grateful to Christian Kleiber, Sonja Michels, Karl Mosler, Andreas Pfingsten,

Friedrich Schmid, Andreas Stich, Mark Trede, Thorsten Ziebach, and an unknown
referee for helpful discussion and comments.
Arnold, B. C. (1987),Majorization and the Lorenz Order: A Brieflntroduction, Springer, Berlin.

Atkinson, A. B. (1970), On the Measurement of Inequality, Journal of Economic Theory, 21,
244-263.
Atkinson, A. B. and F. Bourgignon (1982),The Comparison of Multi-Dimensional Distribution
of Economic Status, Review ojEcorzomic Studies, 49, 183-201.
Berrbi, Z. M. and J. Silber (1987), Dispersion, Asymmetry and the Gini Index of Inequality,
International Economic Review, 28,33 1-338.
Blackburn, M. L. (1989), Interpreting the Magnitude of Changes in Measures of Income In-
equality, Journal of Econometrics, 42,21-25.
Bossert, W. and A. Pfingsten (1989), Intermediate Inequality: Concepts, Indices and Welfare
Implications, Mathematical Social Sciences, 19, 117-134.
Bresciani-Turroni, C. (1910),Di un Indice Misuratore Della Disugualianza Nella Distribuzione
Della Richezza, Studi in Onore di Biagio Brugi, Palermo, 793-812.
Butler, R. J. and J. B. McDonald (1987), Interdistributional Income Inequality, Journal of
Business and Economic Statistics, 5, 13-18.
Chipman, J. (1985), The Theory and Measurement of Income Distribution, Advances in Econo-
metrics, 4, 135-165.
Cowell, F. A. (1980), On the Structure of Additive Inequality Measure, Review ofEconomic
Studies, 47,521-531.
Cowell, F. A. (1988), Inequality Decomposition: Three Bad Measures, Bulletin of Economic
Research, 40,309-312.
Cowell, F. A. and F. Mehta (1982), The Estimation and Interpolation of Inequality Measures,
Review of Economic Studies, 44,273-290.
Dagum, C. (1987), Measuring the Economic Affluence between Populations of Income Re-
ceivers, Journal of Business und Economic Statistics, 5, 5-12.
Dalton, H. (1920), The Measurement of Inequality of Incomes, Economic Journal, 30, 348-
361.
Dasgupta, P., A. K. Sen, and D. Starrett (1973), Notes on the Measurement of Inequality,
Journal of Economic Theory, 180-187.
Ebert, U. (1988), A Family of Aggregative Compromise Inequality Measures, International
Economic Review, 29,363-376.
Eichhorn, W. (1988), On a Class of Inequality Measures, Social Choice and Welfare, 5, 171-
177.
Eichhorn, W., H. Funke, and W. Richter (1984), Tax Progression and Inequality of Income
Distribution, Journal of Mathematical Economics, 13, 127-131.
MEASUREMENT
OF INEQUALITY 59
Eichhorn, W. and W. Gehrig (1982), Measurement of Inequality in Economics, in B. Korte

(ed.), Modern Applied Mathematics, North-Holland, Amsterdam, 657-693.
Elteto, 0.and E. Frigyes (1968),New Income Inequality Measures as Efficient Tools for Causal
Analysis and Planning, Econometrica, 35,383-396.
Fellman, J. (1976), The Effect of Transformations on the Lorenz Curve, Econometrica, 44,
823-824.
Finkelstein, M. 0. and R. M. Friedberg (1967), The Application of an Entropy Theory of
Concentration to the Clayton Act, Yale Law Journal, 76,677-717.
Foster, J. E. (1983),An Axiomatic Characterization of the Theil Measure of Income Inequality,
Journal of Economic Theory, 31, 105-121.
Gastwirth, J. L. (1972), The Estimation of the Lorenz Curve and Gini Index, Review of Eco-
nomics and Statistics, 54,306-316.
Gastwirth, J. L. (1975), On Estimating a Family of Measures of Economic Inequality, Journal
of Econometrics, 3,61-70.
Gastwirth, J. L. and M. Glauberman (1976), The Interpolation of the Lorenz Curve and Gini
Index From Grouped Data, Econometrica, 44,479-483.
Giogi, G. M. (1990),Bibliographic Portrait of the Gini Concentration Ration, Metron, 48,183-
221.
Giogi, G. M. and A. Pallini (1987), About a General Method for the Lower and Upper Distribu-
tion-Free Bounds on Gini’s Concentration Ratio From Grouped Data, Statistica, 47,
171-184.
Hardy, G. H., J. E. Littlewood, and G. Polya (1934),Inequalities, Cambridge University Press,
London.
Kakwani, N. C. (1980), Income, Inequality and Poverty, Oxford University Press, Oxford.
Kakwani, N. C. (1984), Welfare Ranking of Income Distributions, Advances in Econometrics,
3,191-213.
Kolm, S. Ch. (1976), Unequal Inequalities I, Journal of Economic Theory, 12,416-442.
Kolm, S. Ch. (1977), Multidimensional Egalitarianism, Quarterly Journal of Economics, 91,
1-13.
Kondor, Y. (1971),An Old-New Measure of Income Inequality, Econometrica, 39,1041-1042.
Koshevoy, G. and K. Mosler (19%),The Lorenz Zonoid of a Multivariate Distribution, Journal
of the American Statistical Association, 91, 873-882.
Kuznets, S. (1955),Six Lectures on Economic Growth, The Free Press, Glencoe.
Lambert, P. J. (1993), The Distribution and Redistribution of Income, Manchester University
Press, Manchester.
Lerman, L. and S. Yitzhaki (1984), A Note on the Calculation and Interpretation of the Gini-
Index, Economics Letters, 15, 363-368.
Maasoumi, E. (1986),The Measurement and Decomposition of Multi-Dimensional Inequality,
Econometrica, 54,991-997.
Marshall, A. W. and I. Olkin (1979), Inequalities: Theory of Majoritation and Its Applications,
Academic Press, New York.
McDonald, J. B. and M. Ransom (1981), An Analysis of the Bounds for the Gini Coefficient,
Journal of Econometrics, 17, 177-188.
Mehran, F. (1975), Bounds on the Gini-Index Based on Observed Points of the Lorenz Curve,
Journal of the American Statistical Association, 70,64-66.
Mehran, F. (1976), Linear Measures of Income Inequality, Econometrica, 44,805-809.
60 KRAMER
Mosler, K. (1994), Majorization in Economic Disparity Measures, Linear Algebra and Its Ap-
plications, 199,91-114.
Moyes, P. (1987), A New Concept of Lorenz Domination, Economics Letters, 23,203-207.
Newbery, D.A. (1970), Theorem on the Measurement of Inequality, Journal of Economic The-
ory, 2,264-266.
Pfingsten, A. (1988a), New Concepts of Lorenz Domination and Risk Aversion, Methods of
Operations Research, 59, 75-85.
Pfingsten, A. (1988b),Progressive Taxation and Redistributive Taxation: Different Labels for
the Same Product?, Social Choice and WeZjare, 5,235-246.
Pietra, G. (1914/1915), Della Relazione Tra Gli Indice di Variabilita, Atti del Realo Zstituto
Veneto di Scienze, 775-792; 793-804.
Pigou, A. C. (1912), Wealth and Welfare, Macmillan, London.
Pizzetti, E. (1955),Osservazioni Sul Calculo Aritmetica del Rapport0 di Concentrazione, Studi
in onore di Guetano Pietra, Capelli, Bologna.
Pyatt, G. (1976), On the Interpretation and Disaggregation of Gini-Coefficients, Economic
Journal, 86,243-255.
Rasche, R. H., J. Gaffney, A. Y. C. Koo, and N. Obst (1980), Functional Forms for Estimating
the Lorenz Curve, Econometrica, 48, 1061-1062.
Ricci, U. (1916), L'indice di Variabiliti e la Curva d e Redditi, Giornali Degli Economisti e
Reivsta de Statistica, 3,177-228.
Rietveld, P. (1990), Multidimensional Inequality Comparisons, Economics Letters, 32, 187-
192.
Rinott, Y. (1973), Multivariate Majorization and Rearrangement Inequalities with Some Ap-
plications to Probability and Statistics, Israel Journal ofklathematics, 1 5 , 6 0 4 7 .
Russell, R. (1985), A Note on Decomposable Inequality Measures, Review of Economic Stud-
ies, 52,347-352.
Schader, M. and F. Schmid (1994), Fitting Parametric Lorenz Curves to Grouped Income
Distributions-A Critical Note, Empirical Economics, 19, 361-370.
Schrag, H. and W. Kramer (1993), A Simple Necessary and Sufficient Condition for the Con-
vexity of Interpolated Lorenz Curves, Stntistica, 53, 167-170.
Schutz, R. R. (1951), On the Measurement of Income Inequality, American Economic Review,
41,107-122.
Seidl, Ch. (1988),Poverty Measurement: A Survey, Welfare and Efficiency, in D. Bos, M. Rose,
and Ch. Seidl (eds.), Public Economics, Springer, Berlin.
Sen, A. (1976), Poverty: An Ordinal Approach to Measurement, Econometrica, 44,219-231.
Sendler, W. (1979),On Statistical Inference in Concentration Measurement, Metrika 26, 109-
122.
Sheshinski, E. (1972), Relation between a Social Welfare Function and the Gini Index of
Income Inequality, Journa2 of Economic Theory, 4,98-100.
Shorrocks, A. F. (1980), The Class of Additively Decomposable Inequality Measures, Econo-
metrica, 48,613-625.
Shorrocks, A. F. (1983), Ranking Income Distributions, Economica, SO, 3-17.
Shorrocks, A. F. (1984), Inequality Decomposition by Population Subgroups, Econometrica,
52,369-1385.
Shorrocks, A. F. (1988), Aggregation Issues in Inequality Measurement, in W. Eichhorn (ed.),
Economics, Physica, Heidelberg.
MEASUREMENT
OF INEQUALITY 6I
Shorrocks, A. F. and J. E. Foster (1987), Transfer Sensitive Inequality Measures, Review of

Economic Studies, 54,486-497.
Slottje, D. J. (1990), Using Grouped Data for Constructing Inequality Indices-Parametric vs.
Nonparametric Methods, Economics Letters, 32, 193-197.
Theil, H. (1967), Economics and Information Theory, North-Holland, Amsterdam.
Thistle, P. D. (1989), Ranking Distributions with Generalized Lorenz Curves, Southern Eco-
nomic Journal, 56, 1-2.
von Bortkiewicz, L. (1931),Die DisparitatsmaBe der Einkommensstatistik, Bulletin de l’lnstitut
International de Statistique, 25, 189-298.
Wilfling, B. and W. Kramer (1993). The Lorenz Ordering of Singh-Maddala Income Distribu-
tions, Economics Letters, 43, 53-57.
Yitzhaki, S. (1983), On a n Extension of the Gini-Index, International Economic Review, 24,
617-628.
Poor Areas
Martin Ravallion
World Bunk, Washington, D.C.
As China’s economic miracle continues to leave millions behind, more and more
Chinese are expressing anger over the economic disparities between the flour-
ishing provinces of China’s coastal plain and the impoverished inland, where 70
million to 80 million people cannot feed or clothe themselves and hundreds of
millions of others are only spectators to China’s economic transformation. The
New York Times, December 27, 1995, p. 1.
China is not unusual; almost all countries have their well-recognized “poor areas,” in
which the incidence of absolute poverty is unusually high by national standards. In
China, there is high poverty incidence in rural areas of the southwest and northwest
(the “inland” areas referred to in the quotation). Similar examples in other countries
include some of the eastern Outer Islands of Indonesia, parts of northeastern India,
northwestern and southern rural areas of Bangladesh, much of northern Nigeria, the
rural Savannah in Ghana, the northeast of Brazil, and many other places.
We would hope, and under certain conditions expect, that the growth process
will help these poor areas catch up. But that does not appear to be happening in some
countries. Figure 1 illustrates the divergence over time between the relatively well
off and more rapidly growing coastal areas of China and the lagging inland areas.
The figure plots the aggregate rate of consumption growth at county level in southern
China 1985-1990 against the initial county mean wealth. The data cover 119 coun-
ties spanning a region from the booming coastal province of Guangdong through to
the poor inland areas of Guizhou.* There is a positive regression coefficient, sug-
*The figure is reproduced from Ravallion and Jalan (1996).It is based on a panel of farm-household level
data for rural areas in four provinces of southern China. The data cover 4700 households living in 119
counties. The consumption measure is comprehensive, in that it includes imputed values (at local market
63
64 RAVALUON
Change in log county mean consumption per capita, 1985-90
0.4 0 0
0 0 0
0.2
0
0 0
-0.2 0 0 0
0
0
-0.4 I
-0.6
5.6 5.8 6.0 6.2 6.4 6.6 6.8 7.0 7.2 7.4 7.6 7.8 8.0
Log of county wealth per capita, 1985
Figure I Consumption growth by county in rural South China.
gesting divergence, and it is significant (at the 1% level). Initially wealthier counties
tended to have higher subsequent rates of consumption growth.
Nor is China the only country in which poor areas appear to persist in spite
of robust economic growth; for example, the eastern Outer Islands of Indonesia ap-
pear to have shared rather little in that country’s sustained (and generally pro-poor)
economic growth since 1970. It seems that there is a degree of persistence in the
economic geography of poverty; indeed, a generation or more ago, the above list of
“poor areas” by country would probably have looked pretty similar.
As the opening extract suggests, there are widespread concerns about poor
areas, particularly when they persist amidst robust aggregate economic growth. In
assessing the social impact of economic growth or growth-oriented economic reform,
economists have traditionally focused on the impact on one or more measures of
prices) of consumption from own production plus the current service flows from housing and consumer
durables. The data also include a seemingly complete accounting of all wealth including valuations of
all fixed productive assets, cash, deposits, housing, grain stock, and consumer durables. The data are
discussed at length in Chen and Ravallion (19%).
POORAREAS 65
social welfare, including various measures of aggregate poverty. Yet, it appears that
impacts are typically diverse among the poor, and in the society as a whole; some lose
and some gain from economy-wide changes. This can be important to know, if only to
better understand the political economy of growth and reform, though policymakers
may well also make the (normative) judgment that a premium should be attached to
more “balanced” growth. Studying these diverse impacts may also hold important
keys to our understanding of the growth process itself and to a variety of questions
often asked by policymakers, such as what is the best policy response to the problem
of why some subgroups are lagging.
So why do some people, and in particular some regions, do so much better than
others in a growing economy? It turns out that most of the standard tools of analy-
sis used in studying poverty, distribution, and growth are ill-equipped to answer this
question. After reviewing those tools, this chapter suggests some new tools of empiri-
cal analysis that may offer a better chance of answering it. We look at the dynamics of
the geography of poverty from a microlevel to help understand the way various initial
conditions and exogenous shocks impinge on household-level prospects of escaping
poverty over time. While we note the links to various strands of theoretical and em-
pirical economics, the chapter is not a survey. Rather it tries to be forward-looking
on a set of seemingly important research questions, to explore how future research
might better address them.
This is also an issue of considerable relevance, as the chapter will emphasize.
The empirical approach outlined here would appear to entail a substantial expan-
sion in the number of policy-relevant variables which are included in microempiri-
cal models of poverty. Past interventions in poor areas are amongst those variables.
Faced with lagging regions amidst overall growth, governments and donors are regu-
larly called upon to do something about these lagging poor areas. Area-based inter-
ventions are now found in most countries.* How much impact do such interventions
have on living standards? To answer this we must be able to assess what would have
happened to living standards in the absence of the interventions. It should not be as-
~~
*For example, on recognizing the problem of lagging rural areas, China introduced a large antipoverty
program in 1986 which declared that 272 (rural) counties were “national-poor counties,” and targeted
substantial aid to those counties. The extra aid took the form of‘subsidized credit for village-level projects
(provided at well below market rates of interest), funding for public works projects (under “food-for-work”
programs), and direct budgetary support to the county government. This national poor area program is
the main direct intervention in the government of China’s current poverty reduction policy (Leading
Group 1988, World Bank 1992, Riskin 1994).Again China is not unusual. The World Bank has assisted
over 300 area development projects sinre the early 1950s spread over all regions; most of these projects
were designed to develop a selected rural area for the benefit of poor people. Other agencies, such as the
International Fund for Agricultural Development, also provide substantial support for such programs
(Jazairy et al. 1992).There has been a recent resurgence of interest in such programs in the World Bank
and elsewhere.
66 RAVALLION
sumed that such schemes will even entail net gains to poor people; by acting against
the flow of factor mobility from low to high productivity areas, it could be argued that
such interventions actually make matters worse in the longer term. Depending on
how the economy works in the absence of intervention-the nature of the technol-
ogy, preferences, and any constraints on factor mobility-a poor area program could
entail either a net benefit or net cost to poor people.
The paper argues that the geographic variation in both initial conditions and
the evolution of living standards over time offers scope for disentangling the effects of
poor-area programs from other factors. Even within poor countries, geographic areas
differ widely in their endowments of various aspects of “geographic capital,” includ-
ing locally provided public services and access to area-specific subsidies. These dif-
ferences are both geoclimatic and the outcomes of past policies and projects. There
is typically also a spatial variance in poverty indicators. The spatial variation in both
the incidence of poverty and in area characteristics offers hope of better understand-
ing why we see poor areas, what can be done to help them, and how well past efforts
have performed. By exploiting this spatial variation, we should be in a better position
to understand what role the lack of geographic capital plays in creating poor areas,
versus other factors including residential differentiation, whereby people who lack
“personal capital” end up being spatially concentrated.
The following section explains the motivation for studying the problem of “poor
areas.” Section I1 discusses the “standard” empirical tools found in practice, rely-
ing on either static micromodels or aggregate dynamic models. Section I11 discusses
a micromodeling approach and its links with recent work in economics on the de-
terminants of economic growth. Some potential lessons for policy are described in
Section IV.
1. MOTIVATION
A. W h y Do We See Unusually Poor Areas?
The starting point for assessing the pros and cons of poor-area policies is an under-
standing of why we see poor areas in the first place. Among economists and poli-
cymakers, a common explanation of poverty is based on an individualistic model in
which poverty arises from low household-level endowments of privately held produc-
tive resources, including human capital, albeit with important links to the regional
and macroeconomy, notably through wages and prices. This view is epitomized in
the familiar human-capital earnings functions. The dynamic version is the stan-
dard neoclassical growth model, the microfoundation of which assumes atomistic
agents linked only through trade at common prices. If one believes this model, then
poor areas presumably arise because people with poor endowments tend to live to-
gether. Area differences in access to (for example) local public goods might still be
allowed in such a model, but as long as there is free mobility they will not mat-
POORAREAS 67
ter to the welfare of an individual household, which will depend on its a-spatial
exogenous attributes. The types of antipoverty policies influenced by this model
emphasize raising the endowments of poor people, such as by enhanced access to
schooling.
Regional divergence is still possible in such a model. If there are increasing re-
turns to scale in private production inputs, then initially better-off areas, with better
endowments of private capital, will tend to see subsequently higher rates of growth.
This is the essence of the view of regional growth that one finds in the writings of
Myrdal (1957), Hirschman (1958), and others since (see the review in Richardson
and Townroe 1986). By this interpretation, persistently poor areas, and divergence
from wealthier areas, reflect the nature of the technology and the geography of nat-
ural resource endowments. There may still be a case for targeting poor areas, but it
would be a redistributive case, and it would imply a trade-off with the overall rate of
economic growth.
The individualistic model does not attach any causal significance to man-made
spatial inequalities in geographic capital-the set of physical and social infrastruc-
ture endowments held by specific areas. Indeed, with free mobility, the individu-
alistic model predicts that household welfare will only depend on private, mobile,
endowments, and other exogenous attributes of the household. Against this view,
one can postulate a geographic model in which individual poverty depends heavily
on geographic capital and mobility is limited. By this view, the marginal returns to a
given level of schooling, or a loan, depend substantially on where one lives, and lim-
ited factor mobility entails that these differences persist. Relevant geographic factors
might include local agroclimatic conditions, local physical infrastructure, access to
social services, and the stock of shared local knowledge about agroclimatic condi-
tions and about the technologies appropriate to those conditions. It is not implausible
that some or all of these geographic factors alter the returns to investments in private
capital. As I argue later, it is likely that they will also entail increasing returns to ge-
ographic capital when there are nonincreasing returns to private production inputs.
Thus it might well be that people are being left behind by China’s growth process
precisely because they live in poor areas; given their private endowments, they would
do better in China’s coastal areas.
If this model is right, then the policies called for will entail either public invest-
ment in geographic capital or (under certain conditions, discussed below) proactive
efforts to encourage migration, and such policies need not entail a trade-off with the
overall rate of growth. That will depend on the precise way in which differences in
geographic capital impact on the marginal products of private capital and, hence,
the rate of growth. That is an empirical question.
Neither model provides a complete explanation for poor areas. The individ-
ualistic model begs the questions of why individual endowments differ persistently
and why residential differentiation occurs. The geographic model begs the questions
of why community endowments differ and why mobility is restricted. But, as I will
68 RAVALLION
argue, knowing which model is right, or what the right hybrid model looks like, can
provide valuable information for policy.
8. PastWork
What do we already know about poor areas that might throw light on which of these
models is most relevant? There has been a vast amount of empirical research testing
the individualistic model, including human-capital earnings functions and similarly
motivated income and consumption determination models estimated on microdata
(Section 1I.B). This research has often assumed that the individual, “private capi-
tal,” model holds. While sometimes spatial variables are added, this is done in an
ad hoc way. At the same time, there is also a large, but mostly independent, liter-
ature on economic geography and regional science which has emphasized the im-
portance of spatial effects on the growth process (for a survey see Richardson and
Townroe 1986). The individualistic model has not been tested rigorously against the
geographic model, in an encompassing framework which would allow the two models
to fight it out.
But there is evidence of spatial effects in the processes relevant to creating
and perpetuating poverty. The evidence of spatial effects comes from a variety of
sources, including the following.
1. “Poverty profiles” (decompositions of aggregate poverty measures by sub-
groups of a population, including area of residence) typically contain evidence of
seemingly significant spatial differences in poverty incidence or severity. However,
typical poverty profiles do not allow one to say whether it is the individualistic model
or the geographic model that is producing these spatial effects. In the (far fewer)
cases in which suitable controls were used, spatial effects did appear to persist (van
de Walle 1995, Jalan and Ravallion 1996, Ravallion and Wodon 1997).
2. In some of the settings in which there are persistently poor rural areas
there does not appear to be much mobility among rural areas (though more so from
rural to urban areas). In some cases (such as China) mobility has been deliberately
restricted, but intrarural mobility seems uncommon elsewhere (such as in much of
South Asia, though exceptions exist, such as seasonal migration of agricultural la-
bor). Then the individualistic model immediately seems implausible; for how did the
residential differentiation come about with rather little mobility?
3. The literature on the diffusion process for new farm technologies has em-
phasized local community factors, including the demonstration effect of the presence
of early adopters in an area, and there is some supportive evidence for India in Foster
and Rosenzweig (1995).
4. There is also evidence for India that areas with better rural infrastructure
grow faster and that infrastructure investments tend to flow to areas with good agro-
climatic conditions (Binswanger et al. 1993).The type of data used has not, however,
allowed identification of external effects (discussed further in Section 111).
POORAREAS 69
5. Gains from the geographic concentration of some industries (arising from

scale economies and limited factor mobility; see Krugman 1991)could also entail
spatial effects in a growth process. This could be magnified by intersectoral link-
ages. For example, there is evidence for China (Sengupta and Lin 1995)that higher
non-farm-sector growth-possibly policy induced-brings external benefits to the
traditional farm sector through improved technologies and management. There is
evidence of positive external effects of higher density of economic activity on pro-
ductivity across states of the United States (Ciccone and Hall 1996).
6. There is also some evidence of human-capital spillover effects. In the
United States, the neighborhood where a child was raised appears to influence her
schooling performance and adult wages (Borjas 1995, Datcher 1982, Wilson 1987,
Case and Katz 1991).
7. Informal risk-sharing arrangements within poor communities entail that
individual consumption depends in part on the community’s aggregate consumption;
there is supportive evidence for India in Townsend (1991), though also see Ravallion
and Chaudhuri (1997).
8. Preferences, including discount rates, may well be formed within commu-
nities, entailing spatial effects on (among other things) the growth process. This is
hard to test, but there is evidence for Indonesia of spatial autocorrelation in consumer
demand behavior (Case 1991).
9. There is also some evidence of spatial autocorrelation in growth processes
at the country level; the higher the average growth rate of a country’s neighbors the
higher is its own growth rate ceteris paribus (e.g., Easterly and Levine 1995).
All this is suggestive, but we are still a long way from a good understanding
of why poor areas exist and persist. The answers could have great bearing on devel-
opment policy. If the process of escaping poverty involves strong spatial effects then
there may be large benefits from policies and projects which are targeted to poor ar-
eas, even if they are not targeted to households with poor endowments per se. It may
also niean that, without (possibly substantial) extra resources, or greater mobility,
the poor may be caught in a spatial poverty trap. To have any chance of success, an
antipoverty policy may have to break the community-level constraints on escaping
poverty, by public investment or encouraging migration.
C. Poor-Area Policies
What are the policy options in assisting poor areas? Two broad types of area-based
policy intervention can be identified which are aimed (explicitly or implicitly) at
poverty reduction: one is geographic targeting of subsidies, taxes, or public invest-
ments, and the other is migration policy.
Geographic targeting of antipoverty schemes has been popular, though so far
the assessments of poverty impacts have largely ignored dynamic effects. The at-
traction of this policy option for targeting stems from the existence of seemingly
70 RAVALLION
substantial regional disparities in living standards in many developing countries.

Place of residence may thus be a useful indicator of poverty. Local governments pro-
vide an administrative apparatus. However, some assessments of the potential for
this policy instrument to have more than a minor impact on aggregate poverty have
not been encouraging. While regional poverty profiles for LDCs typically show large
geographic disparities, regional targeting still entails a leakage of benefits to the non-
poor in poor regions, and a cost to the poor in rich regions. And even with marked
regional disparities, these effects can wipe out a large share of the aggregate gains to
the poor.*
A deeper analysis of why poor areas exist could have a number of implications
for all such geographically targeted development policies and projects. The case for
poor-area interventions depends on precisely why we observe poor areas. Here it can
be important to know just how much of the poverty one sees is attributable to area-
specific attributes versus personal attributes which may best be dealt with through a-
spatial programs. If poor areas arise from residential differentiation through mobility
(as the individualistic model would suggest), then such mobility will clearly limit the
scope for targeting on the basis of where people live. On the other hand, if place of
residence does matter even when one controls for personal characteristics then the
case for area-based programs and investments could be greatly strengthened.
Past assessments of geographic targeting have been essentially static, though
the limitations of this view have been noted by Ravallion (1993). Yet without suc-
cessful intervention, the competitive equilibrium in the geographic-capital model
would be unlikely to achieve a Pareto optimum, given the pervasive externalities.
The geographic model could thus imply dynamic efficiency gains from investing in
geographic capital; that will depend on the way in which geographic capital affects
the marginal returns to investment in private capital. If borne out by the evidence,
the geographic model may thus lead one to question any presumption that targeting
poor areas would necessarily have an aggregate growth cost.
A deeper understanding of why we see poor areas is also needed to inform
choices about the specific types of poor-area programs needed. For example, a com-
mon debating point in formulating poor-area programs is the priority to be given to
basic health and education versus credit and physical infrastructure. Should policy
be focusing on education or should it be roads? Is a package on interventions called
for, as in the “Integrated Rural Development Programs”? This is also a matter of the
*For India, Datt and Ravallion (1993) consider the effects on poverty of pure (nondistortionary) trans-
fers among states, and between rural and urhan areas. They find that the qualitative etfect of reducing
regional/sectoral disparities in average living standards generally favors the poor. However, the quanti-
tative gains are small. For example, the elimination of regional disparities in the means while holding
intraregional inequalities constant, would yield only a small reduction in the proportion of persons below
the poverty line, from an initial 33% to 32%. Also see Ravallion (1993) for Indonesia.
POORAREAS 71
nature and extent of the interaction efSects among area characteristics as they affect
living standards and their evolution over time.
There are also potentially important implications for economic evaluations
of the dynamic gains from area-specific interventions to reduce poverty. With lit-
tle mobility, living in a designated poor area can be taken as exogenous to house-
hold choices. However, the existence of spatial externalities may well entail that the
growthpath of future household living standards is dependent on the same area char-
acteristics which influence the public decision to declare the community poor. The
problem is essentially one of omitted-variable bias when there is state dependence in
the growth process. For example, a low endowment of local public goods may simul-
taneously induce a lower rate of growth and a higher probability of the community
being declared poor. Unless this is accounted for, the value to households of living in
an area which is targeted under a poor-area program will be underestimated (Jalan
and Ravallion 1996).
Assessing the case for all such interventions requires a deeper understanding
of how poor areas came to exist. That understanding can also inform other areas of
policy. The case for proactive migration policies may be strengthened if one finds that
there are strong geographic factors in the creation and perpetuation of poor areas.
That will depend in part on the precise nature of those factors. If the geographic effect
is largely explicable in terms of physical infrastructure endowments (as at least the
proximate cause), then the case for migration policies will be strengthened; migrants
who go to better-endowed areas will gain, and those left behind will also gain if there
is less crowding of the existing infrastructure in the poor area. But if the geographic
factors are largely social (to do with social capital, or the spillover effects of local
endowments of human capital), then the migration policy may make matters even
worse for those left behind. (This is often said about the effects of suburbanization
or inner-city areas in the United States.)
Motivated by the above discussion, the following sections will discuss various
approaches to empirical modeling which might prove fruitful in understanding the
economic geography of poverty so as to inform these difficult policy issues.
II. STATIC MICROMODELS AND DYNAMIC

AGGREGATE MODELS
A. Static Poverty Profiles
Standard empirical models of poverty are estimated on single cross-sectional sample
surveys. To illustrate, suppose we have a single cross-sectioned sample of house-
holds giving consumption C;, for household i at date t . If one regresses this (or its
log) against a set of location dummy variables d l , d,, . . . , d, for rn regions, then
one will retrieve an estimate of what can be termed the “unconditional geographic
poverty profile”:
72 RAVALLION
where a and Pj, are parameters to be estimated and uit is an unobserved error term.
If we were to repeat this for other years, then we could also see how the geographic
poverty profile changes over time.
This type of geographic poverty profile is found in (for example) almost every
poverty profile found in the World Bank’s Country Poverty Assessments (the regres-
sion may not be run, but in this case a bivariate cross-tab of the poverty measure by
region is just another way of running the regression). While useful for some purposes
(including geographic targeting), this type of poverty profile tells us nothing about
why there is more poverty in one place than another. It may be, for example, that
people with little education tend to live in certain places. Then if one controlled for
education, the regional dummy variables would become insignificant.
Extending this logic, it is becoming common practice to estimate more com-
plex multivariate models in which a set of household characteristics are added, rep-
resented by the vector xi, giving the augmented regression
If the geographic dummy variables remain significant in this augmented regression,

and one has controlled for all relevant household characteristics, then we can con-
clude that there are location-specific factors at work independently of personal char-
acteristics. Of course, the personal characteristics may well also be a function of
where one lives, and if one “solves out” this effect one would be back to (1). The
interest in (2) is in testing if there are locational effects which appear to have little
or nothing to do with private endowments; people in some area may tend to be less
well educated (a personal characteristic) or (because of poor local infrastructure, for
example) they may be unable to obtain a good return to their education, even when
they are as well educated as people elsewhere. The aim in estimating (2) is to identify
the latter effect (Ravallion and Wodon 1997).
If one also has community-level data, giving the stocks of physical and human
infrastructure and locations of any area-based interventions, then one can use these
data to try to explain the (conditional and unconditional) regional poverty maps im-
plied by the above regressions. This can either be done using a two-step estimator
or (probably more efficiently) in one step, by replacing the area-dummy variables in
(1)and (2) by geographic characteristics to give
The specification in (2) and (3) imposes additive separability between the
regional effects and household-level effects. However, this can be readily relaxed
POORAREAS 73
by adding interaction effects (so, for example, returns to education may depend on
where one lives). All parameters can be allowed to vary locationally by estimating a
separate regression on household characteristics for each region.
If one has access to repeated cross-section samples representing the same pop-
ulation then these static models can be repeated to see how the identifiable regional
effects (both conditional and unconditional) have evolved over time. Does one find,
for example, that poor regions are catching up over time, or are they diverging? This
can be addressed by studying how the (conditional and unconditional) data-specific
coefficients on the area dummies evolve.
So far the discussion has focused on a single welfare indicator. But this is
almost surely too restrictive. More generally one can postulate a set of indicators
aiming to capture both “income” and “non-income” dimensions of well-being. In
addition to consumption of market goods and services one could include indicators
of attainments in terms of basic capabilities, such as being healthy and well nour-
ished. The aim here is not to make a long and overlapping list of such indicators but to
capture the aspects of welfare that may not be convincingly captured by consumption
or income as conventionally defined (Ravallion 1996).So indicators of child nutri-
tional status or morbidity would be compelling since conventional household-level
aggregates may be weak indicators of distribution within households.
B. Interpreting Static Micromodels

The interpretation of these regressions is often difficult, given that it is often unclear
how one would motivate them from economic theory. Under special conditions (per-
fect credit markets, rational expectations, quadratic instantaneous utility function)
one can interpret these equations as models of permanent income, in which case the
right-hand-side variables are the various lifetime assets and their rates of return, or
determinants of these. More generally there is a more complex intertemporal model,
probably involving liquidity constraints such that current income and holdings of
liquid assets are the key explanatory variables one should be looking to account for
with whatever data can be observed in the current cross-sectional survey. Fuzziness
in the link ftom observed household and geographic data to the relevant assets adds
to the difficulties of interpretation. This discussion will focus on some of the special
problems posed by the geographic variables.
Insignificant regression coefficients on the regional dummy variables (/3;t s in
(1)and (2)) would suggest that area characteristics have no independent effect on
living standards, either because there is free mobility-so that households with the
same x can achieve the same standard of living everywhere-or there is no mobility,
but area characteristics are fundamentally irrelevant to welfare. Provided one has
fully captured the relevant household characteristics, significant values for the p J t s ,
on the other hand, are inconsistent with free mobility. If, however, there are impor-
tant omitted household characteristics-a possibility that one would be unwise to
74 RAVAUION
dismiss-and the equilibrium under free mobility entails residential differentiation

according to those characteristics, then this could produce significant p j l s .
The interpretation of the coefficients on area dummy variables or area charac-
teristics as “pure” geographic effects thus depends critically on having a complete
set of the x variables to control for household characteristics. The geographic dum-
mies might be picking up an omitted a-spatial household characteristic in a resi-
dentially differentiated location equilibrium. As always, it is a matter of judgment
to what extent omitted variable bias of this sort is a problem. However, this con-
cern does speak to the need for rich integrated microdata sets in which a very wide
range of data on individual and household characteristics are collected for the same
sampled households.
Consistent estimation by OLS requires that the regional dummies and the
household characteristics are uncorrelated with the error term. Another source of
bias is migration. There are ways of dealing with this, such as by estimating a switch-
ing regression which determines which region or sector the household is located. The
estimated probabilities of being in a given region or sector can then be used to cor-
rect for selectivity bias (Maddala 1986 surveys this class of econometric models). As
always, the problem of identification will arise in that the same variables determin-
ing levels of living at a given location will presumably also influence location choice.
A regional switching regression could also be plagued by endogeneity problems of
its own; is being better educated the cause or the effect of living in urban areas, say,
where schools are better and more accessible? This is a case where the econometric
cure could be worse than the disease.
The importance of these considerations should be judged on a case-by-case
basis. The extent to which endogeneity of household location is a serious concern
in many developing-country applications is unclear. As a stylized fact, the cost of
moving the whole household can be considerable; a rural farm-household typically
does not abandon its land to move in its entirety to urban areas (say) but exports
surplus workers, who retain the right (or even obligation) to return. Intrarural mi-
gration of whole farm-households also appears to be uncommon. In many settings it
may be plausible to identify locational effects for a (possibly large) subset of house-
holds, such as farm households but not others (relatively mobile landless laborers
and urban workers, for example.)
C. Estimating Static Micromodels in the Form of a

“Poverty Regression”
Static “poverty regressions” have become a standard tool in poverty analysis. The
most common approach assumes that the poverty measure is the headcount index,
given by probability of living below the poverty line. One postulates that real con-
sumption or income Ci is a function of a (column) vector of observed household char-
+
acteristics xi, namely Ci = px; E ; , where p is a (row) vector of parameters and ii
POORAREAS 75
is an error term; this can be termed the “levels regression.” A now common method
in poverty analysis is not to estimate the levels regression but to define the binary
variable h; = 1 if z 2 C ; , and hi = 0 otherwise. The method then pretends not to
observe the yis, acting as if only hi and the vector of characteristics x; is observed.
The probability that a household will be poor is P = Prob[C < z I x] = Prob[& <
t - 8x1 = F ( t - px), where F is the cumulative distribution function specified
for the residuals in the levels regression. A probit or logit is then usually estimated,
depending on the assumption one makes about the distribution of the error term &i.*
(One could also use a semiparametric estimator which allows the distribution of the
error to be data determined.) One can also generalize this procedure to other (“higher
order”) poverty measures and estimate censored regression models.
However, this common practice is difficult to defend since-unlike the usual
binary response model-here the “latent” variable is fully observed. So there is
no need for a binary response estimator if one wants to test impacts on poverty of
household characteristics. The parameters of interest can be estimated directly by
regressing Ci on xi. The relevant information is already contained in the levels re-
gression which is consistently estimable under weaker assumptions about the errors.
Measurement errors at extreme Cs may prompt the use of probits, though there are
almost certainly better ways of dealing with such problems, which do not entail the
same loss of information, such as by using more robust estimation methods for the
levels regression.
Nor is the “poverty regression” method necessary if one is interested in cal-
culating poverty measures conditional on certain household characteristics. Subject
to data availability, Prob[C < z I x] can be estimated directly from sample data.
When the number of sampled households with a specific vector of characteristics of
interest, xl say, is too small to reliably estimate Prob[C < z 1 XI]from a subsample,
one can also turn to regression methods for out-of-sample predictions. But these pre-
dictions can also be retrieved from the levels regression, though one must then know
the distribution of the errors. (For example, if the errors are normally distributed with
zero mean and a variance o’,then the probability of being poor is F [ ( z - Bx)/a],
where F is standard normal.) There is nothing gained from using a binary-response
estimator, so the econometric sophistication of “probits” and so on buys us very little
in this case.
Poverty regressions may make more sense if one wants to test the stability of
the model for poverty across a range of potential poverty lines. Suppose, for example,
that one of the regressors is the price of food, and that very poor people tend to be net
*The earliest example that I know of is Bardhan (1984),who used a logit regression of the probability of a
household being poor against a range of household and community characteristics using sample survey
data for rural West Bengal. Other examples include Gaiha (1988), Grootaert (1994),Foley (1995),and
World Bank (1995).
76 RAVALUON
consumers of food, while those who are somewhat better off tend to be net producers.
Then the distributional shift with a change in the price of food will not entail first-
order dominance, as assumed by the standard levels regression. Instead one might
want to specify a set of regression functions the parameters of which vary according
to the segment of the distribution one is considering. One way of estimating such
a model is by assuming that the segment-specific error terms are of the logit form,
entailing a multinomial logit model (Diamond et al. 1990).
All the above models are essentially static; some welfare indicator at a single
date is modeled as a function of a range of individual and geographic data. Such mod-
els cannot distinguish effects on the growth rates of consumption (or other welfare
+
metric) from effects on its level. The true model could be Cj, = (j3 y t ) x i E ; ~ im-
, +
plying that x influences the growth rate of consumption as well as its level. However,
static data cannot distinguish B from y .
D. Aggregate Models
Another increasingly common tool of analysis is a regional (or country)-level time-

series model. Here the dynamics can be readily introduced, by allowing for the ef-
fects of past outcomes and other variables on current outcomes. Among the models
that might be postulated, regional growth regressions are being seen increasingly;
by this approach, the growth rate over time for the region as a whole is modeled
as a function of initial conditions, exogenous shocks and policy changes over the
period.*
These aggregate dynamic models also have their limitations for understanding
the economic geography of poverty. Clearly the distributional analysis of policy and
other impacts on living standards which is impossible with only aggregate data.?
Working from the micromodel allows one to better understand the distributional im-
plications of the aggregate growth process.
When attempting to assess the welfare impacts of area-based interventions,
aggregate models can also be particularly vulnerable to bias arising from endogene-
ity of program placement. The political decision on program placement may itself be
a function of observed aggregate poverty levels. Knowledge of how programs were
assigned and the history of area characteristics can help avoid this problem (as in
Pitt et al. 1995).The problem is not of course confined to aggregate models. With mi-
cro (household-level) data it may be plausible that program placement is exogenous
to the extent that no individual household has much influence on the placement de-
*See Barro and Sala-i-Martin (1995) for an overview of'these models, with developed country applications;
-L
examples for developing countries include Cashin and Sahay (1995)and Datt and Ravallion (1996).
1 Datt and Ravallion (19%) use a long time series of' repeated cross-sectional surveys (for India) to relax
this attribute of standard growth models.
POORAREAS 77
cision. However, bias can still arise from either migration (discussed above) or omit-
ted variables which simultaneously influence household-level welfare outcomes and
program placement.
Another disadvantage of aggregate models is that they do not allow one to dis-
tinguish internal and external effects on production and welfare. This can matter to
policy. By using the household as the unit of observation one can identify external
effects of geographic capital, including local public goods, on production processes
at household level (Ravallion and Jalan 1996). Consider Figure 1, based on aggre-
gate (county-level) data for China. Such evidence tells us nothing per se about the
spatial effects in a growth process. The highly aggregated form of such data does not
allQw one to distinguish two possible ways in which initial conditions may influence
the growth process at the microlevel. One way is through effects of individual condi-
tions on the individual growth process, and this is a common interpretation given to
nonzero values of the regression coefficient on initial income in a growth regression;
declining marginal product of capital would suggest a tendency for convergence; by
this interpretation, the type of divergence depicted in Figure 1 suggests increasing
returns toprivate capital. If this is right, then regional divergence, and the existence
of persistently poor areas, is to be expected when the rate of growth is at its maximum.
Conversely, under these conditions, governmental attempts to shift the allocation of
investment in favor of poor areas will entail a growth cost, though a policymaker
may still be willing to pay that cost to achieve a more balanced (and possibly more
sustainable) growth path.
But there is another way in which the divergence in Figure 1 can arise, even
with declining marginal products with respect to own capital at the microlevel and
constant returns to scale in private inputs. The microgrowth process might be driven
by intraregional externalities; individual growth prospects may be better in an ini-
tially better-off region through positive local spillover effects. Quite generally, the
marginal product of capital will depend on area characteristics. (Only with rather
special separability assumptions will this not be true.) There may well be declining
marginal products with respect to “own capital” but increasing marginal products to
geographic capital. Indeed, if there is constant returns to scale in the private inputs,
and geographic capital is productive, then there must be increasing returns overall.*
That may well be why we see the aggregate divergence in Figure 1,with the external
effect dominating. But the aggregation hides the difference. If in fact the regional
divergence is really due to the external effect of differences in geographic capital,
*Let output be F(K, G ) ,where K is a vector of private inputs (“own capital”) and G is a vector of pub-
lic inputs (“geographic capital”), and consider any A > 1. By constant returns to K , F(AK, AG) =
AF(K, AG) > AF(K, G ) , since geographic capital has a positive marginal procinct. Thus I;‘ exhibits
increasing returns to scale overall.
78 RAVALLION
then successful interventions to reduce the inequality in geographic capital need not
entail any cost to the overall rate of growth.
This second “external” channel through which area characteristics can alter a
growth process has received relatively little attention in empirical work on the deter-
minants of economic growth, though the possibility has been recognized in some of
the theoretical literature (notably Romer 1986 and Lucas 1988). The reason why this
external channel has been relatively neglected in growth empirics is undoubtedly
that the level of aggregation in past work has meant that-even if one was aware of
the possibility-the genuinely spatial effects of intraregional spillover effects could
not possibly be identified empirically.
To encompass both the “internal” (individualistic) and “external” (geographic)
channels through which initial conditions can affect a growth process one needs
to model that process at the microlevel. The growth rate for each household will
be a function of both its own initial conditions, characteristics of the area in which
the household lives, and external shocks during the period. The areawide growth
relationships (such as depicted in Figure 1) can then be interpreted as (approximate)
averages formed over the underlying microgrowth processes; but in the averaging one
loses the ability to distinguish the internal from the external effects (Ravallion and
Jalan 1996).
The recurrent problem in aggregate models is that the economic theory which
motivates them is typically a microeconomic model. So tests using aggregate data
always beg the question of whether one is testing the micromodel or the aggregation
assumptions.
111. THEORY AND ESTIMATION FOR DYNAMIC

MICROMODELS
A. Theoretical Foundations for a Microempirical Model of Growth
A theoretical model capable of motivating an empirical analysis of a number of the is-

sues raised above can be formulated by a reinterpretation of the Romer (1986)model
of endogenous growth under increasing returns to scale.* Analogously to the dis-
tinction between firm-specific knowledge and economy-wide knowledge in Romer’s
model, one can conjecture that output of the farm household is a concave function
*A number of versions of the classic Ramsey model-in which an intertemporal utility integral is max-
imized subject to flow constraints and product ion functions-have been proposed which can yield a
nonzero solution for the rate of consumption growth which will be a function of initial human and phys-
ical assets as well as preference and production parameters. For surveys of the theories of endogenous
growth see Grossman and Helpman (1991), Hammond and Rodriguez-Clare (1993),and Barro and Sala-
i-Martin (1995).
POORAREAS 79
of various privately provided inputs, but that output also depends positively on the
level of geographic capital. On fully accounting for all private inputs (all profits being
reckoned as payments for those inputs), there will then be constant returns to scale to
the privately provided inputs, but increasing returns to scale over all inputs, includ-
ing geographic capital. With the farm-household maximizing an intertemporal utility
sum-with instantaneous utility depending on current consumption, which must be
partly forgone to ensure future output-one can derive an endogenous consump-
tion growth rate which depends on the initial endowments of both private capital
and geographic capital. With this reinterpretation, the results on existence and wel-
fare properties of equilibrium in Romer (1986)model can be applied to the present
problem.
The key intertemporal equilibrium condition from such a model equates the
intertemporal marginal rate of substitution with the marginal product of “own capi-
tal,” which is a decreasing function of the initial endowment of own capital and in-
creasing in the amount of geographic capital, taken as exogenous at the microlevel.
With appropriate functional forms, the farm-household’s consumption growth rate
over any period is then a decreasing function of its endowment of private capital and
an increasing function of the level of geographic capital.
Past growth empirics have relied on country or regional aggregates. The trans-
lation of this approach to the microlevel is straightforward; one is simply undoing
the aggregation conditions used to go from the microgrowth theory to the aggregate
regional or country data. The translation is even more straightforward when it is
noted that many of the households in the world’s poor areas are farm households
who jointly produce and consume, rather than economies in which separate con-
sumers and producers interact through trade. But this is largely a matter of inter-
pretation; the separation of an economy into households (which consume) and firms
(which produce) is not essential in theoretical growth models.* In the present setting,
the farm-household can be thought of as a small open economy, trading with those
around it.
It should also be recognized that a poor area may have become poor due to a
location-specific transient shock (a local drought, for example). There may also be
lags in the growth process of consumption. By explicitly modeling these features of
the data generation process, it should be possible to identify longer-term impacts in
panels of sufficient length. (Averaging prior to estimation is not an efficient way of
dealing with these data features.)
Motivated by the above considerations, an empirical approach can be sug-
gested which entails consistently estimating a dynamic model of consumption growth
*Standard endogenous growth models postulate separate households and firms, but an equivalent formu-
lation is possible in which households both consume and produce (Barro and Sala-i-Martin 1995).
80 RAVALLION
at the household level using panel data. The model allows one to test the dynamic im-
pact over the length of the panel of a wide range of initial conditions at both household
and community levels. The proposed approach differs from the usual “fixed-effects”
method. While it is common to model the variables of interest in first-difference form,
or as deviations from their time means, this is typically done in the context of a static
model in levels, for which the time slope is a constant and there are unobserved fixed
effects. Clearly such a formulation is of little interest here since it does not allow ini-
tial conditions-including area-specific policies and projects-to affect the growth
path of the variable of interest.
B. Toward an Estimable Model

Unlike standard single cross-sectional sample surveys of households, here one needs
panel data in which the same households are observed over time. (Later I will dis-
cuss possible approaches using repeated cross-sectional surveys.) Such data sets
are, however, becoming more common, and panels of varying lengths are available
for a number of developing and developed countries. Notice also that estimating a
microgrowth model will require less data in the second survey than the first. At a
minimum, a second reading of the household consumption or income level will do,
though information on relevant demographic or other “shocks” would be desirable.
So this approach is already feasible and will probably become more widely applica-
ble in the future.
To see how an empirical model capable of addressing the questions posed here
can be constructed, let us assume (following the discussion above) that the long-run
household-level consumption growth rate is determined in part by a vector of exoge-
nous initial conditions, comprising both internal (within the household) and external
(community) endowments of physical and human assets, as well as any exogenous
household characteristics influencing the discount rate, liquidity constraints, tastes,
and production functions. The importance of the internal factors can arise from the
dependence of equilibrium growth rates on the initial human and physical capital
stocks in household-operated production processes. The external effects can arise
from the existence of local public goods or differing agroclimatic conditions. The di-
rection of the effects of all such initial conditions are difficult to predict on a priori
grounds, and will depend on the nature of the technology (for example, if there are
increasing returns then divergent effects are possible whereby higher initial wealth
can result in higher future growth), how well markets work (credit market imperfec-
tions, for example, can entail that liquidity-constrained households cannot realize
the same growth potential as others) and the political economy of local public policy
(it has been argued that higher initial inequality, for example, may promote policy
choices which inhibit growth).
In carrying such a formulation to data, it would be unrealistic to assume that
the growth rate actually observed at any date is the steady-state value as implied by
POORAREAS 81
a standard growth-theoretic model. One would want to allow for deviations from the
underlying steady-state solution, due to shocks and/or adjustment costs. It is thus
better to postulate an autoregressive distributed lag structure for the growth process,
augmented by exogenous shocks and unobserved effects. This will permit a more
powerful test of the impact of initial conditions on the evolution of living standards
than is possible by only modeling the long-run average growth rates.
Thus one can postulate an econometric model of the growth rate in living stan-
dards at the household level as a function of ( 1 ) initial conditions at the household
level, (2) initial conditions at the local community level, and (3) exogenous time-
varying factors (“shocks”) at both levels. The variance in household-level growth
rates due to the second set of variables could in principle be “explained” by a com-
plete set of area-dummy variables. However, by collating the micro (household-level)
data with geographic data bases on agroclimatic variables, and stocks of physical
and social infrastructure, it will be possible to obtain a far more illuminating speci-
fication in which specific attributes of the local area enter explicitly. As a check for
omitted-variable bias, one can then compare the results with a model in which the
geographic variation is picked up entirely by dummy variables. This may also sug-
gest idiosyncratic regional effects, such as due to local political factors, that might
be best studied on an ad hoc (case study) basis.
On introducing dynamics and both time-invariant and time-varying unobserved
effects, a suitable dynamic model could take the form
for household i (=I 1, . . . , N ) at date t (= 1 , . . . , T ) , where C;, is consumption

by i at date t , xi, is a 1 x k vector of time-varying explanatory variables, z; is a p -
dimensional vector of initial conditions, and qi is a time-invariant household-level
fixed effect. The vector z; comprises both area-specific factors (such as initial values
of indicators of physical and social infrastructure ) and household-specific charac-
teristics (such as age of the head of the household and education levels). Both x;,
and z; include interaction effects (including between individual and areas character-
istics).*
The problem of estimating this model is different from the usual “within” es-
timator for panel data. It is known that the ordinary least-squares estimator of an
autoregressive fixed effects model is not consistent for a typical panel where the
number of periods is small and where the asymptotics are driven by the number of
cross sections going to infinity (Hsiao 1986).The inconsistency arises because of the
potential correlation between the lagged endogenous variables and the residuals in
the transformed model.
*One might also hypothesize that area-mean consumption entprs (4), but this effect is generally not iden-
tifiable; see Manski (1993).
82 RAVALLION
Thus consistent estimation of the above model does present a more difficult
problem than either the static micro- or dynamic macromodels. But a solution is
available (Jalan and Ravallion 1996). First notice that the error term in (4) has two
components: an unobserved individual specific time-invariant fixed effect, qi, and
the standard innovation error term, UiL. Let us assume that the unobserved individual-
specific effect q; is correlated with the regressors, i.e., E ( q i z ; ) , E ( q i , xit), and
E(qiCil-l) are nonzero.* The error uit is however serially uncorrelated and thus
satisfies the orthogonality conditions:
These conditions ensure that suitably lagged values of C;, and xit can be used as
instruments. In order to get consistent estimators, the unobserved fixed effects qi
need to be eliminated. This can be done by taking the first differences of (4) to obtain
the transformed “growth model”:?
There are various options for estimating such a model. GMM methods appear
to offer the best approach (Arellano and Bond 1991). Given that the uits are serially
uncorrelated, the GMM estimator is the most efficient one within the class of in-
strumental variable (IV) estimators. In estimating (6),Cit-2 or higher lagged values
(wherever feasible) are valid instrumental variables. Heteroscedasticity-consistent
standard errors can be computed using the residuals from a first-stage regression to
correct for any kind of general heteroscedasticity. Inferences on the estimated pa-
rameter vector are appropriate provided the moment conditions used are valid. Tests
for overidentifying restrictions can be implemented to test the null hypothesis that
the instruments are optimal (i.e., the instruments and the error term are orthogo-
nal); see Sargan (1958,1988) and Hansen (1982). In addition, a second-order serial
correlation test (the test statistic will be normally distributed) can be constructed
given that the consistency of the GMM estimators for the first-differenced model de-
pends on the assumption that E(Au;, A u ; , - ~ = ) 0.z Tests for spatial correlation
in the errors-arising from omitted geographic effects-can also be performed (fol-
lowing Frees 1995), though they will need to be adapted to the present problem.
*Bhargava and Sargan (1983)offer a dynamic random-effects model where it is assumed that some of the
regressors are uncorrelated with the unobserved individual specific effect.
Various transformations can he used to eliminate the nuisance parameters, though the estimation pro-
cedures used are similar to the one proposed here.
*There may be some first-order serial correlation; i.e., E(Aui, A u i l - l ) may not he equal to zero since
Auit are the first differences of serially uncorrelated errors. Alternatively, if uil is a random walk, then
there should not be any serial correlation in the first differenced Aujl.
POORAREAS 83
If corrective action is called for, then one can try introducing more geographic data,
or more geographic structure to the error process.*
C. Further Specification Issues

Questions are often asked about the prioritization of physical versus social infra-
structure development. Should one even assign a priority, or is a balanced “inte-
grated” approach needed? Does the answer depend on the “stage” of development?
It has often been argued that it is the combination of certain physical and/or human
infrastructure endowments that matters. One at a time they may not help much. The
returns to irrigation for example may depend on education (as van de Walle 1995
found for Vietnam). There are other possible interaction effects such as the possibil-
ity that greater ethnic cohesion in an area increases the chances of cooperation and
hence the returns to investments in geographic capital.
In principle, all the explanatory variables in the models described above could
enter the model in highly nonlinear ways, and this should be tested. However, the
above considerations suggest that identifying interaction effects could be of special
interest in this context.
Another issue concerns nobility and the possible endogeneity of a household’s
area characteristics. The plausibility of a free-mobility equilibrium in the settings
considered here is questionable; even with no governmental restrictions on mobility,
migration within India over a long period has responded little to regional disparities
(Datt and Ravallion 1996). Nonetheless, the existence of even limited mobility raises
questions about the possible endogeneity of area characteristics in the microgrowth
process. How might one test for effects of area characteristics on the spatial distri-
bution of the population?
The significance of area fixed effects in the levels of living standards, after con-
trolling for a-spatial household characteristics, would be suggestive that mobility is
imperfect; if there was free mobility then any two households with the same personal
(mobile) characteristics should be able to achieve the same level of welfare. This is
directly testable.
There are other tests that may also help. A theoretically consistent and em-
pirically tractable approach to introducing mobility, allowing for adjustment costs,
can be proposed following the approach outlined in Ravallion (1982, 1984) in the
related context of local public finance in which (under certain conditions) mobility
can reveal preferences over local public goods. This entails modeling the spatial
distribution of specific population subgroups as functions of, inter alia, area charac-
teristics, allowing for dynamic effects in the adjustment process to a free-mobility
*Potential approaches include Froot (1989) and Conley (19%);also see the special issue of Regional
Science and Urban Economics, September 1992, on “Space and Applied Econometrics.”
84 RAVALLION
equilibrium. In both estimating the poverty regressions (with “social externalities”)

and in deriving aggregate welfare (including poverty) impacts, this model could then
be used to endogenize the population shares.
The dynamic model estimated on panel data suggests a further test. One of
the main sources of attrition in a panel is outmigration. Thus standard tests for attri-
tion bias and corrective actions (see, for example, Hsiao 1986, Chap. 8)can also be
interpreted as a means of dealing with migration responses to area characteristics.
Household mobility is not the only way in which endogeneity of area-based in-
terventions can bias results from the types of models described above. Another prob-
lem is that program placement may be a function of variables which influence welfare
impacts (Pitt et al. 1995). A version of this problem arises in the dynamic micromod-
els described above when program placement is determined by initial conditions
which also influence the future evolution of living standards (Jalan and Ravallion
1996). The best solution to this problem is to find out what area characteristics in-
fluenced program placement and include those characteristics as explanatory vari-
ables. Since program placement must have been a function of observable area at-
tributes suitable controls should be available in practice.
D. Alternative Methods Using Repeated Cross Sections

The attractions of panel data for the types of analysis described above are clear.
Though more of these data sets are emerging, panel data sets are still far less common
than repeated cross-sectional surveys. Is there any way of estimating spatial effects
on the microgrowth process using repeated cross sections? If the answer is yes, then
this would open up a wide range of potential applications in setting in which panel
data are unavailable.
One approach is by the analysis of demographic cohorts.* For each wave of
a set of cross-sectional surveys one can calculate mean household consumption or
income for persons in a given cohort defined by initial age and (in this case) place
of residence. One can also do this for other household characteristics. One can then
construct a model that looks like an individual-level model but is for cohorts. In
effect, one takes cohort averages of the household-level model. Thus one can still
identify the internal (cohort averages of household characteristics) and externaZ (area
characteristics) effects on the evolution of the poverty indicator.
However, there may well be a better approach. We want to see how (say) con-
sumption growth from time t to t +
1 is affected by time t characteristics of the
household and its area of residence. With panel data we simply regress the change
+
in (log) consumption from t to t 1 on (inter alia) household is Characteristics at t .
*This approach has showed promise in research on other topics, such as intertemporal coiisumption be-
havior and inequality (Deaton and Paxson 1994).
POORAREAS 85
+
With cross sections we do not know consumption at t 1. But if we know the future
values of one or more predictors of consumption then these can be used as instru-
ments. One first models time t + 1 consumption as a function of variables observed
+
in time t 1 but also at time t . Then one uses that model to predict the consumption
at time t+ 1 of each household in the time t sample and, hence, estimate its rate
+
of consumption growth from t to t 1. This can then be regressed on the individual
and area characteristics at time t .
Many cross-sectional surveys do obtain information about likely future char-
acteristics which can be used as instruments. For example, the Rural Household
Surveys for China collect both beginning and end-of-year data on financial and phys-
ical wealth in each round. So the end-of-year data can be used to predict the next
period’s consumption along with other time-invariant variables. There are other po-
tential instruments; the next period’s demographic composition (number of persons
by age groups) of the household can be predicted from the current period’s composi-
tion. R2 will be lower, but consistent estimates should still be possible under regular
conditions. Estimators are available for dynamic models of this sort using repeated
cross sections (see Moffitt 1993 and references therein). The performance of these
methods could be studied using the panel data, but treating it as repeated cross sec-
tions; it would also be of interest to try the method out on the original samples (prior
to panel construction) so as to assess effects of panel attrition.
IV. CONCLUSIONS AND POTENTIAL LESSONS

FOR POLICY
When confronted with the reality of extreme poverty in remote rural areas with poor
natural resources, observers often ask: “Why don’t these miserably poor people just
move out?” Those who claim that outmigration is the answer often also argue against
public investment in these areas. “This will just be at the expense of more profitable
investments elsewhere” the argument goes.
This chapter has questioned this reasoning, but it certainly has not refuted
it. That will be a matter of future empirical research. But some points can be made
now. One is that we should understand the nature of the incentives and constraints on
outmigration from these areas. You need some money to start up elsewhere, you need
some basic skills, and you need information. All are generally lacking, but more so
for some people than others. The reasons they are lacking can be traced to market
imperfections of one sort or another and how these interact with poverty. The outside
non-farm-labor-market options are typically thin or nonexistent for someone who
is illiterate, reflecting a lack of substitution possibilities with moderately educated
labor in even quite labor-intensive manufacturing. Credit market failures mean that
there is little chance of borrowing to finance the move. There is highly imperfect
information about prospects elsewhere, and sizable uninsured risk.
86 RAVALLION
The process of outmigration may be a mixed blessing for a poor area at least
initially. Those who have the money, skills, and information will naturally tend to
be the relatively better off. Their departure is likely to put upward pressure on the
incidence of poverty in the poor area. This comes about in various ways. As a purely
statistical proposition, most measures of poverty will rise when the nonpoor leave.
But there are more subtle dynamic effects through “ghettoization.” The local skill
base is likely to have external effects on the local growth process. It follows that
the outmigration of the better educated workers entails an erosion of local resource
base with adverse longer-term growth consequences. Results from research on poor
areas of southwest China have suggested that there exist strong external effects of
physical and human infrastructure on the returns to private investment and (hence)
the prospects of escaping poverty there (Jalan and Ravallion 1997). These effects
will be mitigated to some extent by remittances and reduced pressure on the land.
All this suggest that one of the best ways that government can help is by in-
vesting in the schooling, health, and nutrition of the children of the poor in these
areas. Public assistance with credit (to cover search costs for poor outmigrants), and
information will complete the package.
Should we also be investing in the land and physical capital of these areas?
What should be the balance between those investments and human resource de-
velopment? Here one could proceed on an ad hoc basis; if the investment passes a
standard (distribution-unweighted) cost-benefit test, then it should be done. But the
“anti-investment” argument would maintain that private capital flows would already
have found such opportunities.
One response is that, unless there is perfect factor mobility (which nobody
seems to consider plausible), there may still be an equity case for such investments
up to some point. Then they are part of a redistributive policy, exploiting the possi-
bilities for geographic targeting (Lipton and Ravallion 1995).That is fine. However,
for the same reasons that there may be too little outmigration, there may also be too
little investment in these areas from an efficiency point of view as well. Credit mar-
ket imperfections can entail that there are unexploited opportunities for investing in
the land and physical capital of these areas. The liquidity constraints that make it
hard to finance outmigration will also make it hard to finance otherwise profitable
local investments. And asymmetric information and supervision costs deter outside
investors.
The argument that investing in poor areas would entail lower overall growth in
the economy also breaks down as soon as one introduces local public goods and other
forms of “geographic capital” into the analysis, i.e., goods which cannot be supplied
efficiently by markets and which alter the rate of return to private investment. Poor
rural infrastructure in these areas could then be the underlying reason for low private
investment; better infrastructure would then encourage private capital inflow.
However, much of this is conjecture, based on little more than casual obser-
vations and common sense. The chapter has suggested some econometric methods
POORAREAS 87
which might be used to address these issues more rigorously. But is it an ambitious
research agenda, both in terms of the types of data needed, and the level of economet-
ric sophistication needed to convincingly disentangle these effects. So it is important
to ask: What will we have learned from such research that can reliably inform the
above policy choices about poor areas? Three types of potential lessons for policy
can be identified.
The first set of policy lessons concern the economic case for area-based in-
terventions. Should such interventions only be viewed as a specific kind of redis-
tributive policy with probable costs to the overall rate of growth? This view derives
from a growth model in which the existence of persistently poor areas, and regional
divergence more generally, are traced to the natural resource endowments and tech-
nologies, notably the (claimed) existence of increasing returns to private production
inputs. Private investment flows to the areas with the highest returns which are also
(according to this model) the initially richer areas. This still begs a number of policy
questions. For example, we need to know more about what the best indicators are for
this type of redistributive policy when the aim is to help individuals escape poverty
in the future.
But maybe we will find that the empirical results from the type of research
proposed here will reject this model at a fundamental level in favor of one which
says, in effect, that poor areas and divergence reflect spatial inequalities in access
to credit, and publicly provided social and physical infrastructure, and have rather
little to do with increasing returns to private capital, residential differentiation, and
so on. That conclusion could well dramatically alter the policy dialogue on poor-
area interventions and shift the emphasis to the task of redressing these preventable
spatial inequalities. If that conclusion is borne out by the data, then such policies will
be good for growth and good for equity. Or the results may point to a more complex
and mixed picture, possibly with a degree of country, and even regional, specificity.
A second set of broad policy lessons stem from the fact that the approach pro-
posed here allows one to measure spatial externalities. This can throw light on, for
example, how much of the welfare gain from schooling is transmitted though the
internal effects on earnings and so on, and how much is external, arising from the
(presumably positive) neighborhood effects of better education. This will have impli-
cations for the priority one attaches to efforts at finely targeting education subsidies
and for the policy arguments often made about how much basic education needs to
be subsidized on the grounds of its external benefits.
A third set of policy implications will be more specific to the types of projects
that should be recommended for dealing with persistently poor areas. In the process
of addressing these broad questions, empirical models can include explanatory vari-
ables of more or less direct policy relevance. One set of such variables is the very
existence of poor-area interventions. Is the subsequent rate of growth in living stan-
dards of poor people higher when a poor-area program is in place than would other-
wise have been the case, controlling for both household and community-level initial
88 RAVALLION
conditions and time-varying exogenous shocks? What were the longer-term welfare
gains? How do they compare to the budgetary outlays on such programs? There will
be other explanatory variables of policy relevance, such as the initial stocks of var-
ious components of publicly provided social and physical infrastructure, for which
all of the same questions apply, though here of course it may not always be easy to
account fully for their historical costs (though costs of new facilities will often be
known). What priority should be attached to social services versus physical infra-
structure or credit, and how does this vary with other factors? This should allow a
deeper understanding of what the complementarities are among these various types
of publicly provided inputs; we may learn, for example, how much access to one type
of infrastructure alters returns to another, or how much poor agroclimatic conditions
affect returns to different types of publicly provided inputs.
This chapter has argued that the long-standing problem of lagging poor areas
in growing economies, and more generally the diversity in prospects of escaping
poverty that one finds, are explicable with the right empirical tools and data. This
offers hope for better informing a number of difficult public choices on appropriate
responses to poor areas.
ACKNOWLEDGMENTS
I have had many useful discussions on this topic with Jyotsna Jalan. For their help-
ful comments on an earlier draft, I am also grateful to the Handbook’s referee, and
to Hans Binswanger, Ken Chomitz, Klaus Deininger, Lionel Demery, Bill Easterly,
Paul Glewwe, Emmanuel Jimenez, Aart Kraay, Valerie Kozel, Peter Lanjouw, Andy
Mason, Branko Milanovic, Lant Pritchett, Martin Rama, Zmarak Shalizi, Lyn Squire,
Dominique van de Walle, Mike Walton, and Quentin Wodon.
REFERENCES
Arellano, M. and S. Bond (1991), Some Tests of Specification for Panel Data: Monte-Carlo
Evidence and An Application to Employment Equation, Review o j Economic Studies,
58,277-298.
Bardhan, P. K. (1984), Land, Labor and Rural Poverty: Essays in Development Economics,
Columbia University Press, New York.
Barro, R. and X. Sala-i-Martin (1995), Economic Growth, McGraw-Hill, New York.
Bhargava, A. and J. D. Sargan (1983), Estimating Dynamic Random Effects Models from Panel
Data Covering Short Time Periods, Econometrica, 51, 1635-1659.
Binswanger, H. S. R. Khandker, and M. Rosenzweig (1993), How Infrastructure and Financial
Institutions Affect Agricultural Output and Investment in India, Journal of Develop-
ment Economics, 41,337-366.
POORAREAS 89
Borjas, G. J. (1995), Ethnicity, Neighborhoods, and Human-Capital Externalities, American

Case, A. (1991), Spatial Patterns in Household Demand, Econometrica, 59, 953-65.
Case, A. and L. F. Katz (1991), Company You Keep: The Effects of Family and Neighbor-
hood on Disadvantaged Youths, Working Paper 3705, National Bureau of Economic
Research, Washington, D.C.
Cashin, P. and R. Sahay (199S), Internal Migration, Center-State Grants and Economic Growth
in the States of India, IMF Working Paper WP/95/66.
Chen, S. and M. Ravallion (1996), Data in Transition: Assessing Rural Living Standards in
Southern China, China Economic Review, 7,23-56.
Ciccone, A. and R. E. Hall (1996), Productivity and the Density of Economic Activity, Amer-
ican Economic Review, 86, 54-70.
Conley, T. (1996), Econometric Modelling of Cross-Sectional Dependence, mimeo, Northwest-
ern University.
Datcher, L. (1982), Effects of Community and Family Background on Achievement, Review of
Economics and Statistics, 64. 32-41.
Datt, G. and M. Ravallion (1993), Regional Disparities, Targeting, and Poverty in India, in M.
Lipton and J. van der Gaag (etls.), Iizcludirig the Poor, World Bank, Washington, D.C.
Datt, G. and M. Ravallion (1996), Why Have Some States of India Performed Better than
Others in Reducing Absolute Poverty?, Economicci, forthcoming.
Deaton, A. and C. Paxson (1994), Intertemporal Choice and Inequality, Journal of Political
Economy, 102,437-467.
Diamond, C. A., C. J. Simon, and J. T. Warner (1990), A Multinomial Probability Model of the
Size Distribution of Income. Journal oJ’Econometrics, 43, 43-61.
Easterly, W. and R. Levine (199S), Africa’s Growth Tragedy. A Retrospective, 1960-89, Policy
Research Working Paper 1503, World Bank, Washington, D.C.
Foley, M. C. (1995), Poverty in Russia: Static and Dynamic Analyses, in J. Klugman (ed.),
Poverty in Russici during the Transition,,World Bank, Washington, D.C.
Foster, A. D. and M. Rosenzweig (199S), Learning by Doing and Learning from Others: Hu-
man Capital and Technical Change in Agriculture, Journal of Political Economy, 103,
1176-1209.
Frees, E. (1995), Assessing Cross-Sectional Correlation in Panel Data, Journal of Economet-
rics, 69,393-414.
Froot, K. (1989), Consistent Covariance Matrix Estimation with Cross-Sectional Dependence
and Heteroskedasticity in Financial Data, Journal of Financial and Quantitative
Analysis.
Gaiha, R. (1988), On Measuring the Risk of Poverty in Rural India, in T. N. Srinivasan, and
P. K. Bardhan (eds.), Rurul Poverty in South Asia, Oxford University Press, Oxford.
Grootaert, C. (1994), The Determinants of Poverty in C6te d’Ivoire, mimeo, World Bank, Wash-
ington D.C.
Grossman, G. M. and E. Helpman (1991), Innovution and Growth in the Global Economy, MIT
Press, Cambridge, MA.
Hammond, P. J. and A. Rodriguez-Clare (1993), On Endogenizing Long-Run Growth, Scan-
dinavian Journal of Economics, 95, 391-425.
Hansen, L. P. (19829, Large Sample Properties of Generalized Method of Moments Estimators,
Econometrica, SO, 1029-1054.
90 RAVALLION
Hirschman, A. 0. (1958), The Strategy ofEconomic Development, Yale University Press, New
Haven, CT.
Hsiao, C. (1986),Analysis of Panel Data, Cambridge University Press, New York.
Jalan, J. and M. Ravallion (1996), Are There Dynamic Gains from a Poor Area Development
Program?, Journal of Public Economics, forthcoming.
Jalan, J. and M. Ravallion (1997). Spatial Poverty Traps? Policy Research Working Paper,
World Bank, Washington, D.C.
Jazairy, I., M. Alamgir, and T. Panuccio (1992), The State of World Rural Poverty: An Inquiry
into its Causes and Consequences, New York University Press for the International Fund
for Agricultural Development, New York.
Krugman, P. (1991), Geography and Trade, MIT Press, Cambridge MA.
Leading Group (1988), Outlines of Economic Development in China’s Poor Areas, Office of
the Leading Group of Economic Development in Poor Areas under the State Council,
Agricultural Publishing House, Beijing.
Lipton, M. (1995), and M. Ravallion, Poverty and Policy, in J. Behrman and T. N. Srinivasan
(eds.), Handbook of Development Economics, vol. 3, North-Holland, Amsterdam.
Lucas, R. E. (1988), On the Mechanics of Economic Development, Journal of Monetary Eco-
nomics, 12,3-42.
Maddala, G. S. (1986), Disequilibrium, Self-Selection, and Switching Models, in Z. Griliches,
and M. D. Intriligator (eds.), Handbook of Econometrics, North-Holland, Amsterdam.
Manski, C. F. (1993), Identification of Endogenous Social Effects: The Reflection Problem,
Review of Economic Studies, 60, 531442.
Moffitt, R. (1993), Identification and Estimation of Dynamic Models with a Time Series of
Repeated Cross-Sections, Journal of Econometrics, 59,99-123.
Myrdal, G. (1957), Economic Theory and Underdeveloped Regions, Duckworth, London.
Pitt, M., M. Rosenzweig, and D. Gibbons (1995), The Determinants and Consequences of
the Placement of Government Programs in Indonesia, in D. van d e Walle and K. Nead
(eds.), Public Spending and the Poor Theory and Evidence, Johns Hopkins University
Press.
Ravallion, M. (1982), The Welfare Economics of Local Public Spending: An Empirical Ap-
proach, Economica, 4 9 , 4 9 4 1 .
Ravallion, M. (1984), The Social Appraisal of Local Public Spending Using Revealed Fiscal
Preferences, Journal of Urban Economics, 16,46-64.
Ravallion, M. (1993), Poverty Alleviation through Regional Targeting: A Case Study for In-
donesia, in K. Hoff, A. Braverman, and J. Stiglitz (eds.) The Economics of Rural Orga-
nization, Oxford University Press, Oxford.
Ravallion, M. (1996), Issues in Measuring and Modelling Poverty, Economic Journal, 106,
1328-44.
Ravallion, M. and S. Chaudhuri (1997), Risk and Insurance in Village India: Comment, Econo-
metrica, 65, 171-184.
Ravallion, M. and J. Jalan (1%6), Growth Divergence due to Spatial Externalities, Economics
Letters, 53,227-232.
Ravallion, M. and Q. Wodon (1997), Poor Areas, or Only Poor People? Policy Research Work-
ing Paper 1363, World Bank, Washington, D.C.
Richardson, H. W. and P. M. Townroe (1986), Regional Policies in Developing Countries, in
P. Nijkamp (ed.), Handbook of Regional and Urban Economics, North-Holland, Am-
sterdam.
POORAREAS 91
Riskin, C. (1994), Chinese Rural Poverty: Marginalized or Dispersed?, American Economic

Review, Papers and Proceedings, 84,281-284.
Romer, P. (1986), Increasing Returns and Long-Run Growth, Journal of Political Economy,
94,1002-1037.
Sargan, J. D. (1958),The Estimation of Economic Relationships Using Instrumental Variables,
Sargan, J. D. (1988), Testing for Misspecification after Using Instrumental Variables, in E.
Massoumi (ed.), Contributions to Econometrics: John Denis Sargan, vol. I, Cambridge
University Press, Cambridge.
Sengupta, J. L. and Bo Q. Lin (1995), Recent Rural Growth in China: The Performance of the
Rural Small-Scale Enterprises, mimeo, University of California, Santa Barbara, CA.
Townsend, R. (1991), Risk and Insurance in Village India, Econometrica, 62,539-591.
van d e Walle, D. (1995), Rural Poverty in an Emerging Market Economy: Is Diversification
into Non-Farm Activities in Rural Viet Nam the Solution?, mimeo, Public Economics
Division, Policy Research Department, World Bank, Washington, D.C.
Wilson, W. J. (1987), The Truly Disadvantaged, University of Chicago Press, Chicago, IL.
World Bank (1992), China: Strategiesjor Reducing Poverty in the 1990s, World Bank, Wash-
ington, D.C.
World Bank (1995), Understanding Poverty in Poland, World Bank, Washington, D.C.
The Demand for Health Services in a
Developing Country
The Role of Prices, Service Quality, and
Reporting of Illnesses
Anil B. Deolalikar
University of Washington, Seattle, Washington
1. INTRODUCTION
As a result of the debt crises of the 1980s and the ensuing structural adjustment
and stabilization programs, many less-developed countries (LDCs) have had to cut
back social spending, including spending on government health programs (Cornea,
Jolly, and Stewart 1987). As a result, these countries have been forced to explore
alternative means of financing health services, including greater recovery of (re-
current) costs in the government health sector via user fees. Proponents of greater
cost recovery base their recommendations on the findings of several empirical stud-
ies that suggest that the demand for health care in LDCs is price inelastic (Akin
et al. 1987, Jimenez 1987, World Bank 1987). On the other hand, opponents of
the cost recovery argument contend that raising fees will reduce access to care,
especially by the poor, and consequently adversely affect health status (Cornea,
Jolly, and Stewart 1987, Gilson 1989).
Unfortunately, the empirical bases on which both arguments are made are
weak. The relatively few empirical studies of health-care demand for LDCs are flawed,
largely because of their failure to recognize (1)the role of quality of health services
in influencing demand and (2) the effect of health-care prices on utilization of health
services via their effect on the reporting of illnesses by individuals. The most orb-
vious reason for the lack of control for quality is that observable and quantifiable
data on quality are rarely available. But, since the price charged for medical care
93
94 DEOMLIKAR
often reflects the quality of care provided, the lack of control for quality confounds
quality with price effects and biases estimated price effects toward zero (as price
and quality influence demand in opposite directions). In addition, health-care de-
mand functions that are conditioned on reported morbidity can greatly understate
the total effect of health-care prices on the utilization of health services, since they
ignore the potentially adverse effect that these prices can have on the reporting of
morbidity.
II. PREVIOUS STUDIES
A number of studies have previously attempted to estimate the demand for health
services in LDCs. Unfortunately, the existing literature in this area offers confusing
evidence regarding the price response of health-services utilization to user fees. One
strand of literature suggests that prices are not important determinants of health-care
utilization. Heller (1981), Akin et al. (1984, 1986), Birdsall and Chuhan (1986),
and Schwartz et al. (1988) all report very small and sometimes positive price ef-
fects, most of which are statistically insignificant. Another strand of work by Mwabu
(1986), Gertler et al. (1987), Alderman and Gertler (1988), and Gertler and van der
Gaag (1990)conclude that prices are important. The results of the first group of stud-
ies contrast sharply with most studies on the demand for medical care in developed
countries which report price elasticities ranging from -0.2 to as high as -2.1 (Ros-
set and Huang 1973, Goldman and Grossman 1978, Newhouse and Phelps 1974,
Manning et al. 1987). This divergence between the literature on developed and de-
veloping countries is paradoxical, since one would expect prices to be more impor-
tant in determining utilization in developing than in developed countries for two
reasons: first, income levels are substantially lower in the developing countries; sec-
ond, medical insurance, which is almost universal in developed countries, is virtually
nonexistent in most developing countries.
The paradox may be explained by the fact that most previous studies on health-
services utilization in developing countries are flawed in three respects. First, the
treatment of the price of health services in much of the previous work has been far
from satisfactory. While some studies have used expenditures per medical visit re-
ported by consumers as the relevant price, other studies have used standard fee
schedules, as reported by providers. Both methods are incorrect and can cause mis-
leading results. The amount paid by a consumer per provider visit (namely, the “unit
value”) depends not only on the price charged by that provider for a standard treat-
ment but also on the type of treatment and quality of service chosen by the patient.
For example, a visit for a common cold will necessarily cost less than a visit for a
more serious problem. In addition, health providers, like other suppliers of goods and
services, can typically provide a range of treatments of varying quality (and price) for
the same ailment. To calculate the true price of health services, the disease-specific
technological effect and the consumer-chosen quality effect need to be purged from
SERVICES IN A DEVELOPING
HEALTH COUNTRY 95
observed unit values. Much of the previous work on health-care demand has con-
founded these price, quality, and disease-specific variations.*
The use of established or official fees as price constructs does not solve the
problem either. Indeed, this procedure introduces another set of biases in the esti-
mates of demand functions. For example, in estimating the demand for health ser-
vices, Gertler and van der Gaag (1990)assume that the price of obtaining health ser-
vices from government medical establishments in Peru and the Ivory Coast is zero,
since such establishments do not have user fees in principle.? However, a number of
recent surveys in developing countries suggest that there may be a wide discrepancy
between officially established fees for medical visits and payments actually made by
patients (World Bank 1992a, Deolalikar and Vashishta 1992). Individuals may be
able to obtain speedier service and higher-quality treatments by paying for services,
even when such services are officially free of charge. Imposing the assumption that
prices do not vary in the sample (when, if fact, they do) can reduce the efficiency of
price elasticity estimates and incorrectly lead to the result that the price elasticity
of demand for health services is not significantly different from zero.
The second major problem with previous studies is that they estimate the de-
mand for health services, conditional on an illness episode being reported by an indi-
vidual or household. To the extent that health-care prices can affect morbidity (i.e.,
the probability of an individual experiencing an illness episode) and the reporting of
morbidity by individual respondents, the price effects obtained from a conditional
health-care demand model are partial. To complicate matters, health-care prices are
likely to have opposing effects on morbidity (positive), reporting of morbidity (nega-
tive), and health-care utilization (negative), so it is impossible to infer the total effect
of prices on health-services utilization from the conditional (and partial) demand es-
ti mates.
The third major problem in studies of health-care demand is the omission of
food prices. Within a general behavioral model of health determination, the demand
for food and medical care are jointly determined, since nutrition and medical care
are (possibly substitutable) inputs in the “production” of health status. This means
that the demand for health care is influenced not only by the price of health services
but also by food prices, in much the same way as the demand for different foods is
determined by food and health-care prices.$ Of course, the omission of food prices
from health-care demand functions will not necessarily bias the estimated effects of
health-care prices on health-care demand unless food prices are correlated with the
price of health services.
*See Deaton (1988)for a discussion of a somewhat similar problem in the analysis of food demand in
LDCs.
tThey assume that all of the price variation occurs in the form of variation in distance traveled to
providers.
$See Behrman and Deolalikar (1988).
96 DEOLALIKAR
This chapter attempts to address all of these shortcomings of existing research.

A model is developed which separates interhousehold and spatial variation in house-
hold-reported unit values of health services. Under the assumption that interhouse-
hold variations in unit values within a geographical area or cluster reflect quality
variations and spatial variations reflect true price Variation, cluster-specific prices
of health services that are purged of quality and disease-specific variations are de-
rived and used in a multinomial logit treatment choice model. In order to assess
the effect of health-care prices on the reporting of illness episodes, the utilization
of curative health services is estimated both conditionally and unconditionally on
the probability of an illness being reported. Finally, prices of important foods (also
purged of quality variations) are included as determinants of the demand for health
services.
111. T H E MODEL
Since the theory of demand for medical care is well-developed,* there is no need
here to develop an elaborate model of individual health determination. If it is as-
sumed that individuals maximize a utility function having health status and other
consumption as its argument, subject to a budget constraint and a health production
function that includes food and medical care as inputs, the resulting reduced-form
derived demand functions for food and medical care will include as their arguments
the prices of food and medical care, household income, and socio-demographic indi-
vidual and household characteristics. As noted earlier, the major empirical problem
in estimating such a reduced-form demand system is that health-care prices are not
directly observed; what are observed instead are the (endogenous) unit values. The
latter need to be purged of disease-specific technological and household-specific
quality variations before they can be treated as health-care prices. This is a problem
that Deaton (1988)has dealt with in the context of food prices.
We assume that (1)for a given type of health provider (e.g., private physician
versus a public health clinic), interhousehold variation in expenditure per illness
episode that is explained by individual and household characteristics, such as sex,
age, marital status, household income, household size and composition, and traits
(e.g., age, schooling and occupation) of the household head, reflects variation in the
quality of health services, and (2) that the spatial (intercluster) component of the
unexplained variation in unit values reflects true (quality-constant) price variation.
In other words, it is assumed that when individuals with high income and better
schooling spend a larger amount on treating the same ailment from the same type
of health provider, they are in effect buying higher quality of health care. However,
*See Grossman (1972).Behrman and Deolalikar (1988) also develop a generic model of health determi-
nation for an LDC.
HEALTH
SERVICES I N A DEVELOPING
COUNTRY 97
even after controlling for income, education, and other characteristics, if individuals
in one location spend more than consumers in another location for treating the same
ailment from the same type of provider, that difference reflects true price variation in
the cost of health services across the two locations. While this is a strong assumption,
it does not appear to be unreasonable.* Given that the quality of health care selected
by households is unobserved, an assumption of this type is required to identify prices
from observed unit values.
Controlling for disease-specific effects in the estimation of health-care prices
is much more straightforward, since the ailments for which individuals obtain med-
ical care are observed in the data.
An individual’s decision not to seek treatment for an illness or to seek treat-
ment from a traditional healer or modern provider is modeled as a multinomial logit
problem. The probability of seeking no, tradition, or modern care is
+ b k In p r + dk In x;4- ekzi; -k
exp(ak In p y pi;)
P ( M ; = k)
E,exp(n, In p y + 6, In pJF + d, In K; + e,Zi; + pi;)
(1)
k = 0,2, z=l,2
where i = index of the individual
j = index of the location or cluster of residence
M = choice of medical care
k = 0 for no treatment, 1 for a traditional healer, and 2 for a “modern” provider
(private physician, health clinic, or hospital)
p M = vector of health-care prices (i.e., the price of services obtained from tra-
ditional healers and modern health providers), derived below
p F = vector of food prices, derived below
Y = household income
Z = vector of individual and household characteristics, including age, educa-
tion, family size and composition, etc.
p = disturbance term
The only problem in estimating relation (1) is that quality-constant health-care

prices, p M , are unobserved. What are observed instead are the unit values of health
services obtained from various providers. The derivation of prices from unit values is
based on the assumption, stated earlier, that interhousehold variations in unit values
reflect quality variations, while spatial variations in unit values reflect true price vari-
ation. The unit value can be thought of as the product of the true (quality-constant)
price of health services and the quality of health services purchased, i.e.,
*This assumption is similar to that made by Deaton (1988)to separate the effect of true prices from the
effect of quality variations on consumer food demand.
98 DEOIALIKAR
where qM is the unit value and Q is the quality of care obtained from a particular
provider.* Note that the true price charged by that provider, pM,does not have an i
subscript, since it is assumed to vary only spatially. Taking the logs of both sides of
(2), we have
If it is further assumed that the quality of care purchased, QM,is a function of

household income and a vector of household characteristics ( Z ) ,namely
where v is a disturbance term, then (3)can be rewritten as
Equation (5) is a standard cluster fixed-effects model of unit values, in which In p y

are the cluster fixed effects (or intercepts).
In order to be consistent, food prices, p'; are derived in an identical manner,
even though the control for variations in food quality is not central to this chapter.
IV. BACKGROUND, DATA, AND EMPIRICAL MODEL
With a total population estimated at 178.2 million in mid-1990, Indonesia is the

fourth most populous country in the world (World Bank 1992b). The Indonesian
economy has enjoyed rapid economic growth during the last two decades; for ex-
ample, between 1965 and 1990, Indonesia achieved an annual growth rate of per
capita GNP of 4.5%, a rate that few developing countries could match. Available es-
timates also suggest an impressive reduction in infant mortality in Indonesia during
the same period, from 128 infant deaths per 1000 live births in 1965 to 61 in 1990
(World Bank 199213). Despite this performance, Indonesia has one of the highest
infant and maternal mortality rates among Southeast Asian countries. For example,
Indonesia has higher infant and maternal mortality levels than the Philippines, Thai-
land, or even Vietnam. Anemia is the major cause of maternal mortality (Gopalan
*The true price and the quality of' service are easier to interpret if the unit value (or observed price) is
modeled as a product (as opposed to another function) of the true price and quality. In this case, the true
price,y,".can be thought of as the amount (in Indonesian Rupiahs) paid per provider visit of standardized
quality, while the quality variable, QY, can be regarded as the ratio of a visit of standardized quality to
an actual visit.
?In the absence of any priors on the functional form for the quality-of-service function, I have chosen the
log-linear form for reason of convenience and tractability.
HEALTH COUNTRY 99
1988),while immunizable diseases (particularly tetanus), diarrhea, and acute res-

piratory infections are thought to be the leading causes of infant and child mortality
(Government of Indonesia-UNICEF 1989, p. 40).
The data for this study come from the 1987 round of the National Socioeco-
nomic Survey (SUSENAS), which is a nationally representative survey of Indone-
sia that is undertaken periodically. The 1987 round, conducted in January, covered
roughly 250,000 individuals residing in 50,000 households. While focusing on the
health status of individuals and the choice of health providers for curative care, the
1987 SUSENAS survey obtained detailed information on household consumption
expenditures and income as well. Using both one-week and three-month recall tech-
niques, the health module collected data on perceived illnesses (occurrence, type,
and length) and the choice of provider for any treatment obtained.
The other data source used is the Village Potential (Potensi Desa) module of
the Economic Census 1986, a census of all the villages in Indonesia. The Economic
Census reports, among other things, information on the environmental hygiene con-
ditions of villages. Although, in principle, it is possible to merge the SUSENAS
household data with the village-level information from the Economic Census, the
SUSEN AS data tapes identify only the district (kabupaten)-not the village-of res-
idence of households in order to protect their confidentiality. Hence, the Economic
Census Information has been aggregated at the level of kabupatens before merging
it with the household-level SUSENAS data.
For the purposes of this study, there are two major drawbacks of the 1987
SUSENAS health module. First, as no clinical diagnosis was performed in assessing
morbidity, all measures of morbidity are respondent-reported and as such subject
to measurement error and respondent biases in illness perception.* For this reason,
as previously mentioned, it is even more important not to condition the demand for
health services on an illness episode being reported, as there is a strong likelihood of
health-care prices influencing the reporting of morbidity by individual respondents.
The second problem of the 1987 SUSENAS data is that information was collected
for each illness episode, not for each visit to a health provider.? Much of the dis-
cussion in the previous section is based on survey information being available on
expenditures per visit. To the extent that individuals may have made multiple visits
to the same or different health providers for treating an illness episode, observed
household expenditures on medical care are unlikely to be unit values. In turn, this
would contaminate the derivation of health-care prices from expenditure data. Fortu-
nately, however, the 1987 SUSENAS obtained information on health-care visits and
expenditures for two reference periods-one week and three months preceding an
*Of course, most household health surveys are plagued by this problem, since the collection of objective
indicators of ill health is difficult and expensive.
?Indeed, no information is available on how many visits an individual made to the same or different health
providers in treating an illness episode.
100 DEOMLIKAR
interview. It is much less likely that individuals would have made multiple visits to
a health provider within a one-week period than in a three-month period. Therefore,
the one-week recall data are likely to provide more meaningful and reliable results.
In addition, the one-week recall data are more dependable because three months is
generally too long a period to recall an illness episode with much accuracy. There-
fore, only the one-week recall data are used in the analysis that follows.
The 1987 SUSENAS reports eight treatment alternatives for illnesses reported
during the past week and the past three months: no treatment, self-treatment, tra-
ditional healer, private physician, hospital, public health center, polyclinic, and
paramedic. Since each treatment alternative adds a total of 34 parameters to be es-
timated in the multinomial logit model, the eight treatment choices have been col-
lapsed into three broad alternatives: no treatment (with the dependent variable as-
suming a value of zero), self-treatment, or treatment from a traditional healer (value
of one), and treatment from a modern health provider (which includes all the remain-
ing choices) (value of 2). With the one-week recall data, the percentage of adults re-
porting an illness episode who sought these treatment methods were 31.3%, 5.4%,
and 63.3%, respectively (Table 1). For children below five years of age, the corre-
sponding figures are 27.6%, 4.3%, and 68.1%. The illness episodes of children aged
5-17 years appear to receive the least attention. For this age group, nearly 37.2%
of reported illness episodes were not treated, and fewer than 60% were treated by a
modern health provider.
The household characteristics (2) included in the unit value equations and
the treatment choice model are size and demographic composition of the household,
age, and schooling years of the household head, whether the household head is a
salaried employee, and whether the household is covered by any health insurance.
The presence of health insurance can often dramatically affect both the choice of
quality of health care as well as the choice of health providers for curative services.
In Indonesia, the only households typically covered by health insurance are those
with a member working in the public sector; in the current sample, only 1.5%of the
households belong to this category. Two cluster characteristics are included in the
unit value and treatment choice models: the proportion of villages in the kabupaten
having organized garbage collection and those having piped (and hence presumably
clean) drinking water. Since both of these environmental hygiene variables affect the
production of health status from food and medical care inputs, they are expected to
influence the demand for medical care. Finally, three individual characteristics are
included in the treatment choice model; these are sex, age, and schooling years.*
There are two econometric problems that warrant discussion. First, since
household income includes labor earnings, which are likely to be affected adversely
*Since the unit values are observed only at the household level, the unit-value equations do not include
individual characteristics as explanatory variables.
HEALTH
COUNTRY 10I
by an illness in the family and by the treatment choices (including no treatment)

made for that illness, there may be a simultaneous-equations bias introduced by the
inclusion of household income in the multinomial logit treatment choice model. To
address this problem, I use an instrumental variable for log of household income,
with the instruments being all of the variables in vector Z (see next paragraph) and
nonlabor income.* The latter variable thus provides identification, since the vector
Z already appears in the treatment choice model.
Second, the estimates of the cluster fixed effects ( l n p y in (5))contain error,
which can lead to biased estimates of the price effects in the treatment choice model.
Fortunately, the Potensi Desa dqta providc kabupaten-level information on a large
number of environmental and physic 1 &r:;rastructurevariables, which can be used
as instruments for the set of estimate 1 fixed effects. Thus, the parameters of (1) are
estimated using instrumental-variable methods to correct for errors in variables in
the measurement of In py and In p;, with kabupaten characteristics as identifying
instruments. These characteristics include the proportion of villages in a kabupaten
having a primary school, a lower secondary school, a higher secondary school, a
college, a resident health worker, a physician, a hospital, another health facility, a
library, piped water, organized garbage collection, and an all-weather road; the pro-
portion of villages in a kabupaten that are urban centers, coastal, and accessible only
by water; average village population, population density, cultivated area per person,
and the proportion of cultivated area that is double-cropped and that is irrigated;
average age and schooling of the village head; and the proportion of village heads (in
a kabupaten) that are female.
V. EMPIRICAL RESULTS
A. Unit Value Regressions
Cluster fixed-effects estimates of the unit value regressions are reported in Table 2.t
Since the unit value regression for traditional providers is estimated with very few
observations, only the empirical results relating to modern-provider unit values pro-
viders are discussed here. The estimates indicate that these unit values depend
*Usually, the problem in using nonlaljor income in most developing-country data sets is that it provides
very little additional information, since few households report any asset income. However, this is not
the case with the 1987 SUSENAS survey, in which over 94%of households reported nonzero values of
annual nonlabor income. Transfers comprise a large fraction of nonlabor income.
'ISince the true prices of medical care are derived from the estimated cluster effects, it is not possible
to control additionally for household fixed effects in the unit value regressions. If individuals within
a household have correlated error, not taking into account the error structure may lead to incorrect
standard errors. However, this is not a major problem in the current sample. Fewer than 10%of the
sample observations for the unit value regressions are accounted for by multiple individuals residing in
the same household.
Table I Descriptive Statistics, Indonesia, 1987
Household statistics Children under 5 years Children 5 1 7 years Adults over 17 years
Variable Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
Whether illness episode reported

during preceding week? 0.114 0.318 0.054 0.227 0.074 0.262
Proportion of reported illness episodes treated by
no one or self 0.276 0.504 0.372 0.533 0.313 0.547
traditional health provider 0.043 0.204 0.036 0.189 0.054 0.238
modem health provider 0.681 0.520 0.593 0.531 0.633 0.556
Percentage distribution of reported illness episodes
tuberculosis-type symptoms 0.7% 8.3% 1.1% 10.4% 4.w0 19.7%
diarrhea-type symptoms 6.3% 24.2% 2.4% 15.2% 1.7% 13.1%
malaria-type symptoms 14.8% 35.6% 25.2% 43.4% 16.3% 36.9%
measles-type symptoms 6.3% 24.2% 3.6% 18.7% 0.7% 8.1%
liver-related ailments 5.1% 22.wo 6.3% 24.3% 13.99’0 34.6%
other single symptoms 6 1.4% 48.7% 57.3% 49.5% 57.w0 49.5%
multiple symptoms 5.5% 22.8% 4.1% 19.99’0 6.4% 24.5%
“Unit value” (in Rupiahs per reported illness episode) of
Traditional health services 568
Modem health services 1.218
Log total household income 11.514 0.725
Log nonlabor household income 8.547 3.432
Age of household head 43.012 13.802
Schooling of household head 5.032 3.978
Whether household head employee? 0.386 0.487
Whether health insurance coverage? 0.015 0.120
Age of individual 2.018 1.366 10.671 3.713 37.040 14.930
Whether individual male? 0.513 0.500 0.516 0.500 0.488 0.500
Schoolingof individual 5.219 3.978
Household size 4.788 2.235
Proportion in total household size of"
Males 0 4years 0.055 0.108
Males 5-14 years 0.118 0.153
Males 15-24 years 0.090 0.160
Males 25-34 years 0.085 0.143
Males 35-44 years 0.052 0.106
Males 45-54 years 0.039 0.098
Males 55-59 years 0.016 0.069
Males 60 years and over 0.035 0.111
Females (PIyears 0.052 0.105
Females 5-14 years 0.110 0.146
Females 15-24 years 0.104 0.159
Females 25-34 years 0.081 0.125
Females 35-44 years 0.051 0.108
Females 45-54 years 0.049 0.124
Females 55-59 years 0.017 0.084
Proportion of villages in district (kubupaten) having
Organized trash collection 0.159 0.186
Piped drinking water 0.089 0.170
"Excluded category is females 60 years of age and over.
104 DEOLALIKAR
Table 2 Cluster (kabupaten) Fixed-Effects Estimates of Health-Care Log “Unit Values”

(i.e., expenditures per illness episode), Indonesian Households, 1987
Unit values for:
Traditional providers Modern providers
Independent Variable Parameter T ratio Parameter T ratio
Whether household head employee? -0.222 -1.4 0.030 0.8

Household size 0.023 0.4 -0.074 -5.8.
Log total household incomen 0.550 1.7 0.804 11.3
Age of household head -0.014 -1.4 0.003 1.5
Schooling of household head -0.012 -0.3 0.021 2.8
Whether health insurance coverage? .- 0.7 16 -0.7 0.434 3.1
Proportion in total household size ogh
Males 0-4 years -0.526 -0.5 -0.597 -2.7
Males 5-14 years - 1.353 -1.5 -0.357 -1.9
Males 15-24 years -2.467 -2.6 -0.117 -0.6
Males 25-34 years -2.069 -1.9 -0.265 -1.1
Males 35-44 years - 1.035 -0.9 -0.213 -0.8
Males 45-54 years -2.791 -2.3 -0.225 -0.9
Males 55-59 years - 1.634 -1.1 -0.105 -0.3
Males 60 years and over 0.050 0.0 -0.249 -1.0
Females 0-4 years -1.151 -1.2 -0.424 -1.9
Females 5-14 years -1.526 -1.9 -0.209 -1.1
Females 15-24 years -1.060 -1.3 -0.200 -1.0
Females 25-34 years -2.289 -2.2 -0.053 -0.2
Females 35-44 years -0.978 -0.8 -0.135 -0.6
Females 45-54 years -0.046 -0.1 -0.081 -0.4
Females 55-59 years -1.118 -0.7 0.318 1.2
Distribution of illness episodes
in household during past week‘
tuberculosis-type symptoms -1.744 -4.0 0.524 6.8
diarrhea-type symptoms -0.736 -1.8 -0.102 -1.3
malaria-type symptoms -0.045 -0.2 -0.208 -4.6
measles-type symptoms -0.057 -0.2 -0.126 -1.4
liver-related ailments 0.254 1.3 0.197 4.4
multiple symptoms 0.094 0.3 0.092 1.4
F-ratio 1.750 2,107
R-square 0.239 0.535
No. of obs. 178 51,248
aInstrumental variable. Identifying instrument is log nonlahor household income.
‘Excluded category is females aged 60 years and over.
‘Excluded category is ‘‘Other symptoms.”
HEALTH
COUNTRY 105
significantly on the type of illness and symptoms for which treatment is sought. For
example, treatment for liver-related diseases costs 19.7% more, and that for tuber-
culosis costs 52.4% more, than treatment for the excluded category of “other ill-
nesses.” In contrast, malaria treatment costs 20.8% less than the excluded category.*
Controlling for the disease-specific and the spatial (or cluster) effects, household in-
come has a strong positive impact on both modern- and traditional-provider unit
values, with the income elasticity for the former being 0.80 and the latter, 0.55. As
would be expected, individuals with health insurance coverage spend significantly
(43.4%) more per illness episode than individuals without insurance coverage. The
unit value regressions also imply that individuals from large households buy lower
levels of health-care quality (i.e., spend less per illness episode), while those from
households with better-schooled heads purchase higher quality. Finally, the house-
hold composition effects indicate that additional numbers of children aged 0-4 years
and boys aged 5-14 years significantly reduce household expenditure per illness
episode.
B. Treatment Choice Model

Table 3 shows the conditionalt and unconditional multinomial logit estimates of the
treatment choice model for children under S years of age, while Table 4 and 5 show
the corresponding estimates for children aged S-17 years and adults aged 18 and
over. Ten major empirical findings are discussed.
1. The conditional estimates, which are comparable to the estimates reported
by most previous studies, indicate that while the demand for traditional healer ser-
vices is significantly responsive to price (with own price elasticities of -0.10, -0.45,
and 0.60t for individuals aged 0 4 , 5-17, and 18 and over, respectively), the de-
mand for health care from modern providers is not significantly influenced by price.
However, the unconditional estimates imply significant own-price responsiveness of
demand for modern health services, with estimated elasticities of -0.13, -1.64, and
-0.82 (all significantly different from zero at the 5% level) for the three age groups.
The estimated price responses are large, especially for children aged 5-17 but also
for adults. Indeed, such large own-price elasticities have rarely been reported in the
literature. The difference in own price elasticity estimates for modern health care
between the conditional and unconditional provider choice models suggests that the
price of care in the modern sector strongly reduces the reporting of morbidity by
*The large coefficient on the tuljerculosis dummy variable in the unit-value regression for traditional
providers is not reliable, since only a couple of individuals in the sample obtained treatment for tuber-
culosis from a traditional provider.
?These are conditioned on an individual reporting an illness during the one-week reference period.
Unless otherwise stated, all elasticities reported are evaluated at sample means.
Table 3 Multinomial Logit Estimates of the Probability of Seeking Treatment from a Traditional or Modern Health Provider for a Reported Illness
Episode Experienced during the Preceding Week, Children under 5 Years of Age, Indonesia, 1987
Conditioned on an illness being reported Unconditioned on an illness being reported
Traditional health provider Modem health provider Traditional health provider Modem health provider
Independent variable Parameter T ratio Elasticity Parameter T ratio Elasticity Parameter T ratio Elasticity Parameter T ratio Elasticity
Intercept - 19.369 -1.5 2.461 0.5 -23.597 -2.1 -6.14% -2.1

Whether household head employee? -0.786 -2.8 -0.128 0.248 2.5 0.008 -0.778 -3.0 -3.088 0.226 4.1 0.188
Household size -0.028 -0.3 0.045 -0.092 -2.4 -0.003 -0.026 -0.3 1.173 -0.089 -4.2 -0.071
Log total household income" 0.150 0.3 -0.007 0.205 1.o 0.0o0 -0.023 0.0 -0.643 0.186 1.6 0.039
Age of household head -0.026 -1.4 -0.131 0.003 0.4 0.008 -0.033 - 1.9 -3.349 -0.004 -0.9 0.203
Schooling of household head 0.020 0.4 -0.023 0.054 2.7 0.001 -0.010 -0.2 0.094 -0.015 - 1.4 -0.006
Whether health insurance coverage? -32.234 0.0 -3.962 -0.181 -0.5 0.241) - 16.746 0.0 -51.979 0.146 0.7 3.157
Age of child -0.171 -2.1 -0.036 -0.015 -0.4 0.002 -0.229 -3.2 -0.939 -0.077 -4.2 0.057
Whether child male? -0.095 -0.3 0.015 -0.213 -1.5 -0.001 0.076 0.3 0.417 -0.059 -0.7 -0.025
Price of
Traditional health servicesb -0.899 -4.0 -0.103 -0.067 -0.7 0.006 -0.688 -3.3 -2.2.38 0.040 0.9 0.136
Modern health servicesb 0.812 2.1 0.129 -0.235 -1.5 -0.008 0.305 0.9 2.186 -0.406 -4.8 -0.133
Domestic rice 1.755 1.4 0.016 1.625 3.1 -0.001 1.039 1.o 1.659 0.500 1.8 -0.101
Cassava -0.252 -0.6 -0.025 -0.049 -0.3 0.002 -0.035 -0.1 -0.276 0.055 0.5 0.017
Onion -0.so5 -0.6 0.006 -0.350 -1.7 0.0o0 -0.220 -0.5 -0.549 -0.041 -0.4 0.033
Canned milk - 1.365 -1.8 -0.173 0.032 0.1 0.010 -0.883 -1.4 -3.200 0.157 0.9 0.194
Meat soya 0.945 2.0 0.096 0.171 0.8 -0.006 0.807 1.9 1.243 0.403 3.4 -0.076
Sugar 1.327 0.7 0.463 -2.420 -2.8 -0.028 2.344 1.3 9.1,SO -0.630 -1.3 -0.556
Proportion in total household size of
Males 0-4 years 6.225 1.8 0.146 -1.530 -1.3 -0.009 6.247 1.8 3.001 -0.541 -0.9 -0.182
Males 5-14 years 4.748 1.4 0.079 - 1.597 -1.5 -0.005 4.618 1.4 2.027 -1.052 -1.8 -0.123
Males 15-24 years 3.453 1.0 0.026 -1.210 -1.1 -0.002 3.386 1.o 0.558 -0.650 -1.0 -0.034
Males 25-34 years 0.954 0.3 0.039 -1.787 -1.5 -0.002 1.658 0.5 0.721 -0.622 -1.0 -0.044
Males 3 5 4 4 years 1.937 0.5 0.033 -3.561 -2.9 -0.002 2.8% 0.8 0.736 -1.640 -2.4 -0.045
Males 45-54 years 6.922 1.7 0.02 1 -2.626 -1.8 -0.001 7.797 2.0 0.550 -1.435 -1.8 -0.033
Males 55-59 years -3%.746 0.0 -0.201 1.693 0.7 0.012 -178.350 0.0 -2.513 -0.597 -0.5 0.153
Males 60 years and over 8.985 1.9 0.013 -2.019 -1.1 -0.001 9.377 2.2 0.312 -0.829 -0.9 -0.019
Females 0 4 years 6.246 1.8 0.140 -1.763 -1.5 -0.008 6.392 1.9 3.047 -0.795 -1.3 -0.185
Females 5-14 years 4.836 1.4 0.076 -1.285 -1.2 -0.005 5.151 1.6 1.989 -0.689 -1.2 -0.121
Females 15-24 years 3.682 1.0 0.057 - 1.070 -0.9 -0.003 4.217 1.2 1.128 -0.078 -0.1 -0.069
Females 25-34 years 1.284 0.3 0.02s -0.656 -0.5 -0.002 1.570 0.4 0.542 -0.1% -0.2 -0.033
Females 3- years 2.840 0.7 0.015 -1.789 -1.3 -0.001 3.115 0.8 0.366 -0.713 -1.0 -0.022
Females 45-54 years 4.981 1.2 0.010 -0.202 -0.1 -0.001 5.635 1.5 0.219 0.578 0.8 -0.013
Females 55-59 years -1.619 -0.2 -0.002 2.319 1.2 O.OO0 -2.661 -0.4 -0.046 0.342 0.4 0.003
Proportion of vilhges in district having
Organized trash collection -0.852 -0.7 -0.023 0.374 0.8 0.001 - 1.914 -1.8 -0.641 -0.679 -3.0 0.039
Piped drinking water -0.451 -0.3 -0.017 1.023 2.2 0.001 0.148 0.1 -0.292 1.138 5.2 0.018
No. of obs. 2,767 24,202
Log-likelihood ratio -1.986 -7,010
All elasticities evaluated at sample means.

'Instrumental variable. Identifying instrument is log nonlabor household income.
bInstrumental variable. Identifying instruments are various district (kabupaten)characteristics. See text for a list of these characteristics.
Table 4 Multinomial Logit Estimates of the Probability of Seeking Treatment from a Traditional or Modem Health Provider for a Reported Illness
Episode Experienced during the Preceding Week, Children Aged 5-17 Years of Age, Indonesia, 1987
Traditional health provider Modem health provider Traditional health provider Modern health provider
Intercept -27.732 -2.8 -4.503 -1.2 -35.104 -4.0 - 13.391 -5.5

Whether household head employee? -0.348 - 1.6 -0.266 0.056 0.7 0.017 -0.197 -1.0 -5.492 0.115 2.4 0.354
Household size -0.015 -0.2 0.112 -0.041 -1.7 -0.007 -0.043 -0.7 3.213 -0.071 -4.4 -0.207
Log total household incomea -0.244 -0.6 -0.310 0.226 1.6 0.020 -0.063 0.2 -6.473 0.305 3.3 0.418
Age of household head -0.022 - 1.6 -0.797 0.007 1.4 0.052 -0.034 -2.5 -24,440 -0.002 -0.7 1.577
Schooling of household head 0.004 0.1 -0.201 0.062 4.2 0.013 -0.066 -1.7 -6.871 0.006 0.7 0.443
Whether health insurance coverage? -35.475 0.0 -23.329 -0.042 -0.2 1.512 -15.222 0.0 -271.515 0.208 1.4 17.517
Age -0.029 -1.1 -0.044 -0.022 -2.1 0.003 -0.067 -2.6 -0.204 -0.066 -9.9 0.013
Whether male? 0.085 0.4 0.056 -0.001 0.0 -0.004 0.076 0.4 1.270 0.004 0.1 -0.082
Price of
Traditional health servicesb -0.584 -3.2 -0.44% 0.097 1.4 0.029 -0.471 2.7 -9.720 0.081 2.1 0.627
Modem health servicesb 1.351 4.1 0.972 -0.126 -1.1 -0.063 0.751 2.3 25.433 -0.695 -9.7 -1.641
Domestic rice 2.81 1 2.6 0.401 2.202 5.5 -0.026 2.058 2.2 15,510 1.177 5.1 -1.001
Cassava -0.743 -2.3 -0.500 0.0 15 0.1 0.032 -0.610 -2.1 - 13.260 0.144 1.6 0.855
Onion -0.617 -1.5 -0.290 -0.176 -1.2 0.019 -0.492 -1.3 - 10.701 0.116 1.2 0.690
Canned milk -1.322 -2.2 -0.790 -0.122 -0.5 0.051 -0.742 - 1.3 -20.678 0.433 3.1 1.334
Meat soya -0.007 0.0 -0.218 0.324 2.2 0.014 -0.4% -1.4 -13.283 0.259 2.6 0.857
Sugar 5.602 3.6 5.369 -2.554 -4.1 -0.348 6.%2 4.8 139.206 -0.949 -2.4 -8.981
Proportion in total household sue of
Males 0-4 years 0.626 0.3 -0.018 1.217 1.5 0.001 0.868 0.4 0.011 0.852 1.6 -0.001
Males 5-14 years -0.061 0.0 -0.165 1.182 1.6 0.011 -0.310 -0.2 -2.047 0.269 0.5 0.132
Males 15-24 years 1.107 0.5 0.004 1.038 1.4 0.o00 1.095 0.6 0.916 0.526 1.0 -0.059
Males 25-34 years 0.057 0.0 -0.072 2.331 2.7 0.005 -0.662 -0.3 - 1.099 0.866 1.5 0.071
Males 35-44 years 0.385 0.2 -0.067 2.120 2.5 0.004 -0.367 -0.2 -0.887 0.495 0.9 0.057
Males 45-54 years 2.158 0.9 0.015 1.599 1.9 -0.001 2.127 1.o 1.158 0.572 1.o -0.075
Males 55-59 years 2.362 0.8 0.004 1.742 1.6 0.o00 2.253 0.8 0.315 0.601 0.8 -0.020
Males 60 years and over 2.225 0.7 0.010 1.072 1.o -0.001 2.184 0.7 0.538 0.169 0.2 -0.035
Females 0-4 years 1.191 0.5 0.016 0.636 0.8 -0.001 1.531 0.8 0.784 0.388 0.7 -0.051
Females 5-14 years 0.049 0.0 -0.171 1.418 1.9 0.011 -0.209 -0.1 -2.266 0.481 1.o 0.146
Females 15-24 years 0.074 0.0 -0.071 1.528 2.0 0.005 -0.147 -0.1 -0.967 0.515 1.0 0.062
Females 25-34 years 0.040 0.0 -0.047 0.976 1.1 0.003 0.236 0.1 -0.653 0.794 1.4 0.042
Females 35-44 years 1.580 0.7 0.030 0.7% 0.9 -0.002 1.518 0.7 0.935 0.652 1.2 -0.060
Females 45-54 years -0.668 -0.3 -0.024 0.521 0.6 0.002 -0.185 -0.1 -0.423 0.535 0.9 0.027
Females 55-59 years -1.651 -0.5 -0.003 -0.993 -0.9 0.o00 -0.178 -0.1 0.059 -0.654 -0.8 -0.004
Proportion of villuges in district having
Organized trash collection -0.226 -0.3 -0.034 0.130 0.4 0.002 -0.721 -0.9 -1.145 -0.338 -1.8 0.074
Piped drinking water -3.434 -2.4 -0.190 0.163 0.5 0.012 -2.337 -1.8 -4.846 0.522 2.7 0.313
No. of obs. 4,218 77,573
Log-likelihood ratio -3,262 -11,150

aInstrumental variable. Identifying instrument is log nonlabor household income.
'Instrumental variable. Identifying instruments are various district (kabupaten)characteristics. See text for a list of these characteristics.
Table 5 Multinomial h g i t Estimates of the Probability of Seeking Treatment from a Traditional or Modem Health Provider for a Reported Illness
Episode Experienced during the Preceding Week, Adults 18 Years of Age and Over, Indonesia, 1987
~~
Traditional health provider Modem health provider Traditional health provider Modem health provider
Intercept -37.223 -6.4 -4.617 - 1.7 -45.595 -8.9 -14.149 -8.8

Whether household head employee? -0.158 -1.3 -0.171 -0.012 -0.2 0.014 -0.061 -0.5 -1.818 0.066 2.0 0.149
Household size -0.033 -0.8 -0.225 042 0.1 0.019 -0.053 -1.3 -2.955 -0.016 -1.5 0.243
Log total household income' 0.245 1.o 0.006 0.239 2.4 -0.001 0.129 0.6 0.534 0.092 1.5 -0.044
Age of household head 0.002 0.4 -0.079 0.004 1.3 0.007 -0.005 -0.8 -0.661 -0.w -2.2 0.054
Schooling of household head -0.044 -1.4 -0.392 0.030 2.3 0.033 -0.072 -2.5 -5.221 -0.003 -0.3 0.429
Whether health insurance coverage? -0.947 -0.9 -0.931 -0.155 -0.6 0.078 - 1.333 -1.3 -13.953 -0.358 -2.6 1.147
Age of individual -0.002 -0.4 -0.165 0.001 0.7 0.014 0.02 1 4.7 -2.738 0.027 20.6 0.225
Whether individual male? 0.225 1.9 0.42 1 -0.133 -2.5 -0.035 0.369 3.4 5.656 -0.026 -0.8 -0.465
Schooling (years) of individual -0.078 -2.7 -0.556 0.036 3.2 0.046 -0.121 -4.5 -8.800 -0.003 -0.4 0.724
Price of
Traditional health servicesb -0.438 -4.0 -0.597 0.070 1.4 0.050 -0.369 -3.8 -6.152 0.060 2.3 0.506
Modern health services' 0.840 4.5 0.915 0.062 0.8 -0.076 0.147 0.8 10.018 -0.553 -11.2 -0.824
Domestic rice 2.871 4.3 1.687 1.437 5.1 -0.141 2.438 4.2 26.345 0.597 3.6 -2.166
Cassava -0.192 -0.8 -0.156 -0.059 -0.6 0.013 0.098 0.5 -0.768 0.152 2.5 0.063
Onion -0.559 -2.2 -0.043 -0.523 -4.9 0.004 -0.159 -0.7 -2.556 0.019 0.3 0.210
Canned milk 0.568 1.6 0.511 0.134 0.8 -0.043 0.697 2.2 6.713 0.228 2.4 -0.552
Meat soya 0.501 2.0 0.487 0.087 0.8 -0.041 0.403 1.7 3.534 0.156 2.3 -0.291
Sugar 2.104 2.2 3.7,M - 1.088 -2.4 -0.313 3.497 4.0 38.382 0.816 3.0 -3.156
Proportion in total household size of
Males 0-4 years 0.268 0.3 0.018 -0.067 -0.2 -0.002 0.098 0.1 0.202 -0.187 -0.9 -0.017
Males 5-14 years 0.427 0.7 0.057 0.006 0.0 -0.005 0.104 0.2 0.657 -0.284 - 1.7 -0.054
Males 15-24 years 0.208 0.3 0.054 -0.254 -0.9 -0.004 0.099 0.2 0.686 -0.341 - 1.9 -0.056
Males 25-34 years 1.877 2.7 0.157 -0.060 -0.2 -0.013 1.612 2.4 2.277 -0.436 -2.1 -0.187
Males 35-44 years 1.311 1.5 0.053 0.267 0.7 -0.004 0.604 0.8 0.697 -0.432 -1.8 -0.057
Males 45-54 years 1.849 2.5 0.082 0.268 0.8 -0.007 1.303 1.9 0.922 -0.207 -0.9 -0.076
Males 55-59 years 1.521 1.7 0.028 0.220 0.5 -0.002 0.685 0.8 0.249 -0.336 -1.3 -0.020
Males 60 years and over 0.480 0.7 0.024 0.095 0.3 -0.002 0.449 0.7 0.227 -0.003 0.0 -0.019
Females 0-4 years 1.177 1.6 0.069 -0.1 18 -0.3 -0.006 1.154 1.6 0.751 0.031 0.1 -0.062
Females 5-14 years 0.260 0.4 0.071 -0.321 - 1.2 -0.006 -0.059 -0.1 0.730 -0.520 -3.0 -0.W
Females 15-24 years 0.543 0.9 0.079 -0.088 -0.3 -0.007 0.574 1.o 0.978 -0.024 -0.1 -0.080
Females 25-34 years -0.647 -0.9 -0.053 -0.006 0.0 0.004 -0.423 -0.6 -0.667 0.206 1.0 0.055
Females 35-44 years 0.191 0.3 0.017 -0.102 -0.3 -0.001 0.229 0.3 0.265 -0.137 -0.7 -0.022
Females 45-54 years 0.374 0.6 0.009 0.245 0.9 -0.001 0.399 0.7 0.121 0.232 1.4 -0.010
Females 55-59 years 0.862 1.3 0.015 0.189 0.6 -0.001 0.575 0.9 0.204 -0.292 -1.3 -0.017
Proportion of villages in district having
Organized trash collection 0.466 1.o 0.007 0.430 2.0 -0.001 0.106 0.2 0.418 -0.055 -0.4 -0.034
Piped drinking water -1.315 -2.3 -0.128 -0.089 -0.4 0.011 -0.598 -1.1 -1.608 0.445 3.5 0.132
No. of obs. 8,512 114,559
Log-likelihood ratio -6,778 -22,922
OInstrumental variable. Identifying instrument is log nonlabor household income.
'Instrumental variable. Identifying instruments are various district (kabupaten) characteristics. See text for a list of these characteristics.
I12 DEOLALIKAR
respondents, with the result that health-care demand models estimated only on the
sample of individuals self-reporting an illness episode are likely to significantly un-
derstate the adverse effect of health-care prices on the demand for care in the modern
health sector.
The fact that the unconditional own-price elasticity of demand for curative
treatment is small and insignificant for the 0-5 years age group suggests that very
young children in households are relatively more protected than other members
against price changes. On the other hand, children aged 5-17 years are most vulner-
able to price increases, with their modern provider visits falling by 1.6%for a 1%
increase in the price of modern care.
2. The estimated own-price elasticities of demand for traditional health ser-
vices also differ significantly between the conditional and the unconditional model.
However, all elasticities for traditional health services tend to be unusually large be-
cause the sample mean of the unconditional probability of using traditional health
services, which is the point at which the elasticities are evaluated, is extremely small.
3 . Household size has a strong negative effect on both the conditional and
unconditional demand for modern care among children (0-1 7 years) but not among
adults. Surprisingly, household income is generally not significant in determining
the demand for either traditional or modern care. The only case where income is
significant is in influencing the demand for modern care among children aged 5-17
years (with an income elasticity of 0.42). However, two facts are relevant in putting
the nonsignificance of income i n perspective: (1)the significant negative effect of
household size implies that the demand for modern health providers does increase
with household income per capita, especially for children aged 0 - 4 and 5-17 years
of age; and (2) as noted earlier, household income has strong positive effects on the
demand for quality (i.e., unit values) of both modern and traditional health services.
4. The cross-price elasticities of demand for health care-estimated by few
previous studies for developing countries-are positive for all three age groups (al-
though not significantly so for children under 5), indicating that traditional and mod-
ern health services are (gross) substitutes for each other. The estimated elasticity
of demand for modern care with respect to the price charged by traditionai health
providers is 0.63 for children aged -5-17 years and 0.51 for adults. The correspond-
ing cross-price elasticities for traditional health services are extremely large (25.43
and 10.02 for persons aged 5-17 and 18 and over, respectively), again reflecting the
extremely low levels of usage of traditional healers. These estimates indicate that an
increase in the price of modern health services can cause a significant shift away
from the modern to the traditional health sector.
S. The effects of schooling of the individual* and that of the household head
on health-care demand also differ between the conditional and unconditional mod-
*An individual’s own schooling was inc:luded in the demand equations only for adults over 17 years of
age, since it is likely to be enclogrnoiis for rhildrrn.
HEALTH
COUNTRY I 13
els. While the former suggest that the demand for modern health care is strongly
(positively) influenced by an individual's schooling or the schooling of the household
head, the unconditional estimates shown neither or these effects to be significant.*
The latter result reflects the fact that the unconditional estimates confound the effects
of schooling on morbidity (which are likely to be negative), the reporting of morbid-
ity (positive), and the treatment of illness episodes from modern health providers
(positive).
6. Food prices, which have not been included in the demand for health care
by previous studies, are observed to generally have very strong effects on the de-
mand for both traditional and modern health care. Most of the estimated food price
effects on the demand for traditional care are positive, implying that food and tra-
ditional health care inputs are viewed as substitutes. However, the probability of
choosing a modern provider declines with most food prices, implying that food and
modern health-care inputs are complementary. Thus, the effect of rising food prices,
especially the price of rice and sugar, appears to be to shift people from modern to
traditional health providers.
7. Among all three age groups, age is associated with an increased (uncon-
ditional) probability of using modern providers. Since age has generally no effect
on the conditional probability of choosing a modern provider,i this suggests that ei-
ther true morbidity or the reporting of morbidity increases with age. Among adults,
morbidity is likely to increase with age; however, it is likely that, among children
under S, it is the perception and reporting of morbidity-not morbidity itself-that
increases with age.
8. Health insurance coverage and employee/self-employed status of the
household head have strong influences on the demand for modern medical care. The
magnitudes of the estimated health insurance coverage effects are especially large
(with insured individuals aged 04,5-17, and 18and over having 315.7%, 1751.7%,
and 114.7%, respectively, greater demand for modern care than uninsured individ-
uals). However, only the estimate for adults over 17 is significant. A possible reason
for the nonsignificance of the insurance variable is that it is highly correlated with
the employee status variable, since all government employees in Indonesia as well
as most salaried individuals employed in the formal business sector are covered by
health insurance schemes. Likelihood-ratio tests showed that the employee status
and insurance coverage variables were jointly significant in determining the choice
of modern health providers for all three age groups.
*For adults, both own schooling and the household head's schooling do significantly reduce the con-
ditional and unconditional proba1)ility of using a traditional health provider. For children aged 5-17
years, schooling of the household head is associated with a significantly lower unconditional probability
of choosing a traditional healer.
?For children aged 5-17 years, age does have a significant, although numerically small, effect on the
conditional probahility of using a modern provider.
I14 DEOULIKAR
9. Relatively few of the household composition effects are significant in in-

fluencing the choice of providers. An increase in the proportion of males aged 5-14,
35-44, and 45-54 in the household significantly reduces the use of modern health
providers for children aged 0-5 years. For children aged 5-14, there are no signifi-
cant household composition effects. Finally, an increase in the population of males
aged 5 4 4 and females aged 5-14 years in the household significantly reduces the
use of modern health providers by adults.
10. The environmental hygiene variables generally have significant and pos-
itive effects on the (unconditional) probability of using modern providers. Both the
availability of safe (piped) drinking water and organized trash collection serve to in-
crease the demand for modern health providers among persons of all ages,* implying
that environmental hygiene and modern medical care are viewed by households as
complementary inputs in the production and maintenance of health status.
VI. CONCLUDING REMARKS
The empirical results in this chapter clearly show that the demand for both tradi-
tional and modern health services in Indonesia is highly responsive to own price
when it is not conditioned on the reporting of morbidity and when quality variations
are purged from the price of health services. Indeed, unlike most previous studies for
LDCs that show very small price effects on health-care demand, the price elasticities
estimated here are large in magnitude (-0.13, -1.64, and -0.82 for modern care
and -2.2, -9.7, and -6.2 for traditional care for children under 5, children aged
5-17 and adults over 17, respectively).t The fact that the estimated own-price elas-
ticities for modern health providers are large and significant in the unconditioned
probability model but small and insignificant in the conditioned probability model
suggests that the price of care in the modern sector strongly reduces the reporting
of morbidity by respondents, so that health-care demand models estimated only on
the sample of individuals self-reporting an illness episode are likely to significantly
understate the total effect of health-care prices on the demand for care in the modern
health sector.
Thus, these findings cast doubt on the assumption, commonly maintained in
the literature, that increasing user fees at government health facilities will have lim-
ited effects on the utilization of health services and thus enable governments to raise
health-sector revenues. The empirical results clearly indicate that the utilization of
modern health services, especially by children aged 5-1 7 and adults, will decline
*The single exception is the effect of garbage c.ollec.tion on the demand for modern providers among
adults, which is not statistically signific-ant.
?The demand elasticities for traditional healer services are particularly large because the sample mean
level of usage of traditional healers is very low.
HEALTH
COUNTRY I 15
sharply if prices in the modern health sector are raised without simultaneously im-
proving the quality of health services.
The empirical results also indicate that a number of other prices-in partic-
ular, the price for health services charged by traditional healers and the price of
food staples (namely rice and sugar)-have significant effects on the demand for
modern health services. An increase in user fees charged by traditional healers
is associated positively, while an increase in the price of rice and sugar is asso-
ciated inversely, with the utilization of modern health services. Interestingly, the
availability of piped drinking water and organized trash collection in the commu-
nity, both of which are health-improving public interventions, serves to increase the
use of modern health providers. These results thus imply that individuals view tra-
ditional and modern health services as substitutes, modern medical care and food
inputs as complements, and modern health care and environmental hygiene inter-
ventions as complements, in the production and maintenance of health status. There
are obvious implications here for food price policy. Likewise, the empirical find-
ings indicate that there may be important positive externalities (in the form of in-
creased use of modern health services) to environmental hygiene and sanitation
interventions.
The empirical estimates also imply a massive increase in the demand for mod-
ern health services with an expansion in health insurance coverage in the Indone-
sian population. Currently, a very small fraction of households-only those having
a member in government or formal-sector wage employment-are covered by health
insurance schemes. However, like many other LDCs, Indonesia has been experi-
menting with expanded health insurance coverage for a much larger proportion of
the population.
Finally, the empirical results suggest that income growth alone is unlikely to
increase the utilization of modern health services in Indonesia. With increasing in-
comes, however, individuals are likely to purchase higher quality of health services
(i.e., spend more per illness episode).
ACKNOWLEDGMENTS
I would like to thank Jeffrey Hammer and Martin Ravallion, with whom I have had
numerous discussions on this topic. This chapter is a product of a World Bank re-
search project, “Determinants of Nutritional and Health Outcomes in Indonesia and
Implications for Health Policy Reforms,” in which both of them participated. I am
grateful to the World Bank Research Committee for funding the research. I would also
like to acknowledge the useful comments of an anonymous referee and of seminar
participants at Harvard, Yale, and the University of Pennsylvania, where an earlier
version of this work was presented. The responsibility for all errors rests entirely
with me.
I16 DEOMLIKAR
REFERENCES
Akin, J. S., N. Birdsall, and D. de Ferranti (1987), Financing Health Services in Developing
Countries, The World Bank, Washington, D.C.
Akin, J. S., C. Griffin, D. K. Guilkey, and B. M. Popkin (1984), The Demand for Primary
Health Care in the Third World, Littlefield, Adams, Totowa, NJ.
Akin, J. S., C. Griffin, D. K. Guilkey, and B. M. Popkin (1986), The Demand for Primary
Health Care in the Bicol Region of the Philippines, Economic Development and Cul-
tural Change, 34, 755-782.
Alderman, H. and P. Gertler (1988), The Substitutability of Public and Private Medical Care
Providers for the Treatment of Childrens' Illnesses in Urban Pakistan, Living Standards
Measurement Study Working Paper, World Bank, Washington, D.C.
Behrman, J. R. and A. B. Deolalikar (1988), Health and Nutrition, in Hollis Chenery and
R. N. Srinivasan (eds.), Handbook o j Development Economics, vol. 1, North-Holland,
Amsterdam.
Birdsall, N. and P. Chuhan (1986), Client Choice of Health Treatment in Rural Mali, mimeo,
Population, Health and Nutrition Department, World Bank, Washington, D.C.
Cornea, A. P., R. Jolly, and F. Stewart (eds.) (1987), Adjustment with a Human Face, vol. I,
Clarendon Press for UNICEF, Oxford.
Deaton, A. (1988), Quality, Quantity and Spatial Variation of Price, American Economic Re-
view, 78,418-430.
Deolalikar, A. B. and P. Vashishta (1992), The Utilization of Government and Private Health
Services in India, mimeo, University of Washington, Seattle.
Gertler, P., L. Locay and W. Sanderson (1987), Are IJser Fees Regressive? The Welfare Im-
plications of Health Care Financing Proposals in Peru, Journal of Econometrics, 33,
67-88.
Gertler, P. and J. van der Gaag (1990), The Willingness to Payfor Medical Care: Evidence from
Two Developing Countries, Johns Hopkins University Press, Baltimore.
Gilson, L. (1988), Government Health Care Charges: Is Equity Being Abandoned? EPC Pub-
lication No. 15, London School of Hygiene and Tropical Medicine.
Goldman, F. and M. Grossman (1978), The Demand for Paediatric Care: A Hedonic Appraisal,
Journal ojPolitica1 Economy, 86,259-280.
Government of Indonesia and UNICEF (1989), Situation Analysis of Children and Women in
Indonesia, Jakarta.
Grossman, M. (1972), On the Concept of Health Capital and the Demand for Health, Journal
of Political Economy, 80.
Heller, P. (1981), A Model of the Demand for Medical and Health Services in Peninsular
Malaysia, Social Science and Medicine, 16, 267-284.
Jimenez, E. (1987), Pricing Policy in the Socid Sectors: Cost Recovery for Education and
Health in Developing Countries, Johns Hopkins for The World Bank, Baltimore.
Manning, W. G., J. P. Newhouse, N. Daan, E. Keeler, B. Benjamin, A. Leibowitz, M. S. Marquis,
and J. Zwanziger (1987), Health Insurance and the Demand for Medical Care, mimeo,
The Rand Corporation, Santa Monica.
Mwabu, G. (1986), Health Care Decisions at the Household Level: Results of Health Survey
in Kenya, Social Science arid Medicine, 22, 313-319.
Newhouse, J. P. and C. E. Phelps (1974), Price and Income Elasticities for Medical Services,
in M. Perlman (ed.), The Econ,omic.s oj'Hecdth and Medical Care, Wiley, New York.
HEALTH
COUNTRY I 17
Rosset, R. N. and L. F. Huang (1973), The Effect of Health Insurance on the Demand for
Medical Care, Journal of Political Economy, 81,281-305.
Schwartz, J. G., J. S. Akin, and B. M. Popkin (1988), Price and Income Elasticities of Demand
for Modern Health Care: The Case of Infant Delivery in Indonesia, The World Bank
World Bank (1987), Health Center Expenditure in Indonesia, Background Paper I11 for Health
Planning and Budgeting, Washington, D.C.
World Bank (1992a), Viet Nam: Population, Health and Nutrition Sector Review, Report No.
10289-VN, Country Department I, East Asia Regional Office, The World Bank.
World Bank (1992b), World Development Report 1992, Oxford University Press for The World
Bank, New York.
On Mobility
Esfandiar Maasoumi
Southern Methodist University, Dallas, Texas
1. INTRODUCTION
Mobility in any social hierarchy is an indication of opportunity. It is therefore a mea-

sure of fairness in any economic system. In a real sense, mobility should be of greater
concern to policymakers and analysts than such other important concerns as inequal-
ity. It may be understandable and socially tolerable for the young, say, to have less
assets or income, but the likelihood and the opportunity to move up and earn in rela-
tion to effort is surely a decidedly desirable social goal. What matters in this context
is “lifetime equity” rather than instantaneous equality.
There are essentially two types of movements that constitute mobility. One is
a “growth” or “structural” mobility which may arise from a general economic move-
ment up, or down. The other is the “lateral type,” or “exchange mobility” which
obtains when individuals, households, or groups of such units move from one state
to another. Since there are at least some temporal limits to such upward movements,
mobility, particularly of the lateral type, must be “equalizing.” Thus welfare com-
parison between any two mobility states necessarily requires functionals that reflect
degrees of social preference for equity, exchange mobility, and growth (new oppor-
tunities).
There are two important questions that need attending to. One is how does
one characterize economic status? The other is, how does one evaluate and compare
different mobility situations? The latter question poses the issue of why is mobility
socially desirable and, if it is, what kinds of social welfare functions (SWFs) rep-
resent our preferences for mobility? The first question is all too often resolved by
using “income” as a proxy. Such a proxy is evidently more meaningful the more mar-
ket oriented an economy and a culture, and the less is the level of in-kind and other
transfer payments and entitlements.
I19
I20 MAASOUMI
The reliance on SWFs for evaluating and comparing welfare situations has
provided a scientifically useful tool that provides for economy of thought, as well as
discipline, since it forces a declaration of principles that are too often implicit. In the
instant case, the desirability of both types of mobility suggests functionals that are
increasing in incomes (for the growth component), and ultimately equality preferring
(for the exchange component). While changes in earnings and incomes have been
and should be studied with the aim of identifying significant “causal” factors, the
evaluation of an existing degree of mobility requires welfare comparisons.
There are currently at least two complementary lines of analyzing mobility.
The older approach requires specification of transition probabilities between social
states, and a welfare evaluation of the transition matrices that are estimated from the
existing data. This clearly requires detailed data, with large panel data being the best
source for sufficient cell repetitions which is necessary for reliable statistical infer-
ence. There are numerous mobility indices which are mappings from the transition
matrices to scalars. Shorrocks (1978a) and Geweke et al. (1986)provide systematic
discussions of criteria for sensible mobility indices.
In line with a general direction toward unanimous partial ordering in the liter-
ature on inequality, useful welfare ran king relations have been developed for tran-
sition matrices which rekindle the essential role played by Lorenz and generalized
Lorenz criteria. While the SWF evaluation has begun to be emphasized in this ap-
proach, the task of devising consensus SWFs over matrices remains a challenge
reminiscent of that faced in the multidimensional analysis of inequality. I provide
an account of this line of inquiry in Section 111.
The second approach was initiated by Shorrocks (1978b) and generalized by
recasting mobility as a multidimensionality question as in Maasoumi (1986); e.g.,
see Maasoumi and Zandvakili (1986, 1989, 1990). In this approach inequality in-
dices are computed for multiperiod incomes, that is, a type of “permanent income”
measured over more than a single period, and mobility indices with a profile of
equalization over time are obtained. The latter are directly related to and inter-
pretable by the familiar classes of increasing and equality preferring (Schur con-
cave) SWFs. It will be seen that the more recent welfare theoretic development of
the transition matrices approach is converging to the same welfare comparisons,
similar notion of long-run or “permanent” income, and thus similar preferred mo-
bility indices! Quite general mobility indices are proposed and empirically imple-
mented by Maasoumi and Zandvakili (1989, 1990). This approach is also demand-
ing of data and, like the first approach, it is ideally implemented with plentiful
micro data. But the Maasoumi-Shorrocks-Zandvakili (MSZ) measures may be ad-
equately estimated with data grouped by age, income, education, etc. This is use-
ful since much data is made available in this aggregated form, and the approach is
particularly focused on “equalization” between income groups in a way that allows
controlling for sources of heterogeneity among individuals and households. Such
controls with aggregate data appeal to some who may wish to tolerate some diver-
ON MOBILITY 12 1
sity due to such things as age or education. I present an account of this approach
in Section 11.
Our survey makes clear that both approaches share a concern for the distri-
bution of a welfare attribute as well as its evolution. And the convergence in both
approaches to rankings by Lorenz-type curves is of econometric significance. There
is now a well developed asymptotic inference theory for empirically testing for or-
der relations such as stochastic dominance, Lorenz dominance, and concentration
curves. This type of testing is a crucial first step since comparing mobility indices
(statistically or otherwise) is of questionable value when, for instance, generalized
Lorenz curves cross. Section IV contains a brief account of the statistical tools that
are available for inference on both the indices of mobility and order relations.
A loosely related strand of econometric research seeks to specify statistical
models of earnings or income “mobility.” While it is true that such models are fo-
cused on explaining “change” and “variation” which are not as meaningful as “mo-
bility,” they can shed light on significant explanations of earnings changes, as well as
account for heterogeneity. This is consequential for policy analysis. Further, there are
econometric models that seek to fit transition probabilities. Such studies are directly
useful for not only estimating transition matrices, but for explaining the estimated
probabilities. We do not delve into this empirically substantive area in this survey.
Section V concludes with several empirical applications of the “inequality reduc-
ing” mobility measures. An insightful survey of income mobility concepts is given
in Fields and Ok (1995). Absolute measurement of income mobility and partial or-
dering of absolute mobility states is treated in Mitra and Ok (1995).
II. INEQUALITY REDUCING MOBILITY INDICES

A. The MSZ indices
The most common approach to comparing welfare situations is based on indices.
For example, two income vectors depicting the respective distribution of households
at two points in time, or in different regions, are explicitly and implicitly ranked
by scalar measures of such things as inequality, mobility, or poverty. The welfare
theoretic underpinnings and limitations of index-based comparisons are now well
understood. Notwithstanding these serious limitations, the need for “measurement”
has led to the development of many “sensible” indices, supplemented with statistical
tools which are based upon the asymptotic distributions of the indices; see Maasoumi
(1996) for a recent survey. This “index-based” approach is limited because of a lack
of unanimity on the acceptable index even when there is broad consensus about cer-
tain normative principles. But the main difficulty is a consequence of the difficulties
of welfare “comparability” as well as the appropriate cardinalization of the class of
admissible criterion functions.
122 MAASOUMI
In the case of mobility indices, Shorrocks (197833) and Maasoumi and Zand-
vakili (1986)argued that a more reliable measure of individual welfare and the dis-
tribution of incomes is to be obtained by considering the individual or household’s
“long-run” income. They argued that such a measure of income computed over in-
creasingly longer periods of time would remove transitory and some other life cycle
related movements which are picked up by year-to-year comparisons of income dis-
tributions. The annual “snapshots” are incapable of accounting for mobility and re-
turns to investments and/or human capital. The effects of seniority alone may make
the notion of income inequality meaningless.
These authors are therefore concerned with dynamics of income distribution,
and are thus accounting for, and challenged by, a fundamental lack of homogeneity
among households. The natural labelling of individuals at different points in their
life cycle is an essential form of heterogeneity which contradicts the common as-
sumption of symmetry (anonymity) which plays a crucial role in much of the welfare
theory that underlies inequality analysis. Shorrocks (1978b) proposed the simple
sum of incomes over T periods as the aggregate income. Maasoumi and Zandvakili
(1986, 1989,1990) proposed more general measure of “permanent income” encom-
passing the simple average and sum. This author recognized the essential multidi-
mensionality of the mobility analysis and proposed its treatment on the basis of the
techniques and the concepts developed in Maasoumi (1986). Consequently, general
functions for “permanent incomes” where developed which are maximum entropy
(ME) aggregators.
The next step in this development is to analyze the inequality in the perma-
nent incomes and compare with single period inequalities. A weighted average of the
latter was used by these authors to represent “short-run” inequality over any desired
number of periods. Clearly, a distribution of permanent incomes is being compared
and ranked with a reference short run income distribution. In view of this, all of the
rich welfare theory supporting Lorenz and generalized Lorenz (second order stochas-
tic dominance) dominance relations, as well as the convex inequality measures con-
sistent with such relations, comes at the disposal of the analyst for evaluating mo-
bility profiles and dynamic evolution of the income distribution. At first sight, this
appears an unnecessarily restrictive setting for defining the welfare value of mobil-
ity. Further reflection suggests that this is not so. Indeed, we will see that the recent
development of a welfare theory basis for the alternative of transition matrices has
made clear a certain inevitability for the role of the same welfare criteria and, there-
fore, the same types of mobility measures as the Maasoumi-Shorrocks-Zandvakili
indices!
Let X;,be the income of the ith individual in the tth state (period, say); i E
[ 1, n], and t E [ 1, T]. Let Si (Xi;
a,#?)denote the ith individual’s permanent income
(living standard?) over a number of periods k = 1 , 2 , . . . , T , and S = (S1, Sz,
. . . , S,) denote the vector of such incomes for the n households or individuals. The
inequality measures which are consistent with a set of axioms to be described below
ON MOBILITY I23
are represented by I,(.). Let X denote the welfare matrix with the typical element
Xit and denote its ith row by X ; and its tth column by X‘. The latter is the income
vector in the tth periodktate.
The k-period long-run inequality is given by Z,(S), and short run inequality
may be represented by I; = atZ,(X‘), for k = 1 , 2 , . . . , T . The vector a =
( a l ,a2, . . . , a ~represents
) the weights given to the income distribution in different
periods, such as the ratio of the mean income in the period, p,, and the overall mean
of incomes in all the k periods under analysis.
Shorrocks (197813) proposed S ; = ELl X i t , and p l / p ,the ratio of means just
described, as weights, and Maasoumi and Zandvakili (1986)generalized S; and the
weight functions, suggesting the following index of mobility:
with
R,k = -, as a measure of stability

‘yk
where it is to be noted that S ( - ) is measured over the same k periods. When S is
quasi-concave R E [0,1] follows from the convexity of the inequality measures con-
sidered. A “mobility profile” is generated by depicting R as k increases over its
range. Both Shorrocks (1978b) and Maasoumi and Zandvakili (1986, 1989, 1990)
report empirical studies of the US data from the PSID. Section V describes some of
these studies.
Bounds on R t may be established using the decompositions of multidimen-
sional measures of inequality given in Maasoumi (1986,Propositions 1 and 2). PT =
(Rl,RB,. . . , R T )is a stability profile which can also reveal the effect of increasing
smoothing of the income variable starting from R I = 1. Other than this smooth-
ing out of the short-run effects, R, is capable of revealing durable “mobility” toward
equalization in a way that may be obscured by looking at each I , ( X t ) , t = 1, . . . , T .
This is most clearly seen by considering a situation in which only a permutation of
the income vector has occurred between two periods. As is well known, this leaves
our anonymous (symmetric), relative inequality measures unchanged, but the dis-
tribution of the aggregated incomes over the two periods will be changed, possibly
dramatically (unless there is perfect equality to begin with!). Consequently, I , (S)
and R, will measure any mobility over the two periods.
Given that the indices I,(.) are supported by Schur-concave SWFs, ordering
mobility states by the class of measures in R is consistent with unanimous partial
orderings conducted with generalized Lorenz curves and related orderings. For this
reason, and to help in the discussion of Section 111, it is useful to have a brief account
of the relevant welfare theory.
124 MAASOUMI
I. The FundamentalWelfare Axioms

Kolm (1969) and Atkinson (1970) provided the modern and influential formalizations
of the relationship between SWFs and inequality measures. The need for establish-
ing this important relationship is now recognized widely when analyzing poverty and
mobility. However, the axiomatic SWF approach does not by itself identify unique
indices even when a particular set of normative properties are consented to by a ma-
jority. To appreciate this point, consider the Atkinson family of inequality measures
for income x having mean p x :
where F is the c . d . f . of income. Similarly, the generalized entropy (GE) family of

indices is
Ordinally Ir is equivalent to the coefficient of variation and the Herfindahl in-

dex, and includes the variance of logarithms and Theil’s first and second measures,
10 and 1-1, respectively. Also, up to a monotonic transformation, there is a unique
member of GE corresponding to each member of the Atkinson family. v = - y is the
degree of aversion to relative inequality; the higher its absolute value the greater is
the sensitivity of the measure to inequality (transfers) in the tail areas of the distri-
but ion.
Employing the functional theory first developed in “information theory” for
identifying appropriate measures of divergence between distributions, (Maasoumi
1993),and noting that inequality and many other indices are similar measures of
divergence, one puts down an explicit set of normative properties (axioms) which
inequality indices and/or SWFs must satisfy. Using these axioms as explicit con-
straints on the function space one then obtains the appropriate inequality index. To
exemplify, let us follow Bourguignon (1979) or Shorrocks (1980, 1984) in their dis-
cussion of the “fundamental welfare axioms” of symmetry (anonymity), continuity,
principle of transfers, and additive decomposability. Combined with homogeneity,
these axioms identify members of the GE as the desirable family of relative inequal-
ity measures.
Axiom I (Anonymity). The inequality index (function) is symmetric in in-
comes.
Axiom 2 (Homogeneity). Invariance to scalar multiplication.
Axiom 3. Continuity.
ON M O B I L I ~ 125
Axiom 4. Principle of transfers.

This requires that inequality decrease if we redistribute from a single richer
individual to a poorer one, leaving their respective ranking and all the other individ-
uals’ incomes unchanged.
Axiom 2 is a serious limitation as it restricts attention to “relative” inequality.
This is so since this requirement implies mean invariance-doubling everyone’s
income would leave inequality unchanged.
The class of functionals satisfying Axioms 1 - 4 is large! Also, any further ax-
ioms are less likely to command unanimity. In fact, any further requirements must be
justified by plausible considerations of such things as, heterogeneity amongst house-
holds and/or individuals, policy, empirical necessity, and practical interest. The most
commonly invoked of such requirements is:
Axiom 5. Additive decomposability.
This requirement, later strengthened as an “aggregation consistency’’ axiom
by Shorrocks (1984), says that total inequality must be the sum of a “between-
group” component, obtained over group means, and an additive component which
is a weighted sum of “within-group” inequalities. This kind of decomposability is
very useful for controlling and dealing with heterogeneity of populations, and as a
means of unambiguously identifying the sources of inequality and those that are af-
fected by it. In the context of mobility analysis, this property further serves to identify
the contributions of each time interval under consideration.
If the additive decomposability requirement of Axiom 5 is imposed, such mea-
sures as Gini, the correlation coefficient, and variance of logarithms must be ex-
cluded. The latter measures provide ambiguous decompositions of overall inequal-
ity by population subgroups (Shorrocks 1984). For the GE family, for instance, a
discretized (estimation) formula that helps to demonstrate its decomposability is
where Z; is the between group inequality computed over the group mean incomes,
I‘ is the inequality within the rth group, r = 1 , 2 , . . . , R , n, is the number of units
in group I, and X, is total income of the rth group.
2. Aggregate or Permanent Income

As in Maasoumi (1986),we consider the following weighted generalized entropy
measure of divergence between S, on the one hand, and X‘,t = 1 , 2 , . . . , T , we
have
126 MAASOUMI
where a,s are the weights attached to each period. Minimizing Dg with respect to S ;
such that C
Si = 1, produces the following “optimal” aggregate income functions:
These are, respectively, the hyperbolic, the generalized geometric, and the
weighted means of the incomes over time. Noting that the “constant elasticity of
+
substitution” -0 = 1/(1 p), these functional solutions include many of the well-
known utility functions in economics, as well as some arbitrarily proposed aggregates
in empirical applications. For instance, the weighted arithmatic mean subsumes the
simple total income discussed earlier, and a popular “composite welfare indicator”
based on the principal components of X , when at’s are the elements of the first eigen-
vector of the X’X matrix (Maasoumi 1989a).The “divergence measure” I),(.) forces
a choice of an aggregate income vector S = (Sl , S2, . . . , S,) with a distribution that
is closest to the distributions of its constituent variables. This is desirable when the
goal of our analysis is the assessment of income distribution and its dynamic evolu-
tion. The entropy principle establishes that any other S would be extra distorting of
the objective information in the data matrix X . The distribution of the data reflect the
outcome of all optimal allocative decisions of all agents in the economy (Maasoumi
1986).
The next step in constructing general mobility indices as proposed by Maa-
soumi and Zandvakili (1986) is the selection of a measure of inequality. The GE
index described above was computed for the Si functions just obtained. It is instruc-
tive to analyze this measure in the discrete case:
Io(S) = C ST log , (z) Theil’s first index
1-1 (S) = C p ; log ($) , Theil’s second index
where pi is the ith unit’s population share (typically = l/n), and Sr is S; divided by
the total K + S;.
ON MOBILITY I27
These inequality indices are normalized isoelastic transformations of the ag-

gregate income functions S;. As such they are “symmetric,” “homogeneous,” and
consistent with the Lorenz criterion. They will be homogeneous with respect to ev-
ery X t , the tth columdperiod in X, if in all the above one works with the matrix of
shares, x = (xi,). While this will not change the functional solutions given above,
it requires a rather unusual assertion that individual well-being depends on rela-
tive incomes. Alternatively, one can impose a general form of scale invariance at the
outset (Maasoumi 1996b).
Useful decomposability properties are possessed by these measures both in
population groups and in the time directions.
Theorem I (Decornposability of GE). Let xit = Xit/Tt, Tt = xi Xit, =
TJK, Z,(XL) = the GE inequality in the tth period, and 6 , = a,/ af.Then
+
(i) r f l y = -B, we have
(ii) rfthe marginal distributions are identical-i.e., x t = x k , Vt and k-we

have
Z,(S) = Zy(Xt), any t E [l,T ] (14)
(iii) For Theil’sfirst and second measures (p = 0, -l), we have

m
Zo(S) = C t Zo( XL)- D-1(x, s*;C)

;= 1
Z-l(S) = c T
t= 1
stz-1 ( X t ) - Do(x,s*;6)
where by application of L’H6pital’s rule to Dg(.)defined earlier, we have
DO(S*,x; 6) = 6, s; log
t i
and Si defined at B = 0 and
and Si defined at B = -1.

128 MAASOUMI
Proof: See Maasoumi (1986, Propositions 1-2).

In view of the nonnegativity of the D(.)terms in part (iii) it is clear that multi-
period inequality is no more than the weighted average of inequalities in the single
periods. This is due to the intertemporal “substitution” effects and a consequence of
the convexity of the inequality measures.
B. Welfare Properties of R, Indices

In the spirit of the “axiomatic approach” discussed above for inequality measures,
Shorrocks (1992) lays down a set of desirable properties that any mobility index
may satisfy. These properties are quite suggestive and also useful in any discussion
including that of mobility indices based on transition matrices. The desirability of
some of the properties may be evident while that of others may be less obvious or
compelling. In fact some of the proposed properties are not consistent, so one or
more has to be abandoned.
The domain of income structures X over which the measure is well defined is
given as follows.
(A I ) Universal domain. Suppose the feasible set of n units and T periods for
positive incomes is denoted
X n =~{ X I dim X = n x T ; X;l > 0)
A mobility measure should be well defined for all X E X , where
However, there are some types of income structures that are eliminated from
this set. An example is situations in which, in every period t , all individuals receive
the same income pCLt.
The MSZ indices satisfy this requirement. Other indices may not; see below.
(A2) Continuiv. The degree of mobility varies continuously with the incomes
in X.
The MSZ indices satisfy this property.
(A3) Population symmetry. If X’ = nX for some permutation matrix l7,then
X and X’ are equally mobile. This requires “rank invariance” to be acceptable since
X and X‘ may not given the same ranking of individuals.
The term “population symmetry” is used because in the analysis of mobility,
unlike inequality, we can consider permutation of the time period distributions (the
columns of X ) as well as permutations of the individual income profiles (the rows of
X ) . One may thus define time symmetry separately:
ON MOBILITY I29
(A3’) Time symmetry. A mobility measure is time symmetric if X and X‘ are

equally mobile wherever X = X l 7 for some permutation matrix l7.
Broadly speaking, time symmetry suggest that we care about the distribution
of an individual’s income receipts over time, but not about the time sequence of those
receipts. This is more than time symmetry implies, however, since a permutation l7
swaps all the incomes in two periods s and t , not just X i , with X ; , for a single person
i. But even the weaker idea of time symmetry, as stated, is objectionable since we
may not be indifferent to the time sequence of incomes as, for instance, between the
situation in which incomes are originally different and then become equal, and the
time symmetric equivalent structure in which incomes are initially equal, and then
become different.
The MSZ indices satisfy time symmetry, but this may or may not be a desirable
quality.
(A4) Population replication invariance. X and X’ are equally mobile whenever
X’ is a population replication of X .
(A4’) Time replication invariance. A mobility measure is time replication in-
variant if X and X’ are equally mobile whenever X’ is a time replication of X .
X’ E Xrn,T is apopulation replication of X E XnT if r is a positive integer
and Xj, = Xi, wherever j = k n + i for some integer k 2 0. That is, X’ is the
aggregate income structure for subpopulations each having the income structure X .
Similarly, X’ E X n , k ~ is a time replication of X E X n if~ k is a positive integer
and X’ = [ X , X , . . . , X I . Invariance with respect to population replication is the as-
sumption typically used in inequality measurement to compare income distributions
for different sized populations.
The MSZ index with additive aggregate income functions exhibit replication
invariance with respect to time and the population. However, replication invariance
with respect to time may in any case be regarded as a suspect property, for much the
same reason as time symmetry. It implies that the degree of mobility is unchanged if
the pattern of incomes received in the first T period, say, is exactly repeated for all
individuals in the next T periods, and so on. “But this does not take into account the
fact that the distribution of income in period T m a y be radically diferent f r o m that
in period 1, so moving f r o m period T to period T i- 1 (and hence back to period 1
incomes again) m a y be quite a jolt. The desirability of time replication invariance is
therefore less than transparent” (Shorrocks 1992).
(AS) Normalization. Mobility is a minimum whenever X is completely im-
mobile.
Definition. A structure X is completely immobile if and only if XiS/ps =
Xit/pt for all i , s, and t .
Completely immobile income structures perform a role similar to that played
by completely equal distributions in inequality measurement. But it is implicit in
130 MAASOUMI
the above definition that relative mobility is measured since one is only looking at
changes in relative incomes. This rules out pure exchange mobility. It is a type of
rank invariance property that goes beyond requiring that a structure in which all
individuals incomes are constant over time is completely immobile.
The MSZ indices satisfy this property. They also satisfy the following stronger
normalization property:
(A6) Strong normalization. Mobility is a minimum if and only if X is com-

pletely immobile.
The other benchmark of interest is perfect mobility. Ideally we would define
such a state as one in which the probability of achieving an income level in pe-
riod t + 1 is independent of the income received in period t. The concept of per-
fect mobility is difficult to formulate in terms of observed income structures X . A
plausible parallel that falls short of “independence” is to require that not only are
X , and X , uncorrelated, but also any, arbitrarily transformed vectors @,(X,) and
& ( X t ) , where @,(X,) -= ( 4 , ( X l L ) ,@ l ( X 2 , ) , . . . , 4 t ( X n ~ ) )for
’ some real-valued
function @ t . Thus:
Definition. An income structure X = [ X , , . . . , X T ] is perfectly mobile if and

only if $,(A’,) and # L ( X , ) are uncorrelated for all s, t , and all real functions $,, 4,.
An example is given by Shorrocks (1992).Suppose there are J income levels
yll, y21, . . . , yj1 at a time 1 and K income levels y12, y22, . . . , yK2 at time 2, and let
X consist of the J K income profiles (51, yk2) for j = 1 , . . . , J and k = 1, . . . , K .
Now consider
Then the sample covariance of the observed income profile can be shown to
be zero.
The associated property is defined as follows:
(A7) Perfect mobility. Mobility is a maximum whenever X is perfectly mobile.

Note that this requires that all perfectly mobile structures have the same index
value as well as this common value being a maximum. Not all members of the MSZ
family satisfy this requirement. See below for a member that partially satisfies (A7).
But none of the members satisfy a strong version of (A7) defined as follows:
(A8) Strong perfect mobility. Mobility is a maximum if and only if X is per-

fectly mobile.
(AS) Unit interval range. Mobility index should be in [0, 13.

This is convenient and satisfied by MSZ and many other indices.
ON MOBILITY 13 I
(AIO) Scale invariance. X and X’ are equally mobile if X’ = h X , for any

scalar h > 0.
And,
(A I I ) Intertemporal scale invariance. X and X’ are equally mobile whenever
X’ = X A , for any positive diagonal matrix A .
Shorrocks’ index based on sum of incomes satisfies (A10) but not ( A l l ) . But
a generalization to a weighted sum aggregate proposed in Maasoumi and Zandvakili
(1986, 1990) does satisfy both (A10)and (A1 1).
How do mobility indices (welfare functions) rank intermediate situations be-
tween perfect mobility and total immobility? This requires a careful consideration of
transfer sensitivity of mobility measures and has implications for the class of welfare
functions. Smoothing transfers that are generally considered as equalizing (therefore
preferred) conflict with normalization and change the mean of incomes at the point of
transfer. It is difficult to consider them as mobility reducing. Compensating smooth-
ing transfers, on the other hand, preserve the mean incomes ( p t )but can change
cross section distribution (Shorrocks 1992). Thus the following type of “switches”
are considered which can be seen to be mobility enhancing:
Definition. The income structure X’ is obtained for X by a “simple switch”
if, for some i and j ,
Then
(A I 2) Atlunson-Bourguignon condition (for two periods). The income structure
X’ is more mobile than X whenever X’ is obtained from X by a simple switch.
This condition implies that if income profiles of the two persons i and j are
initially rank correlated, then a switch of incomes in either period enhances mobility.
This condition has not been generalized to T 2 2. The MSZ indices satisfy this
condition when income aggregates are weighted averages.
Shorrocks (1992) studied a particular member of the MSZ family which sat-
isfies all of the above properties except (A8),strong perfect mobility. This member
is obtained from the measures defined by Maasoumi and Zandvakili (1986, 1989)
where aggregate income is E, a t X i t , a,= l / p t ,and the short run inequality is rep-
resented by a,,utZ,(Xt), and Z,(.) is the coefficient of variation, a 2 / p .He refers
to this as an “ideal” index. Shorrocks (1992) also looked at another mobility index
attributed to Hart. The latter index is more conveniently described in relation to
indices defined over transition matrices to which we now turn.
132 MAASOUMI
111. TRANSITION MATRICES
As was argued before, income distributions change over time under the effect of dif-
ferent transition mechanisms. Transition mechanisms affect social welfare by chang-
ing the income distribution. Two societies with the same income distribution at a
point in time may have different levels of social welfare depending on the mobility
of the populations. This requires welfare functions defined over an expanding time
dimension.
In a Markov chain model of income generation, Dardanoni (1993) consid-
ers how economic mobility influences social welfare by following the approaches
of Atkinson (1970), Markandya (1984),and Kanbur and Stiglitz (1986).He consid-
ers the welfare prospects of individuals in society by deriving the discounted stream
of income distributions which obtain under different mobility structures. He pro-
poses a class of SWFs over the aggregates of these welfareprospects, and derives some
necessary and sufficient conditions for unambiguous welfare rankings. Since these
aggregates are the discounted stream of incomes, a special case of the aggregates
proposed by this author and described in the previous section, the two approaches
of this section and the previous one converge when the same welfare functions and
the same measures of “permanent” income are used.
The fundamental inequality theorem states that the Lorenz curve gives the
normatively significant ordering of equal mean income distributions. Inequality in-
dices are difficult to interpret when Lorenz curves cross. In a similar vein, Dardanoni
(1993) derives a partial order of social mobility matrices which can be considered
as the natural extension of the Lorenz ordering to mobility measurement. The de-
rived ordering may provide conditions for an unambiguous welfare recommendation
without employing a specific mobility measure.
Summary mobility measures induce a complete order on the set of mobility
matrices and have the advantage of providing intuitive measurements and firm rank-
ings. However, it is clear that there are substantial problems in trying to reduce a ma-
trix of transition probabilities into a single number. This is very much the problem
of multidimensional inequality measurement addressed by Maasoumi (1986),Ebert
(1995a), and Shorrocks (1995). Dardanoni (1993) offers the following example of
three mobility matrices
PI = [ .35 B
.6 .4
.35 :
] P2 =
[ .6 .3 . l ]
.3 .5 .2 & =
[ ]
.6 .4 0
.3 .4 .3
.05 .25 .7 . 1 .2 .7 .1 .2 .7
The rows denote current state and columns denote future state. Suppose we
use some common summary immobility measures as proposed and discussed by,
for instance, Bartholomew (1982) and Conlisk (1990).Consider the second largest
eigenvalue modulus, the trace, the determinant, the mean first passage time, and
Bartholomew’s measure. These indices are defined below. Any of the three matrices
ON MOBILITY 133
may be considered the most mobile depending on which immobility index is chosen.
This is illustrated in the following table, which shows the most mobile mobility matrix
according to the different indices.
Indices Eigenvalue Trace Determinant

Most Mobile P2 P I ,P i PI
Different mobility rankings obtain depending on the mobility measure adopted.

The welfare-based partial order is similar to the axiomatic welfare analysis exempli-
fied above as it aims to clarify the situation depicted in this example in a sound
fashion.
A. The Welfare Ranking of Mobility Matrices

Consider a discrete Markov chain process for income. Let there be n income states.
Let P = [pij],such that pi; 2 0 and xi
pi; = 1, be the n x n transition matrix,
assumed regular (i.e., Pk is strictly positive for sufficiently large integer k ), so that
the strictly positive steady state probability vector n exists and is the unique solu-
tion to n’ = n’P.The element pi; is the probability that an individual in state i
will be in state j in the following period. n,+l = n L Pdenotes the vector whose ith
element gives the fraction of the population which is in state I at time t + 1. It is
assumed that transitions are independent across individuals and P is constant over
time. Strictly speaking, P need not be square. Restricting attention to this case ex-
ploits the properties of bistochastic matrices. Lastly, income states are in ascending
order.
For a given transition matrix, P , we may derive the implied distribution of
expected lifetime welfare for the individuals who live in the society whose mobility
is governed by P . Consider a society, assumed in equilibrium, consisting of identical
individuals who are born simultaneously and live exactly f o r t periods. The transition
mechanism may be either intra-generational or intergenerational.
Let U = (u1, u2, . . . , U,)’ denote a vector of instantaneous utilities, where U ;
denotes the utility value of income in state i, and VP = (Vl , V2, . . . , V,)’ denotes a
vector of expected discounted lifetime utilities. The typical element V,’ denotes the
expected lifetime discounted utility of an individual beginning life in income class
+ + + +
i, given by the ith element of the vector V” = U p P u p2P2u . - . p t P t u ,
where 0 5 p < 1 denotes the discount factor. This typical element is comparable
with the S ; functions of the previous section. VP will in general depend on the vector
U , on the transition matrix P , on the discount factor p, and on the time period t.As
t + 00, V p = [ I - pP]-’u which, for convenience, may be normalized as
v p = (1 - p)[Z - pP]-’u = P( p ) u (21)

134 MAASOUMI
say. The typical element of P ( p ) , p l i ( p ) , may be interpreted as the average dis-

counted “lifetime” probability of moving from the initial state i to state j .
Dardanoni suggests that transition matrices be ranked according to real-valued
SWFs defined over the vector of lifetime expected utilities VP. This is comparable
with the “inequality reducing” rankings of the previous section where Schur-concave
welfare functions correspond to inequality measures. There is more a priori structure
imposed on the data here by assuming a Markov transition process.
Mobility can occur through general growth in equilibrium income distribu-
tions. This is known as “structural mobility” or the “growth component.” But there
is some intertemporal movement of individuals among the different social classes,
for a given equilibrium distribution of the number of individuals in each class; this
latter effect is defined as “exchange” or “pure” mobility. Dardanoni (1993)isolates
the pure mobility effect by assuming societies with identical steady-state income
distributions. In other words two societies have within each period an identical spot
income distribution, but individuals may move between income states differently
under the two transition mechanisms. Note that under the stated assumptions the
distribution of individuals in each state will be given by the equilibrium vector n,
with the typical element ni indicating the proportion of people in income state i.
This procedure is the dynamic counterpart to the usual static inequality analysis
(e.g., Atkinson 1970), where to isolate the pure inequality effect on social welfare
one considers societies with equal mean incomes. Lorenz rankings would suffice in
that situation.
Allowing for different mean incomes requires consideration of growth. This
would be similar to the analysis of inequality on the basis of generalized Lorenz (GL)
curves, proposed by Shorrocks (1983),and second order stochastic dominance. This
is the extension considered by in the context of mobility matrices.
Following Dardanoni (1993), take two regular transition matrices P and Q
with equal steady-state income distribution vector n’= n’P = n’Q. He considers
the class of symmetric and additively separable (i.e., linear) SWFs xi qp n; which
adds up, for a given U and p, the expected lifetime utility of the individuals in the
population. This is equivalent to the SWFs considered by Atkinson (1970)and Kolm
(1969) for the inequality ranking of income distributions. Noting that n ’ P ( p ) =
n’Q(p), we have
7ri q p = n’vp = n’P(p)u = n’u = n’Q(p)u = n’VQ = yQ (22)

It;
i i
Therefore, given a vector U and a p, any two transition matrices with equal
steady-state income distributions will be indifferent. This result is given by Atkinson
(1983)and Kanbur and Stiglitz (1986)and indicates that we are not ranking mobility
as such, but the social welfare implications of each mobility matrix. The symmetric
additive social welfare functional implies that movement between income states is
ON MOBILITY 135
irrelevant. What is important is the spot distribution at each period since additive
separable lifetime welfares remove any influence that exchange mobility may have
on intertemporal social welfare. Thus additive SWFs take inadequate account of fair-
ness considerations. Under the stated assumptions, the equilibrium Lorenz curve of
the distribution of income will look identical each period under any transition matrix
with equal steady-state distribution, so that any (additive or otherwise) symmetric ex
post SWF defined on the vector of realized utilities will rank the matrices as indiffer-
ent. Yet, under different transition matrices the composition of people in each income
state will be different in each time period. For example, under the identity transition
matrix each individual in the population remains in the same income group as in
the initial situation; on the other hand, if transition is governed by a matrix in which
each entry is equal to l / n , each individual will have the same probability of belong-
ing to any of the n income groups regardless of the initial state. Therefore, though the
equilibrium ex post Lorenz curves associated with each of these matrices could look
identical for each period, social welfare may well be considered different if we take
account of several periods in terms of the position of each individual in the past.
Clearly the natural labelling of welfare units in the context of mobility requires
a relaxation of the “symmetry” assumption, such as (A3), which are replaced by ad-
ditional assumptions on “comparability.” These assumptions are discussed in, for
example, Sen (1970) and Atkinson and Bourguignon (1987). Here the natural “la-
bel” for each individual is hidher starting position in the income ranking. Thus one
restricted SWF would be the weighted sum of the expected welfares of the individ-
uals, with greater weights to the individuals who start with a lower position in the
society. That is, W ( V p ,A) = xi nihiyP,where h = ( h l h 2 , . . . , An)’ denotes a
nonincreasing nonnegative vector of weights. This is a step toward cardinalization
(Maasoumi 1996b). Furthermore, this asymmetric treatment makes sense only if it
is a disadvantage to start at a lower position. With no restriction on the mobility
matrices, this is not necessarily a disadvantage. There could be a transition matrix
such that the lower states are the preferred starting point in terms of lifetime expected
utility. Therefore the additional assumption:
Assumption. Transition matrices are monotone.

A transition matrix is called monotone if each row stochastically dominates
the row above it (Conlisk 1990). In an intergenerational mobility context, a monotone
mobility matrix implies that each child at time t is better off, in terms of stochastic
+
dominance, by having a parent from state i 1 than a parent from state i. In an intra-
generational mobility context, a monotone mobility matrix implies that an individual
+
who at time t is in state i 1 faces a better lottery, in terms of stochastic dominance,
than an individual who is in state i. If w e let y be a n x 1 vector it may be shown
that P y is nondecreasing for all nondecreasing y if and only if P is monotone. Since
P ( p ) is monotone when P is monotone, the expected lifetime utility vector will be
136 MAASOUMI
nondecreasing. Estimated transition matrices are often either exactly monotone or

within sampling errors from being monotone.
Considering two extreme cases, hl = 1 and hi = 0 for all i > 1, which is
“Rawlsian,” and hi = 1 for all i, which is the symmetric case, it is seen that there
is a need for exploring necessary and sufficient conditions on transition matrices for
the unanimous ranking of W ( V p ,A) = xir r i A i y P for all nonincreasing positive A.
Theorem 2. !,it P and Q be two monotone regular transition matrices such
that nf = n’P = n’Q,and for a given p, thefollowing conditions are equivalent:
Proof: See Dardanoni (1993, Theorem 1).

T is the triangular summation matrix with its inverse, T-’, as the first “dif-
ferencing” matrix. For instance, P T transforms each row of P to the cumulative
distribution function. l 7 is a block-diagonal matrix with the typical block being the
n-vector.
This result is an extended horizon version of the first order stochastic domi-
nance relations obtained by Atkinson (1983). Further, if we denote by n ( n )the set
of regular monotone transition matrices, condition (lii) induces (iff) a partial order-
ing ? M , that is reflexive, antisymmetric, and transitive.
If one further assumes monotonicity of the reverse chain-i.e., at each time t
an individual in state i has faced a stochastically dominant lottery than an individual
in state i - 1, the above result would hold for all 0 5 p < 1.
An interesting result concerns the effect of transfers such as smoothing and
“simple switches” considered in an earlier section. Dardanoni considers the follow-
ing dynamic Pigou-Dalton (DPD) transfers: Given integers 0 < i , j , s , k < n, with
+ +
i k < n and j s < n, let us decrease the probabilities of the event, “initial state
+ +
illifetime state j,” and the event, “initial state (i k)/lifetime state ( j s),” by a
quantity 0 5 h 5 1. Simultaneously, increase by the same h the probabilities of the
+ +
events, “initial class illifetime class ( j s),” and the “initial class (i k)/lifetime
class j,” in such a manner as to not violate monotonicity. This transfer is mobility
enhancing as it would leave the row and column sums unchanged and improve the
lifetime status of a poorer individual. Finally noting that for a more mobile situa-
tion there will be smaller covariance between the initial and the lifetime status, the
following general result is established:
Theorem 3. Let P and Q be two transition matrices in n ( n )and let p be
given. Then the following conditions are equivalent:
(i) P(P> P M Q ( p ) -
(ii) The Lorenz curve of permanent income for P lies nowhere below that of Q
V nondecreasing income vectors.
ON MOBILITY 137
(iii)The covariance between initial status and lifetime status is greater under
Q for any nondecreasing score (rank) vectors.
(iv) P ( . ) can be derived f r o m Q ( . ) by afinite sequence of DPD exchanges.
Formby et al. (1995) extend this result by relaxing the assumption of identical
steady states. They note that T ’ n P ( p )y is the generalized Lorenz vector of “perma-
nent incomes,” and show the following result:
Theorem 4. Let P and Q be two monotone transition matrices with a given
discount factor p. Then the following conditions are equivalent:
(li)w(v“,A) 3 W ( V ~A).,
(lii)T ’ n [ ( P ( p )- Q ( p ) l T I0.
(liii)The generalized L0ren.z curve of permanent income for P lies nowhere
below that f o r Q f o r all nondecreasing income vectors y .
(liv) P ( p ) can be derived f r o m Q ( p ) by afinite sequence of DPD exchanges
and simple increments.
Proof By noting that the assumption of steady-state income distribution is not
crucial in proving Dardanoi’s results as well as in Theorem 1, Formby et al. (199s)
prove the equivalence among conditions (li), (lii), and (liii). Also, (liv) implies
(lii). The converse can be similarly proved. Note that each DPD or simple increment
leaves all elements other than the (2, j)th of T ’ n [ ( P ( p )- Q ( p ) ] T unchanged.
Formby et al. (1995) demonstrate a further result which can be useful in em-
pirical testing based on the generalized concentration curves:
Theorem 5. Let P and Q be two monotone transition matrices. Then the fol-
lowing conditions are equivalent:
(li) P(p> Q(P).

(lii) T ’ n P y 5 T ’ n Qyfor all nondecreasing income vectors y, i.e., the “snap-
shot”genera1ized concentration curvefor P lies nowhere below thatf o r Q
f o r all nondecreasing income vectors, if the condition of the monotonicity
of the reverse Mnrkov chain isfiLrther assumed.
This result establishes that if the generalized concentration curve for P lies
nowhere below that for Q, then the generalized Lorenz curve of the “lifetime” for P
also lies above that for Q. If one adopts Dardanoni’s assumption of identical steady
states (no growth mobility), the above result reduces to the ordinary concentration
curve dominance which can be used as an intuitive tool in ranking “pure” mobility.
This result is empirically significant as it suggests the concentration curve as
the dynamic counterpart of the standard Lorenz curve in terms of social welfare.
If the generalized concentration curve of P dominates that of Q, the social welfare
of permanent income under P will be no less than under Q for all Schur-concave
symmetric SWFs.
138 MAASOUMI
Two main points emerge. One is that, both the inequality reducing and transi-
tion matrix measures are supported by the same type of welfare functions in terms of
“lifetime incomes.” Second, the empirical techniques for testing Lorenz type domi-
nance, or stochastic dominance, are all made available. For instance, Bishop, Chow,
and Formby (1994)show that matched pairs of estimates of generalized Lorenz and
concentration curve ordinates have a joint normal distribution and this sampling
property is distribution free. In general they have to be applied to the same statistics
which are, however, defined over the types of aggregate income functions proposed
by Maasoumi and Zandvakili (1986, 1989, 1990). Further description of some sta-
tistical techniques is given in the next section.
B. Some Mobility Indices
We close this section with a definition of some measures of immobility based on

transition matrices. The first is attributed to Prais and is called the trace index:
trace(P) - 1
Trace =
n-1
This index has been criticized for ignoring the off-diagonal transition prob-
abilities. It also violates transfer properties such as (A12), for instance when DPD
transfers increase any element along the diagonal of P.
Determinant = IPJ (24)

This is an index of immobility such as 1-Ry.Unfortunately, this index obtains
its perfectly mobile limit whenever any two rows or columns of P are identical! It also
violates the ranking condition of the above theorems.
Second largest eigenvalue modulus IA.2 I (25)
This measure of immobility is in the unit interval (A9) and measures the speed
of escape from the initial state (Theil 1972). But it too is incoherent with the ranking
conditions for DPD-type exchanges:
1
Bartholomew’s = ___ nipijli - j~
n-1
i j
This immobility index measures the expected number of crossings between

periods in the steady state. It satisfies the normalization and the unit interval condi-
tions. This measure is also coherent with DPD rankings.
The mean first passage time = d M P n (27)
where M P is the mean first passage matrix. This index measures the expected num-
ber of periods before the “first” individual reaches the state of the second individual.
This measure of immobility is not coherent with DPD rankings.
As we have seen, all of the MSZ measures in the previous section, or indeed
any reasonable measure corresponding to Schur-concave SWFs, and based on “per-
manent income” is coherent with DPD rankings and the results of the above theo-
rems. Shorrocks (1992) considered an interesting measure due to Hart (1976a). Let
the Galtonian model of income evolution be written in terms of the geometric mean
incomes, rn, at time t , as follows:
This model can be used to analyze both income movements over time as well
as the effect of mobility on the distribution of income. pt measure the extent to which
incomes regress toward the geometric mean. The case of a “unit root” corresponds to
Gibrat’s law of proportional effect: changes in relative incomes are independent of
current income. This simple model of mobility may be extended by further modeling
the E , in terms of individual specific characteristics and/or time varying effects. An
example is Lillard and Willis (1978) where panel data are used. Of course, using
the techniques of limited dependent variable models, transition probabilities can
be similarly modeled in terms of individual specific and time varying components.
A survey of several applications is given in Creedy (1985).Alternative models of
diffusion describing the evolution of income have been proposed which derive its
steady-state distribution forms. An example is Sargan (1957). Interestingly, the focus
seems to have shifted to analyzing the properties of the equilibrium distribution and
instantaneous inequality and poverty, rather than the dynamic welfare implications
of the evolution mechanism. Mobility analysis is thus a return to first principles.
Shorrocks defined the following function of Hart’s mobility index (Hart 1976a,
1976b, 1981, 1983),which is derived from the Galtonian model above:
This is related to the simple correlation coefficient. In the Galtonian model,

/It < 1 will reduce inequality while o ( E , ) would increase it. Shorrocks (1992) shows
that H ( - ) is consistent with (A2)-(AS), (AlO)-(A12), and partially with (A7) and
(A9). It fails the universal domain, strong normalization, and strong perfect mobility
properties. Also, it satisfies time symmetry but not necessarily time replication in-
variance. It is coherent with DPD-type rankings. As noted above, the MSZ indices
are consistent with almost all of (Al)-(A12) properties and thus are superior to H ( . ) .
Recently, Chakravarty (1 995) proposed the Kullback’s minimum discrimina-
tion statistic as a measure of mobility. This is given by
where stands for proportion of population at time t in state i.

140 MAASOUMI
This is a natural entropy measure of divergence between two distributions.

Indeed it may be extended so that we may measure the divergence between the dis-
tribution of the “permanent income” and the distribution of income at any desired
point. Examples of the latter are the short-run income distribution of MSZ indices,
perfect mobility distribution, and complete immobility distributions.
The first property to note about K is that it is not a metric as it violates the
triangularity rule. There are other entropy measures that have similar properties and
are “metric” (Maasoumi 1993).Chakravarty (1995)notes that K satisfies many of the
useful properties discussed in Shorrocks (1978a) and summarized above. But it fails
to satisfy the “monotonicity” property. We note that the ordering relation discussed
by Dardanoni (1993) is coherent with perfect mobility, as is K , but does not imply
monotonicity.
Chakravarty (1995)points out that the well-known asymptotic x2 distribution
of 2 K may be used to test some very interesting hypotheses about mobility.
Shorrocks (1976) gives a good account of the properties of mobility indices
based on transition matrices. Dardanoni’s (1993) account of the same also leaves
one with the conclusion that at least some members of the MSZ family are “ideal.”
IV. STATISTICAL INFERENCE
There have been several significant advances in the development of statistical infer-
ence tools in the area of income inequality. These are generally applicable to infer-
ence on mobility indices and on ranking distributions.
For mobility indices such as the MSZ the connection is rather immediate. In-
equality indices are estimated by the method of moments (MM) estimators since they
are functions of population moments. Explicit formulae are derived for derivatives
that are required in the delta method which extends the well-known theory of MM
asymptotic distributions to that of inequality indices. This is surveyed in Maasoumi
(1996a) which contains an extensive citation to original sources. The extension to
mobility indices requires thinking in terms of long-run incomes and the inequality
in their distributions. Trede (1995)gives the details for the asymptotic distribution of
some of the MSZ measures, such as those based on the Atkinson family and Theil’s
inequality indices, but where aggregate income is the simple sum of incomes ana-
lyzed by Shorrocks (1978b). Extension to weighted sum function is immediate, but
the statistical theory for the more complicated aggregate functions is developed in
Maasoumi and Trede (1997). Trede (1995) also gives the asymptotic distributions of
the mobility indices that are based on transition matrices. Some of these measures
were discussed earlier. An application to German data is reported in Trede (1995)
which analyzes earnings mobility for different sexes. A program written in GAUSS
code is made available by Trede.
ON MOBILITY 141
But inference about indices derived in either of the two approaches described
in this paper may be inconclusive, if not bewildering, when Lorenz-type curves cross.
Therefore it is desirable to test for rank relations of the type described above and in
Maasoumi (1996a, 1996b). This type of testing is now possible and is inspired by
testing for inequality restrictions in econometrics and statistics. A brief account of
some of the available techniques for stochastic dominance would be helpful and
follows.
A. Tests for Stochastic Dominance

Stochastic dominance (SD) relations and comparisons of distributions on the basis of
their Lorenz and GL curves are intended to avoid the “index choice” problem. As we
have seen, the SWF rankings of mobility structures can essentially follow the same
path as that of static inequality analysis but in terms of lifetime incomes. In practice,
however, numerical SD rankings often encounter a predictable difficulty since many
distributions and (Lorenz) curves cross, making it impossible to be decisive. But the
realization that all such comparisons are based on sample-based estimates of dis-
tribution functions (or curves) suggests that such comparisons should be conducted
statistically and tested accordingly. The statistical approach is both sound and able
to deliver more clear-cut statistical decisions!
The basic characteristic of tests for rankings is that of ordered populations and
inequality restrictions. Starting with the work of Lehmann (1959)and Bartholomew
(1959), likelihood ratio and Wald-type tests have been and are being developed for
such hypotheses. These tests supplement other well-known procedures based on
one-sided Wilcoxon rank, and the multivariate versions of the Kolmogorov-Smirnov
tests. See Maasoumi (1996a) for a recent selective survey.
In the area of income distributions and tax analysis, initial developments fo-
cused on tests for Lorenz curve comparisons as in Beach and Davidson (1983),
Bishop, Formby, and Thistle (1989). In practice, a finite number of ordinates of
the desired curves or functions are compared. These ordinates are typically rep-
resented by quantiles and/or conditional interval means. Thus, the distribution the-
ory of the proposed tests are typically derived from the existing asymptotic theory
for ordered statistics and quantiles. Recently Beach, Davidson, and Slotsve (1995)
outlined the asymptotic distribution theory for cumulative/conditional means and
variances which are useful for statistically comparing Lorenz and GL curves. This
theory is particularly useful for third-order stochastic dominance (TSD) ranking of
crossing GL curves when a “transfer sensitivity” condition is assumed; see the def-
inition of TSD below. To control for the size of a sequence of tests at several points
the union intersection (UI) test and Studentized maximum modulus technique for
multiple comparisons is generally favored in this area.
Some alternatives to these multiple comparison techniques have been sug-
gested which are typically based on Wald-type joint tests of equality of the same or-
142 MAASOUMI
dinates, see Bishop et al. (1994)and Anderson (1994).These alternatives are some-
times problematic when their implicit null and alternative hypotheses are not a sat-
isfactory representation of the inequality (order) relations that need to be tested.
Xu et al. (1995) and Xu (1995) take proper account of the inequality nature of
such hypotheses and adapt econometric tests for inequality restrictions to testing
for FSD and SSD, and to GL dominance, respectively. Their tests follow the work in
econometrics of Gouieroux et al. (1982), Kodde and Palm (1986), and Wolak (1988,
1989), which complements the work in statistics exemplified by Perlman (1969),
Robertson and Wright (1981),and Shapiro (1988).The asymptotic distributions of
these x2 tests are mixtures of chi-squared variates with probability weights which
are generally difficult to compute. This leads to bounds tests involving inconclusive
regions and conservative inferences. In addition, the computation of the X 2 statis-
tic requires Monte Carlo or bootstrap estimates of covariance matrices, as well as
inequality restricted estimation which requires optimization with quadratic linear
programming.
In contrast, Maasoumi et al. (1996) propose a direct bootstrap approach that
bypasses many of these complexities while making less restrictive assumptions about
the underlying processes. They offer an empirical application for ranking U S . in-
come distributions from the CPS and the PSID data. Their chosen statistic is the
Kolmogorov-Smirnov (KS) as characterized by McFadden (1989), Klecan et al. (1991),
and Kaur et al. (1994).
McFadden (1989)and Klecan, McFadden, and McFadden (1991) have pro-
posed tests of first- and second-order “maximality” for stochastic dominance which
are extensions of the Kolmogorov-Smirnov statistic. McFadden (1 989) assumes i.i.d.
observations and independent variates, allowing him to derive the asymptotic distri-
bution of his test, in general, and its exact distribution in some cases (Durbin 1973,
1985). Klecan et al. generalize this earlier test by allowing for weak dependence
in the processes both across variables and observations. They demonstrate with an
application for ranking investment portfolios. The asymptotic distribution of these
tests cannot be fully characterized, however, prompting Monte Carlo and bootstrap
methods for evaluating critical levels. In the following section some definitions and
results are summarized which help to describe these tests.
I. Definitions and Tests

Let X and Y be two income variables at either two different points in time, or two
lifetime income vectors. Let X 1 , X 2 , . . . , X , be the not necessarily i.i.d observations
on X , and Yl , Y2, . . . , Y, be similar observations on Y . Let U1 denote the class of all
utility functions U such that U‘ 1 0 (increasing). Also, let U 2 denote the class of all
utility functions in U1 for which U” 5 0 (strict concavity). Let X ( i ) and Y(i)denote
the ith order statistics, and assume F ( x ) and C ( x ) are continuous and monotonic
cumulative distribution functions (cdfs) of X and Y , respectively.
ON MOBILITY 143
Quantiles q , ( p ) and qr(p) are implicitly defined by, for example, F[X 5
qx(p)l = P -
Definition (FSD). X first-order stochastic dominates Y , denoted X FSD Y , if
and only if any one of the following equivalent conditions holds:
(1) E [ u ( X ) ]2 E [ u ( Y ) ]for all U E U , , with strict inequality for some U .

(2) F ( x ) 5 C ( x ) for all x in the support of X , with strict inequality for some
N
Definition (SSD). X second-order stochastic dominates Y , denoted X SSD Y ,

if and only if any of the following equivalent conditions holds:
(1) E [ u ( X ) ]2 E [ u ( Y ) ]for all U E U 2 , with strict inequality for some U .

(2) F(t) dt 5 i!, G ( t ) d t for all x in the support of X and Y , with
strict inequality for some x.
(3) 1: q x ( t ) d t 2 Af q r ( t ) d t , for all 0 5 p 5 1, with strict inequality for
some value(s) p .
The tests of FSD and SSD are based on empirical evaluations of conditions
(2) or (3). Mounting tests on conditions (3) typically relies on the fact that quantiles
are consistently estimated by the corresponding order statistics at a finite number
of sample points. Mounting tests on conditions (2) requires empirical cdfs and com-
parisons at a finite number of observed ordinates. Also, from Shorrocks (1983) or Xu
(1995) it is clear that condition ( 3 ) of SSD is equivalent to the requirement of GL
dominance. FSD implies SSD.
Noting the usual definition of the Lorenz curve of, for instance, X as L,(x) =
( l / p x )m
J: X x d F ( t ) , and its GL (x) = p,L,(x), some authors have developed
tests for Lorenz and GL dominance on the basis of the sample estimates of conditional
interval means and cumulative moments of income distributions; e.g., see Bishop
et al. (1989), Bishop et al. (1991), Beach et al. (1995), and Maasoumi (1996a) for a
general survey of the same. The asymptotic distributions given by Beach et al. (1995)
are particularly relevant for testing for third-order stochastic dominance (TSD). The
latter is a useful criterion when Lorenz or GL curves cross at several points and the
investigator is willing to adopt “transfer sensitivity” of Shorrocks and Foster (1987),
that is a relative preference for progressive transfers to poorer individuals. When
either Lorenz or generalized Lorenz curves of two distributions cross unambiguous
ranking by FSD and SSD is not possible. Whitmore (1970) introduced the concept of
TSD in finance. Shorrocks and Foster (1987) showed that the addition of the “transfer
sensitivity” requirement leads to TSD ranking of income distributions. This require-
ment is stronger than the Pigou-Dalton principle of transfers and is based on the
class of welfare functions U:3 which is a subset of U2 with U”‘ 2 0. TSD is defined
as follows:
144 MAASOUMI
Definition (TSD). X third-order stochastic dominates Y , denoted X TSD Y, if

and only if any of the following equivalent conditions holds:
(1) E[u(X)] 3 E[u(Y)] for all U E U s , with strict inequality for some U.
(2) Jf, l:,[F(t) - G ( t ) ]dt dv 5 0, for all x in the support, with strict in-
equality for some x,with the endpoint condition S_+,”[F(t) - G ( t ) ]dt 5
0.
(3) When E[X] = E[Y], X TSD Y iff h:(qi) 5 h;(qi), for all Lorenz curve
+
crossing points i = 1 , 2 , . . . , ( n 1);where h;(q;) denotes the “cumu-
lative variance” for incomes up to the ith crossing point (Davies and Hoy
1995).
When n = 1, Shorrocks and Foster (1987) showed that X TSD Y if (a) the
Lorenz curve of X cuts that of Y from above, and (b) Var(X) 5 Var(Y). This situation
revives the coefficient of variation as a useful statistical index for ranking distri-
butions.
Kaur et al. (1994)assume i.i.d observations for independent prospects X and
Y. Their null hypothesis is condition (2) of SSDf o r each x against the alternative of
strict violation of the same conditionf o r all x. The test of SSD then requires an appeal
to union intersection technique which results in a test procedure with maximum
asymptotic size of a if the test statistic at each x is compared with the critical value
2, of the standard normal distribution.
McFadden offers a definition of “maximal” sets, as follows:
Definition (Moximality). Let /E = {Xl , Xz, . . . , XK} denote a set of K dis-

tinct random variables. Let Fk denote the cdf of the kth variable. The set /E isfirst-
(second-) order maximal if not variable in LE is first- (second-) order weakly domi-
nated by another.
Let X , n = ( x l n , X Z ~ ., . . , XK,,),n = 1,2, . . . , N , be the observed data. As-
sume X.n is strictly stationary and a-mixing, and assume F;(Xi), i = 1,2, . . . , K ,
and exchangeable random variables, so that resampling estimates of the test statis-
tics converge appropriately. This is less demanding than the assumption of indepen-
dence which is not realistic in many applications (as in mobility analysis, and before-
and after-tax scenarios). In general F k is unknown and estimated by the empirical
distribution function F k , v ( X k ) . Finally, if we adopt Klecan et al.’s mathematical reg-
ularity conditions pertaining to von Neumann-Morgenstern (VNM) utility functions
that generally underlie the expected utility maximization paradigm, the following
theorem defines the tests and the hypotheses being tested:
Theorem 6. Given the mathematical regularity conditions;
(a) The variables in /E arejirst-order stochastically maximal; i.e.,

(1) d = mini#; max,[F,(x)-FJ(x)] > 0, ifandonlyifforeach i and j . There
exists a continuous increasing function U such that Eu(Xi) > Eu(X;).
ON MOBILITY 145
The variables in E are second-order stochastically maximal; i.e.,

S = mini#; max, l:n3[Fi(p) - F/(p)]d p > 0, ifand only iffor each
i and j , there exists a continuous increasing and strictly concavefunction
U such that Eu(Xi) > Eu(Xj).
Assuming the stochastic process X,n,n = 1, 2 , . . . , to be strictly stationary
and a-mixing with a ( j ) = O(j-'), for some S > 1, we have d 2 + ~ d,
and S ~ -+
N S,where d2N and S ~ are N the empirical test statistics defined
as
d2N = mini+; max,[FiN(x) - ~ N ( X ) and,
]
SPN = mini#; max, l;[FiN ( p )- F / N ( p ) d p
Proof See Theorems 1 and 5 of Klecan et al. (1991).

The null hypotheses tested by these two statistics is that, respectively, E is
not first- (second-) order maximal-i.e., Xi FSD(SSD) X;, for some i and j . We reject
the null when the statistics are positive and large. Since the null hypothesis in each
case is composite, power is conventionally determined in the least favorable case
of identical marginals Fi = FJ. Thus, as is shown in Kaur et al. (1994) and Klecan
et al. (1991),tests based on dzN and S2N are consistent. Furthermore, the asymptotic
distribution of these statistics are nondegenerate in the least favorable case, being
Gaussian (Klecan et al. 1991, Theorems 6-7).
The statistic s 2 N has, in general, neither a tractable distribution, nor an asymp-
totic distribution for which there are convenient computational approximations. The
situation for d 2 is~ similar except for some special cases; see Durbin (1973, 1985)
and McFadden (1989), who assume i.i.d. observations (not crucial), and indepen-
dent variables in E (consequential). Unequal sample sizes may be handled as in
Kaur et al.
Klecan et al. (1991)suggest Monte Carlo procedures for computing the signif-
icance levels of these tests. This forces a dependence on an assumed parametric dis-
tribution for generating MC iterations, but is otherwise quite appealing for very large
iterations. Maasoumi et al. (1996)employ bootstrap methods to obtain the empirical
distributions of the test statistics and confidence intervals. They report an empiri-
cal examination of the U S . income distribution based on the CPS and PSID data.
Their methods are directly applicable to ranking of mobility structures described
previously.
V. SOME EMPIRICAL EXAMPLES
Creedy (1985)contains detailed descriptions of empirical studies which implement

the transition matrix and other model-based techniques. Shorrocks (1976, 1978a)
and Lillard and Willis (1978) also implement the transition matrix method using
some of the same U S . panel data which I will describe below.
146 MAASOUMI
The MSZ index method has been implemented by Shorrocks, Maasoumi, Zand-
vakili, Trede, and others. Trede’s work is based on German panel data and, as men-
tioned earlier, reports statistical tests of significant change in mobility. The first three
authors use the Michigan Panel data. We now exemplify some of these latter studies:
A. Mobility and Gender

The MSZ family of mobility measures was investigated by Maasoumi and Zandvakili
(1990)using the Michigan Panel Study of Income Dynamics (PSID). These measures
are decomposed in order to learn about components that are due to differences in
gender and income groups, on the one hand, and within group components which are
free of such group characteristics. Several aggregator functions were used to com-
pute the “Permanent income” variable. Their justification and role in robustifying
inferences was investigated.
“Household” income data for the period 1969-1981 were taken from the PSID.
Household’s income (head and spouse, if any) consists of the following: income from
wages, salaries, rents, dividends, interest, business, bonuses, commissions, profes-
sional practice, aid to dependent children, social security, retirement pay, pension or
annuities, unemployment compensation, child support, and other transfer payments.
Real total income is obtained using the current consumer price index. Income is ad-
justed for family size (in 1975) to provide a better measure of family income since
family members effectively pool their incomes. We refer to this adjusted income as
the per capita family income PCFI).
In computing the permanent incomes three different schemes were used in
order to weight income at different times. These aLweights are (i) equal weights
for all years, (ii) the ratio of mean income at time t to the mean income over the
entire T periods (MIW in tables), and (iii) the normalized elements of the eigenvector
corresponding to the first principal component of the X’X matrix. We did not find any
qualitative differences in our results between these three cases, and thus report only
the computations based on ratio-of-means weights. The other two cases are reported
in Zandvakili (1987).
In our computations the substitution parameter /3 is restricted by the relations
+
- y = 1 p. We computed four different aggregator functions corresponding to four
inequality measures with - y = zi = ( 2 , 1 , .5, .O). U = 0.0 and 1.0 correspond
to Theil’s first and second inequality measures, respectively combined with the lin-
ear and the Cobb-Douglas forms of the aggregator function. Tables 1-3 are from
Maasoumi and Zandvakili (1990)which provide, respectively, the annual short-run
inequalities, the inequalities in the aggregated (long-run) incomes, and the income
stability measure R,. Decomposition of each based on gender is also given. Note that
as one moves toward 1981 the number of periods over which Si, I,(S) and R, are
calculated is increasing from one to 13. The results for every other year are reported
to save space.
ON MOBILITY 147
Table I 1969-1981 per Capita Family Income: Short-Run

Inequalities
Overall Between Within Men Women
Degree of inequality aversion = 2.0

1969 1.109 0.014 1.095 0.793 1.676
1971 1.049 0.016 1.033 0.729 1.604
1973 1.002 0.019 0.982 0.636 1.600
1975 1.636 0.023 1.613 0.815 3.015
1977 1.569 0.038 1.530 0.981 2.195
1979 1.895 0.041 1.854 1.211 2.580
1981 2.441 0.047 2.394 1.520 3.292

1969 0.430 0.013 0.417 0.340 0.676
1971 0.466 0.014 0.452 0.375 0.709
1973 0.464 0.017 0.446 0.362 0.727
1975 0.531 0.02 1 0.510 0.422 0.808
1977 0.578 0.033 0.545 0.467 0.808
1979 0.613 0.035 0.579 0.498 0.849
1981 0.706 0.039 0.667 0.578 0.964

1969 0.375 0.012 0.363 0.306 0.601
1971 0.407 0.014 0.394 0.341 0.618
1973 0.404 0.017 0.387 0.331 0.635
1975 0.456 0.019 0.436 0.380 0.693
1977 0.494 0.030 0.463 0.419 0.688
1979 0.505 0.032 0.473 0.427 0.707
1981 0.571 0.036 0.535 0.490 0.782

1969 0.367 0.012 0.355 0.304 0.606
1971 0.402 0.013 0.388 0.344 0.615
1973 0.395 0.016 0.380 0.332 0.631
1975 0.448 0.018 0.430 0.383 0.693
1977 0.492 0.029 0.463 0.429 0.682
1979 0.483 0.030 0.453 0.417 0.688
1981 0.549 0.034 0.515 0.481 0.753
148 MAASOUMI
Table 2 1969-1981 per Capita Family Income (MIW): Long-Run

Inequality

1969 1.109 0.014 1.095 0.793 1.676
1%9-71 0.949 0.016 0.933 0.665 1.430
1969-73 0.889 0.018 0.871 0.599 1.36.5
1969-75 0.951 0.020 0.931 0.588 1.533
1969-77 0.974 0.025 0.950 0.602 1.507
1%9-79 1.035 0.030 1.005 0.653 1.509
1969-81 1.100 0.034 1.067 0.690 1.563
Degree of inequality aversion = 1.O
1%9 0.430 0.013 0.417 0.340 0.676
1969-71 0.414 0.014 0.400 0.325 0.651
1969-73 0.408 0.016 0.392 0.316 0.648
1969-75 0.414 0.018 0.396 0.319 0.654
1969-77 0.426 0.02 1 0.404 0.327 0.662
1969-79 0.434 0.025 0.409 0.336 0.654
1969-81 0.455 0.029 0.425 0.352 0.674
1%9 0.375 0.012 0.363 0.306 0.601
1969-71 0.364 0.014 0.350 0.296 0.579
1%9-73 0.360 0.01s 0.345 0.291 0.577
1969-75 0.363 0.017 0.347 0.294 0.579
1969-77 0.373 0.020 0.353 0.302 0.585
1969-79 0.381 0.024 0.357 0.310 0.579
1%9-81 0.396 0.027 0.369 0.323 0.591
1969 0.367 0.012 0.355 0.304 0.606
1969-71 0.355 0.013 0.342 0.295 0.578
1969-73 0.351 0.014 0.337 0.292 0.572
1969-75 0.354 0.016 0.339 0.295 0.573
1969-77 0.366 0.019 0.347 0.306 0.578
1969-79 0.371 0.023 0.349 0.312 0.566
1969-81 0.386 0.026 0.361 0.326 0.573
ON MOBILITY 149
Table 3 1969-1981 per Capita Family Income (MIW): Income

Stability

1969 1.ooo 0.013 0.987 1.000 1.000
1969-7 1 0.916 0.016 0.900 0.907 0.918
1969-73 0.885 0.017 0.868 0.873 0.887
1969-75 0.828 0.018 0.811 0.822 0.823
1%9-77 0.786 0.020 0.767 0.772 0.786
1%9-79 0.744 0.02 1 0.723 0.723 0.747
1969-8 1 0.692 0.02 1 0.671 0.670 0.700
1969 1.ooo 0.030 0.970 1.000 1.000
1969-7 1 0.928 0.032 0.896 0.912 0.949
1969-73 0.903 0.035 0.869 0.881 0.930
1969-75 0.877 0.038 0.839 0.851 0.907
1969-77 0.855 0.043 0.812 0.824 0.887
1%9-79 0.832 0.049 0.783 0.801 0.860
1969-8 1 0.813 0.052 0.761 0.779 0.841

1%9 1.ooo 0.033 0.967 1.000 1.ooo
1969-71 0.932 0.035 0.896 0.917 0.955
1969-73 0.911 0.038 0.873 0.891 0.942
1969-75 0.885 0.041 0.844 0.862 0.920
1%9-77 0.867 0.047 0.820 0.841 0.904
1969-79 0.855 0.054 0.801 0.830 0.885
1969-8 1 0.840 0.058 0.782 0.813 0.870

1969 1.ooo 0.033 0.967 1.000 1.ooo
1969-71 0.927 0.034 0.892 0.915 0.950
1969-73 0.904 0.037 0.868 0.889 0.933
1969-75 0.879 0.039 0.840 0.861 0.911
1969-77 0.864 0.045 0.819 0.846 0.895
1969-79 0.852 0.052 0.800 0.83.5 0.871
1969-8 1 0.839 0.056 0.783 0.822 0.855
150 MAASOUMI
Short-run inequality in Table 1 has generally increased. As expected, inequal-

ity is greater with larger degrees of relative inequality aversion (‘U).There are 1776
male and 529 female headed households in the sample. The “within-group” com-
ponent of short-run inequalities is dominant. The column heading “within” refers
to weighted average of within group inequalities as in the formulae previously dis-
cussed in this chapter. The absolute value of the “between-group” component, how-
ever, has increased over the 13 years. For both sexes annual inequalities have a rising
trend (less uniformly so for ‘U = 2). But short-run inequality amongst female headed
households is always greater than among men. These annual values, however, con-
tain many transitory components which are partially removed from the aggregated
values in Table 2.
Table 2 values exhibit much less volatility. After a decline in the initial years,
Z,(S) has increased back to about its original value. Also, long-run inequality is
always smaller than the corresponding short-run inequality. Once again, inequality
among women is greater than among men, and within-group inequality is several
times the between-group component. These relative values are somewhat sensitive
to the family size adjustment of incomes. For instance, the between-group component
increases to 15-25% of overall inequality for unadjusted incomes; see Zandvakili
(1987).This is partly due to a larger proportion of two income earners being among
the male headed families.
The corresponding stability measures are presented in Table 3. Again, seven
of the 13 possible values are reported without any qualitative loss. 0 < R , < 1 in
all cases. The following may be concluded from Table 3:
(i) There is a tendency for the profiles to fall and then level off as the num-
ber of aggregated years increases from one to 13.
(ii) The profiles for households headed by men fall faster and further than
those of women headed households.
(iii) These patterns are robust with respect to the choice of aggregation func-
tion, family size adjustment, and inequality measure.
The fact that the profiles are becoming flatter is an indication that, although
there have been some transitory movements in the size distribution of income, there
is a lack of any permanent equalization. Further, while some equalization has taken
place within each group of households, inequality between men and women headed
households has increased in absolute value.
B. Mobility and Income Level

Maasoumi and Zandvakili (1989) give inequality and mobility decompositions by
age, education, and race. Similar decompositions by income level can reveal the
aggregate impact of all such non-income characteristics (including gender). It is
ON MOBILITY I 5 I
anticipated that if the major causes of variation in incomes are transitory in nature,
the length of time spent in any income class will be short. “Permanent” income in-
equality changes will be very revealing in this context.
The total sample is divided into seven income groups (Gl-G7). The assign-
ment to groups is one a one-time basis and according to the simple arithmetic mean
income of the individual household over the 13-year period. These real income lev-
els begin with mean incomes of less than $4999, and increasing in increments of
$5000. The last group contains mean incomes of $35,000 or more.
Short-run inequalities and their decompositions based on income level are
given in Table 4. All the tables and figures in this section are taken from Maasoumi
and Zandvakili (1990). There are several recognizable patterns. The between-group
inequality has increased steadily over this period. The within-group component of
inequality fluctuates around a relatively constant mean value. The observed pat-
terns suggest that the nonincome differences do contribute to the increase in between
group inequality. Over 70% of women-headed households earn less than $15,000. Of
course, this is confounded by the differential impact of inflation on different income
groups (we use real incomes).
In Table 5 long-run inequality levels have risen after an initial decline. De-
composition by income level shows that the between-group component of I,(S) has
increased uniformly. At the same time the within-group inequality has decreased
steadily. This change has been dramatic so that in the later years the between-group
component is larger than the within component. These changes include the well-
known life-cycle and human capital effects, and are non inconsistent with the cu-
mulat ive effects predict ed by discrimination theories.
The long-run within-group inequalities reveal a falling trend for each of the
seven income groups. This is anticipated since transitory components are smoothed
out and individual incomes have approached group mean incomes in the long run.
These long-run grouping observations are somewhat sensitive to the family size.
Within-group aggregate income inequalities are noticeably smaller when income is
not adjusted for family size, and there is generally less inequality within the higher-
income groups.
Table 6 reports the stability profiles which reveal much higher degrees of per-
manent equalization within income groups than was observed for the gender groups
of the last section. Note that as the stability profiles of the whole sample flatten, the
corresponding within group profiles continue to fall. In our view some equalization
has occurred, but this is mostly confined to within income groups.
On the basis of the approximately 2300 households which remained in the
Michigan panel over the period 1969-1981. Maasoumi and Zandvakili (1990)con-
clude that (i) there is not a great deal of inequality between the men and women-
headed households; (ii) the dominant within-group component of inequality is either
increasing over this period or, when incomes are smoothed by time aggregation,
152 MAASOUMI
Table 4 1969-1981 per Capita Family Income: Short-Run Inequalities
Overall Between Within G1 G2 G3 G4 G5 G6 G7
1969 1.109 0.151 0.958 0.879 1.429 0.668 0.608 0.282 0.293 0.209
1971 1.049 0.224 0.824 0.672 0.560 0.626 0.725 0.281 0.265 0.450
1973 1.002 0.310 0.692 0.524 0.415 0.430 0.479 0.217 0.220 0.212
1975 1.636 0.412 1.225 0.950 0.539 0.459 0.449 0.274 0.236 0.242
1977 1.569 0.524 1.045 0.633 0.487 0.479 0.365 0.290 0.178 0.242
1979 1.895 0.624 1.271 0.639 0.683 0.544 0.460 0.227 0.232 0.211
1981 2.441 0.725 1.716 0.722 0.985 0.679 0.525 0.570 0.310 0.299
1969 0.430 0.114 0.316 0.429 0.413 0.421 0.385 0.236 0.240 0.179
1971 0.466 0.161 0.306 0.413 0.334 0.356 0.411 0.231 0.227 0.224
1973 0.464 0.205 0.259 0.352 0.284 0.302 0.336 0.190 0.199 0.190
1975 0.531 0.259 0.272 0.354 0.324 0.312 0.314 0.224 0.204 0.211
1977 0.578 0.316 0.262 0.364 0.310 0.322 0.273 0.208 0.165 0.218
1979 0.613 0.346 0.267 0.353 0.374 0.335 0.286 0.194 0.193 0.183
1981 0.706 0.384 0.322 0.393 0.490 0.394 0.342 0.252 0.227 0.225
1969 0.375 0.103 0.272 0.396 0.356 0.389 0.353 0.233 0.243 0.179
1971 0.407 0.144 0.264 0.388 0.299 0.325 0.373 0.228 0.230 0.219
1973 0.404 0.180 0.224 0.336 0.263 0.279 0.314 0,188 0.202 0.192
1975 0.456 0.225 0.231 0.326 0.292 0.291 0.294 0.224 0.205 0.212
1977 0.494 0.275 0.219 0.337 0.282 0.303 0.260 0.205 0.167 0.224
1979 0.505 0.293 0.212 0.332 0.331 0.310 0.269 0.191 0.192 0.183
1981 0.571 0.322 0.249 0.364 0.432 0.361 0.315 0.234 0.218 0.221
1969 0.367 0.096 0.270 0.402 0.339 0.391 0.349 0.243 0.263 0.187
1971 0.402 0.134 0.268 0.398 0.290 0.323 0.367 0.235 0.246 0.228
1973 0.395 0.165 0.230 0.348 0.261 0.272 0.313 0.192 0.214 0.202
1975 0.448 0.207 0.242 0.333 0.284 0.290 0.293 0.237 0.215 0.223
1977 0.492 0.254 0.238 0.343 0.274 0.306 0.263 0.214 0.176 0.245
1979 0.483 0.266 0.218 0.345 0.320 0.311 0.271 0.196 0.200 0.191
1981 0.549 0.293 0.256 0.382 0.424 0.362 0.311 0.236 0.222 0.233
ON MOBILITY 153
Table 5 1969-1981 per Capita Family Income (MIW): Long-Run Inequality
1969 1.109 0.151 0.958 0.879 1.429 0.668 0.608 0.282 0.293 0.209
1969-71 0.949 0.208 0.741 0.610 0.708 0.544 0.601 0.230 0.240 0.261
1969-73 0.889 0.255 0.634 0.486 0.504 0.425 0.506 0.208 0.222 0.222
1969-75 0.951 0.311 0.640 0.490 0.404 0.344 0.426 0.188 0.208 0.200
1969-77 0.974 0.365 0.609 0.444 0.341 0.303 0.322 0.166 0.188 0.186
1969-79 1.035 0.425 0.610 0.422 0.313 0.270 0.274 0.142 0.160 0.172
1969-81 1.100 0.481 0.619 0.401 0.308 0.244 0.247 0.147 0.146 0.155
1969 0.430 0.114 0.316 0.429 0.413 0.421 0.385 0.236 0.240 0.179
1969-71 0.414 0.1M 0.270 0.354 0.311 0.346 0.374 0.198 0.204 0.176
1969-73 0.408 0.170 0.238 0.309 0.261 0.282 0.328 0.181 0.193 0.166
1969-75 0.414 0.200 0.213 0.275 0.225 0.242 0.290 0.166 0.176 0.159
1969-77 0.426 0.233 0.192 0.259 0.200 0.210 0.233 0.150 0.158 0.154
1969-79 0.434 0.264 0.171 0.241 0.181 0.180 0.192 0.129 0.135 0.143
1969-81 0.455 0.294 0.161 0.232 0.179 0.167 0.180 0.122 0.122 0.130
1969 0.375 0.103 0.272 0.396 0.356 0.389 0.353 0.233 0.243 0.179
1969-71 0.364 0.128 0.236 0.334 0.274 0.320 0.342 0.199 0.208 0.177
1969-73 0.360 0.150 0.210 0.294 0.236 0.262 0.304 0.182 0.198 0.169
1969-75 0.363 0.175 0.188 0.260 0.205 0.225 0.269 0.168 0.179 0.163
1969-77 0.373 0.204 0.169 0.245 0.182 0.196 0.219 0.157 0 .161 0.160
1969-79 0.381 0.231 0.150 0.229 0.164 0.172 0.185 0.135 0.139 0.150
1969-81 0.396 0.257 0.139 0.219 0.162 0.163 0.175 0.128 0.127 0.138
~~ ~
1969 0.367 0.0% 0.270 0.402 0.339 0.391 0.349 0.243 0.263 0.187
1969-71 0.355 0.117 0.238 0.339 0.261 0.318 0.335 02.07 0.223 0.187
1969-73 0.351 0.137 0.214 0.298 0.226 0.256 0.2% 0.188 0.212 0.179
1969-75 0.354 0.159 0.195 0.263 0.197 0.219 0.262 0.176 0.190 0.173
1969-77 0.366 0.186 0.180 0.246 0.174 0.191 0.214 0.168 0.169 0.173
1969-79 0.371 0.210 0.161 0.230 0.153 0.171 0.182 0.144 0.147 0.161
1969-81 0.386 0.235 0.152 0.220 0.153 0.165 0.173 0.136 0.135 0.152
154 MAASOUMI
Table 6 1969-1981 per Capita Family Income (MIW): Income Stability
Degree of' inequality aversion = 2.0
1969 1.OOO 0.136 0.864 1.000 1.OOO 1.OOO 1.OOO 1.000 1.000 1.OOO
1969-71 0.916 0.201 0.715 0.812 0.857 0.872 0.907 0.857 0.878 0.%7
1969-73 0.885 0.254 0.631 0.754 0.775 0.766 0.857 0.815 0.853 0.825
1969-75 0.828 0.271 0.557 0.712 0.669 0.680 0.794 0.726 0.793 0.780
1969-77 0.786 0.295 0.492 0.676 0.605 0.592 0.670 0.631 0.756 0.744)
1969-79 0.744 0.305 0.439 0.644 0.538 0.507 0.578 0.562 0.675 0.709
1969-81 0.692 0.303 0.390 0.614 0.480 0.437 0.523 0.4% 0.577 0.617
1969 1.000 0.265 0.735 1.000 1.OOO 1.OOO 1.OOO 1.000 1.000 1.000
1969-7 1 0.928 0.323 0.605 0.852 0.861 0.901 0.930 0.880 0.892 0.871
1969-73 0.903 0.376 0.527 0.799 0.785 0.805 0.880 0.844) 0.870 0.833
1969-75 0.877 0.425 0.452 0.745 0.689 0.726 0.821 0.768 0.819 0.791
1969-77 0.855 0.468 0.386 0.713 0.627 0.634 0.713 0.693 0.773 0.755
1969-79 0.832 0.505 0.327 0.670 0.554 0.545 0.611 0.617 0.684 0.716
1969-81 0.813 0.526 0.287 0.637 0.510 0.490 0.572 0.567 0.607 0.639
1969 1.000 0.275 0.725 1.000 1.OOO 1.OOO 1.OOO 1.000 1.WO 1.OOO
1969-7 1 0.932 0.328 0.604 0.864 0.860 0.908 0.936 0.890 0.903 0.885
1969-73 0.911 0.379 0.532 0.812 0.789 0.815 0.887 0.853 0.885 0.853
1969-75 0.885 0.427 0.458 0.758 0.694 0.732 0.825 0.785 0.833 0.812
1969-77 0.867 0.474 0.393 0.722 0.629 0.638 0.721 0.727 0.783 0.781
1969-79 0.855 0.518 0.337 0.685 0.553 0.564 0.629 0.654 0.705 0.750
1969-81 0.840 0.545 0.296 0.647 0.513 0.520 0.595 0.609 0.637 0.675
1969 1.000 0.262 0.738 1.000 1.OOO 1.OOO 1.OOO 1.000 1.000 1.OOO
1969-71 0.927 0.306 0.621 0.861 0.848 0.904 0.932 0.891 0.906 0.894
1969-73 0.904 0.352 0.553 0.806 0.776 0.806 0.879 0.851 0.886 0.860
1969-75 0.879 0.395 0.484 0.752 0.682 0.718 0.811 0.790 0.834 0.822
1969-77 0.864 0.440 0.424 0.712 0.614 0.627 0.708 0.745 0.780 0.800
1969-79 0.852 0.483 0.369 0.673 0.530 0.563 0.623 0.667 0.705 0.767
1969-81 0.839 0.510 0.330 0.633 0.4% 0.528 0.592 0.628 0.649 0.700
ON MOBILITY 155
relatively stable; (iii) this larger within-group component of inequality is due to high
levels of inequality within lower income groups (such as women headed households);
(iv) grouping by real income brackets leads predictably to very large between-group
inequality values. (v) Some equalization of real incomes has occurred over time
within most income groups, but this is very hard to judge by a comparison of an-
nual inequality measures and most clearly revealed by using our “permanent in-
come” distributions; (vi) modest levels of mobility are recorded as the aggregation
interval is expanded, but the corresponding profiles flatten out after about eight
or nine years.
We close this subsection with Fig. 1, which summarize the evolution of the
income distribution for this sample with the graph of the stability profile pt.
C. Mobility by Education, Age, and Race
Maasoumi and Zandvakili (1989)is based on the same data as the previous section,
but the role of years of schooling of the head of household, hisher age, and race were
examined through decomposition of the inequality/mobility measures. Tables 7-1 5
are from that source.
Table 7 is a summary of short run, long run, and the stability values for all the
13years. Tables 8-10 provide decompositions by educational attainment which was
indicated by the years of completed schooling by the head. The increase in overall
short-run inequality is primarily due to increases in within group inequalities. Long-
run inequality is quite stable. The R measure declines over longer periods. This
indicates that while there is much short-run mobility (change), this does not change
permanent income inequality. Note that this phenomenon may be partly due to the
anonymity of our measures which are invariant to short-run switching of positions by
individuals.
It should be noted that education is both a capital good and a provider of
a stream of consumption. It has different values for different individuals. This
heterogeneity effect is here controlled for leading to conditional inferences. For a
discussion of these issues and a multidimensional treatment in which education
is regarded as a distinct attribute (with income and wealth) see Maasoumi and
Nickelsburg (1988).
Tables 8-10 indicate that the greatest inequality is within the group with fewest
schooling years. Indeed, within-group inequality declines steadily with educational
attainment: education is an equalizer (some might argue it is a restraint over un-
usual earnings!). Between groups inequality is rising somewhat over these years but
is about one quarter of total inequality, and declining proportionately.
Long-run inequality is much more stable over time and is smaller than short-
run inequalities. Looking at these figures a policymaker is less likely drawn to quick
reactions to transitory phenomenon, and more likely to focus on stable features, for
156 MAASOUMI
I .oo 9=2.0
-A--
0.95
0.90
Measure 0.85
of income
stability
0.80
0.75 -
0.70 -
c
I
L I 1 I I I I I I I I I
69 I1 73 75 77 79 81
Year
Figure I Measure of income stability PCFI 1969-1981, based on mean of income weights.
Table 7 Per Capita Family Income 1969-1981 Based on Mean of Income Weights
Short-run inequalities
1%9 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981
/2.0(Y) 1.109 0.961 1.049 0.940 1.002 1.190 1.636 1.357 1.569 1.840 1.895 2.072 2.441
Il,o(Y) 0.430 0.443 0.466 0.460 0.464 0.488 0.531 0.548 0.573 0.594 0.613 0.669 0.706
Zo,s(Y) 0.375 0.388 0.407 0.404 0.404 0.423 0.456 0.468 0.494 0.4% 0.505 0.551 0.571
Zo.o(Y) 0.367 0.381 0.402 0.399 0.395 0.415 0.448 0.458 0.492 0.481 0.483 0.541 0.M9
~~
2.0 1.109 0.979 0.949 0.906 0.889 0.896 0.951 0.955 0.974 1.007 1.035 1.060 1.100
1.0 0.430 0.417 0.414 0.412 0.408 0.409 0.414 0.418 0.426 0.429 0.434 0.445 0.455
0.5 0.375 0.364 0.364 0.362 0.360 0.360 0.363 0.368 0.373 0.377 0.381 0.387 0.3%
0.0 0.367 0.355 0.355 0.353 0.351 0.350 0.354 0.359 0.366 0.370 0.371 0.378 0.386
Income stability
V R1 R2 R3 R4 R5 R6 R7 R8 R9 RIO R11 R12 R13
2.0 1.000 0.948 0.916 0.897 0.885 0.860 0.828 0.809 0.786 0.764 0.744 0.717 0.692
1.0 1.000 0.955 0.928 0.915 0.903 0.891 0.877 0.865 0.855 0.843 0.832 0.822 0.813
0.5 1.000 0.954 0.932 0.917 0.911 0.897 0.885 0.877 0.867 0.861 0.855 0.843 0.840
0.0 1.000 0.950 0.927 0.912 0.904 0.890 0.879 0.872 0.864 0.859 0.852 0.841 0.839
158 MAASOUMI
Table 8 1969-1981 per Capita Family Income, Short-Run Inequalities
O-11th 12th Advanced

V Overall Between Within grade grade College degree
2.0 1.112 0.081 1.031 1.124 0.513 0.377 0.224

0.964 0.096 0.868 0.864 0.506 0.407 0.298
1.051 0.102 0.949 0.886 0.689 0.469 0.300
0.942 0.104 0.838 0.760 0.645 0.453 0.252
1.003 0.101 0.902 0.860 0.603 0.448 0.275
1.192 0.110 1.083 0.998 0.775 0.513 0.284
1.641 0.118 1.523 1.478 0.825 0.609 0.283
1.360 0.117 1.243 1.119 0.909 0.610 0.289
1.573 0.109 1.464 1.399 0.901 0.706 0.334
1.844 0.113 1.730 1.529 1.474 0.680 0.406
1.898 0.114 1.785 1.467 1.831 0.807 0.383
2.076 0.141 1.935 1.661 1.380 0.888 0.450
2.447 0.146 2.301 2.018 1.437 0.996 0.572
1.0 0.431 0.079 0.352 0.443 0.294 0.258 0.190
0.444 0.092 0.352 0.431 0.302 0.263 0.234
0.467 0.098 0.369 0.435 0.334 0.297 0.223
0.460 0.099 0.361 0.424 0.323 0.306 0.201
0.464 0.097 0.367 0.445 0.316 0.292 0.204
0.488 0.104 0.384 0.457 0.347 0.307 0.213
0.531 0.112 0.419 0.499 0.370 0.349 0.222
0.549 0.111 0.438 0.532 0.376 0.353 0.229
0.579 0.105 0.474 0.598 0.374 0.370 0.261
0.594 0.109 0.485 0.605 0.402 0.365 0.276
0.613 0.110 0.503 0.632 0.431 0.360 0.259
0.669 0.133 0.536 0.668 0.438 0.417 0.301
0.706 0.138 0.568 0.712 0.473 0.416 0.325
0.5 0.376 0.078 0.297 0.394 0.265 0.243 0.185
0.388 0.091 0.297 0.381 0.276 0.245 0.227
0.408 0.098 0.310 0.383 0.300 0.276 0.208
0.405 0.099 0.306 0.378 0.286 0.287 0.193
0.404 0.097 0.307 0.393 0.285 0.270 0.194
0.423 0.104 0.319 0.404 0.303 0.287 0.200
0.456 0.112 0.345 0.431 0.326 0.321 0.214
0.468 0.111 0.358 0.464 0.324 0.325 0.221
0.494 0.105 0.389 0.532 0.32 1 0.337 0.257
0.495 0.190 0.386 0.521 0.334 0.330 0.260
0.504 0.110 0.394 0.540 0.357 0.317 0.240
0.550 0.133 0.418 0.564 0.358 0.381 0.281
0.571 0.138 0.433 0.592 0.395 0.361 0.296
(continued)
ON MOBILITY 159
Table 8 (Continued)

0.0 0.367 0.079 0.288 0.392 0.260 0.245 0.188

0.381 0.092 0.289 0.373 0.274 0.245 0.233
0.402 0.099 0.303 0.375 0.300 0.277 0.203
0.399 0.100 0.299 0.371 0.280 0.290 0.195
0.396 0.098 0.298 0.383 0.284 0.270 0.193
0.415 0.105 0.310 0.3% 0.295 0.292 0.196
0.448 0.113 0.336 0.416 0.321 0.324 0.217
0.458 0.112 0.347 0.456 0.315 0.327 0.223
0.492 0.107 0.384 0.557 0.311 0.341 0.269
0.481 0.111 0.370 0.514 0.321 0.328 0.263
0.483 0.113 0.370 0.526 0.343 0.310 0.238
0.540 0.135 0.405 0.547 0.334 0.403 0.284
0.549 0.141 0.408 0.567 0.381 0.352 0.297
2299 1096 624 452 127
Table 9 1969-1981 per Capita Family Income, (MIW) Long-Run Inequality

2.0 1.112 0.081 1.031 1.124 0.513 0.377 0.224

0.98 1 0.090 0.891 0.925 0.481 0.361 0.230
0.952 0.096 0.856 0.855 0.520 0.363 0.229
0.908 0.099 0.809 0.789 0.524 0.358 0.217
0.891 0.101 0.790 0.762 0.519 0.353 0.213
0.899 0.105 0.794 0.756 0.531 0.354 0.212
0.053 0.109 0.844 0.805 0.542 0.361 0.211
0.957 0.113 0.844 0.791 0.556 0.364 0.210
0.977 0.117 0.860 0.798 0.568 0.374 0.210
1.010 0.120 0.890 0.81 1 0.617 0.381 0.2 14
1.037 0.123 0.914 0.813 0.684 0.391 0.217
1.061 0.127 0.935 0.825 0.697 0.399 0.218
1.103 0.131 0.972 0.855 0.706 0.412 0.227
(continued)
160 MAASOUMI
Table 9 (Continued)
0-11th 12th Advanced

I/ Overall Between Within grade grade College degree
1.0 0.431 0.079 0.352 0.443 0.294 0.258 0.190

0.418 0.086 0.332 0.413 0.283 0.242 0.192
0.415 0.090 0.325 0.398 0.285 0.240 0.186
0.413 0.093 0.319 0.388 0.283 0.242 0.178
0.409 0.095 0.314 0.382 0.279 0.238 0.173
0.410 0.097 0.312 0.379 0.277 0.237 0.172
0.414 0.101 0.313 0.380 0.278 0.240 0.172
0.419 0.104 0.315 0.381 0.279 0.242 0.173
0.426 0.107 0.319 0.388 0.280 0.246 0.175
0.430 0.109 0.32 1 0.392 0.279 0.247 0.178
0.435 0.110 0.324 0.398 0.282 0.245 0.178
0.445 0.114 0.331 0.410 0.283 0.246 0.181
0.455 0.118 0.337 0.422 0.285 0.248 0.184
0.5 0.376 0.078 0.297 0.394 0.265 0.243 0.185
0.365 0.085 0.279 0.366 0.257 0.226 0.188
0.364 0.090 0.274 0.353 0.259 0.226 0.181
0.362 0.093 0.269 0.345 0.255 0.228 0.172
0.360 0.094 0.266 0.341 0.253 0.225 0.168
0.360 0.097 0.263 0.339 0.249 0.224 0.166
0.364 0.100 0.263 0.338 0.250 0.227 0.167
0.368 0.103 0.265 0.340 0.249 0.231 0.169
0.373 0.105 0.269 0.347 0.249 0.234 0.174
0.378 0.107 0.271 0.354 0.248 0.236 0.178
0.381 0.108 0.273 0.361 0.250 0.234 0.177
0.387 0.112 0.275 0.369 0.248 0.233 0.180
0.396 0.116 0.280 0.381 0.249 0.234 0.185
0.0 0.367 0.079 0.288 0.392 0.260 0.245 0.188
0.356 0.086 0.270 0.359 0.253 0.227 0.193
0.355 0.091 0.265 0.345 0.257 0.226 0.182
0.354 0.093 0.260 0.336 0.252 0.230 0.173
0.351 0.095 0.257 0.331 0.250 0.226 0.169
0.351 0.097 0.254 0.329 0.245 0.225 0.166
0.355 0.100 0.254 0.326 0.245 0.230 0.168
0.359 0.103 0.256 0.329 0.244 0.235 0.170
0.366 0.104 0.263 0.344 0.244 0.239 0.178
0.370 0.105 0.265 0.350 0.242 0.240 0.184
0.372 0.106 0.265 0.356 0.243 0.237 0.182
0.378 0.111 0.267 0.364 0.239 0.239 0.186
0.386 0.116 0.270 0.374 0.241 0.238 0.193
2299 1096 624 452 127
ON MOBILITY I 6 I
Table I 0 1969-1981 per Capita Family Income, (MIW) Income Stability

2.0 1.000 0.073 0.927 1.ooo 1.000 1.000 1.000

0.948 0.087 0.861 0.936 0.945 0.920 0.876
0.916 0.092 0.824 0.899 0.909 0.866 0.833
0.897 0.098 0.799 0.879 0.883 0.834 0.806
0.885 0.101 0.785 0.860 0.875 0.817 0.791
1.860 0.100 0.760 0.832 0.843 0.789 0.780
1.828 0.095 0.734 0.797 0.815 0.757 0.769
1.809 0.096 0.713 0.771 0.791 0.731 0.760
1.786 0.094 0.692 0.737 0.774 0.707 0.737
1.764 0.091 0.673 0.711 0.740 0.694 0.711
0.744 0.088 0.656 0.688 0.711 0.673 0.698
0.717 0.086 0.631 0.663 0.686 0.644 0.664
0.693 0.082 0.610 0.642 0.663 0.622 0.636
1.0 1.000 0.183 0.817 1.ooo 1.ooo 1.ooo 1.000
0.955 0.197 0.758 0.946 0.949 0.926 0.901
0.929 0.202 0.726 0.914 0.919 0.881 0.864
0.915 0.207 0.708 0.898 0.901 0.858 0.839
0.903 0.209 0.694 0.881 0.890 0.840 0.826
0.891 0.212 0.679 0.865 0.866 0.82 1 0.819
0.877 0.214 0.663 0.846 0.846 0.802 0.811
0.865 0.215 0.650 0.826 0.829 0.788 0.805
0.855 0.214 0.640 0.804 0.820 0.776 0.792
0.843 0.213 0.630 0.788 0.799 0.765 0.780
0.832 0.211 0.62 1 0.775 0.785 0.751 0.769
0.822 0.211 0.611 0.768 0.768 0.728 0.752
0.813 0.210 0.603 0.762 0.747 0.715 0.737
0.5 1.000 0.209 0.791 1.ooo 1.ooo 1.ooo 1.ooo
0.954 0.223 0.730 0.945 0.949 0.926 0.909
0.932 0.231 0.701 0.917 0.924 0.886 0.874
0.917 0.235 0.682 0.899 0.904 0.864 0.847
0.911 0.238 0.672 0.887 0.897 0.852 0.840
0.897 0.241 0.656 0.872 0.871 0.831 0.830
0.885 0.244 0.641 0.853 0.851 0.817 0.826
0.877 0.246 0.631 0.836 0.837 0.808 0.823
0.867 0.243 0.624 0.817 0.827 0.799 0.815
0.861 0.243 0.618 0.809 0.813 0.793 0.814
0.855 0.242 0.613 0.802 0.803 0.781 0.803
0.843 0.244 0.599 0.795 0.781 0.751 0.788
0.840 0.247 0.594 0.796 0.766 0.743 0.782
(continued)
162 MAASOUMI
Table I0 (Continued)
~~ ~~ ~ ~~ ~ ~

0.0 1.000 0.215 0.785 1.ooo 1.ooo 1.000 1.ooo

0.9.50 0.229 0.721 0.940 0.947 0.925 0.914
0.927 0.236 0.69 1 0.910 0.923 0.885 0.875
0.9 12 0.241 0.671 0.891 0.902 0.865 0.847
0.905 0.244 0.66 1 0.878 0.895 0.85 1 0.839
0.890 0.246 0.644 0.863 0.865 0.831 0.828
0.879 0.249 0.630 0.842 0.847 0.820 0.827
0.872 0.249 0.623 0.827 0.833 0.816 0.827
0.864 0.244 0.620 0.815 0.824 0.810 0.827
0.859 0.244 0.61.5 0.808 0.810 0.802 0.831
0.852 0.244 0.608 0.802 0.801 0.789 0.815
0.841 0.247 0.594 0.796 0.777 0.762 0.807
0.839 0.252 0.587 0.796 0.765 0.751 0.808
2299 1096 624 452 127
Table I I 1969-1981 per Capita Family Income, Short-Run Inequalities
V Overall Between Within 18-29 30-39 40-49 50-59 260
2.0 1.109 0.004 1.105 1.294 0.667 0.836 1.992 1.059

0.961 0.006 0.956 0.587 0.641 0.982 1SO8 1.083
1.049 0.007 1.042 0.840 0.834 0.920 1.598 1.081
0.940 0.009 0.931 0.643 0.777 0.942 1.332 0.854
1.002 0.013 0.988 0.575 0.751 1.061 1.350 1.019
1.190 0.017 1.173 0.638 0.854 1.676 1.452 0.852
1.636 0.022 1.615 0.788 0.999 1.375 3.657 0.937
1.357 0.027 1.330 0.971 1.298 1.379 1.660 0.916
1.569 0.031 1.537 0.894 1.332 2.110 1.694 0.973
1.840 0.051 1.789 1.937 1.572 1.990 1.912 0.965
1.895 0.061 1.833 0.959 2.309 2.095 1.725 1.012
2.072 0.083 1.989 1.034 2.050 2.499 1.891 1.036
2.441 0.098 2.343 1.267 2.292 3.267 1.914 1.135
(continued)
ON MOBILITY I63
Table I I (Continued)
1.0 0.430 0.004 0.426 0.297 0.373 0.422 0.533 0.510

0.443 0.006 0.437 0.322 0.366 0.427 0.554 0.537
0.466 0.007 0.460 0.341 0.406 0.449 0.570 0.542
0.460 0.009 0.451 0.334 0.416 0.445 0.556 0.4%
0.464 0.012 0.451 0.318 0.391 0.467 0.560 0.508
0.488 0.015 0.473 0.332 0.419 0.484 0.5% 0.518
0.531 0.019 0.512 0.368 0.445 0.520 0.646 0.572
0.548 0.023 0.525 0.382 0.478 0.536 0.646 0.565
0.578 0.027 0.551 0.381 0.495 0.572 0.671 0.619
0.594 0.042 0.552 0.388 0.500 0.573 0.683 0.591
0.613 0.051 0.562 0.380 0.488 0.577 0.709 0.648
0.669 0.067 0.602 0.381 0.553 0.610 0.787 0.644
0.706 0.080 0.626 0.457 0.539 0.656 0.787 0.664
0.5 0.375 0.004 0.371 0.263 0.331 0.369 0.448 0.446
0.388 0.006 0.382 0.289 0.326 0.364 0.472 0.481
0.407 0.007 0.401 0.299 0.362 0.389 0.481 0.492
0.404 0.009 0.396 0.297 0.373 0.384 0.473 0.467
0.404 0.012 0.392 0.286 0.345 0.399 0.472 0.474
0.423 0.014 0.409 0.294 0.371 0.406 0.501 0.497
0.456 0.018 0.438 0.325 0.387 0.434 0.537 0.562
0.468 0.021 0.446 0.334 0.400 0.448 0.546 0.557
0.494 0.025 0.469 0.336 0.418 0.469 0.571 0.639
0.496 0.038 0.457 0.325 0.417 0.464 0.578 0.584
0.505 0.047 0.458 0.331 0.395 0.452 0.606 0.652
0.551 0.061 0.489 0.336 0.466 0.472 0.684 0.632
0.571 0.073 0.498 0.406 0.423 0.500 0.698 0.637
0.0 0.367 0.004 0.362 0.258 0.323 0.359 0.430 0.437
0.381 0.006 0.375 0.283 0.318 0.346 0.462 0.488
0.402 0.007 0.395 0.293 0.361 0.375 0.466 0.506
0.399 0.008 0.390 0.292 0.374 0.37 1 0.458 0.493
0.395 0.012 0.384 0.282 0.388 0.385 0.454 0.501
0.415 0.014 0.401 0.287 0.369 0.391 0.483 0.539
0.488 0.017 0.432 0.323 0.382 0.416 0.521 0.635
0.458 0.020 0.438 0.344 0.384 0.430 0.532 0.634
0.492 0.024 0.468 0.388 0.411 0.452 0.563 0.802
0.481 0.036 0.446 0.318 0.408 0.411 0.565 0.669
0.483 0.044 0.439 0.329 0.376 0.416 0.599 0.768
0.541 0.057 0.484 0.341 0.489 0.435 0.69s 0.718
0.549 0.068 0.481 0.424 0.394 0.461 0.729 0.707
2305 359 525 641 471 309
164 MAASOUMI
instance, the fact that some mobility is experienced in the early part of this period,
but has ceased in the latter part of the sample. Similarly, we note that between-group
long-run inequality has risen consistently, suggesting that the expected returns to
schooling has materialized. In fact the inequality gap between the educational groups
here has widened by 50% over time; see John et al. (1993).
Tables 11-13 have the same structure as before but focus on the impact of
age. These tables suggest that between-group inequality, both short and long run,
has increased dramatically over time. Seniority matters! Within group inequality is
larger the older the group. This is also due to accumulation of returns to different
investments, opportunities, and attainments. Short-run inequalities increase within
groups, while long-run inequalities are stable with a moderate increase toward the
end of this period. Maasoumi et al. (1996)find these trends have continued. Finally
we note that these figures are based on per capita incomes. Since family size and
composition changes over time these figures show greater volatility than the authors
found with total family incomes unadjusted for family size (Maasoumi and Zandvakili
1989, Appendix).
Tables 14-15 provide decompositions by race, noting that the “non-white”
group includes all heads not classified as “white.” This explains the large within-
group inequality. The number of households in each group is given in the last col-
umn. Inequality among non-whites has increased faster than among whites. Short
run inequality has increased within both groups, and somewhat increased in the ag-
gregated incomes. Between-group inequality in the short run distributions declined
somewhat in the first half of the period and increased again in the last 4-5 years
of the sample. For the long-run incomes, between-group inequalities are rather sta-
ble with a slight decline over time. It would appear that within-group characteristics
are controlling of the degree of inequality in this sample and for this decomposition.
Other grouping criteria that are more race specific than “non-white” are known to
indicate greater between group inequality. See John et al. (1991).
Experimentation over the members of the GE family, as well as with different
sets of weights given to incomes at different points in time, represent an attempt to
robustify summary findings. This is an important element of empirical work in this
area since unanimity with respect to weights and degree of aversion to inequality is
not likely. Of course, an interpretation of this “robustification” technique is that it
is an empirical substitute for unanimous ranking by Lorenz-type comparisons over
plausible ranges of parameter values. This is useful when such curves cannot be
statistically ordered when they cross only at extreme parametric values.
Several other applications to U.S. and U.K. data are reported in Shorrocks
(1978b, 1981).A good deal more is now possible given the dominance testing tech-
niques of Section IV, and the asymptotic distribution theory summarized in Maa-
soumi (1996a). The bootstrap alternative appears very promising, as demonstrated
by Mills and Zandvakili (1996).
ON MOBILITY 165
Table I 2 1969-1981 per Capita Family Income, (MIW) Long-Run Inequalities

~
2.0 1.109 0.004 1.105 1.294 0.667 0.836 1.992 1.059

0.979 0.005 0.974 0.859 0.618 0.867 1.683 1.016
0.949 0.005 1.944 0.798 0.648 0.839 1.601 0.968
0.906 0.006 0.900 0.723 0.647 0.830 1.486 0.883
0.889 0.007 0.882 0.667 0.645 0.844 1.416 0.859
0.8% 0.008 0.889 0.634 0.651 0.932 1.359 0.812
0.951 0.009 0.942 0.625 0.667 0.950 1.614 0.784
0.955 0.010 0.944 0.637 0.707 0.985 1.544 0.764
0.974 0.012 0.962 0.635 0.741 1.033 1.482 0.740
1.007 0.014 0.993 0.716 0.782 1.077 1.447 0.735
1.035 0.016 1.018 0.711 0.866 1.117 1.386 0.733
1.060 0.020 1.039 0.712 0.909 1.171 1.331 0.729
1.100 0.025 1.075 0.718 0.963 1.260 1.290 0.725
1.0 0.430 0.004 0.426 0.297 0.373 0.422 0.533 0.510
0.417 0.005 0.4 12 0.277 0.353 0.408 0.528 0.505
0.414 0.005 0.409 0.277 0.352 0.404 0.527 0.492
0.412 0.006 0.406 0.277 0.355 0.402 0.523 0.475
0.408 0.007 0.402 0.271 0.351 0.403 0.517 0.461
0.409 0.007 0.402 0.269 0.352 0.407 0.516 0.455
0.414 0.009 0.405 0.272 0.355 0.412 0.521 0.456
0.418 0.010 0.408 0.275 0.360 0.416 0.521 0.455
0.426 0.012 0.414 0.277 0.369 0.428 0.523 0.454
0.429 0.014 0.415 0.276 0.372 0.433 0.520 0.455
0.434 0.017 0.417 0.275 0.374 0.437 0.518 0.464
0.445 0.021 0.424 0.276 0.380 0.445 0.525 0.471
0.455 0.026 0.429 0.282 0.386 0.453 0.526 0.473
0.5 0.375 0.004 0.371 0.263 0.331 0.369 0.448 0.446
0.364 0.005 0.359 0.246 0.313 0.351 0.447 0.448
0.364 0.005 0.358 0.246 0.314 0.349 0.448 0.442
0.362 0.006 0.356 0.246 0.318 0.347 0.443 0.433
0.360 0.006 0.353 0.243 0.316 0.349 0.440 0.426
0.360 0.007 0.352 0.241 0.315 0.350 0.438 0.425
0.363 0.008 0.355 0.243 0.318 0.353 0.441 0.433
0.368 0.010 0.358 0.247 0.320 0.358 0.444 0.437
0.373 0.011 0.362 0.249 0.325 0.366 0.446 0.442
0.377 0.014 0.364 0.248 0.329 0.371 0.445 0.450
0.381 0.017 0.364 0.247 0.328 0.371 0.447 0.466
0.387 0.021 0.366 0.248 0.330 0.373 0.454 0.472
0.396 0.026 0.370 0.257 0.334 0.378 0.460 0.476
(continued)
166 MAASOUMI
Table I 2 (Continued)

~~ ~ ~ ~
0.0 0.367 0.004 0.362 0258 0.323 0.359 0.430 0.437

0.355 0.005 0.350 0.240 0.304 0.336 0.432 0.444
0.355 0.005 0.350 0.240 0.308 0.334 0.432 0.443
0.353 0.006 0.348 0.241 0.314 0.332 0.427 0.440
0.351 0.006 0.345 0.238 0.311 0.334 0.422 0.436
0.350 0.007 0.343 0.235 0.310 0.335 0.419 0.439
0.354 0.008 0.346 0.238 0.313 0.338 0.423 0.455
0.3.59 0.010 0.349 0.244 0.313 0.343 0.426 0.466
0.366 0.011 0.355 0.248 0.319 0.352 0.429 0.487
0.370 0.013 0.356 0.246 0.323 0.357 0.427 0.499
0.371 0.016 0.355 0.246 0.32 1 0.353 0.429 0.522
0.378 0.02 1 0.358 0.249 0.329 0.352 0.438 0.529
0.386 0.026 0.361 0.264 0.328 0.355 0.447 0.533
2305 359 525 64 1 471 309
Table I3 1969-1981 per Capita Family Income, (MIW) Income Stability
2.0 1.000 0.004 0.996 1.ooo 1.000 1 .ooo 1.ooo 1.ooo

0.948 0.004 0.944 0.927 0.946 0.951 0.967 0.948
0.916 0.005 0.911 0.891 0.905 0.919 0.949 0.903
0.897 0.006 0.891 0.876 0.882 0.901 0.935 0.873
0.885 0.007 0.878 0.873 0.877 0.887 0.927 0.851
0.860 0.007 0.853 0.857 0.857 0.850 0.898 0.830
0.828 0.008 0.821 0.836 0.832 0.829 0.851 0.807
0.809 0.009 0.801 0.814 0.802 0.810 0.833 0.795
0.786 0.010 0.777 0.794 0.780 0.782 0.810 0.759
0.764 0.011 0.753 0.749 0.757 0.763 0.788 0.765
0.744 0.012 0.732 0.746 0.723 0.746 0.763 0.760
0.717 0.014 0.703 0.737 0.695 0.721 0.729 0.748
0.692 0.016 0.677 0.719 0.677 0.693 0.704 0.732
(continued)
ON MOBlLtn 167
1.0 1.000 0.010 0.990 1.ooo 1.000 1.000 1.000 1.000

0.955 0.011 0.944 0.893 0.955 0.961 0.971 0.964
0.928 0.012 0.917 0.864 0.922 0.934 0.955 0.929
0.915 0.013 0.902 0.853 0.907 0.922 0.945 0.913
0.903 0.014 0.889 0.842 0.901 0.912 0.934 0.893
0.891 0.016 0.875 0.832 0.889 0.903 0.918 0.881
0.877 0.018 0.859 0.819 0.878 0.890 0.903 0.867
0.865 0.021 0.844 0.812 0.867 0.878 0.888 0.856
0.855 0.023 0.831 0.801 0.861 0.876 0.872 0.832
0.843 0.027 0.815 0.788 0.852 0.868 0.853 0.829
0.832 0.032 0.800 0.778 0.845 0.860 0.835 0.827
0.822 0.039 0.783 0.774 0.832 0.854 0.187 0.824
0.813 0.046 0.767 0.765 0.831 0.846 0.799 0.813
0.5 1.000 0.011 0.989 1.000 1.000 1.000 1.000 1.000
0.954 0.013 0.94 1 0.890 0.953 0.959 0.970 0.963
0.932 0.013 0.918 0.867 0.925 0.936 0.959 0.934
0.917 0.014 0.903 0.856 0.910 0.922 0.946 0.918
0.911 0.016 0.894 0.850 0.910 0.918 0.940 0.905
0.897 0.018 0.878 0.837 0.895 0.908 0.923 0.893
0.885 0.021 0.865 0.826 0.887 0.896 0.909 0.881
0.877 0.023 0.853 0.823 0.877 0.889 0.897 0.873
0.867 0.027 0.840 0.814 0.873 0.888 0.881 0.846
0.861 0.032 0.830 0.806 0.870 0.887 0.865 0.851
0.855 0.038 0.817 0.798 0.866 0.881 0.852 0.856
0.843 0.046 0.797 0.793 0.846 0.871 0.832 0.850
0.840 0.005 0.785 0.795 0.850 0.868 0.818 0.844
0.0 1.000 0.012 0.988 1.000 1.000 1.000 1.000 1.000
0.950 0.013 0.937 0.887 0.949 0.954 0.967 0.959
0.927 0.014 0.9 13 0.863 0.92 1 0.930 0.954 0.927
0.912 0.014 0.897 0.8.54 0.908 0.915 0.940 0.913
0.904 0.016 0.888 0.846 0.907 0.911 0.933 0.899
0.890 0.018 0.871 0.832 0.890 0.901 0.914 0.886
0.879 0.020 0.859 0.822 0.884 0.890 0.900 0.875
0.872 0.023 0.849 0.824 0.873 0.885 0.889 0.867
0.864 0.026 0.838 0.818 0.872 0.887 0.871 0.842
0.859 0.031 0.828 0.810 0.872 0.887 0.853 0.847
0.852 0.038 0.8 14 0.8ott 0.865 0.877 0.837 0.855
0.841 0.046 0.795 0.800 0.852 0.865 0.817 0.847
0.839 0.056 0.784 0.8 15 0.850 0.862 0.802 0.841
2305 359 525 641 47 1 309
168 MAASOUMI
Table I 4 1969-1981 per Capita Family Irxome, Short-Run Inequalities
V Overall Between Within White Non-whi te
2:'o 1.109 0.087 1.022 0.599 1.099

0.961 0.085 0.877 0.609 0.870
1.049 0.088 0.961 0.718 0.901
0.940 0.085 0.855 0.638 0.810
1.002 0.076 0.925 0.602 0.980
1.190 0.079 1.111 0.890 1.022
1.636 0.074 1.562 0.770 1.891
1.357 0.070 1.286 0.816 1.414
1.569 0.069 1.so0 0.900 1.704
1.840 0.070 1.770 1.008 2.048
1.895 0.066 1.829 1.117 2.085
2.072 0.079 1.994 1.328 2.068
2.441 0.078 2.363 1.356 2.653
1.o 0.430 0.073 0.357 0.322 0.428
0.443 0.071 0.372 0.336 0.444
0.466 0.074 0.393 0.363 0.451
0.460 0.071 0.388 0.354 0.456
0.464 0.065 0.398 0.344 0.507
0.488 0.067 0.421 0.372 0.519
0.531 0.063 0.468 0.410 0.582
0.548 0.060 0.488 0.420 0.623
0.578 0.059 0.5 19 0.477 0.66 1
0.594 0.060 0.533 0.462 0.676
0.613 0.057 0.5Fi6 0.483 0.703
0.669 0.067 0.602 0.540 0.727
0.706 0.066 0.640 0.554 0.810
0.5 0.375 0.068 0.308 0.289 0.388
0.388 0.066 0.322 0.304 0.398
0.407 0.068 0.339 0.326 0.403
0.404 0.066 0.388 0.322 0.410
0.404 0.061 0.343 0.313 0.453
0.423 0.062 0.361 0.366 0.457
0.456 0.059 0.397 0.370 0.500
0.468 0.056 0.411 0.375 0.542
0.494 0.056 0.438 0.402 0.568
0.496 0.057 0.439 0.404 0.566
0.505 0.053 0.45 1 0.414 0.583
0.551 0.062 0.489 0.464 0.591
0.571 0.062 0.510 0.47 1 0.656
(continued)
ON MOBILITY I69
V Overall Between Within White Non-white
0.0 0.367 0.063 0.303 0.284 0.392

0.381 0.062 0.319 0.302 0.395
0.402 0.064 0.338 0.324 0.402
0.399 0.062 0.337 0.321 0.408
0.395 0.057 0.338 0.313 0.453
0.415 0.058 0.357 0.336 0.450
0.448 0.056 0.393 0.371 0.488
0.458 0.053 0.405 0.372 0.547
0.492 0.052 0.439 0.409 0.569
0.481 0.053 0.428 0.399 0.553
0.483 0.050 0.433 0.401 0.567
0.541 0.058 0.482 0.465 0.561
0.549 0.058 0.491 0.460 0.632
2305 1534 77 1
Table I 5 1969-1981 per Capita Family Income, (MIW)

Long-Run Inequality
V Overall Between Within White Non-white
2.0 1.109 0.087 1.022 0.599 1.099

0.979 0.087 0.892 0.565 0.923
0.949 0.089 0.860 0.577 0.858
0.906 0.089 0.817 0.561 0.804
0.889 0.088 0.801 0.545 0.795
0.896 0.089 0.808 0.570 0.784
0.951 0.088 0.862 0.567 0.872
0.955 0.087 0.868 0.566 0.885
0.974 0.086 0.888 0.570 0.916
1.007 0.088 0.919 0.581 0.950
1.035 0.087 0.947 0.595 0.984
1.060 0.087 0.973 0.613 1.012
1.100 0.087 1.013 0.628 1.062
(continued)
170 MAASOUMI
V Overall Between Within White Non-whi te
1.o 0.430 0.073 0.357 0.322 0.428

0.417 0.073 0.345 0.310 0.413
0.414 0.073 0.341 0.310 0.404
0.412 0.073 0.339 0.307 0.402
0.408 0.072 0.337 0.302 0.406
0.409 0.071 0.338 0.301 0.410
0.414 0.071 0.343 0.305 0.420
0.418 0.069 0.349 0.307 0.432
0.426 0.068 0.357 0.3 12 0.446
0.429 0.068 0.361 0.316 0.452
0.434 0.067 0.367 0.320 0.460
0.445 0.067 0.377 0.328 0.475
0.455 0.067 0.387 0.335 0.492
0.5 0.375 0.068 0.308 0.289 0.388
0.364 0.067 0.297 0.279 0.373
0.364 0.068 0.296 0.280 0.365
0.362 0.067 0.294 0.278 0.363
0.360 0.066 0.294 0.275 0.370
0.360 0.065 0.294 0.275 0.372
0.363 0.064 0.299 0.279 0.379
0.368 0.063 0.304 0.282 0.391
0.373 0.062 0.311 0.287 0.403
0.377 0.062 0.3 16 0.291 0.409
0.381 0.061 0.320 0.295 0.416
0.387 0.061 0.326 0.300 0.424
0.396 0.061 0.335 0.307 0.439
0.0 0.367 0.063 0.303 0.284 0.392
0.355 0.062 0.293 0.276 0.372
0.355 0.063 0.292 0.277 0.364
0.353 0.063 0.291 0.276 0.361
0.351 0.061 0.290 0.273 0.369
0.350 0.060 0.290 0.272 0.370
0.354 0.059 0.295 0.278 0.374
0.359 0.058 0.301 0.281 0.390
0.366 0.057 0.309 0.288 0.404
0.370 0.056 0.313 0.292 0.409
0.371 0.055 0.316 0.294 0.415
0.378 0.056 0.322 0.301 0.419
0.386 0.056 0.330 0.307 0.432
2305 1534 77 1
ON MOBILITY 171
ACKNOWLEDGMENTS
I would like t o t h a n k t h e editors, an a n o n y m o u s referee, E. Ok,and s e m i n a r partici-

p a n t s for comments.
REFERENCES
Anderson, G. J. (1994), Nonparametric Tests of Stochastic Dominance in Income Distribu-

tions, Working Paper, University of Toronto.
Atkinson, A. (1970), On the Measurement of Inequality, Journal of Economic Theory, 11,
244-263.
Atkinson, A. (1983), The Measurement of Economic Mobility, in Social Justice and Public
Policy, MIT Press, Cambridge, MA.
Atkinson, A. and F. Bourguignon (1982),The Comparison of Multidimensioned Distributions
in Economic Status, Review ojEconomic Studies, 49, 183-201.
Atkinson, A. and F. Bourguignon (1987), Income Distribution and Differences in Needs,
in G . R. Feiwel (ed.), Arrow and the Foundations of the Theory of Economic Policy,
Macmillan, London.
Atkinson, A., F. Bourguignon, and C. Morrison (1992),Empirical Studies ojEarnings Mobility,
Fundamentals of Pure and Applied Economics, vol. 52, Distribution Section, Harwood
Academic.
Bartholomew, D. J. (1959), A Test of Homogeneity for Ordered Alternatives, Biometrika, 46,
36-48.
Bartholomew, D. J. (1982), Stochastic Modelsfor Social Processes, 3rd ed., Wiley, New York.
Beach, C. M. and R. Davidson (1983), Distribution-Free Statistical Inference with Lorenz
Curves and Income Shares, Review ($Economic Studies, 50, 723-735.
Beach C. M., R. Davidson, and G. A. Slotve (1995), Distribution-Free Statistical Inference
for Inequality Dominance with Crossing Lorenz Curves, Unpublished Paper N95A03,
Universi t6 d'Aix-Marseille.
Beach, C. M. and J. Richmond (1985), Joint Confidence Intervals for Income Shares and
Lorenz Curves, International Economic Review, 26, 439-450.
Bishop, J. A., K. V. Chow, and J. P. Formby (1994), Testing for Marginal Changes in Income
Distributions with Lorenz and Concentration Curves, International Economic Review,
35,479-488.
Bishop, J. A., J. P. Formby, and P. D. Thistle (1989), Statistical Inference, Income Distribu-
tions, and Social Welfare, Research on Economic Inequality, 1 , 4 9 4 2 .
Blackorby, C. and D. Donaldson (1978), Measures of Relative Equality and Their Meaning in
Terms of Social Welfare, Journal ($Economic Theory, 18, 59-80.
Bourguignon, F. (1979), Decomposable Income Inequality Measures, Econometrica, 47,901-
920.
Chakravarty, S.R . (1995), A Note on the Measurement of Mobility, Economics Letters, 48,
33-36.
Conlisk, J. (1989), Ranking Mobility Matrices, Economics Letters, 29,231-235.
Conlisk, J. (1990), Monotone Mobility Matrices, Journal ofMuthematical Sociology, 15, 173-
191.
172 MAASOUMI
Creedy, J. (1985), Dynamics of Income Distribution, Basil Blackwell, Oxford.

Dardanoni, V. (1992), On Multidimensional Inequality Measurement, Unpublished Paper,
University of Perugia.
Davies, J. and M. Hoy (1994), The Normative Significance of Using Third-Degree Stochastic
Dominance in Comparing Income Distributions, Journal of Economic Theory, 64,520-
530.
Davies, J. and M. Hoy (1995), Making Inequality Comparisons When Lorenz Curves Intersect,
American Economic Review, 85,980-986.
Deshpande, J. V. and S. P. Singh (1985), Testing for Second Order Stochastic Dominance,
Communications in Statistics A , 14,887-893.
Dardanoni, V. (1993), On Measuring Social Mobility, Journal of Economic Theory, 61,372-
394.
Durbin, J. (1973), Distribution Free Tests Based on the Sample Distribution Function, SIAM,
Philadelphia, PA.
Durbin, J. (1985), The First Passage Density of a Continuous Gaussian Process to a General
Boundary, Journal of Applied Probability, 22,99-122.
Dykstra, R. L., S. Kochar and T. Robertson (1995), Inference for Likelihood Ratio Ordering in
Two Sample Problems, Journal ofthe American Statistical Association, 90,1034-1040.
Ebert, U. (1995a),Income Inequality and Differences in Household Size, Mathematical Social
Sciences, 30, 1034-1040.
Fields, G. and E. A. Ok (1995), The Meaning and the Measurement of Income Mobility, in J.
Silber (ed.), Income Inequality Measurement: From Theory to Practice, Kluwer, forth-
coming.
Geweke, J., R. C. Marshall, and G. A. Zarkin (1986), Mobility Indices in Continuous Time
Markov Chains, Econometrica, 54, 1407-1423.
Gourieroux, C., A. Holly, and A. Monfort (1982), Likelihood Ratio Test, Wald Test and Kuhn-
Tucker Test in Linear Models with Inequality Constraints in the Regression Parameters,
Hadar, J. and W. R. Russell (1969), Rules for Ordering Uncertain Prospects, American Eco-
nomic Review, 5 9 , 2 5 3 4 .
Hadar, J. and W. R. Russell (1971), Stochastic Dominance and Diversification, Journal of
Economic Theory, 3,288-305.
Hanoch, G. and H. Levy (1%9), The Efficiency Analysis of Choices Involving Risk, Review
of Economic Studies, 36,335-346.
Hart, P. E. (1976a),The Dynamics of Earnings, 1963-1973, Economic Journal, 86,541-565.
Hart, P. E. (1976b),The Comparative Statics and Dynamics of Income Distributions, Journal
of the Royal Statistical Society A , 139, 108-125.
Hart, P. E. (1981),The Statics and Dynamics of Income Distributions: A Survey, in A. Klever-
marken and J. A. Lybeck (eds.), The Statics and Dynamics of Income, Tieto, Clevedon.
Hart, P. E. (1983),The Size Mobility of Earnings, Oxford Bulletin of Economics and Statistics,
45,181-193.
Juhn, C., K., M. Murphy, and B. Pierce (1991), Accounting for the Slowdown in Black-White
Wage Convergence, in M. H. Kosters (ed.), Workers and Their Wages: Changing Pat-
terns in the United States, American Enterprise Institute Press, Washington, DC.
Juhn, C. and K. M. Murphy, and B. Pierce (1993),Wage Inequality and the Rise in Returns to
Skill, Journal of Political Economy, 101,410-442.
ON MOBILITY 173
Kanbur. S. M. R. and J. E. Stiglitz (1986), Intergenerational Mobility and Dynastic Inequality,

Woodrow Wilson Discussion paper No. 111, Princeton University.
Kaur, A., B. L. S.Rao, and H. Singh (1994),Testing for Second-Order Stochastic Dominance
of Two Distributions, Econometric Theory,lO, 849-866.
Klecan, L., R. McFadden, and D. McFadden (1991), A Robust Test for Stochastic Dominance,
Working Paper, Department of Economics, MIT.
Klevermarken, A. and J. A. Lybeck (eds.), The Statics and Dynamics oflncome, Tieto, Cleve-
don.
Kodde, D. A. and F. C. Palm (1986), Wald Criteria for Jointly Testing Equality and Inequality
Restrictions, Econometrica, 50, 1243-1248.
Kolm, S.C. (1969), The Optimal Production of Social Justice, in J. Margolis and H. Guit-
ton (eds.), Public Economics. Also presented at IEA Conference in Public Economics,
Biaritz, 1966.
Kolm, S.C. (1976a), Unequal Inequalities: I, Journal ofEconomic Theory, 12,416442.
Kolm, S. C. (1976a), Unequal Inequalities: 11, Journal ofEconomic Theory, 13,82-111.
Lehmann, E. L. (1959), Testing Statistical Hypotheses, Wiley, New York.
Lillard, L. and R. Willis (1978), Synamic Aspects of Earnings Mobility, Econometrica, 46,
985-1012.
Maasoumi, E. (1986), The Measurement and Decomposition of Earnings Mobility, Economet-
rica, 54,991-997.
Maasoumi, E. (1989a), Composite Indices of Income and Other Developmental Indicators: A
General Approach, Research in Income Inequality, forthcoming.
Maasoumi, E. (1 989b), Continuously Distributed Attributes and Measures of Multivariate In-
equality, Journal of Econometrics, 42, 131-144.
Maasoumi, E. (1991), ed., Measurement oj'weljhre and Inequality, Annals issue of Journal of
Econometrics, 50, 112.
Maasoumi, E. (1993),A Compendium to Information Theory in Economics and Econometrics,
Econometric Reviews, 12, 1-49.
Maasoumi, E. (1996a), Empirical Analyses of Welfare and Inequality, in P. Schmidt and M. H.
Pesaran (eds.), Handbook ojApplied Microeconometrics, Blackwell, Oxford, forthcom-
ing.
Maasoumi, E. (1996b), Multidimensioned Approaches to Welfare Analysis, Chapter 15 in J.
Silber (ed.), Income Inequality Measurement: From Theory to Practice, Kluwer, forth-
coming.
Maasoumi, E., J. Mills, and S.Zandvakili (1996), Consensus Ranking of US Income Distribu-
tions: A Bootstrap Application of Tests for Stochastic Dominance, mimeo, Department
of Economics, Southern Methodist University.
Maasoumi, E. and J. Nickelsburg (1988), Multivariate Measures of Well Being and an Analysis
of Inequality in the Michigan Data, Journal of Business and Economic Statistics, 6,
327-334.
Maasoumi, E. and M. Trede (1997), Statistical Inference for General Mobility Measure, Work-
ing Paper, Department of Economics, Southern Methodist University.
Maasoumi, E. and S. Zandvakili (1986), A Class of Generalized Measures of Mobility with
Applications, Economics Letters, 22,97-102.
Maasoumi, E. and S.Zandvakili (1989), Mobility Profiles and Time Aggregates of Individual
Incomes, Research on Income Inequality, 195-218.
174 MAASOUMI
Maasoumi, E. and S. Zandvakili (1990), Generalized Entropy Measures of Mobility for Differ-
ent Sexes and Income Levels, Journal of Econometrics, 43, 121-133.
Markandya, A. (1984), The Welfare Measurement of Changes in Economic Mobility, Econo-
metrica, 51,457-471.
McFadden, D. (1989), Testing for Stochastic Dominance, in T. Formby and T. K. Seo (eds.),
Studies in the Economics of Uncertainty, Part 11 (in Honor of J. Hadar), Springer-Verlag,
New York.
Mills, J. and S.Zandvakili (19%), Bootstrapping Inequality Measures, Working Paper, De-
partment of Economics, University of Cincinnati.
Mitra, T. and E. A. Ok (1995), The Measurement of Income Mobility: A Partial Ordering
Approach, mimeo, Department of Economics, Cornell University.
Perlman, M. D. (1969), One-sided Testing Problems in Multivariate Analysis, Annals ofMath-
ematical Statistics, 40, 549-562.
Prais, S.J. (1955), Measuring Social Mobility, Journal of the Royal Statistical Society A , 56-
66.
Robertson, T. and F. T. Wright (1981), Likelihood Ratio Tests for and Against Stochastic Or-
dering Between Multinomial Populations, Annals ofStatistics, 9, 1248-1257.
Roberston, T. and F. T. Wright, and R. Dykstra (1982), Order Restricted Statistical Inference,
Wiley, New York.
Sargan, J. D. (1957), The Distribution of Wealth, Econometrica, 25,568-590. Also Reprinted
as Chapter 3 in E. Maasoumi (ed.), Contributions to Econometrics: J. D.Sargan, Vol. 1,
Cambridge University Press, Cambridge.
Sen, A. (1970), Collective Choice and Social Weware, Holden-day, San Francisco, reprinted,
North-Holland, Amsterdam, 1979.
Shapiro, A. (1988),Towards a Unified Theory of Inequality Constrained Testing in Multivariate
Analysis, International Statistical Review, 5 6 , 4 9 4 2 .
Shorrocks, A. F. (1976), Income Mobility and the Markov Assumption, Economic Journal, 86,
566-577.
Shorrocks, A. F. (1978a), The Measurement of Mobility, Econometrica, 46, 1013-1024.
Shorrocks, A. F. (1978b), Income Inequality and Income Mobility, Journal of Economic The-
ory, 19,376-393.
Shorrocks, A. F. (1980), The Class of Adaptively Decomposable Inequality Measures, Econo-
metrica, 48,376-393.
Shorrocks, A. F. (1981), Income Stability in the IJnited States, Chapter 9 in N. A. Klevmarken
and J. A. Kybeck (eds.), The Stutics and Dynamics of Income, Tieto, Clevedon.
Shorrocks, A. F. (1983), Ranking Income Distributions, Economica, 50,s-17.
Shorrocks, A. F. (1984), Inequality Decomposition by Population Subgroups, Econornetrica,
52,1369-1385.
Shorrocks, A. F. and J. Foster (1987), Transfer Sensitive Inequality Measures, Review of Eco-
nomic Studies, 54,485-497.
Shorrocks, A. F. and J. Foster (1992), On the Hart Measure of Income Mobility, in Industrial
Concentration and Income Inequality: Festschrzj f o r Peter Hart, forthcoming.
Shorrocks, A. F. and J. Foster (1995), Inequality and Welfare Evaluation of Heterogeneous In-
come Distributions, Unpublished Paper, Department of Economics, University of Es-
sex.
Theil, H. (1972), Statistical Decomposition Analysis, North-Holland, Amsterdam.
ON MOBILITY 175
Trede, M. (1995), Statistical Inference in Mobility Measurement: Sex Differences in Earnings

Mobility, Unpublished Paper, University of Koln.
U S . Bureau of the Census (Various Years), Current Population Reports: Consumer Income,
Series P-60.
Wolak, F. A. (1988), Duality on Testing Multivariate Hypotheses, Biometrika, 75,611-615.
Wolak, F. A. (1989), Local and Global Testing of Linear and Nonlinear Inequality Constraints
in Nonlinear Econometric Models, Econometric Theory, 5, 1-35.
Wolak, F. A. (1991), The Local Nature of Hypothesis Tests Involving Inequality Constraints
in Nonlinear Models, Econometricn, 5 9 , 9 8 1-995.
Xu, K. (1995), Asymptotically Distribution-Free Statistical Test for Generalized Lorenz Curves:
An Alternative Approach, Department of Economics, Dalhousi University. Paper pre-
sented at the 7th World Congress of the Econometric Society, Japan.
Xu, K., G. Fisher, and D. Wilson (1995),New Distribution-Free Tests for Stochastic Domi-
nance, Working Paper No. 95-02, Department of Economics, Dalhousi University.
Zandvakili, S. (1987), The Measurement ojMobility and Income Stabi1ity:An Information The-
oretic Approach, Ph.D. thesis, Department of Economics, Indiana University, Bloom-
ington.
Aggregation and Econometric
Analysis of Demand and Supply
R. Robert Russell, Robert V. Breunig, and Chia-Hui Chiu

University of California at Riverside, Riverside, California
1. INTRODUCTIONAND SUMMARY
A. Overview
The neoclassical development of producer and consumer theory, culminating in the
use of duality theory and the introduction of flexible functional forms in the 1970s,
focused on the restrictions on demand and supply functions implied by optimizing
behavior of producers and consumers. These restrictions are completely character-
ized by the symmetry and negative semidefiniteness of the (Slutsky) matrix of sub-
stitution effects in consumer theory and the symmetry and positive semidefiniteness
of the Jacobian of the (net) supply functions in the case of producer theory.* They
are important for econometric demand and supply analysis in part because they re-
duce the number of independent parameters to be estimated. A classic example is
development of the linear expenditure system, first estimated by Stone (1954). Be-
ginning with a system of equations in which optimal expenditure on each commodity
is a linear function of income and n prices, Klein and Rubin (1947-1948) showed
that requiring the system to be generated by income-constrained utility maximization
reduces the number of parameters to be estimated from ( n 1)2 to 2n - 1.+
The problem with this nexus between theory and empirical application is that
the estimation of demand and supply systems typically uses aggregate, or per capita,
*In addition, these matrices must have reduced rank hecause of homogeneity properties. The conditions
are most easily derived in the dual from the (easily proved) concavity and homogeneity of expenditure
functions and convexity and homogeneity of profit functions in prices.
I 77
I78 RUSSELL
ET AL.
data. In the case of aggregate supply functions of producers, this poses no problem so
long as all inputs are efficiently allocated, as is the case in competitive equilibrium
with no fixed inputs and no production externalities. In this case, profit maximiza-
tion by all producers on individual technology sets yields the same aggregate net
supply as profit maximization on the aggregate technology set, which is obtained
by simple summation of the individual technology sets. The equivalence of these
two optimization problems is salient in general-equilibrium theory and welfare eco-
nomics (Debreu 1954,1959) and has been elegantly illustrated and aptly referred to
by Koopmans (1957) as the “interchangeability of set summation and optimization.”
The essence of Koopman’s interchangeability principle is that boundary points of
the aggregate technology set are obtained as the sum of boundary points of the in-
dividual technology sets where the supports are equal; but this is equivalent to an
efficient allocation of net outputs, where the aggregate net output vector and the op-
timal vectors of each producer are supported by the same price vector. (The vector
summation of boundary points of individual technology sets with unequal supports
yields interior points of the aggregate technology set.)
From the standpoint of econometric applications (and other applications as
well, especially in macroeconomic theory and international trade), the beauty of the
aggregation result for profit-maximizing, price-taking producers is that no restric-
tions (other than those needed for the existence of an optimum) are required. (In
particular, convexity of technology sets is not required.) This means that there is no
loss of generality in positing the existence of a “representative producer,” which gen-
erates aggregate net-supply functions by maximizing aggregate profit subject to the
constraint that the aggregate net-supply vector be contained in the aggregate tech-
nology set. As a result, the Jacobian of the system of aggregate net supply functions
has the same properties as those of individual producers.
But aggregation on the consumer side is not so simple; the symmetry and neg-
ative semidefiniteness of the substitution matrix does not carry over to aggregate de-
mand systems. In fact, as shown by Debreu (1974), Mantel (1974), and Sonnenschein
(1973),* the only restrictions imposed on aggregate demand functions by individual
optimization are Walras’ law (simply the aggregate budget identity) and homogeneity
of degree zero (in prices and income for income-constrained consumers and in prices
for endowment-constrained consumers). This, in turn, implies that, without further
restrictions, the use of a representative agent in consumer theory is, a fortiori, un-
justified. Essentially, the reason for this discouraging result in the aggregation of
consumer demand systems is that aggregate demand depends on the arbitrary dis-
tribution of incomes or endowments. In fact, if consumer incomes were determined
*See also Mas-Colell and Neuefeind (1977). An excellent survey of these results can be found in Shafer
and Sonnenschein (1982).
ANALYSIS
OF DEMAND
AND SUPPLY 179
optimally by maximizing a (Bergson-Samuelson) social welfare function, a represen-

tative consumer would exist (Samuelson 1956).* (Similarly, if the econometric anal-
ysis of producer demand and supply systems entailed an arbitrary distribution of
idiosynchratic variables, such as fixed capital stocks, the simple aggregation results
would not go through for producers as well.)
There are essentially two ways out of this quandary. One is to impose restric-
tions on individual preferences that imply certain regularity conditions for aggregate
demands. These range from the most restrictive assumption, yielding a representa-
tive agent in aggregate demands, to weaker restrictions implying that the aggregate
demand systems satisfy the weak axiom of revealed preference (equivalently, that
the Jacobian of the aggregate demand system is negative semidefinite).
The first approach to dealing with the consumer aggregation problem predates
the Debreu-Sonnenschein-Mantel results and indeed goes all the way back to the
classic result of Gorman (1953a). He showed that a representative consumer exists
if and only if Engel curves of the consumers are linear and parallel across consumers
(for given prices). Gorman’s result is derived in the dual. The indirect utility function
must be affine in income (for given prices), in which case the (Marshallian) ordinary
demand functions, derived by Roy’s Identity, are also affine in income. Equivalently,
the consumer’s expenditure function is affine in utility (or any monotonic transforma-
tion of utility), in which case the (Hicksian) income-compensated demand functions
are affine in the utility scalar. This structure is commonly referred to as the “Gor-
man polar form,” and has been elucidated in the primal by Blackorby, Boyce, and
Russell(l978) and Blackorby and Schworm (1993). Gorman’s result is discussed in
Section 1I.A.
Gorman’s notion was subsequently generalized by Muellbauer to obtain a
weaker type of aggregation and, concomitantly, weaker form of representative con-
sumer, one in which aggregate commodity shares depend on some (generally non-
linear) function of the distribution of income and (possibly) prices. He called the
required structure “generalized linear,’’ since it is a generalization of Gorman’s “lin-
ear” (actual affine) structure. Muellbauer’s generalized linear structure can be al-
ternatively characterized by the fact that the Engel curves have “rank 2”: they are
contained in a two-dimensional subspace of the commodity space. That is, each can
be obtained as a linear combination of the others. Requiring that the aggregate in-
come be independent of prices yields Muell bauer’s “price-independent generalized
linearity” (PIGL) demand system or his “price-independent generalized logarith-
mic” (PIGLOG) system. Muellbauer’s approach, along with the extension by Deaton
and Muellbauer (1980), the “almost ideal demand system,” is discussed in Sec-
tion 1I.B.
*See also, Jerison (1984b).

I80 RUSSELLET AL.
Section I1 closes with a brief discussion of the tests for the existence of a rep-
resentative agent, including the parametric tests of Christensen, Jorgenson, and Lau
(1975), and others (usingflexible functional forms like the translog) and the nonpara-
metric, nonstochastic (mathematical programming) tests of Varian (1982) and Diew-
ert and Parkan (1985).The power of these nonparametric tests has been assessed
by Bronars (1987) and Russell and Tengesdal (1996).Lewbel (1991) has provided
some evidence in favor of Muellbauer’s PIGLOG specification if we exclude incomes
in the tails of the distribution, but evidence against it if we include the tails.
The existence of a representative consumer, while necessary for many pur-
poses, is not necessary for the existence of aggregate demand functions that re-
quire less-than-complete information about the distribution of incomes. Lau (1977a,
1977b, 1982) and Gorman (1981) spelled out restrictions on individual demand
functions that are necessary and sufficient for the existence of an aggregate demand
function that depends on prices and summary statistics of the income distribution.
This weaker aggregation condition, discussed in Section III.A, is referred to as “exact
aggregation” and is a natural generalization of the (rank 2) Muellbauer conditions.
In his remarkable theorem, Gorman (1981)also showed that the requisite indi-
vidual demand systems are consistent with income-constrained utility maximization
(i.e., satisfy the integrability conditions) if and only if the Engel curves can be con-
tained in a three-dimensional subspace (for given prices)-that is, that the demand
systems have rank no greater than 3. His theorem, presented in Section III.B, also
completely characterizes the class of such functions, which encompasses virtually
all demand systems that have been estimated econometrically. Gorman’s theorem
has been extended and clarified in a series of papers by Heineke (1979, 1993)and
Heineke and Shefrin (1982, 1986, 1987, 1988).
Consumer attributes (like household size, geographical region, age of head,
etc.) have been incorporated into the exact aggregation framework by Lau (1977a,
1977b, 1982) and imp1emente.l Ilsing the translog specification, by Jorgenson, Lau,
and Stoker (1980).A related issde is the recovery of the parameters of individual
demand systems from the estimati, n of the aggregate demand system. The neces-
sary and sufficient conditions for ihis identification property have been developed
by Heineke and Shefrin (1990). This literature is briefly discussed in Section 1II.C.
Additional research by, for example, Jorgenson, Lau, and Stoker (1981), Pol-
lak and Wales (1978, 1980), Stoker (1984), Russell(1983), Jorgenson and Slesnick
(1984), Stoker (1986a, 1986b), Buse (1992), Blundell, Pashardes, and Weber (1993),
and Nicol(l994) have further developed and applied these ideas. A summary of this
literature is presented in Section III.D.*
*The notion of exact aggregation can also be applied to firms with individual characteristics pertaining
to the technology or to fixed inputs or outputs (see Appelbaum 1982, Borooah and Van Der Ploeg 1986,
Gourikroux 1990, Fortin 1991, and Chavas 1993).
OF DEMAND
ANALYSIS AND SUPPLY I 8I
Gorman’s results on the rank of demand systems have been extended through
a series of papers that have characterized well-known-and some lesser-known-
specifications. The key contributions have been made by Lewbel (1987, 1989b,
1990,1991).In a related study, Hausman, Newey, and Powell(l995) provide a cross-
sectional test of the Gorman rank-3 condition.
The rank-2 phenomenon shows up in other studies of aggregation (e.g., Jeri-
son’s (19Ma) results on painvise aggregation with a fixed income distribution) and
as one of the necessary and sufficient conditions for the weak axiom of revealed pref-
erence to hold in the aggregate (Freixas and Mas-Collel 1987). Curiously, the rank-2
condition also emerges in a study of a (proportional) budgeting procedure for an or-
ganization (Blackorby and Russell 1993). These disparate results, which seem to be
crying out for a unifying general theorem, are discussed in Section IV.
The second approach to dealing with the paucity of implications for aggre-
gate consumer demand systems of individual optimization is to restrict the distribu-
tion of incomes or preferences of the population. Hildenbrand (1983,1993), Hardle,
Hildenbrand, and Jerison (1991), and Grandmont (1992) seek additional restrictions
on aggregate demand implied by reasonable restrictions on the distribution of income
or preferences. An early precurser of these results is Becker (1962), who showed
that aggregate demand curves will be downward sloping even if individuals are “ir-
rationally” nonoptimizing, in the sense that they distribute their demand “randomly”
across the budget hyperplane (in particular, according to a rectangular distribution).
The intuition behind this result is fairly obvious: the Giffen paradox will occur only
if a sufficient number of individuals is concentrated in the Giffen portions of their
demand functions, but this will not happen if the distribution of consumers across
the budget plane is rectangular (more generally, “sufficiently” dispersed). This is
the motif for the results of Hildenbrand and others, who show that the Jacobian of
the aggregate demand system will be negative semidefinite (equivalently, that the
weak axiom of revealed preference holds in the aggregate), implying that demand
curves are downward sloping if the distribution of income is nonincreasing. Sim-
ilarly, Grandmont derives the same restriction on aggregate demands by assuming
that (neatly parameterized) preferences are sufficiently heterogeneous. These results
are surveyed in Section V.
As noted above, there is no aggregation problem on the production side of the
economy only if all inputs and outputs are efficiently allocated. If some inputs are
not efficiently allocated, aggregation is not so straightforward. Inefficient allocation
would occur, for example, if some inputs (e.g., capital) are fixed in the short run
(and perfect capital rental markets do not exist). If we take as given the distribution
of fixed inputs among firms, aggregation over the variable inputs is straightforward.
There has been, however, a persistent interest in the question of the existence of an
aggregate amount of fixed inputs, such as an aggregate capital stock that is fixed in
the short run. Interest in this area, sparked in part by the “Cambridge controversy,”
has centered not only on aggregation of fixed inputs across firms, but also across dif-
I82 RUSSELLET AL.
ferent fixed inputs. At the individual level, the existence of such aggregates is known
to be equivalent to certain separability conditions. (See the classic contributions by
Gorman 1959, 1968a and the subsequent expositions by Blackorby, Primont, and
Russell 1978, 1997.) Requiring the existence of such commodity aggregates at the
macro level, however, requires stronger restrictions on individual technologies.
This aggregation problem was, in fact, first posed by Klein (1946a, 1946b),
solved by Nataf (1948), and extended by Gorman (1953b). The Klein-Nataf aggrega-
tion problem assumes that no inputs are efficiently allocated and leads to a very un-
realistic (linear) structure for individual technologies. It was pointed out early, how-
ever, by May (1946) and Pu (1946), and emphasized later by Solow (1964), that the
efficient allocation of some inputs could be used to restrict the admissable allocations
and hence weaken the aggregation conditions for the fixed inputs. These conjectures
turned out to be correct, as rigorously shown by Gorman (1968b).* Blackorby and
Schworm (1984,1988a) provide comprehensive treatments of the problem of obtain-
ing aggregate inputs in aggregate technologies under different assumptions about the
existence of efficiently allocated and fixed inputs. The problem of the existence of ag-
gregate commodities also is relevant to the empirical analysis of consumer demand,
since most studies employ such aggregates; Blackorby and Schworm (1988b) pro-
vide necessary and sufficient conditions for the existence of commodity aggregates
in market demand functions.
These results on aggregation across both agents and commodities are surveyed
in Section VI. Section VII concludes.
B. Caveats
The aggregation literature surveyed in this chapter overlaps with an extensive liter-
ature on the specification of functional form, cross-section (Engel curve) estimation,
and other areas of interest to applied econometricians. While we unavoidably touch
on these subjects, we limit our discussions to our focus on the issue of aggrega-
tion over agents in econometric estimation. For excellent surveys of the literature
on econometric demand analysis, specification of functional form, and econometric
modeling of producer behavior see, respectively, Deaton (1986), Lau (1986), and
Jorgenson (1986).
Another topic that we do not cover is the aggregation of individual preferences
to obtain a social welfare function. There is, of course, a huge (social choice) litera-
ture on this topic, emanating from the classic impossibility theorem of Arrow (1951).
As noted above, aggregate data will be consistent with the existence of a represen-
tative consumer if commodities are allocated by maximizing a Bergson-Samuelson
*See also the series of papers by Fisher (1965, 1968a, 1968b, 1982, 1983).
ANALYSISOF DEMAND
AND SUPPLY I83
social welfare function. This fact, of course, is of little use in the econometric esti-
mation of demand systems in a market economy.
It is important to distinguish the notion of a representative agent that is useful
in econometric studies from the notion of a “representative consumer” commonly
employed in the macro/finance literature regarding the replication of a competi-
tive equilibrium (Constantinides 1982, Aiyagari 1985, Huang 1987, Eichenbaum,
Hansen, and Singleton 1988, and Vilks 1988a, 198833).This concept requires no re-
striction on preferences, which is not surprising, since it generates a representative
consumer only at equilibrium points, but not in even a neighborhood of prices and
incomes; as such, it is not a useful construct for econometric applications.*
Finally, despite our efforts to limit the scope of this survey, we discovered
that the literature on aggregation across agents is even more extensive than we had
expected (and is still burgeoning). Hence, our main focus has been to try to integrate
this large body of research (something that has not been done since the publication
of Green 1964);t we have made no serious attempt to critique it. Even with this
limited objective, we have not been as successful as we would have liked; it seems
to us that there is a need for a monograph on aggregation over agents-one that would
be accessible to econometricians as well as theorists.
II. AGGREGATION AND REPRESENTATIVE AGENTS

A. Gorman Aggregation
The modern theory of consumer aggregation has its genesis in the classic paper by
Gorman (1953a). He posed the following question. Suppose we have H optimizing
consumers, each maximizing utility subject to a budget constraint with fixed prices
and total expenditure. Thus, each consumer has a demand function, d h , generated
bY
& ( p , y h ) = argmax,h{U h (xh I p . x h 5 y h ) E R; (1)
where p E R ;, is the (common) price vector and Uh, xh, and y h are the utility
function, consumption vector, and total expenditure of consumer h. The aggregate
demand vector is given straightforwardly by$
*This characterization is not intended to demean these results: while they cannot be used in economet-
ric modeling, they are relevant to calibration exercises that attempt to simulate a single history of an
economy (cf., the large body of calibration stimulated by Kydland and Prescott 1982).
7 See, however, van Daal and Merkies (1984),who cover some of‘the topics addressed in the survey.
$Notation: A =: B or B := A means B is defined by A.
I84 RUSSELLET AL.
Simple additive aggregation holds if aggregate demand is independent of the distri-

bution of total expenditures-i.e., if there exists a function, D, such that
Theorem. Assume that d h ( y , y h ) > 0for all h.* The necessary and suficient
restriction on the individual demand functions, d", h = 1 , . . . , H , f o r additive ag-
gregation, (3),to hold is that they take the form
Sufficiency of this structure is obvious, since?
h h
Thus, the individual demand functions must be affine in the (idiosynchratic)

total-expenditure variables with a common (price dependent) coefficient on these
variables. In fact, this is a general condition for any set of structural equations to
aggregate consistently as in (3):the individual functions must be affine, with com-
mon coefficients for different households, in all idiosynchratic variables-those that
require an individual identifier index h-and the coefficients can depend on the
common variables (in this case, prices).$ The equations can be made stochastic and
still satisfy the aggregation condition as long as the vector of disturbance terms, 8,
enters additively; e.g.,
d h ( p , y h ) = &p> + B ( p > y h+ E h Vh
in which case the disturbance term of the macro function is Eh e h :
*If d ( p , y h ) is restricted to be nonnegative, but corner solutions are allowed, it is necessary that a h ( p )=
0 for all h in the following restriction.
TSketch of proof of necessity (assuming differentiahility): Substitute (2) into (3) and differentiate with
respect to each y h to obtain
This shows that, for all i, the derivative is identical for all h. Moreover, as the left-hand side (LHS) is
L
independent of y h for # h, the right-hand side (RHS) must be independent of y , which in turn implies
that the LHS is, in fact, independent of y h as well. Integration then yields (4).More generally (eschewing
differentiability), (3) is a system of Pexider equations, whose solution is (4);see Corollary 10 on page 43
of Aczel and Dhombres 1989.
$If one does not require that demands aggregate exactly, as in (3), but only that the expected value of
zh be independent of the distribution of y h , then it is sufficient that the fi coefficients be distributed
independently of xh.
OF DEMAND
ANALYSIS AND SUPPLY I85
The result in the above theorem has been known for a long time; see, for ex-
ample, the papers culminating in Theil (1954) and reviewed by Green (1964). Gor-
man's (1953a) contribution is to characterize this aggregation problem in terms of
the restrictions on consumer preferences.* Formally, the problem is posed as fol-
lows. Roy's identity, combined with (4), yields the following system of (nonlinear)
partial differential equations for consumer h.
where V h is the indirect utility function. Integration of this system yields the required
structure of preferences:?
Theorem (Gorman 19530). Assume interior solutions to the individual opti-
mization problems.$ The aggregation condition ( 3 ) holds $ and only fz consumer
preferences can be represented by expenditure functions with structure
E h ( u ,p ) = n ( p ) u h + Ah(p) Vh (8)
or, equivalently, by indirect utility functions with the structures
*Antonelli (1886) seems to have been the first to notice that homothetic and identical preferences are
necessary and sufficient for consistent aggregation if we require that the conditions hold globally. Re-
quiring that they hold only in a neighborhood, as in Gorman (1953a, l%l),yields a richer structure of
preferences, allowing preference heterogeneity.
?This integration problem can be simplified, since affinity of the ordinary (Marshallian) demand functions
in income implies affinity of the constant-utility (income compensated, or Hicksian) demand functions
in (a particular normalization of) the utility variable:
s h ( u ,p ) = G h ( p ) + B(p>uh Vh
Integration of this system of differential equations yields the consumer expenditure functions given by
(8) below.
$This restriction is inconsequential in most empirical studies. If the restriction is eschewed, the necessary
and sufficient conditions for aggregation is homotheticity of preferences (see Samuelson 1956), in which
case Ah( p ) = 0 in (8) and (9) and, as noted earlier, ah( p ) = 0 for all h in (4)or (6).
§This structure presumes a particular normalization of the utility representation; see the discussion to
come.
I86 RUSSELLET AL.
where nhand Ah are concave and homogeneous of degree 1 (hence, rhis homoge-
neous of degree 0).
Proof of sufficiency is trivial. By Roy’s identity:
so that aggregate demand is given by
The salient feature of (8)or (9)is that they completely characterize, in the dual,
the preferences that generate demands with the affine structure in the one idiosyn-
chratic variable y h . The marginal propensity to consume the ith commodity is
which is independent of h as well as y h , indicating that Engel curves, and income

consumption curves, are linear and parallel across consumers. Hence, transferring
a dollar of total expenditure from one consumer to another will leave total demand
unchanged.
The direct preferences corresponding to (8) and (9) can be further explicated.
The constant-utility (vector valued) demand function is given by applying Shephard’s
lemma to (8):
Thus, the ICCs are linear but not necessarily parallel for different prices (although,
or course, they cannot intersect). Consumption bundles on the base (zero utility)
indifference surface are given by
Note that preferences are well defined for all consumption bundles on or above the
base indifference surface, but may not be well defined below this surface. This is so
because (8)represents consumer preferences only if it is concave in prices. But when
U < 0, the first term in (8)is convex in prices and for sufficiently small values of U
the convexity of this term will dominate the concavity of the second term, violating
the fundamental regularity condition for consumer expenditure functions. Of course,
OF DEMAND
ANALYSIS AND SUPPLY I87
if the base indifference surface does not intersect the positive orthant, preferences
are well defined globally.
The structure (8) has an evocative interpretation. The “intercept” term (p) rh
can be interpreted as the “fixed cost” of obtaining the base indifference surface-or
the base utility level (normalized to be zero)-and I 7( p ) is the marginal price of the
composite commodity “utility.”
The individual structure of consumer preferences that are necessary and suffi-
cient for Gorman aggregation is often referred to as the “Gorman polar form” (GPF),
following Blackorby, Primont, and Russell(1978), but Gorman refers to it as “quasi-
homotheticity,” since it is a generalization of homotheticity. By Shephard’s (1970)
decomposition theorem, homotheticity is characterized in the dual by
E h ( u , p ) = rI(p)uh
in which case the base indifference surface degenerates to a single point-the ori-
gin-and all income consumption curves are rays. An intermediate special case is
“affine homotheticity,” generated by
in which case the base indifference surface degenerates to a point y h = ( y : , . . . , y,),

h
not necessarily the origin. A prominent example of affine homotheticity is the Stone-
Geary structure, generated by (16) and
The direct utility function dual to this structure is a Cobb-Douglas function in affine
transformations of the consumption quantities,
and the demand system is*
*The Stone-Geary structure evolved as the solution to a classic demand-system integrability problem.
Klein and Rubin ( 1 9 4 7 4 ) showed that the unconstrained linear expenditure system
p d , ( y , P) = BlY + caI,P, vi
J
E,
is integrable if and only if the following parameter restrictions hold: B, > 0 V i ; = 1; a,, = y,B1
for j # i; and all = y , ( l - B,) V i , in which case the demand system simplifies to (19). Later, Geary
(1950), using Roy’s identity, solved the partial integration problem to obtain the form of the utility func-
tion and Stone (1954) implemented the system hy estimating the parameters using British consumption
data.
I88 RUSSELL
ET AL.
The Stone-Geary specification was generalized by Brown and Heien (1972),

who replaced (17) with a CES function of prices, in which case (18)becomes a CES
function in affine translations of the consumption quantities. A further generaliza-
tion was estimated by Blackorby, Boyce, and Russell (1978), who specified flexible
functional forms (in particular, Diewert’s (1971) generalized Leontief) for Ah and
l l h . But the Stone-Geary specification was the real workhorse of empirical demand
analysis for two decades, until its replacement by flexible functional forms in the
1970s.*
The principal thrust of Gorman’s aggregation theorem is that aggregate de-
mand systems are consistent with individual utility maximization (individual “ration-
ality”) and simple aggregation if and only if individual preferences have the structure
(8) with interpersonal commonality of the function Il. but more can be said. The total
expenditure of the H optimizing consumers is
h h h
If we interpret U = E h uh as aggregate utility and set A(p) = EhA h ( p ) ,
has the structure of an expenditure function in aggregate utility (since ll and A have
the requisite homogeneity, monotonicity, and curvature properties); indeed, (21) has
the same (GPF) structure as the individual expenditure functions in (8). Correspond-
ing to (21) is the aggregate indirect utility function
which, of course, has the same structure as the individual utility functions (9). Fi-
nally, the application of Roy’s Identity to (22) yields the aggregate demand system
*The Stone-Geary system has been generalized in other directions. Howe, Pollak, and Wales (1979),with
some amendments by van Daal and Merkies (1989),constrained a system that is quadratic in income to
satisfy integrability conditions; their structure is a generalization of the Gorrnan polar form (and hence
does not satisfy his aggregation condition).
ANALBIS OF DEMAND
AND SUPPLY I89
which, with y = Eh
yh, is identical to the aggregate demand system (ll),obtained
by direct aggregation of the individual demand systems. Thus, the aggregate demand
system is rationalized by a utility function defined on aggregate commodity quanti-
ties, x = Ehd . In other words, the aggregate demand system is generated by an
optimizing “representative agent,” with utility U = Ehuh and aggregate income
y = Ehyh. Thus, the econometric estimation of any aggregate demand system with
the structure (23 ) is consistent with individual optimization and aggregation of the
individual demand systems.
The insight in Gorman’s theorem can be further elucidated by a comparison of
the representative-agent phenomenon in production theory and in consumer theory.*
Producers, labeled f = 1 , . . . , F , are assumed to maximize profit, p - zf, on a
technology set, Tf R”,defined as the set of technologically feasible net-output
vectors. The aggregate technology set is T := Ef
Tf.t The net-supply functions,
#f, f = 1 , . . . , F , are defined by
= argmax,j {p . zf I zf E 1 vf
~ . f (24)
These net-supply functions can be trivially aggregated to a macro net-supply func-

tion
But much more can be said, since
@(p) = argmax,{p. z I z E: T} (26)

Thus, the aggregate net-supply function can be generated by maximizing aggregate
profit on the aggregate technology set; i.e., the aggregate net-supply functions are
generated by the optimizing behavior of a “representative producer.” Thus, there is
no aggregation problem on the production side of the economy, so long as all inputs
are efficiently allocated.*
As indicated by Gorman’s theorem, the existence of a representative consumer
is more problematical. To illustrate the problem, recall that expenditure-constrained
utility maximization and utility-constrained expenditure minimization are solved
*Although the discussion is couched in the simplest possible terms, the producer case could be elabo-
rated upon to encompass intertemporal production technologies (see, however, Blackorby and Sehworm
1982), and the consumer case can be adapted to the study of labor supply, asset demand, and other
problems characterized by optimization subject to a preference ordering and a budget constraint (with
fixed endowments).
?This construction, of course, assumes that there are no externalities in production.
$See Section VI.
I90 RUSSELLET AL.
by the same consumption vector, so long as preferences satisfy local nonsatiation.*

Thus, the optimal consumption vectors are given by
;h = argminp { p * xh I xh E N h (ih)}
Vh
where
N h ( i h ) = {xh E R; 1 U h ( x h )>_ Zlh(ih)}
is the no-worse-than-; set of consumer h. The community level set (the Scitovsky
1942 set), the set of aggregate consumption bundles that can be distributed to the
H consumers in ways that place each consumer in the individual no-worse-than-;
set, is given byt
N ” 2 , . . . ,*xH ) = Nh(2)
h
’,
(The boundary of N s ( i . . . , ih)
is a community (Scitovsky) indifference surface.)
Note that, by the interchangeability of set summation and optimization,
= argmin,{p x I x E N ’ ( ; ’ , . . . , ; H > } (30)

giving the appearance of a representative consumer.$ Consider, however, a reallo-
cation of income, adding A to y h and subtracting A from y h . The new aggregate
demand vector, say 2, is not, in general, equal to ;, even though prices and total
income are unchanged. There is, of course, a new Scitovsky set, N S ( 2 ’ , . . . , @)
and
2 = argmin,{p - x I x E N’(?’, . . . , a H > ) (31)
but, in general, the two Scitovsky indifference surfaces intersect. Thus, there does
not exist a preference ordering defined on aggregate consumption vectors that ration-
alizes aggregate demands. If, on the other hand, the income consumption curves of h
and 6were both linear, with identical slopes, in the neighborhoods of and i’,the ih
aggregate demand would be unchanged by this redistribution of income; moreover, if
this condition held globally (or at least in a neighborhood) the Scitovsky sets would
also be unchanged.
As pointed out in Section I, if incomes are optimally allocated by the maxi-
mization of a Bergon-Samuelson (utilitarian) social welfare function, the aggregate
data are rationalized by an aggregate utility function. In this case the social planner’s
optimization problem is
*See, e.g., Debreu (1959) or Varian (1992).

?This construction assumes there are no consumption externalities.
$In fact, this is an example of a “representative agent at a point,” a concept commonly employed in finance
and macro.
ANALYSIS
OF DEMAND
AND SUPPLY 19 I
or, equivalently,
Thus, @ induces an ordering on the aggregate consumption space that rationalizes

the aggregate choices. This construct, however, is not useful for econometric analysis
of demand in a market economy.
To summarize, the problem of aggregation consistency is fundamental to the
theory and econometric estimation of aggregate consumer demand functions, even
under the assumption of efficient allocation of consumption goods, but the problem
does not exist for producers operating in an economy with an efficient allocation of
inputs and outputs.
B. Muellbauer Aggregation
The necessary conditions for Gorman aggregation are quite stringent. This fact in-
spired Muellbauer's (1975,1976)formulation of a weaker form of aggregation-and,
concomitantly, a weaker representative-agent concept. The two fundamental differ-
ences between Gorman aggregation and Muellbauer aggregation are that (i) Muell-
bauer requires consistent aggregation of expenditure shares rather than commodity
demands and (ii) the aggregate income of Muellbauer's representative agent is a pos-
sibly nonlinear function of individual incomes, rather than a simple sum, and can
depend on prices as well as the distribution of income.
Individual commodity shares are defined by
Aggregate commodity shares are then given by
Thus, aggregate shares, like the aggregate demands in (2), depend in general on
the distribution of income and prices. Muellbauer aggregation holds if there exists a
function Y such that
This condition is nontrivial because the function Y is identical for all n aggregate
demand shares. Thus, if Muellbauer aggregation holds, aggregate commodity shares
I92 RUSELLET AL.
'
depend on a common scalar Y (y , . . . , y H ,p ) , interpreted as "aggregate income,"
and the commodity prices.
It is easy to see that Muellbauer aggregation subsumes Gorman aggregation as
a special case. The (GPF) demand system (10)yields the following system of share
equations:
Aggregation over consumers yields
which is the special case of Muellbauer aggregation where
Theorem (Muellbauer I 975, I 976). Assume interior solutions to the individual

optimization problems. Muellbauer aggregation holds $ the individual expenditure
functions have the structure
Eh(& p ) = q h ( u h ,I Y p ) ) n ( p ) Vh
in which case the indirect utilityfunctions a re
Vh
where each qhand each Oh is increasing in itsjrst argument. *
*This structure is necessary as well as sufficient if we add to (40) a function of prices, A h ( p ) ,that van-
ishes when we sum over households: E h A h ( p ) = 0. While this possibility is formally required for the
structure to be necessary as well as sufficient, it is uninteresting-a nuisance term. It does not make a
lot of sense for preferences to satisfy a condition like this. For one thing, it means that individual pref-
erences would have to depend on the number of households H ; otherwise, the sum of these nuisance
terms would not vanish when we changed H . For these reasons, we carry out the analysis of Muellbauer
aggregation ignoring this term (as has been the case in subsequent studies building on Muellbauer's
ideas). (See, however, Blackorby, Davidson, and Schworm 1993 for a different approach.) Note also that
the functions in these representations must satisfy certain homogeneity conditions, but because of the
ways in which they enter the expenditure and indirect utility functions, there are degrees of freedom in
choosing the degrees of homogeneity.
OF DEMAND
ANALYSIS AND SUPPLY 193
To see that this structure is sufficient, apply Roy’s identity to obtain

Sh(P, yh>= v h ( y h ,P I A i tp) + Bi(P> (42)
where
The subscripts on 0 indicate differentiation with respect to the first or second term,
7 and A indicate differentiation with respect to the indicated
and the subscripts on l
price variable. Summation yields
H
= W(yl, * * * 7 y 7 p),p )
where Y ( p ,y l , . . . , y H )is implicitly defined by
(assuming a solution exists*).

Note that setting
q h ( u h ,U p ) ) = $h(uh> +U p ) (48)
in (40) yields the Gorman polar form. Because Muellbauer’s structure is a straight-
forward generalization of Gorman’s “linear” (actually affine) structure, he refers to
(40)or (41) as “generalized linearity.’’
The system of aggregate share equations, along with the implicit definition of
aggregate income in (47)’ is rationalized by an aggregate utility function with the
structure
Application of Roy’s identity yields the si in (46). Inversion in Y (y1 , . . . , y H ,p )

yields
m u , p ) = q t w r(p)>n(p> (50)
*Muellbauer shows that it will exist under certain conditions.

194 RUSSELL
ET AL.
where U is the “utility” of the representative consumer. Thus, as in the case of

Gorman aggregation, the Muellbauer representative consumer has the same pref-
erence structure as that of the individual consumers. Similarly, the preference struc-
ture of Gorman’s representative agent in (8) is obtained by setting *(U, r(p))=
*(U) + r(p).Moreover, just as Gorman’s preference structure is equivalent to the
differential condition (12)’ Muellbauer’s generalized linearity structure is equivalent
to the following differential conditions:*
That is, ratios of marginal effects of income changes on shares are independent of
income levels for both individuals and the representative consumer.
Muellbauer’s generalized linearity is further explicated by rewriting the share
equations in matrix notation:
The n x 2 matrix on the RHS has rank 2 and u h ( y h ,p ) is independent of i. (A

similar identity holds for the representative consumer.) Hence, any share equation
can be obtained as a linear combination of any other, and for given prices, the share
equations lie in a two-dimensional linear subspace. This “rank-2 condition,” which
has been explored in some depth in recent years, is examined further in Sections I11
and IV.
Another interpretation of Muellbauer’s representative consumer in commodity
shares is attributable to Gorman (1976). Note that the existence of a Muellbauer
representative agent can be characterized, using Shephard’s lemma, as the existence
of an aggregate expenditure function, E , such that
and a function U such that
U = U ( u ’ ,. . . , U h , p ) (54)
*See Muellbauer (1975, 1976) for proof of sufficiency of this condition.

OF DEMAND
Gorman first noted that Muellbauer aggregation is trivially satisfied if there are only
two commodities, since determining one aggregate share determines the other and
U therefore is implicitly defined by, e.g.,
This fact is not surprising, since in the two-commodity case there is no integrability
problem. More importantly, it suggests that the solution to Muellbauer’s aggregation
problem might be characterized by the existence of two aggregate pseudocommodi-
ties, and in fact this is an evocative interpretation of the solution given by (40), in
which r(p) and n(p) are interpreted as the “prices,” or “unit costs,” of two “in-
termediate commodities” in the production of utility. A complete explication of this
interpretation requires quite a bit of duality theory, so we refer the reader to Gorman’s
paper for further study. (Compare this interpretation, however, to the interpretation
of the Gorman polar form expenditure function (8).)
Gorman also pointed out a problem with the Muellbauer representative agent:
there is no requirement that Y in (36) or (47) be increasing (or even nondecreasing)
in individual incomes-equivalently, that U defined by (53) be increasing in indi-
vidual utility levels. Thus, the income or utility of the representative agent could be
increasing when the incomes or utilities of the consumers it represents are declining.
This seems to be a consequence of requiring only that the representative consumer
replicate the aggregate shares, but not aggregate demands.
Muellbauer-as well as most follow-up studies-focused on the special case
where Y ( y ’ , . . . , y H ,p ) is independent of p . In this case, u h ( y h ,p ) in (47) must be
independent of p for all h (for arbitrary distributions of incomes, y’, . . . , y h ) :
v h ( y h ,PI = Ch(rh)‘dh (56)

Muellbauer shows that this, in turn, implies that*
c h ( y h )= rhp or plnyh Vh (57)

Thus,
$(p, rh)= Ai(p)rhp + B‘(p) or A i ( p ) lnyh + Bi(p)

(where p has been incorporated into A ’ ( p ) ) and
*To avoid confusion, where the income variable is raised to a power, we move the household indicator
index h to the subscript position.
I96 RUSSELLET AL.
or
The first of these commodity-share equations is called price-independent general-

ized linearity (PIGL), and the second is price-independent generalized logarithmic
(PIGLOG).
Empirical application requires that specific forms of the functions r and l
7 be
specified. Deaton and Muellbauer (1980)adopted a translog specification for and r
the sum of a translog and a Cobb-Douglas for n. They called this specification the
almost ideal demand system, with the (unfortunate) acronym AIDS. This specifica-
tion has been widely applied in empirical demand analysis, and Lewbel(1989a) has
estimated a wider class of functions that nests both the translog and the AIDS spec-
ification. Another generalization of the AIDS specification by Banks, Blundell, and
Lewbel (1993), the quadratic AIDS (QUAIDS), has rank-3 Engel curves and hence
does not satisfy aggregation conditions.
C. Testing the Representative-Consumer Hypotheses

Most tests of the representative-consumer hypotheses have not been designated as
such. Rather they have purported simply to test consumer “rationality”; but since
they have used aggregated data (typically the only data for which consumption-good
prices are available), they have in fact been testing for the existence of a represen-
tative consumer that rationalizes the aggregate data.* As such, these tests have an-
ticipated the widespread use of representative consumers in recent macroeconomic
research.
Two types of tests have been employed, and there is a sharp difference in
the results using these two approaches. Most commonly, researchers have specified
a functional form representing preferences and then tested for parametric restric-
tions consistent with the symmetry of the estimated Slutsky matrix of substitution ef-
fects of price changes. Other tests have eschewed parametric forms, using revealed
preference methods to test for rationalization of the aggregate data by expenditure-
constrained utility maximization.
The first approach began with the Christensen-Jorgenson-Lau (1975) tests for
symmetry of the Slutsky matrix generated by the demand functions derived from
a translog indirect utility function. These tests turn out to be identical to testing
for symmetry of the second-order coefficients in the translog specification. That is,
*See, however, Pollak and Wales (1979) and Shapiro and Braithwait (1979) for direct tests of the GPF
structure. which doesn’t fare at all well.
OF DEMAND
the approach is to estimate the derived demand system with and without symme-
try of these coefficients and use a maximum likelihood ratio or Wald test to test for
symmetry.* There is by now a large literature on this approach, inspired in part by
the competition between proponents of alternative flexible-form specifications, and
we make no attempt to survey it. Suffice it to say that the symmetry conditions are
commonly rejected.
Of course, the weakness of these parametric tests is that rejection can be at-
tributable not to the failure of consistent aggregation to hold, but rather because
the preferences of the representative agent are misspecified. Although, by definition
and design, flexible functional forms can provide a second-order approximation to
an arbitrary “true” specification at a point, they are in fact employed as global spec-
ifications in tests of symmetry (and other properties). The second-order parameters
are estimated from data over the entire sample space.?
The alternative approach, which eschews specification of functional form, is
based on the revealed preference approach to testing the consumer-optimization hy-
pothesis, formulated by Samuelson (1938, 1946-47) and Houthakker (1950) and
first implemented empirically by Houthakker (1963)and Koo (1963). But the mod-
ern formulation has its roots in a remarkable theorem of Afriat (1967, 1972) (which
was nicely elucidated by Diewert 1973). Varian (1982) applied Afriat’s method to
annual U.S. data on nine consumption categories from 1947 to 1978. Remarkably,
he found no violations of the revealed preference axioms. As the data are aggre-
gate U.S. consumption quantities and prices, this result apparently provides strong
support for the existence of a Gorman representative consumer.
Varian raised the possibility, however, that these tests may be plagued by low
power, given the data with which economists have to work. In particular, they will
have low power if the variation in total expenditure over time is large relative to the
variation in relative prices. (If, for example, the budget hyperplanes do not intersect,
the tests have zero power.)
Varian’s concerns were apparently allayed by Bronars (1987): in a Monte Carlo
assessment of the power of Varian’s test against the alternative hypothesis of Becker-
type irrational behavior, he found that these tests have a considerable amount of
power, especially when per capita data are used. Roughly speaking, Becker’s no-
tion of irrational behavior entails a (uniformly) random distribution of consumption
quantities across the budget hyperplane. To implement this notion as an alternative
hypothesis, Bronars took Varian’s price and total expenditure data as given, replaced
the consumption quantities with randomly constructed values (using three different
*Asymmetry of these second-order coefficients is equivalent to violation of Young’s Theorem on the sym-
metry of second-order cross derivatives.
?See White (1980)for a penetrating analysis of the problems with interpreting tests using flexible forms
as tests at “the” point of approximation.
I98 RUSSELLET AL.
algorithms), and applied Varian’s GARP test, alternatively using per capita and ag-
gregate data. He repeated this exercise a large number of times to obtain a measure
of the power of Varian’s test against the Becker alternative hypothesis. The upshot of
his results is that Varian’s test has a substantial amount of power against the Becker
irrationality hypothesis.
Bronars noted that his alternative hypothesis is “rather naive” (page 697). It
is not clear what type of individual behavior, if any, together with aggregation across
consumers, would yield a (uniformly) random allocation of aggregate consumption
quantities across the budget plane. Recently, Russell and Tengesdal (1996) raise
the question of whether Varian’s tests have substantial power against a less naive
alternative hypothesis, predicated on the fact that the null hypothesis of Varian’s
test, given the aggregate data employed, is a compound one, entailing both individual
rationality and aggregation consistency.
Using Monte Carlo methods analogous to Bronars’, Russell and Tengesdal
use actual price and aggregate total expenditure data, but generate aggregate con-
sumption quantity data by (1) specifying heterogeneous individual utility functions
(that do not satisfy Gorman aggregation)* and a distribution of total expenditures
among individuals, (2) generating optimal individual demands, given the prices and
the preferences and total expenditures of individuals, and (3) aggregating demands
across individuals. Using Varian’s algorithm to calculate power indices in terms of
the percentage of the simulations with GARP violations, they find few violations,
despite a large number of sensitivity tests. These results suggest that Varian’s tests
using aggregate U S . consumption data lack power and hence provide little support
for the existence of a Gorman representative consumer.
A recent paper by Fleissig, Hall, and Seater (1994)reinforces this point. They
find that Varian’s GARP tests are violated using both monthly and quarterly US.
consumption data. Another suggestion that these nonparametric (nonstochastic) tests
lack power is the surprising result of Manser and McDonald (1988)’in which it is
shown that the Afriat-Varian tests fail to reject the hypothesis of a homothetic utility
function rationalizing aggregate US. consumption data. But the concomitant impli-
cation of constant budget shares-assuming strictly convex indifference surfaces-
is easily rejected statistically. As Manser and McDonald point out, these two out-
comes are explained by the fact that the rationalizing utility function in Afriat’s the-
orem is piecewise linear, as are the indifference surfaces, so that substantial changes
in budget shares can be consistent with homotheticity.
*In particular, they employ the Stone-Geary specification described in Section II.A, but with heteroge-
neous pi parameters. Gorman aggregation holds for this specification if and only if these parameters are
identical for all consumers. This specification allows a simple parameterization of consumer heterogene-
ity and hence the “degree” of aggregation inconsistency, ranging all the way to “maximal” heterogeneity
(a uniform distribution over the [0,11interval or, alternatively, preferences in which each consumer has
just one positive pi-equal of course to 1).
ANALYSIS OF DEMAND
AND SUPPLY 199
The above studies pertain to the existence of a Gorman representative con-

sumer, defined in terms of the simple sums (or means) or expenditures. Of course,
confirmation of the existence of a Gorman representative consumer implies confir-
mation of the existence of the weaker representative-consumer notion of Muellbauer.
A (statistical) nonparametric study of aggregate consumption by Lewbel(l991) sug-
gests that data in the middle 90% or so of the income distribution are consistent
with Muellbauer’s PIGLOG specification and concomitant representative consumer
in budget shares. The data at both the low and the high end of the income distribution
are inconsistent with Muellbauer’s rank-2 condition. As will be shown in the next
section, this implies that the Engel curves of optimizing consumers have rank 3 (and
hence are incompatible with both forms of representative consumer).
111. “EXACT” AGGREGATION AND HOUSEHOLD

ATTRIBUTES
A. The Exact Aggregation Theorem
As noted at the outset of the previous section, aggregate demand generally depends
on the entire distribution of income. When the necessary and sufficient conditions
for a Gorman representative consumer are satisfied, however, a single statistic-
namely the first moment of the income distribution-suffices to determine aggregate
demand (see (11)). The existence of a Muellbauer representative agent (in budget
shares) is weaker, and the aggregate demand functions accordingly depend on two
income distribution statistics. To see this, note that the aggregate demand functions
implied by (36) are
where t91(y1, . . . , y H ) = Eh y h and 82(y1, . . . , y H ) = Y(yl, . . . , y H ) . (Note that

the second of the identities in (61), being more general than the first, imposes weaker
restrictions on preferences than those needed for a Muellbauer representative agent.)
More specifically, with price-independent aggregate income (58),Muellbauer aggre-
gation requires that the individual demand functions take the following form:
d“p, rh)= Pi1 (P)SI (rh)+ Bi2(p)tz(rh) (62)
= B ; ( p ) / p ; ,for all i,
200 RUSSELLET AL.
for all h. The two cases in (64) generate the PIGL and PIGLOG systems, in which
case individual demands as well as aggregate demands have rank 2 and aggregate
demands are given by
or
A specific case of (65),with p = 1 is that where individual demand functions

are quadratic in income:
in which case aggregate demands are
Thus, in this special case of Muellbauer aggregation, we need information about the
second as well as the first moment of the distribution to determine aggregate demand.
But knowledge of these two statistics allows us to determine aggregate demand ex-
actly, no matter how many individuals are in the economy.
This characterization of Muellbauer's results raises the question of whether
it generalizes to cases where there are more than two aggregator (0,) functions and
where the forms of these functions are arbitrary. rl iis is the motivation for the notion
of exact aggregation, defined by the existence of I' symmetric* functions of individ-
ual incomes, say O,(yl, . . . , y H ), t = 1, . . . , T , where T -= H , such that?
bi(p,y l , . . . , rh)= Q ( P , elc$, . . . , Y h 1, . . . , w y l , . . . , r h ) ) vi (69)
*By symmetry, we mean anonymity-that is, that the values of the aggregator functions are unaffected by
an arbitrary permutation of the y h variables. Without this condition, calculations would be intractible
for large H .
?Obviously, if T 2 H , there is no economy of information and this structure imposes no restrictions on
the individual demand functions.
ANALYSIS
OF DEMAND
AND SUPPLY 20 I
This aggregation notion appears to have been formulated independently by Gorman

(1981) and Lau (1977a, 1977b, 1982);* the necessary and sufficient conditions for
exact aggregation are as follows.
Theorem (Lau I 982, Gorman I 98 I). Given certain regularity cbzditions,? the
aggregate demand functions satisfy (69) if and only if the individual demand func-
tions have thefollowing form:
t= 1
Accordingly, aggregate demands are given by
1=1
Thus, for exact aggregation to hold, individual demand functions must be affine in
functions of income and identical across consumers up to the addition of a function
of prices that is independent of income. Moreover, the identical part of the function
must be multiplicatively separable into a function of prices and a function of income.
Finally, the symmetric functions, O , , t = 1, . . . , T , in the aggregate demand system
must be additive functions of (identical) transformations, ct,
of individual incomes.
Note t at exact aggregation, unlike Muellbauer (hence Gorman) aggregation,
does not irn d y the existence of a representative consumer; that is, without additional
restriction.,. the aggregate demand functions (71) cannot be generated by the max-
imizatior. 2f an aggregate utility function subject to an aggregate budget constraint.
Thus, exact aggregation is weaker than the existence of a representative consumer.
B. Gorman's Theorem on the Maximal Rank of Demand Systems

In matrix notation, the necessary and sufficient condition for exact aggregation is
d h ( p , y h >= ah(p> + B(p)6(rh) =: B h ( p > H y h ) V h (72)
*Gorman uses this aggregation condition as motivation for his theorem on the maximal rank of demand
systems, discussed below. He first presented the paper at a London School of Economics workshop in
January 1977.
?See Lau (1982) and Heineke and Shefrin (1988) for the particulars.
202 RUSSELLET AL.
where B h ( p ) = [ a h ( p ) B ( p ) ] is an n x R matrix, with R = T +

1, and e(yh) is
an R-dimensional vector. This structure, however, is not necessarily consistent with
individual optimization without further restrictions. In a remarkable paper, Gorman
(1981)examined the restrictions on demand functions of type (72) implied by indi-
vidual optimization. Gorman’s theorem is as follows.
Theorem (Gorman I 98 I). Ifthe complete system of demand equations (72) re-
flects well-behaved preferences, then the rank of its coeficient matrix B ( p ) is at most 3.
Moreover, one of thefollowing must hold (where I+ is the set of nonnegative integers):*
Note that the Muellbauer’s PIGLOG and PIGL systems are generated as spe-
cial cases of (i) or (ii), with R = 2 (hence rank 2). PIGLOG is a special case of (i)
with ~1 = 1 and ~2 = 0. PIGL is a special case of (ii) with p2 = 0. Similarly, the
Gorman polar form (quasi-homotheticity) is a special case of (ii) with R = 2 (rank
2), p1 = -1, and p2 = 0. The rank-3 quadratic expenditure system is obtained by
setting p1 = -1, p2 = 0, p3 = 1 in (ii). Homotheticity is rank 1 with p1 = -1
in (ii). In fact, virtually all consumer demand systems that have been estimated be-
long to the class of rank-2 demand systems satisfying the conditions of Gorman’s
theorem. Section IV contains more on rank-2 (and rank-3) demand systems. The key
point here, however, is that exact aggregation is considerably more general than the
existence of a representative agent; the latter is but one way of specifying demand
systems that can be consistently aggregated.
C. Household Attributes
In the Lau-Gorman exact aggregation theorem, heterogeneity of preferences enters
only through the term a h ( p ) in (70), which is independent of income. Additional
heterogeneity can be introduced into the exact aggregation framework by incorpo-
rating an attribute vector ah into B;,(p) and c r ( y h ) ,for each r. (Commonly used
attributes include household size, age of household head, race, and geographical
*As noted by Gorman, the consumption space can divide into subsets with different forms among (i)-(iii)
holding over different regions. Heineke and Shefrin (1987) have shown that, if we require the following
specifications to hold globally, bountledness of budget shares implies that the only integrable system is
the homothetic (rank 1)specification. See Heineke (1979, 1993)and Heineke and Shefrin (1982, 1986,
1987, 1988, 1990) for extensions and clarifications of Gorman’s theorem.
ANALYSIS
OF DEMAND
AND SUPPLY 203
region.) For the sake of symmetry, assume that the heterogeneity of ah is also cap-
tured by these attributes, so that*
where ( p , ah) = a!(p).

Integrability implies that the pr parameters in (73) depend on the attribute
vector uh and exact aggregation implies that the functions separate multiplica-
tively into a function of p and a function of u h , so that?
r= 1
where one of the following holds:
The aggregate demand system then becomes
h r= 1 h
*As heterogeneity of preferences is incorporated entirely through the heterogenous attribute vectors, the
demand functions do not require an h superscript.
?More precisely, these functions separate into a finite sum of multiplicatively separable terms, so that the
number of terms in the following structure may he larger than R; see Heineke and Shefrin (1988),who
refer to this structure as the "finite basis property," for details.
204 RUSSELL
ET AL.
Thus, specifying functional forms for y i r ( p ) and gir(yh, a h ) in (77) and then sum-
ming across consumers of different types, defined by uh, results in an aggregate de-
mand system that is consistent with individual rationality and exact aggregation.
The question that immediately follows is: under what conditions do the pa-
rameters estimated from (77) allow us to recover (uniquely) the parameters of the
microfunctions in (75)? Given a joint distribution of income and attributes, expected
per capita demand is given by
Since the y i r ( p ) terms enter individual and aggregate demand functions in the same
way, the parameters in Yir(P) estimated from (77) will be the parameters in y i r ( P )
of (75). On the other hand, the term E(<ir(uh)gir(yh, a h ) ) in the aggregate demand
function may not include the parameters in the microeconomic demand function.
Heineke and Shefrin (1990) show that (75) is recoverable from the estimation of
(77) if
(i) The p r ( a h )functions in (76) do not depend on any unknown parameters.

(ii) The functions < i r ( a h ) in (76) are linear in the parameters.
Stoker (1984)has shown, for the general problem of microequation recovery
from macroequation estimation, that linearity of the microfunctions in parameters is
a sufficient condition for identification when there are no distributional restrictions
(in this case, on the distribution of incomes and attributes).
As an example, consider Jorgenson, Lau, and Stoker’s (1982) paper on esti-
mating demand systems derived from a translog indirect utility function. The vector
of expenditure shares for household h take the form*
(5,.“) = Q ( p / y h ,a h )
where a! is an n-vector of parameters, Bppand Bba are (appropriately dimensioned)
(79)
matrices of parameters, and
= i’a! + ifBppIn P + i’Bpnah

-
Yh
Note that the parameters are the same for all households, so that preference hetero-
geneity is entirely captured by the different attributes, u h ,h = 1, . . . , H . It follows
from the conditions for exact aggregation that the individual demand functions are
linear in functions of y h and ah . These conditions will be satisfied if Q ( p / y h , a h ) ,
the denominator of (79), does not contain y h and uh. Therefore, the following two
restrictions must be imposed:
*ln(p/yh) is the n-vector of logs of normalized prices and i is an n-vector of ones.

OF DEMAND
i'Bppi =0 and i f B p a= 0 (81)

Thus, Q ( p / y h ,a h ) is reduced to Q ( p ) = i'a L + ~Inp,
B which
~ ~is independent of
h, and the household shares can be rewritten as
Accordingly, the aggregate expenditure shares are
- 1 y h In Yh
-
Q(P)
( a + Bpplnp - Bppe
E h
Eh Yh
+ Bpu-
Ch Yh
Note that (82) satisfies the conditions of exact aggregation; specifically, the
demand equations (given here in expenditure share form) are linear in the parameters
and the microequations (82) contain only parameters that appear in the aggregate
equation (83). The parameters in (83)can be estimated by using both cross-sectional
and time-series data. Thus, the individual expenditure shares can also be obtained
by replacing the unknown parameters in (82) with those estimated from (83).
As shown by Heineke and Shefrin (1990),imposing (81) is equivalent to im-
posing identification conditions; that is, the translog expenditure share system au-
tomatically possesses their restricted finite basis property. The JLS translog system
is in fact a restricted version of (76)(i). As Heineke and Shefrin (1990)have pointed
out, it restricts (76)(i) to being rank 2, when in fact the Gorman theorem allows the
system to be up to rank 3. Also, JLS conflate the rank-:! specification and the re-
quirement that there be only two functions of income and attributes. The restrictions
on rank do not imply that the number of income/attribute functions must be only as
great as the rank of the coefficient matrix Bh(p) in (72).
The JLS translog specification can also be shown to incorporate the following
restrictions on (76)(i):
(i) {il ( a h )must

be linear in the attributes and must have an intercept term
that is constant for all h.*
(ii) c;2(nh)must be constant for all h. There are only two { i r functions, since
the rank of the system is 2 and only two income/attribute functions are
posited.
*In estimation, the attributes considered by JLS can Le divided into discrete categories, allowing for the
use of dummy variables. This implies further restriction on the til functions.
206 RUSSELLET AL.
The translog system is thus a restricted version of (76)(i) and a much wider
class of similar demand functions is in fact compatible with exact aggregation.
D. Empirical Applications
Empirical applications of the theory of exact aggregation are based primarily on
the work of Jorgenson, Lau, and Stoker (1980, 1982).* Their econometric model
incorporates household attributes into the translog indirect utility function with a
set of 18 dummy variables for attributes like family size, age of head, region of res-
idence, race, and type of residence. They impose restrictions on the parameters of
the share equations such that conditions for exact aggregation hold and aggregate
across individuals to obtain the aggregate share demand functions. They estimate
this aggregate demand system by pooling cross-sectional data on expenditures by
individual households and aggregate annual time-series data on expenditure, price
levels, and statistics of the distribution of family income and demographic vari-
ables. In addition to estimating the model, they analyze the welfare changes from oil
price shocks.
Jorgenson, Lau, and Stoker use dummy variables and simple linear equa-
tions for the incorporation of demographic attributes. This method remains the most
commonly used one in applications. On the other hand, Pollak and Wales (1978,
1980) propose two alternative methods for inclusion of demographic variables: de-
mographic translating and demographic scaling. In both demographic translating
and demographic scaling, new parameters are introduced to capture the effects of
demographic heterogeneity. For the demographic translating procedure, n “translat-
ing parameters,” tl,T Z ,. . . , T,,,are added to the model and the original demand
function d? is modified as follows:?
where
*See also Jorgenson, Lau, and Stoker (1981), Stoker (1986a), Jorgenson and Slesnick (1984), and Nicol
(1994).
?This demand is assumed to be derived 1)y utility maximization. The authors investigated the LES, QES,
basic translog (i.e., the translog of Christensen, Jorgenson, and Lau 1975), and generalized translog
demand systems. The primary empirical rrsult of their work is that both the number and age of children
in a household have significant effects on consumption patterns.
OF DEMAND
Thus, we can write the demand as
Similarly, demographic scaling introduces n “scaling parameters,” a 1 , a ~. ,. . ,

on,into the original demand system and the new demand system is
& ( p , y h , a h ) = ai d h; ( p l a i , p n a z , . . . , p n a n ) Vi Vh
* (87)
where
i j
Note that, in (84) and (87), t i and oi are the only parameters in the demand system
that depend on demographic variables. Unfortunately, as Pollak and Wales point out,
their aggregation procedure, unlike that of Jorgenson, Lau, and Stoker, is inappropri-
ate. Moreover, there do not appear to be any applications of demographic translating
and demographic scaling to appropriately aggregable demand systems.
In a related development that accounts for serial correlation, Blundell, Pa-
shardes, and Weber (1993) incorporate household demographic attributes and time-
dependent variables (time-trend and seasonal dummy components) into demand sys-
tems. The aggregate models they derive from this specification seem to perform quite
well compared to models using microlevel data in both forecasting efficacy and eval-
uation of aggregate consequences of public policy changes.
In demand system estimation, most authors include some type of dynamic
specification (commonly an AR(1) process for the error terms) to account for habit
formation. Few, however, include forms that adjust for the distribution of incomes.
In an interesting paper, Stoker (1986b) compares an LES model with the habit for-
mation structure with one that does not have any dynamics, but instead allows for
distributional effects on demand. He shows that the distributional effect is statisti-
cally significant and that it can displace AR(1) dynamics in the widely used models.
His results were confirmed and extended by Buse (1992),who used Canadian rather
than U.S. data and estimated a quadratic expenditure systems (QES) model as well
as an LES model. He concluded that Stoker’s results are widely applicable. Stoker
(1993)provides further extension of this work.
To summarize, “exact” aggregation allows summary statistics of income and
attributes to be used to create aggregate functions that are consistent with aggre-
gation over individual demand functions generated by income-constrained utility
maximization. The conditions for existence of such functions and for their identifi-
ability have been briefly discussed here. Even if the conditions for the existence of
a Gorman representative agent do not hold, “exact” aggregation demonstrates the
conditions under which aggregate data may still be used for econometric estimation
when appropriate distributional information is available.
208 RUSSELLET AL.
IV. ADDITIONAL RESULTS O N AGGREGATION AND

THE RANK OF DEMAND SYSTEMS
A. Tests and Extensions of Gorman Rank Conditions
Several papers have extended Gorman’s (1981) results on the rank of Engel curves
and some testing of the rank conditions has begun. Lewbel (1987) characterized
demand systems that are affine in income and an arbitrary function of income:
4 ( p , y) = a;(PI + bi ( p ) y + ci ( p )lb(y) vi (89)

(For notational convenience, we have temporarily suspended use of the household
index.) This subclass of demand systems encompasses most of those that have been
used to estimate demand taking aggregation into explicit account. Here, demands
are linear in functions of income, but unlike Gorman’s formulation, only one of those
functions is of unspecified form. Lewbel shows that eight different possibilities fully
characterize the demand systems of the form (89): homothetic, quasi-homothetic
(GPF), quadratic (QES), PIGL, PIGLOG, extended PIGL, extended PIGLOG, and
LINLOG.*
Only the QES among all these systems has rank 3. Indeed, it can be seen that
the QES is case (ii) in (73) characterized by Gorman (1981). The rest have rank 2,
except for homothetic demands, which have rank 1. All of the rank-2 systems have
indirect utility functions (or cost functions) that can be written as the generalized
Gorman polar form necessary for two-stage budgeting; see Gorman (1959), Black-
orby, Primont and Russell(1978), and Blackorby and Russell(l997). Lewbel(l990)
characterized full-rank demand systems-those in which the n x R matrix, B h ( p ) ,
of (72) has rank R . It turns out that PIGL and PIGLOG are the only rank-2 demand
systems with full rank. The rank-3 systems had previously been characterized by
Lorman.
Lewbel (1989b) extends the rank result to functions of real income instead of
nominal income. As Gorman himself pointed out, the result on the rank of Engel
curves is rendered less interesting by the fact that the Engel curves are functions of
nominal (money) income and not real income. If we write the income shares for each
consumer as
R
r= 1
*The GPF is described in (lO), the PIGL and PIGLOG were characterized by Muellbauer (58),and the
QES was characterized by Pollak and Wales (64)with the intercept term c r i ( p ) appended). Lewbel in-
troduces three new demand systems: extended PIGL, extended PIGLOG, and LINLOG. The extensions
to PIGL and PIGLOG demands involve the addition of an additive constant term and, in the extended
PIGL case, the A i ( p ) term is multiplied by a differentiaMe function of prices. None of‘ these has been
used in empirical application.
OF DEMAND
where q is the vector of log prices and z is log income, we know that the rank of the
matrix of coefficients [ b r i ] is at most 3. Lewbel shows that if we replace @ r ( z )with
q r ( Z ) , where z is the log of dejlated income, then the system
has at most rank 4. In general, all the other properties that hold for Gorman demand
systems of the form (72) hold for the deflated demand systems of (91).If V ( q , z ) is the
indirect utility function corresponding to a Gorman system, then V ( q , @ ( Z ) ) , where
@ (2) is a nonzero, bounded, continuous differentiable function, is the indirect utility
function for the deflated demand system.
Lewbel(l991) constructed nonparametric tests for the rank of a demand sys-
tem and applied these tests to family expenditure survey data from the U.K. and con-
sumer expenditure surveys from the United States. To minimize income-correlated
demographic variation, Lewbel selected only a subset of households that were fairly
homogeneous in terms of household attributes. (Specifically, only married couples
with two children, where the head of household was employed full-time, were cho-
sen. Note that another approach would be to define rank as the space spanned by
function of “attributes” as well as income. This is the approach explored in Section
1II.C. Here, Lewbel only considered households in one cross section of attributes.)
Using this reduced data set, he found that, in the middle part of the distribution
the data are consistent with rank 2 and that the PIGLOG system gives a fairly
good fit. In the tails of the distribution (the lower and upper S%), however, he finds
that the data appear to be of higher rank and that rank-3 systems would be a
better fit.
Hausman, Newey, and Powell(1995) (HNP) also find that rank-3 demand sys-
tems fit the data better than rank-2 systems. They are primarily concerned with the
problem of estimating systems of demand when there are errors in variables. This
is a common problem with survey data dealing with income and expenditure. Using
the specification for shares,
s;(y) = pi0 + pi1 In y + Bi2(ln y>2 vi (92)
where the Pi’s are constants since the estimation is done at only one price situation,
HNP estimate demands for five categories of goods. They compare the results from
OLS and instrumental variable techniques using a repeated observation as an in-
strument. They find that the inclusion of the quadratic term gives better estimates of
the demand system than a rank-2 specification. To test whether or not a rank-4 spec-
ification gives any additional information, HNP add another term and reestimate the
system. Thus, the specification becomes
si(y) = Boi + Bli 1ny + pi2(Iny)2 + Bi3(lnY)3 vi (93)

2 10 RUSSELLET AL.
If the demand system is rank 3, the ratio Bn//?2 should be constant. (In other words,
the addition of the last term provides no new information.) HNP find that, in fact,
this ratio is almost perfectly constant. Thus, Gorman’s rank condition seems to be
confirmed in the data for this specification.
B. Other Rank-Two Results

In Sections I1 and 111, the question of aggregation was considered without placing
any restrictions on the economy except the usual regularity conditions on individ-
ual consumer demand functions. Several authors have considered weaker forms of
aggregation. In particular, they have asked whether the class of individual demand
functions consistent with representative-agent-type aggregation might be broadened
by placing other restrictions on the economy.
One approach is to restrict the distribution of income. Jerison (1984a) showed
that, when the income distribution is restricted to a set of allowable distributions
(distributions that are relatively open on the unit simplex), a wider class of consumer
demand functions aggregates. Jerison was concerned with a representative consumer
of the Gorman type-where the demand of a group of consumers can be derived from
utility maximization based upon the mean income of the group. Jerison introduced
the concept of pairwise aggregation-a representative consumer existing for each
pair of demand functions. In other words, pairwise symmetry is simply the condition
that mean demand functions for each pair have a symmetric Slutsky matrix. Define
the vector of share differences as
h h’
Ghh’(p,y , y ) = sh(p, y h ) - s q p , yh’) (94)
Jerison defined the demand system as exhibiting “nonconvergence” if y JGhh’(p,y h ,
yh’)l is nondecreasing as y changes. If G h h ’ ( py, h , yh’) does not change direction
(i.e., does not change from positive to negative, or vice versa) for small changes
in mean income, the demand pair is considered to exhibit “local invariance.” If
Ghh’(p,y h , yh’) does not change direction for any changes in mean income, the de-
mand pair exhibits “invariance.”
Jerison derived the relationship between aggregation, pairwise aggregation,
local invariance, nonconvergence, and pairwise symmetry. When the set of allow-
able fixed income distributions is limited to a singleton (where all consumers have a
strictly positive share of total income), pairwise symmetry and nonconvergence imply
consistent aggregation for all pairs of consumers and the existence of a Gorman-type
representative agent for the whole economy. Jerison provided a characterization of
the broadest class of pairwise aggregable demands for the case of a single, fixed
income distribution.
Perhaps of wider interest, however, is the characterization of demand systems
when the income distribution is less restricted:
OF DEMAND
ANALYSIS AND SUPPLY 2II
Theorem (jerison 19840). rfthe number of consumers H is greater than 2, the

income distribution isfixed, and certain regularity conditions hold, * then the demand
system is pairwise symmetric Lfthe consumer budget shares have one of thefollowing
forms:
where B , Bh, C, and Ch satisfy certain regularity conditions. t Furthermore, if the

shares of all pairs of consumers exhibit nonconvergence and $all consumers have
budget shares of the form (i), (ii), or (iii), the demand system exhibits pairwise aggre-
gation.
All of these are rank-2 demand systems. Thus, even when the weaker condi-
tion of pairwise aggregation is considered, the rank-2 condition continues to hold.
Case (iii) is Muellbauer’s generalized linear demand structure.* Jerison’s aggre-
gation question, however, differs from Muellbauer’s in that (i) and (ii) are neither
stronger nor weaker than Muellbauer’s GL demands restricted to satisfy his aggre-
gation condition.
Freixas and Mas-Colell (1987) likewise find that the generalized linear cat-
egory of demands provides an important condition for the weak axiom of revealed
preference (WARP) to hold in the aggregate. Rather than requiring aggregate de-
mand to be rationalizable by a utility maximizing representative consumer, they ask
what conditions must be placed on Engel curves such that aggregate data satisfy
WARP. They assume that preferences are identical across consumers and they put
no restrictions on the income distribution.
Freixas and Mas-Colell introduce two conditions on Engel curves. The first
they call uniform curvature (UC). Uniform curvature essentially requires that goods
are either luxuries or necessities in all ranges of allowable income. No torsion (NT),
the other condition, requires that Engel curves lie in a plane through the origin; in
other words, Engel curves must have rank 2, or the generalized linear structure of
(42). They then prove the following theorem.
*Specifically, demands are analytic (i.e., demands can be represented by a power series arbitrarily closely
at any point) and the income distribution belongs to a convex, open set of income distributions confined
to the H - 1 unit simplex.
ISpecifically, they are analytic with i’H = 0 = i ’ H h and i’C = 0 = i’Ch (where i is the appropriately
dimensioned unit vector) for all consumers h = 1, . . . , U .
$Note, however, that Jerison’s conditions rule out the additive terms that sum to zero in the Muellbauer
necessity conditions (see the footnote to Muellhauer’s theorem on p. 192).
2 12 RUSSELL
ET AL.
Theorem (Freixas and Mas-Colell I 987). r f agents have identical preferences,

the distribution of income has support ( a , b), and the Engel functions satisfy NT and
UC on the interval ( a , b), then the weak axiom of revealed preference holds in the
aggregate.
Again, rank-2 demands play an important role in this type of restriction of

preferences.
Clearly, the results of Freixas and Mas-Colell are important. It is quite desir-
able that the weak axiom of revealed preference be satisfied in the aggregate. In a
sense, this is demanding some minimal rationality from aggregate demand. It also
provides testable implications-though since aggregate income tends to grow over
time while relative prices change only very slowly one would not expect to uncover
violations of WARP in the aggregate very frequently.
Jerison’s results are perhaps less empirically useful. Conceivably, one could
create a test for the presence of pairwise aggregation. However, it is unclear why that
would be a desirable attribute for an economy to have. The restrictions necessary
for pairwise aggregation to imply aggregation are likely to be violated in any actual
economy over time. Thus, pairwise aggregation lends no practical help to solving
the aggregation question, though the results are fascinating from a technical point
oi view.
The rank-2 restriction also arises in the context of proportional budgeting.
Blackorby and Russell(l993) characterize the optimality of the two-stage budgeting
procedure, where a society allocates shares of total income to individual members
independent of the level of aggregate income. (This could also be the two-stage bud-
geting process of a firm.) They use a linear, additive social welfare function, where
individual shares are allowed to depend on social welfare weights and prices. Shares
that do not depend on prices give rise to price-independent proportional budgeting
(PIPB).
For both proportional budgeting and price-independent proportional budget-
ing, it turns out that the individual indirect utility functions must have the general-
ized linear structure. Here we give only the result for the PIPB case.
Theorem (Blackorby and Russell I 993). Price-independent proportional bud-

geting is optimal tff individual indirect utility functions have the form in ( i ) or ( i i ) :
It is straightforward to show that case (i) corresponds to Jerison’s (ii) in (95).

Case (ii) here corresponds to Jerison’s case (i). Jerison’s third possibility for pair-
ANALYSIS
OF DEMAND
AND SUPPLY 2 13
wise aggregation is not sufficient for PIPB. Blackorby and Russell also discuss the
relationship of their results to those of Muellbauer for the more general proportional
budgeting (PB). They show that aggregate share functions have the Muellbauer PIGL
and PIGLOG structure in aggregate expenditure. However, these could not have been
derived by aggregating over individual PIGL or PIGLOG demand structures. The
aggregation rule that emerges from an economy practicing proportional budgeting
is linear. This implies a greater restriction than the results of Muellbauer (see Sec-
tion 11).
V. MARKET DEMAND A N D HETEROGENEITY
While the exact aggregation conditions surveyed in Section 111enable an investigator

to infer values of parameters of individual demand systems, they do not yield useful
restrictions on aggregate demand systems. Moreover, the well-known results of Son-
nenschein (1973, 1974), Mantel (1974), and Debreu (1974) indicate that almost no
aggregate regularities follow from individual optimization. In fact, they proved that
any continuous, aggregate excess demand system that satisfies Walras’ law could
have been generated by a pure exchange economy of consumers with continuous,
strictly convex, and monotone orderings.
Mas-Colell and Neuefeind (1977) extend Debreu’s result, allowing individual
demands to be correspondences instead of functions. Mantel (1979) considers the
case where the excess demand functions are required to be differentiable, with simi-
lar results. Shafer and Sonnenschein (1982) review this literature and its extensions.
Kirman and Koch (1986) consider the Debreu-Sonnenschein-Mantel (DSM) result
under the restriction that all individuals have the same preferences, and show that
even under this strict assumption, the primary negative result-no restrictions on
aggregate demand-still holds. Kirman (1992) discusses how results from estima-
tion of representative agent models that ignore Debreu-Sonnenschein-Mantel can be
misleading.
Individual optimization alone simply provides too little structure to explain
anything about macro behavior except simple adding-up conditions. Properties that
we might wish to find in the aggregate, such as aggregate revealed preference or
downward-sloping aggregate demand curves must depend on something other than
individual optimization. It is also important to be able to explain why such properties
seem to hold in the aggregate despite the results of DSM.
There appear to be three avenues to dealing with the results of DSM. The first
is to take the approach of the literature on the representative consumer in Section
I11 and find restrictions on demand functions that imply and/or are implied by the
desired macro property. The second is to restrict the distribution of income. The
third is to restrict the distribution of preferences. The second and third approaches,
while imposing restrictions on the degree of heterogeneity of incomes or preferences
2 14 RUSSELLET AL.
can be contrasted with the first approach in that they allow preferences to take any
form. Also, such approaches need not be based on any assumptions about individual
rationality or optimization.
A. The Irrational Consumer and Market Demand

The importance of heterogeneity to demand regularity has long been recognized, but
Becker (1962) was one of the first to formalize the concept. Becker hoped to demon-
strate that individual rationality and the existence of a well-defined preference rela-
tionship were not fundamental to aggregate regularity of demand.* Becker imposed
minimumal rationality-that agents consume their entire income, but allowed agents
to be “irrational” in a particular way. Two formulations of agent irrationality were
proposed: (1)agents randomly choose their consumption bundles along the budget
hyperplane; and (2) agents have completely inflexible demands. In the second case,
they demand exactly what they consumed in the last period; if that bundle is no longer
available, they move along the old budget hyperplane to the nearest point that they
can now afford.
Becker showed that, provided the distribution of choices is spread across the
budget hyperplane, aggregate demand will be downward sloping. Thus, regularity
conditions in aggregate demand, which are often modeled as corresponding to the
choice of a “representative consumer,” can be modeled equally well as a collection
of “irrational” agents. In addition to pointing the way toward agent heterogeneity
as a promising line of explanation of aggregate demand regularity, this paper also
provided an interesting rejoinder to critics of the rationality hypothesis in economics.
B. Income and Preference Heterogeneity
The early result by Becker regarding the spread of agents across the budget hyper-
plane leads to the question of what restrictions on aggregate demand might result
from restricting the distribution of income while allowing preference heterogeneity
among agents. For the case where each consumer has a fixed share of total income,
Eisenberg (1961)showed that market demand could be generated by a multiplica-
tive social welfare function with exponential weights equal to each agent’s share of
aggregate income. Though preferences are allowed to differ between agents in this
formulation, they must be homothetic. Of course, any shift in the relative income
distribution would demand a new set of weights to rationalize the allocation.?
*As it turns out, Becker was incorrect in assuming that individual rationality gave downward-sloping
aggregate demand curves. The DSM result 20 years after his work demonstrated this.
?Chipman (1974) clarified this result and proved it in a slightly different manner. See also Shapiro (1977).
OF DEMAND
ANALYSIS AND SUPPLY 2 15
Hildenbrand (1983) seeks conditions that do not involve restrictions on in-

dividual preferences under which the law of demand will be fulfilled in the aggre-
gate-in other words, conditions sufficient for aggregate demand curves to be down-
ward-sloping. Hildenbrand argues that individual demand functions are in fact a
counterfactual construct and that assumptions about individual preferences or
demand functions are inherently untestable.* He proposes to replace such untest-
able assumptions with testable assumptions about distributions of agents and in-
come.?
Consider the vector of individual demands, d h ( p , y h ) = d ( p , y), which satisfy
WARP (or which have been derived from a preference relation) and are identical for
all consumers. Assuming a continuum of consumers, we can write (mean) market
demand as
where p is the income distribution. A sufficient condition for the aggregate law of
demand is that the Jacobian matrix
[y]
be negative definite for all p. By the definition of (mean) market demand,
where 6 is the compensated demand function (identical for all consumers here),
is the mean Slutsky substitution matrix, and TT is the mean Slutsky income-effect
matrix. The T superscript denotes transpose. S is negative semidefinite, since all
the individual substitution matrices must be negative semidefinite. So the law of
demand reduces to the condition that M be positive definite. Hildenbrand shows that
if we restrict the distribution of income of agents such that the density of incomes
is monotone and nonincreasing, then all aggregate partial demand curves will be
downward-sloping:
*See Hildenbrand (1989)and, in particular, Hildenbrand’s (1993) Market Demand.

t Lewbel(l994) critically reviews Hildenhrand‘s approach.
2I 6 RUSSELL
ET AL.
Theorem (Hildenbrand I 983). For every individual demandfunction dh, where

each d h satisfies the weak axiom of revealed preference and f o r every decreasing den-
sity p on R+,the mean demand function D ( p ) is monotone, i.e., for every p and q in
the set of all prices, we have
( p - q ) ( D ( p )- 50 (100)
Notice that restrictions are placed on the distribution of income of agents,
while no restrictions are placed on the form of the preference function. Preferences
here are assumed to satisfy the weak axiom of revealed preference and to be identical
across consumers, but no assumptions are placed on the form of Engel curves.
The restriction that all consumers have identical preferences is certainly an
undesirable one. Let us assume that preferences belong to some allowable set, A .
And let demand vary between consumers, where individual demands are now
d Y P , Y) (101)
where a E A indexes consumers. Market demand is
P
where is the joint distribution of incomes and attributes. The mean Slutsky matrix
is now
Hildenbrand (1989)considers the conditions that ensure the positive definite-

ness of M* when preferences are allowed to vary among agents. He introduces the
assumption of metonymy-that the distribution of preferences, a,conditional on in-
come, does not depend on the level of income.* Without placing any restrictions on
individual preferences, Hildenbrand is able to show that a decreasing density, p ,
and metonymy are sufficient to imply the aggregate law of demand. Here individ-
ual preferences are not assumed to satisfy WARP. In fact, this result holds even if
all individual income effects are negative semidefinite! This is a pretty astounding
result and well demonstrates one of Hildenbrand’s main points: that there is some-
thing qualitatively different about aggregate demand that cannot be captured in a
mere summing of individual demands.
Caution is nonetheless encouraged. In practice, the density function of house-
hold incomes will not usually be decreasing. Thus, this result is not useful in ex-
plaining why aggregate demand data seem to obey general regularity conditions. But
~~~~
*See Hildenbrand (1989)and Hiirdle, Hildenbrand, and Jerison (1991)for technical details.
OF DEMAND
ANALYSIS AND SUPPLY 2 I7
Hildenbrand’s result does show that restricting the income distribution is, at some
level, analogous to restricting the shape of individual Engel curves. It leaves open
the possibility that some combination of restrictions can be shown to hold empirically
in the economy.
Jean-Michel Grandmont (1987) extends Hildenbrand’s result to a slightly
broader class of densities that allow for discontinuities and also allow for unbounded
densities, provided certain convergence conditions are met. Grandmont also shows
that a result similar to Hildenbrand’s can be derived by restricting the distribution
of preferences.
Hardle, Hildenbrand, and Jerism (1991) (HHJ) consider the aggregate data
and the question of whether market demand adheres to the law of demand. They
consider the mean demand function of (102). A sufficient condition of monotonicity
of the mean demand function is that s* be negative semidefinite and M* be positive
semidefinite. We know that S*will be negative semidefinite, since all the individual
substitution effects are negative. Thus, positive semidefiniteness of M* is sufficient
for the law of demand to hold in the aggregate.
Using the assumption of metonymy, HHJ are able to estimate the mean Slutsky
income-effect matrix. Utilizing the symmetricized matrix M ,
M=M*+M*T (104)
(which will only be positive definite when M* is), they estimate a related matrix,
A*,which is identical to M* under the metonymy assumption. They consider nine
commodity aggregates in U.K. expenditure data from 1969 through 1983 and show
that in all cases (for all years and for all commodities) the matrix fi* is positive
definite. Unfortunately, it is hard to judge the robustness of this result since their tests
rely on the eigenvalues of the estimated matrix, and there is no available distribution
theory for these estimated eigenvalues. Using bootstrapping techniques, they give
some idea that their results may be robust.
From nonparametric density estimates presented in their paper, HHJ clearly
show that the distribution of incomes is not decreasing. So decreasing density of
income is not driving the adherance to the law of demand observed in the data. HHJ
identify two possible explanations. One is that, at higher income levels, preferences
seem to be more heterogenous. In other words, Engel curves tend to spread out at
higher income levels. This reduces the possibility of pathological outcomes. It would
also seem to provide some corroboration of Lewbel’s results that at higher income
levels, rank-2 Engel curves do not describe demand adequately (see Section 1V.A).
Also, the slopes of cross-product curves for the estimated fi* matrix are fairly small
compared to the slopes of the own-product curves. Thus own-price effects seem to
be dominating.
Hardle and Hart (1992)develop the asymptotic theory necessary for testing the
positive definiteness of income-effect matrices using a bootstrapping approach. They
2 18 RUSSELL
ET AL.
create a bootstrapped asymptotic distribution for the smallest estimated eigenvalue,

AI,using the assumption that the income-effect matrix is symmetric. This allows for
testing of a null hypothesis of positive semidefiniteness versus the one-sided alterna-
tive of i 1 > 0. They follow the average derivative techniques developed by Hardle
and Stoker (1989).
Grandmont (1992) proposes restricting agent preferences instead of the in-
come distribution. Following Becker's formulation of the idea, one would like to know
if agent preferences are such that individual choices will be spread (enough) along
the budget hyperplane to generate downward-sloping aggregate demand curves. The
problem is the lack of a metric for preferences. How do we measure how far apart
two preference orderings are? How can we measure the dispersion of a distribution
of preferences without such a measure?
Grandmont proposes an algebraic structure on preferences using affine trans-
formations of the commodity space. Given a vector of commodities, x E R",and
any collection of real numbers a = (a1, . . . , a"),a new vector can be generated by
multiplying each commodity x; by the factor ec"i:
x, = ea 8 x = ( e a l x l , . . . , eanx,) (105)
These transforms can then be used to generate a linear structure on preference rela-
tions and/or demand functions. They are affine, since each point in the commodity
space is transformed linearly to a new point based on the values of the a-vector. If
ai = a for every i, then the transformation is said to be homothetic, as each point in
the commodity space is transformed to a point on a ray passing through the original
point.
Grandmont considers one such equivalence class (i.e., one class of demand
functions that are all related to each other as a-transforms). Within that class, any
preference can be represented as an n-dimensional vector, a.If demands in the econ-
omy are all members of one such class, and if the distribution of a is flat enough, then
aggregate demand will be monotone and individual demand curves are downward-
sloping. Aggregate demand also obeys the weak axiom of revealed preference.
What is the meaning of this? The distribution of a represents the degree of
heterogeneity of preferences within any equivalence class. If the distribution of a is
not flat, then preferences are concentrated around some point and aggregate demand
may behave strangely depending on those individual preferences. If the distribution
of a is flat, then preferences are sufficiently heterogenous and aggregate demand is
well behaved.
These results do not depend on any restrictions on individual demands except
that they obey homogeneity and Walras' law. Demands are not even required to be
preference-driven. Thus, individual rationality, as in Becker's formulation, is simply
unnecessary for aggregate demand regularity.
Grandmont does not consider heterogeneity across equivalence classes. Nor
does he consider heterogeneity in both preference and income dimensions. These ar-
OF DEMAND
ANALYSIS AND SUPPLY 2 19
eas seem promising for future research. Given the results of Hildenbrand and Grand-
mont, it seems that some restrictions on heterogeneity across multiple dimensions
should also yield regularity of aggregate demand. Perhaps the unrealistic restric-
tions on the distribution of income (that the density must be nonincreasing) may be
weakened and combined with some restriction on the distribution of preferences to
yield the law of demand.
VI. JOINTAGGREGATION OVER AGENTS AND

COMMODITIES
A. Fixed Inputs or Outputs
In Section I.A, we noted that, so long as all commodities (net outputs in that exposi-
tion) are efficiently allocated, there is no aggregation problem on the production side
of the economy: a “representative producer” exists under the most basic regularity
conditions. But there is a long history of research (and not a little controversy) over
the issue of aggregation in production when inputs are not efficiently allocated. Thus,
some inputs-notably capital (in the absence of perfect capital rental markets)-are
$zed in the proverbial short run. In addition, outputs of, say, regulated firms might
be fixed, in which case they are modeled as minimizing cost subject to an output
constraint. This section surveys the literature on aggregation of fixed and variable
inputs.*
Let vf and zf,f = 1, . . . , F , be the vectors of variable and fixed commodi-
ties, respectively, in the technology set of firm f . The set of feasible variable quanti-
ties is dependent on the quantities of fixed variables and is denoted T f ( z f )for firm
f.Also, define the production functions of the firms, G f , f = 1, . . . , F , by
vf E T f ( z f ) Gf ( v f , zf) 5 0 (106)
Conditional on the quantities of fixed inputs or outputs of all firms, the aggregate
technology set is then defined (assuming no externalities exist) by
T(z’, . .. ,z F ) = cf
Tf(2f)
so that the aggregate production function G is defined by
U E T ( t ’ , .. ., z F ) G ( v , z ’ , . . . ,z F )5 0
*Parts of this section borrow shamelessly from the excellent editorial summaries of Gorman (1953b) (writ-
ten by Bill Schworm) and Gorman (1968b) (written by the editors) in Blackorby and Shorrocks (1995).
220 RUSSELL
ET AL.
Thus, the aggregate technology set and the aggregate production function depend
on the quantity of each fixed input held by each firm (or each fixed output). From an
econometric point of view, this is a nightmare. Thus, it is standard practice (explicitly
or implicitly) to aggregate over these fixed inputs or outputs, in order to reduce the
dimensions of estimation problems to manageable proportions.
The existence of an aggregate fixed input or output for an individual firm is
represented formally by the existence of functions, ef
and Zf such that
GJ(vJ, z f ) = Z.f(z.f)) (109)

Thus, the vector zf is reduced to a scalar Z j ' ( z f ) on the right-hand side. The con-
ditions required for this structure are well known to be separability of the fixed
inputs/outputs from the variable inputs/outputs. That is, technical rates of substi-
tution between fixed inputs/outputs must be independent of the quantities of vari-
able inputs/outputs.*The existence of an aggregator function over fixed inputs in the
economy-wide production function
G(v, zl, . . . , z F ) = e ( v , Z ( z ' , . . . , z F ) ) (1 10)
however, is more problematical.
8. Klein Aggregation
Several ways of simplifying the structure in (108), taking full account of the aggre-
gation problem, have been proposed. The first suggestion was that of Klein (1946a,
1946b), who proposed an aggregation procedure predicated on the assumption that
none of the inputs or outputs is efficiently allocated. He posited multiple types of
labor and capital inputs producing a set of outputs in accordance with the following
production function:
G f ( x f ,n f , z f ) = 0 Vf (111)
where x f , n f , and zf are the vectors of output, labor, and capital quantities, respec-
tively, for firm f.To simplify the analysis of economy-wide production functions,
Klein proposed the existence of functions, c,
X , N , and 2 such that
e ( X ( x ' , . . . , xF), N ( n ' , . . . , n F ) ,Z(z', . . . , z F ) ) = 0
(112)
G f ( x f ,n f , zf) = 0 Vf
Thus, a problem of gargantuan proportions is reduced to one in which there are just
three variables in the economy's production function: aggregate output, aggregate
labor input, and aggregate capital input.
*The equivalence between separability and the functional structure in (109) was developed independently
by Leontief (1947a, 1947b) and Sono (1961) and subjected to a penetrating analysis by Gorman (1%8a).
For a comprehensive treatment and a recent survey, see Blackorby, Primont, and Russell(l978, 1997).
ANALYSIS
OF DEMAND
AND SUPPLY 22 1
The question posed by Klein was answered by Nataf (1948),who showed that
(112)holds if and only if each firm's production function can be written as
Gf ( x f , n s , z f ) = Xf(xf) + N f ( n f ) + Zf (zf) = 0 (113)

That is, there exist aggregators over each of the three types of variables in each pro-
duction function-hence, each set of variables is separable from its complement-
and the technology is linear in these aggregates.* It is easy to see that this structure
is sufficient for the Klein condition (112).Define
X ( x ' , . . . , xF) := Xf(X', . . . , 2)

f
N(n1,. . . , nF) := N f ( n l ,. . . , 2 )
f
Z(t.1, . . . ,z F ) := ~ z q z '. .,. , 2 F )
s
and
e(X(x*,. . . , xF), N ( n l , .. . , n F ) ,Z ( z ' , . . . , z F ) )
:= X ( x 1 , . . . , xF) + N ( n 1 , . . . , nF) + Z ( Z 1 , . . . , 8 )
It is now clear that

X(X1, . . . ,xF) + N ( n l , . . , nF) + Z(z1, . . . , ZF) = 0
*
(116)
X f ( x f ) + N J ( n J )+ ZJ(2f) = 0 vj-
Gorman (1953b) noted that the Klein-Nataf structure is not very useful for
empirical implementation because the aggregator functions for output, labor, and
capital in (112)and (116)depend on the entire distribution of these variables among
firms. He suggested that, to be useful, these aggregate commodities should depend
only on the sum across firms of the component variables: i.e., that the aggregate
production function take the form
He then showed that the necessary and sufficient conditions for this aggregate pro-
duction structure is that the individual production equation have the affine form
*Hence, the variables are, in Gorman's (1968a) terminology, completely separable in Gf:in addition to
each group being separable, arbitrary unions of groups are separable (so that, e.g., technical rates of
substitution between labor and capital inputs are independent of output quantities).
222 RUSSELLET AL.
where a,p, and y are appropriately dimensioned vectors of parameters that are iden-
tical for all firms and Of is an arbitrary, idiosynchratic scalar for firm f.In this case,
the economy production equation is
f f
so that the Klein aggregates have the linear structures
and
C. Aggregation with Both Efficiently Allocated (“Labor”) Inputs and

Fixed (“Capital”) Inputs or Outputs
The restrictions on individual technologies that are necessary and sufficient for the
existence of Klein aggregates are too demanding to be useful in empirical analysis.
Even if we allow the aggregates to depend on the entire distribution of component
variables, as in (112), the individual technologies are linear in the firm-specific ag-
gregates, and if we more realistically require that the economy aggregates depend
only on the sum of component variables, as in (1 17), the individual technologies are
linear in each output and input variable (implying, e.g., linear isoquants and linear
production possibility surfaces).
Subsequent literature on “capital” aggregation has focused on the unreason-
ableness of Klein’s requirement that the allocation of outputs and inputs be en-
tirely arbitrary. In fact, May (1946) and Pu (1946) immediately pointed out that
competitive-equilibrium conditions-in particular, profit maximization by firms-
might be exploited to weaken the requirements for the existence of composite com-
modities in aggregate production technologies. In particular, Solow (1964) showed
how the efficient allocation of labor inputs might facilitate the aggregation of (fixed)
OF DEMAND
capital inputs. His ingenious insight was to note that, under certain conditions, the
efficient allocation of inputs not being aggregated can render efficient an arbitrary
allocation of fixed inputs as well.
The ultimate-and naturally elegant-solution to the problem of aggregating
over fixed inputs in the aggregate technology and in the presence of efficiently allo-
cated inputs was presented by Gorman (1965, 1968b). The existence of efficiently
allocated inputs and outputs allowed him to exploit duality theory to solve the prob-
lem (as in the case of the representative consumer in Section 11). Given the tech-
nology represented in (106), the variable profit function* of firm f,denoted Pf,is
defined by
, := max{p. v f
~ f ( pzf) I ~ f ( v zf f, ) = 01
Vf
where p is the (competitive equilibrium) price vector for variable inputs and outputs.
For the economy, the variable profit function P is defined by
~ ( pzl,, . . . , z F ) = max{p. ‘u I~ ( vz l, , . . . , z F ) = 01
V
where the second identity follows from the existence of a “representative firm” con-
ditional on the allocation of fixed inputs (i.e., the interchangeability of set summation
and optimization), discussed in Section 1I.A. Note that, if outputs are fixed and inputs
are variable, P is the cost function of the firm, showing minimal cost as a function of
input prices and the fixed output quantities.
Theorem (Gorman I 968b). The aggregatorfunction in ( 110) exists $-and only
fi the individual variable projit functions (or cost functions) have the structure
Pf(p, d)= I l ( p ) Z f ( z f ) + A f ( p ) Vf (123)

in which case the variable projit function for the economy is
=: n ( p ) Z ( z ’ ,. . . , z F ) + A(p)
The similarity between this structure and the structure of the expenditure
function that is necessary and sufficient for the existence of a representative con-
sumer in Section I1 is palpable. In fact, as nicely elucidated in the editorial summary
*See, e.g., Diewert (1974) for a discussion of variable profit functions.

224 RUSSEUET AL.
of Gorman (1968b)in Blackorby and Shorrocks (1995),Gorman's proof of the exis-

tence of an aggregate fixed input for the economy constitutes, with little more than a
change in notation, an alternative proof of the existence of a community indifference
map and hence the existence of a Gorman-type representative consumer.
Recall from Section 1I.A that the community (Scitovsky) preference set (the
set of aggregate commodity bundles that makes possible a particular profile of utility
levels of the H consumers) is the sum of the individual level sets:
N(u1, . . . , u h ) = Nh(uh)
h
The individual and community expenditure functions are given by
and
A community preference map, and hence a representative consumer, exists if and

only if the community preference set depends only on a scalar utility (of the repre-
sentative consumer),
N ( u ' , . . . , uh) = R ( U ( u ' , .. . , u h ) ) (128)

in which case the community expenditure function is
E ( u ' , . . . , u h , p ) = E ( U ( u ' , . . . , u h ) ,p ) (129)
Apart from a change in notation (substituting consumer utilities for fixed inputs and
expenditure minimization for profit maximization), this is the same structure as the
above results on the aggregation of fixed inputs. Thus, the individual expenditure
functions must have the form
Eh(& p ) = W p > I l r h b h > A h ( p )+ (130)

which implies that
E ( & . . . , u h ,p > = G ( p ) Ilrh(uh) + Ah(p>

h h (131)
=: I l ( p ) U ( u ' , . . . , u h ) + A(p)
which is (apart from the harmless monotonic transformation of the utility levels) iden-
tical to the structure in the Gorman theorem on representative consumers in Section
11. The parallel nature of these (affine) structures underscores the common thread
running through apparently diverse aggregation problems.
ANALYSIS
OF DEMAND
AND SUPPLY 225
D. Extension to Multiple Fixed Inputs

The Gorman theorem in Section V1.C on the aggregation of fixed inputs in the pres-
ence of some efficiently allocated inputs was essentially worked out in his analysis of
capital aggregation in vintage models for a single firm (Gorman 1965). If we interpret
(106) as representing the technologies for different capital vintages, then the struc-
ture (123) for the vintage-specific variable profit functions is necessary and sufficient
for the existence of an aggregate capital stock, obtained by aggregating over the dif-
ferent capital vintages, as in (1 10). This analysis was extended, and the proof made
much more elegant, in Gorman (1968b). The problem solved in this paper general-
izes the structure in (110)to include multiple aggregate fixed inputs. Thus, rewrite
(106) as
v . f E T . f ( z { ,. . . , Zi) t G f ( v . f, z ( , . . . , Z i ) = 0 Vf (132)
where z;f is the vector of fixed inputs of type r, r = 1, . . . , R , held by firm f.The
aggregate technology set then depends on the distribution of quantities of each type
of fixed input among firms and the question is when can the aggregate technology by
represented by the production equation
- 1 1 F R I
G(v, Z (zl,. . . , z1 ), . . . , Z ( z R , . . . ,z,”)) = 0 (133)
The answer, as elegantly shown by Gorman (1968b), is a straightforward generaliza-
tion of the earlier result: the variable profit functions of the individual firms must
be
and the variable profit function of the economy therefore has the structure
P ( p , Z’(Zi, . . . , zf), . . . , zqz;, . . . ,z,”))
where
Again, the similarity of this structure, which rationalizes the existence of mul-
tiple aggregate inputs in an economy-wide production function, with the structure of
exact aggregation of consumer demand functions with multiple income distribution
statistics, surveyed in Section 111, is palpable. In fact, using the envelope theorem to
derive the net-supply functions of variable netput vectors for firm f,we obtain
226 RUSSELLET AL.
which has a structure that is very similar to that analyzed by Gorman (1981).Aggre-
gate (vector valued) net-supply functions are then given by
=: & ( p , 0,(z1’, . . . , z ; ) . . . , of&;, . . . ,z,“))

This structure of net-demand systems is analogous to that of exact aggregation in
consumer theory, the principal difference being that the 8 functions in the aggregate
consumer demand functions in exact aggregation depend on the H incomes of the
consumers, whereas in the aggregation of fixed inputs, the 0 functions depend on
vectors of fixed inputs of firms.
In fact, in econometric applications, it is standard to represent heterogeneous
technologies of firms through the use of technological “attributes,” analogous to those
representing heterogeneity of consumer preferences in Section 111. This raises prob-
lems that are identical to those posed by the existence of fixed inputs; indeed, one
can think of fixed inputs as one type of technological attribute. These attributes can
be modeled exactly as fixed inputs are modeled in constructing aggregate technolo-
gies. Indeed, one can interpret some of the Z rf variables in (132)as attributes and the
2‘ functions in (133) as aggregate attributes. The approach of Heineke and Shefrin
(1990) can then be used to identify attribute parameters. See Appelbaum (1982),
Borooah and Van Der Ploeg (1986),Chavas (1993),Gourikroux (1990),and Fortin
(1991) for applications of these aggregation notions in producer theory. In particu-
lar, Fortin and Gourikroux find that significant bias results from ignoring aggregation
issues in econometric analysis of production.
Owing to the difference between the optimization problems of consumers and
producers, two caveats need to be emphasized. First, while the structures of exact
aggregation, on the one hand, and aggregation over fixed inputs and technological
attributes are analogous, the proofs of these aggregation results are quite different.
Second, Gorman’s theorem on exact aggregation and the rank of demand systems
does not go through for producers.
E. Other Extensions
Econometric studies using aggregate data also commonly aggregate over variable in-
puts (e.g., different types of labor). Again, at the level of the individual firm the nec-
essary and sufficient condition for the existence of such aggregates is separability,
OF DEMAND
but at the aggregate level more is required. The necessary and sufficient conditions
for aggregation of efficiently allocated inputs are different from, but similar in many
respects to, the conditions for aggregation of fixed inputs. Moreover, as in the case of
fixed inputs, the conditions are less demanding if aggregation is required only over
a subset of the efficiently allocated inputs. In addition, aggregation of variable in-
puts is made more difficult if there exist fixed inputs in the technology. These issues
were analyzed in Gorman (1967, 1982a), and Blackorby and Schworm (1988a) pro-
vide an excellent survey and analysis of the various results obtained under different
assumptions about which subsets of efficiently allocated and fixed inputs are being
aggregated in the economy-wide technology.* To a large extent, the common theme
of affine structures runs through these results, but there are some surprises (which,
in the interest of space, are left to the reader to discover).
The use of common duality concepts in each of these aggregation problems en-
tails the use of strong convexity assumptions regarding the technology. The possible
inadequacy of this approach for long-run analysis led Gorman (1982b) to examine
aggregation conditions under constant returns to scale.
Joint aggregation over commodities and agents is also common in studies of
consumer behavior. As one would expect, the necessary and sufficient conditions,
as shown by Blackorby and Schworm (1988b), are stronger than those required for
either type of aggregation taken separately.
Finally, Gorman (1990)has more recently examined aggregation problems in
which the aggregate inputs for the economy depend on the quantities of all inputs,
perhaps because of externalities. These studies lead to structures that are quite simi-
lar to the affine structures above and, interestingly, reminiscent of the rank conditions
for Engel curves in Sections I11 and IV.
VII. CONCLUDING REMARKS
As Section 1.A summarizes the literature, our concluding remarks will be limited
to two observations. First, when we began research on this survey, we expected the
chapter to be about half theory and half econometrics; the reader has undoubtedly
noted, however, that the theory of aggregation dominates. Thus, the main lesson we
take from this endeavor is that, while the theory of aggregation is fairly well devel-
oped, econometric application is in its infancy. The principal empirical literature-
most notably Jorgenson, Lau, and Stoker (1980, 1981, 1982)-maintains exact ag-
gregation conditions, generates results on demographics and demand, and studies
various issues of welfare economics. Less well developed is the testing of restric-
tions on preferences and on the distribution of incomes or preferences that are nec-
*Also see Fisher (1965, 1968a, 1968b, 1%9, 1982, 1983) and Blackorby and Schworm (1984).
228 RUSSELL
ET AL.
essary or sufficient for various types of aggregation consistency. The most promising
avenues for research in this direction have been forged by Lewbel (1991), Hardle,
Hildenbrand, and Jerison (1991),and Hausman, Newey, and Powell (1995). But it
seems that the potential for testing aggregation conditions has barely been tapped.
Second, just as the theory of aggregation has dominated this literature, it should
be obvious to the reader that one person, Terence Gorman, has in turn dominated the
search for theoretical results. His towering achievements in this area are testimony
to his penetrating intellect. Moreover, while many of his papers are challenging read-
ing, he has always kept his eye on the potential for empirical application. One would
hope that applied econometricians would focus more on testing the aggregation the-
ories promulgated primarily by Gorman.
ACKNOWLEDGMENTS
We are deeply indebted to Chuck Blackorby and Tony Shorrocks for their masterful
editorial effort in making Terence Gorman’s works, many previously unpublished,
available in Volume I of the Collected Works of W M . Gorman; without this towering
volume, including the illuminating editorial introductions to Gorman’s papers, our
task would have been impossible. Nor would the chapter have been possible without
the investment in human capital accruing to the long collaboration between Blacko-
rby and Russell. We also thank Kusum Mundra for comments, and we are especially
grateful to Chuck Blackorby and Bill Schworm for reading the manuscript and for
many valuable suggestions. Needless to say, they are in no way responsible for any
remaining errors.
REFERENCES
Aczel, J. and J. Dhombres (1989), Functional Equations in Several Variables with Applications
to Mathematics, Znjormation Theory and to the Natural and Social Sciences, Cambridge
Afriat, S. N. (1967), The Construction of a Utility Function From Expenditure Data, Znterna-
tional Economic Review, 8,460-472.
Afriat, S. N. (1972), Efficiency Estimation of Production Functions, International Economic
Revieal. 13, 568-598.
A iyagari, S. (1985), Observational Equivalence of the Overlapping Generations and Dis-
counted Dynamic Programming Frameworks for One-Sector Growth, Journal of Eco-
nomic Theory, 35,201-221.
Antonelli, G. B. (1886), Sulla Teoria Matematica della Economiu Politica, Pisa: nella ti-
pografia del Folchetto; English translation in Preferences, Utility, and Demand ( J . S.
Chipman, L. Hunvicz, M. K. Richter, and H. F. Sonnenschein, eds.), Harcourt Brace
Jovanovich, New York, 1971).
ANALYSIS
OF DEMAND
AND SUPPLY 229
Appelbaum, E. (1982), The Estimation of the Degree of Oligopoly Power, Journal of Econo-
metrics, 19, 287-299.
Arrow, K. J. (1951), Social Choice and Zndividual Values, Wiley, New York.
Banks, J., R. Blundell, and A. Lewbel(1993), Quadratic Engel Curves and Welfare Measure-
ment, Working Paper, Institute of Fiscal Studies, London.
Becker, G. S.(1962), Irrational Behavior and Economic Theory, Journal of Political Economy,
70, 1-13.
Blackorby, C., R. Boyce, and R. R. Russell(1978), Estimation of Demand Systems Generated
by the Gorman Polar Form: A Generalization of the S-Branch Utility Tree, Economet-
rica, 46, 345-364.
Blackorby, C., R. Davidson, and W. Schworm (1993), Economies with a Two-Sector Represen-
* tation, Economic Theory, 3, 717-734.
Blackorby, C., D. Primont, and R. R. Russell (1978), Duality, Separability, and Functional
Structure: Theory and Economic Applications, North-Holland/American Elsevier, New
York.
Blackorby, C., D. Primont, and R. R. Russell (1997), Separability, in Handbook of Utility
Theory (S.Barbara, P. J. Hammond, and C. Seidl, eds.), Kluwer, Boston.
Blackorhy, C. and R. R. Russell (1993), Samuelson’s “Shibboleth” Revisited: Proportional
Budgeting among Agents and Rank-Two Demand Systems, in Mathematical Modelling
in, Economics (W. E. Diewert, K. Spremann, and F. Stehling, eds.), Springer-Verlag,
Heidelberg, 3,546.
Blackorby, C. and R. R. Russell (1997), Two-Stage Budgeting: An Extension of Gorman’s
Theorem, Economic Theory, 9, 185-193.
Blackorby, C. and W. Schworm (1982), Aggregate Investment and Consistent Intertemporal
Technologies, Review of’Economic Studies, 49, 595-614.
Blackorby, C. and W. Schworm (1984), The Structure of Economies with Aggregate Measures
of Capital: A Complete Characterisation, Review of Economic Studies, 51,633-650.
Blackorby, C. and W. Schworm (1988a), The Existence of Input and Output Aggregates in
Aggregate Production Functions, Econometricn, S6,613-643.
Blackorby, C. and W. Schworm (1988b),Consistent Commodity Aggregates in Market Demand
Equations, in Measurement in Economics (W. Eichhorn, ed.), Physica-Verlag, Heidel-
berg, 577-606.
Blackorby, C. and W. Schworm (1993), The Implications of Additive Consumer Preferences
in a Multi-Consumer Economy, Review ofEconomic Studies, 60,209-228.
Blackorby, C. and A. F. Shorrocks (1995), Separability and Aggregation: Collected Works of
K M. Gorman, Vol. I, Clarendon Press, Oxford.
Blundell, R., P. Pashardes, and G. Weber (1993), What do We Learn About Consumer De-
mand Patterns from Micro Data? American Economic Review, 83, 570-597.
Borooah, V. K. and F. Van Der Ploeg (1986), Oligopoly Power in British Industry, Applied
Economics, 18,583-598.
Bronars, S.G. (1987), The Power of Nonparametric Tests of Preference Maximization, Econo-
metrica, 55,693-698.
Brown, M. and D. Heien (1972), The S-Branch Utility Tree: A Generalization of the Linear
Expenditure System, Econometrica, 40, 737-747.
Buse, A. (1992), Aggregation, Distribution and Dynamics in the Linear and Quadratic Expen-
diture Systems, Review of Economics and Statistics, 74, 45-53.
230 RUSSELLET AL.
Chavas, J.-P. (1993), On Aggregation and Its Implications for Aggregate Behavior, Ricerche
Economiche, 47,201-214.
Chipman, J. (1974), Homothetic Preferences and Aggregation, Journal of Economic Theory,
8,26-38.
Christensen, L., D. C. Jorgenson, and L. J. Lau (1975), Transcendental Logarithmic Utility
Functions, American Economic Review, 65,367-383.
Constantinides, G. (1982), Intertemporal Asset Pricing with Heterogeneous Consumers and
without Demand Aggregation, Journal of Business, 55,253-267.
Deaton, A. (1986), Demand Analysis, in Handbook of Mathematical Economics, Vol. 3 (K. J.
Arrow and M. D. Intriligator, eds.), North-Holland, Amsterdam, 1767-1839.
Deaton, A. S. and J. Muellbauer (1980), An Almost Ideal Demand System, American Eco-
nomic Review, 70,312-326.
Debreu, G. (1954),Valuation Equilibrium and Pareto Optimum, in Proceedings ofthe National
Academy of Sciences of the USA, 46,588-592.
Debreu, G. (1959), Theory of Value, Wiley, New York.
Debreu, G. (1974),Excess Demand Functions, Journal of Mathematical Economics, 1,15-21.
Diewert, W. E. (1971),An Application of the Shephard Duality Theorem: A Generalized Leon-
tief Production Function, Journal of Political Economy, 79,481407.
Diewert, W. E. (1973), Afriat and Revealed Preference Theory, Review of Economic Studies,
40,419-426.
Diewert, W. E. (1974), Applications of Duality Theory, in Frontiers of Quantitative Economics,
Vol. 2 (M. Intriligator and D. Kendrick, eds.), North-Holland, New York, 106-171.
Diewert, W. E. and C. Parkan (1985), Tests for the Consistency of Consumer Data, Journal of
Econometrics, 30, 127-147.
Eichenbaum, M., L. Hansen, and K. Singleton (1988), A Time Series Analysis of Representa-
tive Agent Models of Consumption and Leisure Choice Under Uncertainty, Quarterly
Journal of Economics, 103,Sl-78.
Eisenberg, B. (1961), Aggregation of Utility Functions, Management Science, 7,337-350.
Fisher, F. (1965), Embodied Technical Change and the Existence of a n Aggregate Capital
Stock, Review of Economic Studies, 32,263-288.
Fisher, F. (1968a),Embodied Technology and the Existence of Labour and Output Aggregates,
Review of Economic Studies, 35, 391-412.
Fisher, F. (1968b), Embodied Technology and Aggregation of Fixed and Movable Capital
Goods, Review of Economic Studies, 35, 41 7-428.
Fisher, F. (1969), The Existence of Aggregate Production Functions, Econometrica, 37, 553-
577.
Fisher, F. (1982), Aggregate Production Functions Revisited: The Mobility of Capital and the
Rigidity of Thought, Review ojEconomic Studies, 49,615-626.
Fisher, F. (1983), On the Simultaneous Existence of Full and Partial Capital Aggregates, Re-
view of Economic Studies, 50, 197-208.
Fleissig, A. R., A. R. Hall, and J. J. Seater (1994),GARP, Separability, and the Representative
Consumer, Working Paper, Department of Economics, North Carolina State University.
Fortin, N. M. (1991), Fonctions de Production et Biais d'Agrkgation, Annales d'kconomie et
de Statistigue, 20121, 41-68.
Freixas, X. and A. Mas-Colell (1987), Engel Curves Leading to the Weak Axiom in the Ag-
gregate, Econometrica, 55,5 15-53 1.
OF DEMAND
ANALYSIS AND SUPPLY 23 I
Geary, R. C. (1950), A Note on “A Constant-Utility Index of the Cost of Living,” Review of

Economic Studies, 1 8 , 6 5 6 6 .
Gorman, W. M. (1953a), Community Preference Fields, Econometrica, 2 1, 63-80 (reprinted
as Chapter 15 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1953b), Klein Aggregates and Conventional Index Numbers, presented to the
Innsbruck Meeting of the Econometric Society (published as Chapter 16 in Blackorby
and Shorrocks 1995).
Gorman, W. M. (19S9), Separable Utility and Aggregation, Econometrica, 27, 469-481 (re-
printed as Chapter 3 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1%1), On a Class of Preference Fields, Metroeconomica, 13,53-56 (reprinted
as Chapter 1 7 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1965), Capital Aggregation in Vintage Model, presented to the First World
Congress of the Econometric Society (published as Chapter 18 in Blackorby and Shor-
rocks 1995).
Gorman, W. M. (1967), Aggregates for Variable Goods, typescript, London School of Eco-
nomics (published as Chapter 18 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1968a), Measuring the Quantities of Fixed Factors, in Vulue, Capital, and
Growth: Essuys in Honor of Sir John Hicks (J. N. Wolfe, ed.), Edinburgh University
Press, Edinburgh, 141-172 (reprinted as Chapter 19 in Blackorby and Shorrocks
1995).
Gorman, W. M. (1%8b), The Structure of Utility Functions, Review of Economic Studies 5,
369-390 (reprinted as Chapter 12 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1976), Muellbauer’s Representative Agent, typescript (published as Chapter
in Blackorby and Shorrocks 1995).
Go. . W. M. (1981), Some Engel Curves, in Essuys in the Theory and Measurement of Con-
sumer Behaviou,. (A. S. Deaton, ed.), Cambridge University Press, Cambridge, 7-29
(reprinted as Chapter 20 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1982a), Aggregation in the Short and Long Run, typescript, Oxford (published
as Chapter 2 5 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1982b), Long-Run Aggregates under Constant Returns, typescript, Oxford
(published as Chapter 26 in Blackorby and Shorrocks 1995).
Gorman, W. M. (1990), More Measures for Fixed Factors, in Measurement and Modeling in
Economics (G. D. Myle?, ed.), North-Holland, Amsterdam (reprinted as Chapter 21 in
I.lackorby and Shorrocks 1995).
GourkIoux, C . (1990), H6tkrog6n6itk 11. Etude des Biais de Reprksentativitk, Annales d ’ t ~ o -
nomie P , de Statistigue, 17, 18.5-204.
Grandmont J. M. (1987), Distributions of Preferences and the Law of Demand, Econometricu,
5‘,, 155-161.
Grandmont, J. M. (1992), Transformations of the Commodity Space, Behavioral Heterogeneity,
and the Aggregation Problem, Journal ‘$Economic Theory, 57, 1-35.
Green, H. A. J. (1964), Aggregation in Economic Analysis: An Introductory Survey, Princeton
IJniversity Press, Princeton.
Hardle, W. and J. Hart (1992), A Bootstrap Test for Positive Definiteness of Income Effect
Matrices, Econometric Theory, 8,276-290.
Hardle, W., W. Hildenbrand, and M. Jerison (1991), Empirical Evidence on the Law of De-
mand, Econometrica, 59, 152.5-1 549.
232 RUSSELLET AL.
Hardle, W. and T. M. Stoker (1989), Investigating Smooth Multiple Regression by the Method
of Average Derivatives, Journal ofthe American Statistical Association, 84,986-995.
Hausman, J. A., W. K. Newey, and J. L. Powell(1995), Nonlinear Errors in Variables Estima-
tion of Some Engel Curves, Journal of Econometrics, 65,205-233.
Heineke, J. M. (1979), Exact Aggregation and Estimation, Economics Letters, 4, 157-162.
Heineke, J. M. (1993), Exact Aggregation and Consumer Demand Systems, Richerche Eco-
nomiche, 47,215-232.
Heineke, J. M. and H. M. Shefrin (1982), The Finite Basis Property and Exact Aggregation,
Heineke, J. M. and H. M. Shefrin (1986),On a n Implication of a Theorem Due to Gorman,
Heineke, J. M. and H. M. Shefrin (1987), On Some Global Properties of Gorman Class Demand
Systems, Economics Letters, 25, 155-160.
Heineke, J. M. and H. M. Shefrin (1988), Exact Aggregation and the Finite Basis Property,
International Economic Review, 29, 525-538.
Heineke, J. M. and H. M. Shefrin (1990), Aggregation and Identification in Consumer Demand
Systems, Journal of Econometrics, 44, 377-390.
Hildenbrand, W. (1983), On the Law of Demand, Econometrica, 51,997-1019.
Hildenbrand, W. (1989), Facts and Ideas in Microeconomic Theory, European Economic Re-
view, 33,251-276.
Hildenbrand, W. (1993), Market Demand: Theory and Empirical Evidence, Princeton Univer-
sity Press, Princeton.
Houthakker, H. S. (1950), Revealed Preference and the Utility Function, Economicu, 17,159-
174.
Houthhakker, H. S. (1963), Some Problems in the International Comparison of Consumption
Patterns, in Les Besoins de Biens de Consommation (R. Mosse, ed.), Centre National de
la Recherche Scientifique, Grenoble.
Howe, H., R. A. Pollak, and T. J. Wales (1979), Theory and Time Series Estimation of the
Quadratic Expenditure System, Econometrica, 47, 1231-1247.
Huang, C. (1987), An Intertemporal General Equilibrium Asset Pricing Model: The Case of
Diffusion Information, Econometrica, 55, 117-142.
Jerison, M. (1984a), Aggregation and Pairwise Aggregation of Demand When the Distribution
of Income Is Fixed, Journal of Economic Theory, 33, 1-31.
Jerison, M. (1984b), Social Welfare and the Unrepresentative Representative Consumer, manu-
script, SUNY Albany.
Jorgenson, D. W. (1986), Econometric Methods for Modeling Producer Behavior, in Handbook
oJ’Mathematica1 Economics, Vol. 3 (K. J. Arrow and M. D. Intriligator, eds.), North-
Holland, Amsterdam, 1841-191 5.
Jorgenson, D. W., L. J. Lau, and T. M. Stoker (1980), Welfare Comparison under Exact Aggre-
gation, American Economic Review, 70, 268-272,
Jorgenson, D. W., L. J. Lau, and T. M. Stoker (1981), Aggregate Consumer Behavior and In-
dividual Welfare, in Macroeconomic Analysis (D. Currie, R. Nobay, and D. Peel, eds.),
Croom-Helm, London, 35-61.
Jorgenson, D. W., L. J. Lau, and T. M. Stoker (1982), The Transcendental Logarithmic Model
of Aggregate Consumer Behavior, in Advances in Econometrics, Vol. 1( R . Baseman and
G. Rhodes, eds.), JAI Press, Greenwich, 97-238.
OF DEMAND
ANALYSIS A N D SUPPLY 233
Jorgenson, D. W. and D. T. Slesnick (1984), Aggregate Consumer Behaviour and the Measure-
ment of Inequality, Review of Economic Studies, 51, 369-392.
Kirman, A. (1992), Whom or What Does the Representative Individual Represent? Journal
of Economic Perspectives, 6, 117-136.
Kirman, A. and K. Koch (1986), Market Excess Demand Functions in Exchange Economies:
Identical Preferences and Collinear Endowments, Review of Economic Studies, 53,
457-463.
Klein, L. (1946a),Macroeconomics and the Theory of Rational Behaviour, Econometrica, 14,
93-108.
Klein, L. R. (1946b),Remarks on the Theory of Aggregation, Econometrica, 14,303-312.
Klein, L. R. and H. Rubin (1947-1948), A Constant Utility Index of the Cost of Living, Review
of Economic Studies, 15,84-87.
Koo, Y. C. (1963), An Empirical Test of Revealed Preference Theory, Econometrica, 31,646-
664.
Koopmans, T. C. (1957), Three Essays on the State of Economic Science, McGraw-Hill, New
York.
Kydland, F. E. and E. C. Prescott (1982), Time to Build and Aggregate Fluctuations, Econo-
metrica, 50, 1345-1370.
Lau, L. J . (1977a), Existence Conditions for Aggregate Demand Functions: The Case of a
Single Index, IMSSS Technical Report 248, Stanford University.
Lau, L. J. (1977b),Existence Conditions for Aggregate Demand Functions: The Case of Mul-
tiple Indexes, IMSSS Technical Report 249, Stanford University.
Lau, L. J. (1982), A Note on the Fundamental Theorem of Exact Aggregation, Economics Let-
ters, 9, 119-126.
Lau, L. J. (1986), Functional Forms in Econometric Model Building, in Hundbook ofEcono-
metrics, Vol. 111 (Z. Griliches and M. D. Intriligator, eds.), North-Holland, Amsterdam.
Leontief, W. W. (1947a), Introduction to a Theory of the Internal Structure of Functional Re-
lationships, Econometrica, 15, 361-373.
Leontief, W. W. (1947b), A Note on the Interrelation of Subsets of Independent Variables
of a Continuous function with Continuous First Derivatives, Bulletin ofthe American
,‘Mathematical Society, 5 3 , 343-350.
Lewbel, A. (1987), Characterizing Some Gorman Engel Curves, Econometrica, 55, 1451-
1459.
Lewbel, A. (1989a),Nesting the AIDS and Translog Demand Systems, Internationul Economic
Review, 30,349-356.
Lewbel, A. (1989b),A Demand System Rank Theorem, Econometrica, 57, 701-705.
Lewbel, A. (1990), Full Rank Demand Systems, International Economic Review, 31, 289-
300.
Lewbel, A. (1991), The Rank of Demand Systems: Theory and Nonparametric Estimation,
Econometrica, 59, 71 1-730.
Lewbel, A. (1994), An Examination of Werner Hildenbrand’s Market Demand, Journal of
Economic Literature, 32, 1832-1841.
Manser, M. E. and R. J. McDonald (1988), An Analysis of Substitution Bias in Measuring
Inflation, 1959-8.5, Econometricu, 56,909-930.
Mantel, R. (1974),On the Characterization of Aggregate Excess Demand, Journal ofEconomic
Theory, 7, 348-353.
234 RUSSELLET AL.
Mantel, R. (1979),Homothetic Preferences and Community Excess Demand Functions, Jour-

nal of Economic Theory, 12, 197-201.
Mas-Colell, A. and W. Neuefeind (1977),Some Generic Properties of Aggregate Excess De-
mand and an Application, Econometric, 45, 591-599.
May, K. (1946),The Aggregation Problem for a One-Industry Model, Econometrica, 14,285-
298.
Muellbauer, J. (1975),Aggregation, Income Distribution, and Consumer Demand, Review of
Economic Studies, 62,269-283.
Muellbauer, J. (1976),Community Preferences and the Representative Consumer, Economet-
rica, 44,525-543.
Nataf, A. (1948),Sur la Possibilitk de la Construction d e Certains Macromodeles, Economet-
rica, 16,232-244.
Nicol, C. J. (1994),Identifiability of Household Equivalence Scales through Exact Aggrega-
tion: Some Empirical Results, Canadiun Journal of Economics, 27, 307-328.
Pollak, R. A. and T. J. Wales (1978),Estimation of Complete Demand Systems from House-
hold Budget Data: The Linear and Quadratic Expenditure Systems, American Economic
Review, 68,348-359.
Pollak, R. A. and T. J. Wales (1979),Welfare Comparisons and Equivalent Scales, American
Economic Review, 69,216-22 1.
Pollak, R. A. and T. J. Wales (1980),Comparison of the Quadratic Expenditure System and
Translog Demand Systems with Alternative Specifications of Demographic Effects,
Pu, S.(1946),A Note on Macroeconomics, Econometrica, 14,299-302.
Russell, R. R. and M. 0. Tengesdal (1996), On the Power of Non-Parametric Tests of the
Representative-Consumer Hypothesis, Discussion Paper, University of California,
Riverside.
Russell, T. (1983),On a Theorem of Gorman, Economics Letters, 11,233-224.
Samuelson, P. A. (1938),A Note on the Pure Theory of Consumer’s Behavior, Economica, 5,
61-71.
Samuelson, P. A. (1946-47),Consumption Theory in Terms of Revealed Preference, Econom-
ica, 15,243-253.
Samuelson, P. A. (1956),Social Indifference Curves, Quarterly Journal of Economics, 70, 1-
22.
Scitovsky, T. (1942),A Reconsideration of the Theory of Tariffs, Review of Economic Studies,
9,89-110.
Shafer, W. and H. Sonnenschein (1982),Market Demand and Excess Demand Functions, in
Handbook of Mathematical Economics (K. Arrow and M. D. Intrilligator, eds.), North-
Holland, Amsterdam, 671-692.
Shapiro, P. (1977), Aggregation and the Existence of a Social Utility Function, Journal of
Economic Theory, 16,475-480.
Shapiro, P. and S.Braithwait (1979),Empirical Tests for the Existence of Group Utility Func-
tions, Review of Economic Studies, 46,653-665.
Shephard, R. W. (1970),Theory ofCost and Production Functions, Princeton University Press,
Prince ton.
Solow, R. (1964), Capital, Labor, and Income in Manufacturing, in The Behavior of Income
Shares, Studies in Income and Wealth, National Bureau of Economic Research 27,
Princeton University Press, Princeton, 101-128.
ANALYSIS
OF DEMAND
AND SUPPLY 235
Sonnenschein, H. (1973), Do Walras' Identity and Continuity Characterize the Class of Com-
munity Excess Demand Functions? Journal of Economic Theory, 6,345-354.
Sonnenschein, H. (1974), Market Excess Demand Functions, Econometrica, 40, 549-563.
Sono, M. (1%1),The Effect of Price Changes on the Demand and Supply of Separable Goods,
International Economic Review, 2,239-27 1.
Stoker, T. M. (1984), Completeness, Distribution Restrictions, and the Form of Aggregate
Functions, Econometrica, 52,887-907.
Stoker, T. M. (1986a),The Distributional Welfare Effects of Rising Prices in the United States:
The 1970's Experience, American Economic Review, 76,335-349.
Stoker, T. M. (1986b), Simple Tests of Distributional Effects on Macroeconomic Equations,
Journal of Political Economy, 94,763-795.
Stoker, T. M. (1993), Empirical Approaches to the Problem of Aggregation over Individuals,
Journal of Economic Literature, 31, 1827-1874.
Stone, R. (1954), Linear Expenditure Systems and Demand Analysis: An Application to the
Pattern of British Demand, The Economic Journal, 64,511-527.
Theil, H. (1954),Linear Aggregation ofEconomic Relations, North-Holland, Amsterdam.
Van Daal, J. and A. H. Q. M. Merkies (1984),Aggregation in Economic Research, D. Reidel,
Dordrecht .
Van Daal, J. and A. H. Q. M. Merkies (1989), A Note on the Quadratic Expenditure Model,
Varian, Hi. R. (1982), The Nonparametric Approach to Demand Analysis, Econometrica, 50,
945-973.
Varian, H. R. (1992),Microeconomic Analysis, W. W. Norton, New York.
Vilks, A. (1988a), Approximate Aggregation of Excess Demand Functions, Journal of Eco-
nomic Theory, 45,417-424.
Vilks, A. (1988b), Consistent Aggregation of a General Equilibrium Model, in Measurement
in Economics (Wolfgang Eichhorn, ed.), Physica-Verlag, Heidelberg, 691-703.
White, H. (1980), Using Least Squares to Approximate Unknown Regression Functions, In-
ternational Economic Review, 21, 149-170.
Spatial Dependence in Linear
Regression Models with an
Introduction to Spatial Econornetrics
Luc Anselin
West Virginia University, Morgantown, West Virginia
Anil K. Bera
University oflllinois, Charnpaign, lllinois
1. INTRODUCTION
Econometric theory and practice have been dominated by a focus on the time di-
mension. In stark contrast to the voluminous literature on serial dependence over
time (e.g., the extensive review in King 1987), there is scant attention paid to its
counterpart in cross-sectional data, spatial autocorrelation. For example, there is
no reference to the concept nor to its relevance in estimation or specification test-
ing in any of the commonly cited econometrics texts, such as Judge et al. (1982),
Greene (1993), or Poirier (1995), or even in more advanced ones, such as Fomby
et al. (1984), Amemiya (1985), Judge et al. (1995), and Davidson and MacKinnon
(1993) (a rare exception is Johnston 1984). In contrast, spatial autocorrelation and
spatial statistics in general are widely accepted as highly relevant in the analysis of
cross-sectional data in the physical sciences, such as in statistical mechanics, ecol-
ogy, forestry, geology, soil science, medical imaging, and epidemiology (for a recent
review, see National Research Council 1991).
In spite of this lack of recognition in “mainstream” econometrics, applied
workers saw the need to explicitly deal with problems caused by spatial autocorrela-
tion in cross-sectional data used in the implementation of regional and multiregional
econonietric models. In the early 1970s, the Belgian economist Jean Paelinck coined
the term “spatial econometrics” to designate a field of applied econometrics dealing
237
238 ANSELINAND BERA
with estimation and specification problems that arose from this. In their classic book
Spatial Econometrics, Paelinck and Klaassen (1979) outlined five characteristics of
the field: (1) the role of spatial interdependence in spatial models; (2) the asymmetry
in spatial relations; (3) the importance of explanatory factors located in other spaces;
(4) differentiation between ex post and ex ante interaction; and (5) explicit model-
ing of space (Paelinck and Klaassen 1979, pp. 5-1 1; see also Hordijk and Paelinck
1976, Paelinck 1982).In Anselin (1988a, p. 7), spatial econometrics is defined more
broadly as “the collection of techniques that deal with the peculiarities caused by
space in the statistical analysis of regional science models.” The latter incorporate
regions, location and spatial interaction explicitly and form the basis of most recent
empirical work in urban and regional economics, real estate economics, transporta-
tion economics, and economic geography. The emphasis on the model as the starting
point differentiates spatial econometrics from the broader field of spatial statistics,
although they share a common methodological framework. Much of the contributions
to spatial econometrics have appeared in specialized journals in regional science
and analytical geography, such as the Journal of Regional Science, Regional Science
and Urban Economics, Papers in Regional Science, International Regional Science
Review, Geographical Analysis, and Environment and Planning A. Early reviews of
the relevant methodological issues are given in Hordijk (1974, 1979), Bartels and
Hordijk (1977), Arora and Brown (1977), Paelinck and Klaassen (1979), Bartels and
Ketellapper (1979), Cliff and Ord (1981), Blommestein (1983), and Anselin (1980,
1988a, 1988b). More recent collections of papers dealing with spatial econometric
issues are contained in Anselin (1992a), Anselin and Florax (1995a), and Anselin
and Rey (1997).
Recently, an attention to the spatial econometric perspective has started to ap-
pear in mainstream empirical economics as well. This focus on spatial dependence
has occurred in a range of fields in economics, not only in urban, real estate, and
regional economics, where the importance of location and spatial interaction is fun-
damental, but also in public economics, agricultural and environmental economics,
and industrial organization. Recent examples of empirical studies in mainstream
economics that explicitly incorporated spatial dependence are, among others, the
analysis of U S . state expenditure patterns in Case et al. (1993), an examination of
recreation expenditures by municipalities in the Los Angeles region in Murdoch
et al. (1993), pricing in agricultural markets in LeSage (1993), potential spillovers
from public infrastructure investments in Holtz-Eakin (1994),the determination of
agricultural land values in Benirschka and Binkley (1994), the choice of retail sales
contracts by integrated oil companies in Pinkse and Slade (1995), strategic interac-
tion among local governments in Brueckner (1996), and models of nations’ decisions
to ratify environmental controls in Beron et al. (1996) and Murdoch et al. (1996).
Substantively, this follows from a renewed focus on Marshallian externalities, spa-
tial spillovers, copy-catting, and other forms of behavior where an economic actor
I N LINEARREGRESSION
SPATIAL DEPENDENCE MODELS 239
mimics or reacts to the actions of other actors, for example in the new economic
geography of Krugman (1991), in theories of endogenous growth (Romer 1986), and
in analyses of local political economy (Besley and Case 1995).Second, a number
of important policy issues have received an explicit spatial dimension, such as the
designation of target areas or enterprise zones in development economics and the
identification of underserved mortgage markets in urban areas. A more practical
reason is the increased availability of large socioeconomic data sets with detailed
spatial information, such as county-level economic information in the REIS CD-
ROM (Regional Economic Information System) of the U.S. Department of Commerce,
and tract -level data on mortgage transactions collected under the Housing Mortgage
Disclosure Act (HMDA) of 1975.
From a methodological viewpoint, spatial dependence is not only important
when it is part of the model, be it in a theoretical or policy framework, but it can
also arise due to certain misspecifications. For instance, often the cross-sectional
data used in model estimation and specification testing are imperfect, which may
cause spatial dependence as a side effect. For example, census tracts are not housing
markets and counties are not labor markets, but they are used as proxies to record
transactions in these markets. Specifically, a mismatch between the spatial unit of
observation and the spatial extent of the economic phenomena under consideration
will result in spatial measurement errors and spatial autocorrelation between these
errors in adjoining locations (Anselin 1988a).
In this chapter, we review the methodological issues related to the explicit
treatment of spatial dependence in linear regression models. Specifically, we focus
on the specification of the structure of spatial dependence (or spatial autocorrela-
tion), on the estimation of models with spatial dependence and on specification tests
to detect spatial dependence in regression models. Our review is organized accord-
ingly into three main sections. We have limited the review to cross-sectional settings
for linear regression models and do not consider dependence in space-time nor mod-
els for limited dependent variables. Whereas there is an established body of theory
and methodology to deal with the standard regression case, this is not (yet) the case
for techniques to analyze the other types of models. Both areas are currently the sub-
ject of active ongoing research (see, e.g., some of the papers in Anselin and Florax
1995a). Also, we have chosen to focus on a classical framework and do not consider
Bayesian approaches to spatial econometrics (e.g., Hepple 1995a, 1995b, LeSage
1997).
In our review, we attempt to outline the extent to which general econometric
principles can be applied to deal with spatial dependence. Spatial eeonometrics is
often erroneously considered to consist of a straightforward extension of techniques
to handle dependence in the time domain to two dimensions. In this chapter, we
emphasize the limitations of such a perspective and stress the need to explicitly
tackle the spatial aspects of model specification, estimation, and diagnostic testing.
240 ANSELINAND 6ERA
II. THE PROBLEM OF SPATIAL AUTOCORRELATION
We begin this review with a closer look at the concept of spatial dependence, or its
weaker expression, spatial autocorrelation, and how it differs from the more familiar
serial correlation in the time domain. While, in a strict sense, spatial autocorrela-
tion and spatial dependence clearly are not synonymous, we will use the terms inter-
changeably. In most applications, the weaker term autocorrelation (as a moment of
the joint distribution) is used and only seldom has the focus been on the joint density
as such (a recent exception is the semiparametric framework suggested in Brett and
Pinkse 1997).
In econometrics, an attention to serial correlation has been the domain of time-
series analysis and the typical focus of interest in the specification and estimation
of models for cross-sectional data is heteroskedasticity. Until recently, spatial auto-
correlation was largely ignored in this context, or treated in the form of groupwise
equicorrelation, e.g., as the result of certain survey designs (King and Evans 1986).
In other disciplines, primarily in physical sciences, such as geology (Isaaks and
Srivastava 1989, Cressie 1991) and ecology (Legendre 1993), but also in geogra-
phy (Griffith 1987, Haining 1990) and in social network analysis in sociology and
psychology (Dow et al. 1982, Doreian et al. 1984, Leenders 1995),the dependence
across “space” (in its most general sense) has been much more central. For example,
Tobler’s (1979) “first law of geography” states that “everything is related to every-
thing else, but closer things more so,” suggesting spatial dependence to be the rule
rather than exception. A large body of spatial statistical techniques has been devel-
oped to deal with such dependencies (for a recent comprehensive review, see Cressie
1993;other classic references are Cliff and Ord 1973,1981, Ripley 1981,1988,Up-
ton and Fingleton 1985, 1989). Useful in this respect is Cressie’s (1993)taxonomy
of spatial data strucures differentiating between point patterns, geostatistical data,
and lattice data. In the physical sciences, the dominant underlying assumption tends
to be that of a continuous spatial surface, necessitating the so-called geostatistical
perspective rather than discrete observation points (or regions) in space, for which
the so-called lattice perspective is relevant. The latter is more appropriate for eco-
nomic data, since it is to some extent an extension of the ordering of observations on
a one-dimensional time axis to an ordering in a two-dimensional space. It will be the
almost exclusive focus of our review.
The traditional emphasis in econometrics on heterogeneity in cross-sectional
data is not necessarily misplaced, since the distinction between spatial heterogene-
ity and spatial autocorrelation is not always obvious. More specifically, in a single
cross section the two may be observationally equivalent. For example, when a spatial
cluster of exceptionally large residuals is observed for a regression model, it cannot
be ascertained without further structure whether this is an instance of heteroskedas-
ticity (i.e., clustering of outliers) or spatial autocorrelation (a spatial stochastic pro-
cess yielding clustered outliers). This problem is known in the literature as “true”
SPATIAL DEPENDENCE MODELS 24 I
contagion versus “apparent” contagion and is a major methodological issue in fields

such as epidemiology (see, e.g., Johnson and Kotz 1969, Chapter 9, for a formal dis-
tinction between different forms of contagious distributions). The approach taken in
spatial econometrics is to impose structure on the problem through the specifica-
tion of a model, coupled with extensive specification testing for potential departures
from the null model. This emphasis on the “model” distinguishes (albeit rather sub-
tly) spatial econometrics from the broader field of spatial statistics (see also Anselin
1988a, p. 10, for further discussion of the distinction between the two). In our re-
view, we deal almost exclusively with spatial autocorrelation. Once this aspect of
the model is specified, the heterogeneity may be added in a standard manner (see
Anselin 1988a, Chap. 9, and Anselin 1990a).
In this section, we first focus on a formal definition of spatial autocorrelation.
This is followed by a consideration of how it may be operationalized in tests and
econometric specifications by means of spatial weights and spatial lag operators.
We close with a review of different ways in which spatial autocorrelation may be
incorporated in the specification of econometric models in the form of spatial lag
dependence, spatial error dependence, or higher-order spatial processes.
A. Defining Spatial Autocorrelation
Spatial autocorrelation can be loosely defined as the coincidence of value similarity

with locational similarity. In other words, high or low values for a random variable
tend to cluster in space (positive spatial autocorrelation), or locations tend to be sur-
rounded by neighbors with very dissimilar values (negative spatial autocorrelation).
Of the two types of spatial autocorrelation, positive autocorrelation is by far the more
intuitive. Negative spatial autocorrelation implies a checkerboard pattern of values
and does not always have a meaningful substantive interpretation (for a formal dis-
cussion, see Whittle 1954).The existence of positive spatial autocorrelation implies
that a sample contains less information than an uncorrelated counterpart. In order
to properly carry out statistical inference, this loss of information must be explicitly
acknowledged in estimation and diagnostics tests. This is the essence of the problem
of spatial autocorrelation in applied econometrics.
A crucial issue in the definition of spatial autocorrelation is the notion of “lo-
cational similarity,” or the determination of those locations for which the values of
the random variable are correlated. Such locations are referred to as “neighbors,”
though strictly speaking this does not necessarily mean that they need to be collo-
cated (for a more formal definition of neighbors in terms of the conditional density
function, see Anselin 1988a, pp. 16-17; Cressie 1993, p. 414).
More formally, the existence of spatial autocorrelation may be expressed by
the following moment condition:
242 ANSELIN
AND BERA
where yi and yj are observations on a random variable at locations i and j in space,

and i, j can be points (e.g., locations of stores, metropolitan areas, measured as lati-
tude and longitude) or area1 units (e.g., states, counties or census tracts). Of course,
there is nothing spatial per se to the nonzero covariance in (1). It only becomes spatial
when the pairs of i, j locations for which the correlation is nonzero have a meaningful
interpretation in terms of spatial structure, spatial interaction or spatial arrangement
of observations.
For a set of N observations on cross-sectional data, it is impossible to es-
timate the potentially N by N covariance terms or correlations directly from the
data. This is a fundamental problem in dealing with spatial autocorrelation and ne-
cessitates the imposition of structure. More specifically, in order for the problem
to become tractable, it is necessary to impose sufficient constraints on the N by
N spatial interaction (covariance) matrix such that a finite number of parameters
characterizing the correlation can be efficiently estimated. Note how this contrasts
with the situation where repeated observations are available, e.g., in panel data
sets. In such instances, under the proper conditions, the elements of the covari-
ance matrix may be estimated explicitly, in a vector autoregressive approach (for a
review, see Liitkepohl 1991)or by means of so-called generalized estimating equa-
tions (Liang and Zeger 1986, Zeger and Liang 1986, Zeger et al. 1988, Albert and
McShane 1995).
In contrast, when the N observations are considered as fixed effects in space,
there is insufficient information in the data to estimate the N by N interactions. In-
creasing the sample size does not help, since the number of interactions increases
with N 2 , or, in other words, there is an incidental parameter problem. Alternatively,
when the locations are conceptualized in a random-effects framework, sufficient con-
straints must be imposed to preclude that the range of interaction implied by a par-
ticular spatial stochastic process increases faster than the sample size as asymptotics
are invoked to obtain the properties of estimators and test statistics.
Two main approaches exist in the literature to impose constraints on the in-
teraction. In geostatistics, all pairs of locations are sorted according to the distance
that separates them, and the strength of covariance (correlation) between them is ex-
pressed as a continuous function of this distance, in a so-called variogram or semi-
variogram (Cressie 1993, Chap. 2). A s pointed out, the geostatistical perspective is
seldom taken in empirical economics, since it necessitates an underlying process
that is continuous over space. In such an approach, observations (points) are consid-
ered to form a sample from an underlying continuous spatial process, which is hard
to maintain when the data consist of counties or census tracts. A possible excep-
tion may be the study of real estate data, where the locations of transactions may be
conceptualized as points and analyzed using a geostatistical framework, as in Dubin
(1988, 1992). Such an approach is termed “direct representation” in the literature,
since the elements of the covariance (or correlation) matrix are modeled directly as
functions of distances.
SPATIAL I N LINEARREGRESSION
DEPENDENCE MODELS 243
Our main focus in this review will be on the second approach, the so-called
lattice perspective. For each data point, a relevant “neighborhood set” must be de-
fined, consisting of those other locations that (potentially) interact with it. For each
observation i, this yields a spatial ordering of locations j E S; (where S; is the neigh-
borhood set), which can then be exploited to specify a spatial stochastic process. The
covariance structure between observations is thus not modeled directly, but follows
from the particular form of the stochastic process. We return to this issue below. First,
we review the operational specification of the neighborhood set for each observation
by means of a so-called spatial weights matrix.
B. Spatial Weights
A spatial weights matrix is a N by N positive and symmetric matrix W which ex-
presses for each observation (row) those locations (columns) that belong to its neigh-
borhood set as nonzero elements. More formally, wlJ = 1 when i and j are neighbors,
and wlJ = 0 otherwise. By convention, the diagonal elements of the weights matrix
are set to zero. For ease of interpretation, the weights matrix is often standardized
such that the elements of a row sum to one. The elements of a row-standardized
weights matrix thus equal ws = w l J / c / w,,. This ensures that all weights are be-
‘J
tween 0 and 1 and facilitates the interpretation of operations with the weights matrix
as an averaging of neighboring values (see Section 1I.C). It also ensures that the spa-
tial parameters in many spatial stochastic processes are comparable between mod-
els. This is not intuitively obvious, but relates to constraints imposed in a maximum
likelihood estimation framework. For the latter to be valid, spatial autoregressive pa-
rameters must be constrained to lie in the interval 1/mmln to l/mm,,, where mm,n and
U,,, are respectively the smallest (on the real line) and largest eigenvalues of the ma-
trix W (Anselin 1982). For a row-standardized weights matrix, the largest eigenvalue
is always +1 (Ord 1975), which facilitates the interpretation of the autoregressive
coefficient as a “correlation” (for an alternative view, see Kelejian and Robinson
1995). A side effect of row standardization is that the resulting matrix is likely to
become asymmetric (since E, w,, # Cl w,,), even though the original matrix may
have been symmetric. In the calculation of several estimators and test statistics, this
complicates computational matters considerably.
The specification of which elements are nonzero in the spatial weights matrix
is a matter of considerable arbitrariness and a wide range of suggestions have been
offered in the literature. The “traditional” approach relies on the geography or spa-
tial arrangement of the observations, designating area1 units as “neighbors” when
they have a border in common (first-order contiguity) or are within a given distance
of each other; i.e., wlJ = 1 for d, 5 6, where d, is the distance between units i
and j , and 6 is a distance cutoff value (distance-based contiguity). This geographic
approach has been generalized to so-called Cliff-Ord weights that consist of a func-
tion of the relative length of the common border, adjusted by the inverse distance
244 ANSELIN
AND BERA
between two observations (Cliff and Ord 1973, 1981). Formally, Cliff-Ord weights
may be expressed as:
where bi; is the share of the common border between units i and j in the perimeter of i
(and hence bi; does not necessarily equal bji), and a! and are parameters. More gen-
erally, the weights may be specified to express any measure of “potential interaction”
between units i and j (Anselin 1988a, Chap. 3). For example, this may be related
directly to spatial interaction theory and the notion of potential, with wi, = l/d: or
wij = e - p d i j , or more complex distance metrics may be implemented (Anselin 1980,
Murdoch et al. 1993). Typically, the parameters of the distance function are set a
priori (e.g., a = 2, to reflect a gravity function) and not estimated jointly with the
other coefficients in the model. Clearly, when they are estimated jointly, the resulting
specification will be highly nonlinear (Anselin 1980, Chap. 8, Ancot et al. 1986,
Bolduc et al. 1989, 1992, 1995).
Other specifications of spatial weights are possible as well. In sociometrics,
the weights reflect whether or not two individuals belong to the same social network
(Doreian 1980).In economic applications, the use of weights based on “economic”
distance has been suggested, among others, in Case et al. (1993). Specifically, they
suggest to use weights (before row standardization) of the form wij = l/lxi - xi),
where xi and x; are observations on “meaningful” socioeconomic characteristics,
such as per capita income or percentage of the population in a given racial or ethnic
group.
It is important to keep in mind that, irrespective of how the spatial weights
are specified, the resulting spatial process must satisfy the necessary regularity con-
ditions such that asymptotics may be invoked to obtain the properties of estimators
and test statistics. For example, this requires constraints on the extent of the range of
interaction and/or the degree of heterogeneity implied by the weights matrices (the
so-called mixing conditions; Anselin 1988a, Chap. 5). Specifically, this means that
weights must be nonnegative and remain finite, and that they correspond to a proper
metric (Anselin 1980).Clearly, this may pose a problem with socioeconomic weights
when xi = x; for some observation pairs, which may be the case for poorly chosen
economic determinants (e.g., when two states have the same percentage in a given
racial group). Similarly, when multiple observations belong to the same area1 unit
(e.g., different banks located in the same county) the distance between them must
be set to something other than zero (or l / d ~+ 00). Finally, in the standard estima-
tion and testing approaches, the weights matrix is taken to be exogenous. Therefore,
indicators for the socioeconomic weights should be chosen with great care to en-
sure their exogeneity, unless their endogeneity is considered explicitly in the model
specification.
Operationally, the derivation of spatial weights from the location and spatial
arrangement of observations must be carried out by means of a geographic informa-
tion system, since for all but the smallest data sets a visual inspection of a map is im-
practical (for implementation details, see Anselin et al. 1993a, 1993b, Anselin 1995,
Can 1996).A mechanical construction of spatial weights, particularly when based
on a distance criterion, may easily result in observations to become “unconnected”
or isolated islands. Consequently, the row in the weights matrix that corresponds to
these observations will consist of zero values. While not inherently invalidating es-
timation or testing procedures, the unconnected observations imply a loss of degrees
of freedom, since, for all practical purposes, they are eliminated from consideration
in any “spatial” model. This must be explicitly accounted for.
C. Spatial Lag Operator

In time-series analysis, values for “neighboring” observations can be easily ex-
pressed by means of a backward- or forward-shift operator on the one-dimensional
time axis, yielding lagged variables yt-k or yt+k, where k is the desired shift (or lag).
By contrast, there is no equivalent and unambiguous spatial shift operator. Only on
a regular grid structure is there a potential solution, although not as straightforward
as in the time domain. Following the so-called rook criterion for contiguity, each grid
cell or vertex on a regular lattice, ( i , j ) , has four neighbors: (i+ 1, j ) (east), ( i - 1, j )
+
(west), (i, j 1) (north), and ( i , j - 1) (south). Corresponding to this are four spa-
tially shifted variables: yi+1,, , y ~ -1, j , y~,;+l,and yi,;- 1, each of which may require
its own parameter in a spatial process model. However, the rook criterion is not the
only way spatial neighbors may be defined on a regular lattice, nor does the number
of neighbors necessarily equal 4. For example, following the queen criterion, each
observation has eight neighbors, yielding eight spatially shifted variables; the four
for the rook criterion, as well as y~-1,;+1, yi-l,;-l, yi+l,j+l and yi+l,;-i, again each
possibly with its own parameter. This notion of a spatial shift operator on a regular
lattice has received only limited attention in the literature, mostly with a theoretical
focus and primarily in statistical mechanics, in so-called Ising models (for details,
see Cressie 1993, pp. 425-426).
On an irregular spatial structure, which characterizes most economic appli-
cations, this formal notion of spatial shift is impractical, since the number of shifts
would differ by observation, thereby making any statistical analysis extremely un-
wieldy. Instead, the concept of a spatial lug operator is used, which consists of a
weighted average of the values at neighboring locations. The weights are fixed and
exogenous, similar to a distributed lag in time series. Formally, a spatial lag operator
is obtained as the product of a spatial weights matrix W with the vector of observa-
tions on a random variable y, or Wy. Each element of the resulting spatially lagged
variable equals c/ wijyj, i.e., a weighted average of the y values in the neighbor set
Si,since wij = 0 for j 4 Si. Row standardization of the spatial weights matrix en-
246 ANSELINAND BERA
sures that a spatial lag operation yields a smoothing of the neighboring values, since
the positive weights sum to one.
Higher-order spatial lag operators are defined in a recursive manner, by ap-
plying the spatial weights matrix to a lower-order lagged variable. For example, a
second-order spatial lag is obtained as W ( W y ) , or W2y. However, in contrast to
time series, where such an operation is unambiguous, higher-order spatial operators
yield redundant and circular neighbor relations, which must be eliminated to ensure
proper estimation and inference (Blommestein 1985, Blommestein and Koper 1992,
Anselin and Smirnov 1996).
In spatial econometrics, spatial autocorrelation is modeled by means of a func-
tional relationship between a variable. y, or error term, E , and its associated spatial
lag, respectively Wy for a spatially lagged dependent variable and WE for a spatially
lagged error term. The resulting specifications are referred to as spatial lug and spa-
tial error models, the general properties of which we consider next.
D. Spatial Lag Dependence

Spatial lag dependence in a regression model is similar to the inclusion of a serially
autoregressive term for the dependent variable ( ~ ~ - 1 in) a time-series context. In
spatial econometrics, this is referred to as a mixed regressive, spatial autoregressive
model (Anselin 1988a, p. 35). Formally,
where y is a N by 1 vector of observations on the dependent variable, Wy is the

corresponding spatially lagged dependent variable for weights matrix W , X is a N
by K matrix of observations on the explanatory (exogenous) variables, E is a N by
1 vector of error terms, p is the spatial autoregressive parameter, and B is a K by
1 vector of regression coefficients. The presence of the spatial lag term Wy on the
right side of (3)will induce a nonzero correlation with the error term, similar to the
presence of an endogenous variable, but different from a serially lagged dependent
variable in the time-series case. In the latter model, yt-l is uncorrelated with E ~ in ,
the absence of serial correlation in the errors. In contrast, (Wy); is always correlated
with E L , irrespective of the correlation structure of the errors. Moreover, the spatial
lag for a given observation i is not only correlated with the error term at i , but also with
the error terms at all other locations. Therefore, unlike what holds in the time-series
case, an ordinary least-squares estimator will not be consistent for this specification
(Anselin 1988a, Chap. 6). This can be seen from a slight reformulation of the model:
y = (I - p W ) - ' X p + (I .- pW)-'& (4)

The matrix inverse (I - p W ) - ' is a full matrix, and not triangular as in the time-
series case (where dependence is only one-directional), yielding an infinite series
that involves error terms at all locations, ( I + p W + p 2 W 2 + p 3 W w ” + - . .)E. It therefore

readily follows that (Wy)i contains the element ~ i as , well as other ~ j j , # i. Thus,
E[(Wy);E;] = E[{W(I - pW)-lE}iEi] #0

The spatial dynamics embedded in the structure of the spatial process model (3)
determine the form of the covariance between the observations at different locations
(i.e., the spatial autocorrelation). For the mixed regressive, spatial autoregressive
model this can easily be seen to equal ( I - pW)-’SZ(I - p W ’ ) - ’ , where SZ is the
variance matrix for the error term E (note that for a row-standardized spatial weights
matrix, W # W’). Without loss of generality, the latter can be assumed to be diagonal
and homoskedastic, or, $2 = a21,and hence, Var[y] = a‘([ - p W ) - ’ ( I - pW’)-’
The resulting variance matrix is full, implying that each location is correlated with
every other location, but in a fashion that decays with the order of contiguity (the
powers of W in the series expansion of (I - p W ) - ’ ) .
The implication of this particular variance structure is that the simultaneity
embedded in the Wy term must be explicitly accounted for, either in a maximum
likelihood estimation framework, or by using a proper set of instrumental variables.
We turn to this issue in Section 111. When a spatially lagged dependent variable
is ignored in a model specification, but present in the underlying data generating
process, the resulting specification error is of the omitted variable type. This implies
that OLS estimates in the nonspatial model (i.e., the “standard” approach) will be
biased and inconsistent.
The interpretation of a significant spatial autoregressive coefficient p is not
always straightforward. Two situations can be distinguished. In one, the significant
spatial lag term indicates true contagion or substantive spatial dependence, i.e., it
measures the extent of spatial spillovers, copy-catting or diffusion. This interpreta-
tion is valid when the actors under consideration match the spatial unit of observation
and the spillover is the result of a theoretical model. For example, this holds for the
models of farmers’ innovation adoption in Case (1992), state expenditures and tax
setting behavior in Case et al. (1993)and Besley and Case (1995),strategic inter-
action among California cities in the choice of growth controls in Brueckner (1996),
and in the median voter model for recreation expenditures of Murdoch et al. (1993).
Alternatively, the spatial lag model may be used to deal with spatial autocorrelation
that results from a mismatch between the spatial scale of the phenomenon under
study and the spatial scale at which it is measured. Clearly, when data are based
on administratively determined units such as census tracts or blocks, there is no
good reason to expect economic behavior to conform to these units. For example,
this interpretation is useful for the spatial autoregressive models of urban housing
and mortgage markets in Can (1992), Can and Megbolugbe (1997), and Anselin and
Can (1996).Since urban housing and mortgage markets operate at a different spatial
scale than census tracts, positive spatial autocorrelation may be expected and will
in fact result in the sample containing less information than a truly “independent”
248 ANSELINAND BERA
sample of observations. The inclusion of a spatially lagged dependent variable in the

model specification is a way to correct for this loss of information. In other words,
it allows for the proper interpretation of the significance of the exogenous variables
in the model (the X ) , after the spatial effects have been corrected for, or filtered out
(see also Getis 1995 for a discussion of alternative approaches to spatial filtering).
More formally, the spatial lag model may be reexpressed as
(I -pW)y = X p + E (6)
where (I - p W ) y is a spatially filtered dependent variable, i.e., with the effect of
spatial autocorrelation taken out. This is roughly similar to first differencing of the
dependent variable in time series, except that a value of p = 1is not in the allowable
parameter space for (3) and thus p must be estimated explicitly (Section 111).
E. Spatial Error Dependence
A second way to incorporate spatial autocorrelation in a regression model is to spec-

ify a spatial process for the disturbance terms. The resulting error covariance will be
nonspherical, and thus OLS estimates, while still unbiased, will be inefficient. More
efficient estimators are obtained by taking advantage of the particular structure of
the error covariance implied by the spatial process. Different spatial processes lead
to different error covariances, with varying implications about the range and extent
of spatial interaction in the model. The most common specification is a spatial au-
toregressive process in the error terms:
y=xp+E (7)
i.e., a linear regression with error vector E , and
where h is the spatial autoregressive coefficient for the error lag WE (to distinguish
the notation from the spatial autoregressive coefficient p in a spatial lag model),
and 6 is an uncorrelated and (without loss of generality) homoskedastic error term.
Alternatively, this may be expressed as
y = X/3 3- ( I - hW)-'C$ (9)

From this follows the error covariance as
E[&&']= a2(I- hW)-'(I - hW')-' = a2[(I- hW)'(I - hW)]-l (10)

a structure identical to that for the dependent variable in the spatial lag model.
Therefore, a spatial autoregressive error process leads to a nonzero error covari-
ance between every pair of observations, but decreasing in magnitude with the or-
der of contiguity. Moreover, the complex structure in the inverse matrices in (10)
SPATIAL DEPENDENCE
IN LINEARREGRESSIONMODELS 249
yields nonconstant diagonal elements in the error covariance matrix, thus inducing
heteroskedasticity in E , irrespective of the heteroskedasticity of (an illuminating
numerical illustration of this feature is given in McMillen 1992). We have a much
simpler situation for the case of autocorrelation in the time-series context where the
model is written as ct = h t - l + el. Therefore, this is a special case of (8)with
w=wT=j
0 0 0 ~
III
~
o1 10 0 0 ~ ~0 01
*
~
0 0
0 0
0 0 0 * * 1 0
where each observation is connected to only its immediate past value. As we know,
for this case, Var(&,) = a2/(1 - h2)for all t. That is, autocorrelation does not induce
heteroskedasticity. In a time-series model, heteroskedasticity can come only through
t l given the above AR(1) model.
A second complicating factor in specification testing is the great degree of
similarity between a spatial lag and a spatial error model, as suggested by the error
covariance structure. In fact, after premultiplying both sides of (9) by (I - hW)and
moving the spatial lag term to the right side, a spatial Durbin model results (Anselin
1980):
y = hWy + xg - hwxg + $I (1 1)
This model has a spatial lag structure (but with the spatial autoregressive parameter
h from (8))with a well-behaved error term 6. However, the equivalence between (7)-
(8)and (1 1) imposes a set of nonlinear common factor constraints on the coefficients.
Indeed, for (11) to be a proper spatial error model, the coefficients of the lagged ex-
ogenous variables WX must equal minus the product of the spatial autoregressive
coefficient h and the coefficients of X , for a total of K constraints (for technical de-
tails, see Anselin 1988a, pp. 226-229).
Spatial error dependence may be interpreted as a nuisance (and the parameter
h as a nuisance parameter) in the sense that it reflects spatial autocorrelation in
measurement errors or in variables that are otherwise not crucial to the model (i.e.,
the “ignored” variables spillover across the spatial units of observation). It primarily
causes a problem of inefficiency in the regression estimates, which may be remedied
by increasing the sample size or by exploiting consistent estimates of the nuisance
parameter. For example, this is the interpretation offered in the model of agricultural
land values in Benirschka and Binkley (1994).
The spatial autoregressive error model can also be expressed in terms of spa-
tially filtered variables, but slightly different from (6). After moving the spatial lag
variable in (11) to the left hand side, the following expression results:
(I - h W ) y = (I - hW)XB +6 (12)
250 ANSELINAND BERA
This is a regression model with spatially filtered dependent and explanatory vari-
ables and with an uncorrelated error term 6, similar to first differencing of both y
and X in time-series models. As in the spatial lag model, h = 1 is outside the pa-
rameter space and thus h must be estimated jointly with the other coefficients of the
model (see Section 111).
Several alternatives to the spatial autoregressive error process (8) have been
suggested in the literature, though none of them have been implemented much in
practice. A spatial moving average error process is specified as (Cliff and Ord 1981,
Haining 1988, 1990):
E=yWt+C (13)
e
where y is the spatial moving average coefficient and is an uncorrelated error term.
This process thus specifies the error term at each location to consist of a location-
specific part, Ci (“innovation”), as well as a weighted average (smoothing) of the
errors at neighboring locations, W t . The resulting error covariance matrix is
E[&&’]= 0 2 ( I + )/W)(I + YW’) = o‘[I+ )/(W + W’) + y 2 W W ’ ] (14)
Note that in contrast to (lO), the structure in (14)does not yield a full covariance ma-
+
trix. Nonzero covariances are only found for first-order ( W W’) and second-order
(WW’) neighbors, thus implying much less overall interaction than the autoregres-
sive process. Again, unless all observations have the same number of neighbors and
identical weights, the diagonal elements of (14)will not be constant, inducing het-
eroskedasticity in &, irrespective of the nature of 6.
A very similar structure to (13)is the spatial error components model of Kele-
jian and Robinson (1993, 1995),in which the disturbance is a sum of two indepen-
dent error terms, one associated with the “region” (a smoothing of neighboring errors)
and one which is location-specific:
&=Wti-@ (15)
with 6 and @ as independent error components. The resulting error covariance is
E [ € € / ]= a;r + a;WW’ (16)
where a$ and o2 6 are the variance components associated with respectively the
location-specific and regional error parts. The spatial interaction implied by (16)
is even more limited than for (14), pertaining only to the first- and second-order
neighbors contained in the nonzero elements of WW’. Heteroskedasticity is implied
unless all locations have the same number of neighbors and identical weights, a sit-
uation excluded by the assumptions needed for the proper asymptotics in the model
(Kelejian and Robinson 1993, p. 301).
In sum, every type of spatially dependent error process induces heteroskedas-
ticity as well as spatially autocorrelated errors, which will greatly complicate spec-
ification testing in practice. Note that the “direct representation” approach based
SPATIAL DEPENDENCE
MODELS 25 I
on geostatistical principles does not suffer from this problem. For example, in Du-
bin (1988, 1992), the elements of the error covariance matrix are expressed di-
rectly as functions of the distance dij between the corresponding observations, e.g.,
E[E;E= ~ ] y ~ e ( - ~ i j with ) , and y2 as parameters. Since e-dii’Y2 = 1, irrespec-
/ ~ ~y1
tive of the value of y2, the errors E will be homoskedastic unless explicitly modeled
otherwise.
F. Higher-Order Spatial Processes

Several authors have suggested processes that combine spatial lag with spatial error
dependence, though such specifications have seen only limited applications. The
most general form is the spatial autoregressive, moving-average (SARMA) process
outlined by Huang (1984). Formally, a SARMA(p, q ) process can be expressed as
Y = Pl WlY + p2w2y + + p p & y +

* . * & (17)
for the spatial autoregressive part, and
for the moving-average part, in the same notation as above. For greater generality, a
regressive component X/3 can be added to (17) as well. The spatial autocorrelation
pattern resulting from this general formulation is highly complex. Models that imple-
ment aspects of this form are the second-order SAR specification in Brandsma and
Ketellapper (1979a) and higher-order SAR models in Blommestein (1983, 1985).
A slightly different specification combines a first-order spatial autoregressive
lag with a first-order spatial autoregressive error (Anselin 1980, Chap. 6; Anselin
1988a. pp. 60-65). It has been applied in a number of empirical studies, most no-
tably i n the work of Case, such as the analysis of household demand (Case 1987,
1991), of innovation diffusion (Case 1992), and local public finance (Case et al.
1993, Besley and Case 1995). Formally, the model can be expressed as a combi-
nation of (3) with (8),although care must be taken to differentiate the weights matrix
used in the spatial lag process from that in the spatial error process:
After some algebra, combining (20) and (19)yields the following reduced form:
y =pwly+~w2y-hpw2wly+X/3-hw~Xq+~ (21)
i.e., an extended form of the spatial Durbin specification but with an additional set of
nonlinear constraints on the parameters. Note that when Wl and W2 do not overlap, for
example when they pertain to different orders of contiguity, the product W2Wl = 0
252 ANSELINAND BERA
and (21) reduces to a biparametric spatial lag formulation, albeit with additional
constraints on the parameters. On the other hand, when Wl and W2 are the same,
the parameters p and h are only identified when at least one exogenous variable is
included in X (in addition to the constant term) and when the nonlinear constraints
are enforced (Anselin 1980, p. 176). When Wl = W2 = W , the model becomes
y =(p + h)Wy - hpW2y + xg - hwxg + 6 (22)

Clearly, the coefficients of W y and W2y alone do not allow for a separate identi-
fication of p and A. Using the nonlinear constraints between the B and -hp (the
coefficients of X and W X )yields an estimate of h, but this will only be unique when
the constraints are strictly enforced. Similarly, an estimate of h may result in two
possible estimates for p (one using the coefficient of W y , the other of W 2 y )unless
the nonlinear constraints are strictly enforced. This considerably complicates esti-
mation strategies for this model. In contrast, a SARMA(1, 1) model does not suffer
from this problem.
In empirical practice, an alternative perspective on the need for higher-order
processes is to consider them to be a result of a poorly specified weights matrix rather
than as a realistic data generating process. For example, if the weights matrix in a
spatial lag model underbounds the true spatial interaction in the data, there will be
remaining spatial error autocorrelation. This may lead one to implement a higher-
order process, while for a properly specified weights matrix no such process is needed
(see Florax and Rey 1995 for a discussion of the effects of misspecified weights). In
practice, this will require a careful specification search for the proper form of the
spatial dependence in the model, an issue to which we return in Section IV. First, we
consider the estimation of regression models that incorporate spatial autocorrelation
of a spatial lag or error form.
111. ESTIMATING SPATIAL PROCESS MODELS
Similar to when serial dependence is present in the time domain, classical sam-
pling theory no longer holds for spatially autocorrelated data, and estimation and
inference must rely on the asymptotic properties of stochastic processes. In essence,
rather than considering N observations as independent pieces of information, they
are conceptualized as a single realization of a process. In order to carry out mean-
ingful inference on the parameters of such a process, constraints must be imposed
on both heterogeneity and the range of interaction. While many properties of esti-
mators for spatial process models may be based on the same principles as developed
for dependent (and heterogeneous) processes in the time domain (e.g., the formal
properties outlined in White 19%, 1994), there are some important differences as
well. Before covering specific estimation procedures, we discuss these differences
in some detail, focusing in particular on the notion of stationarity in space and the
SPATIAL DEPENDENCE
IN LINEARREGRESSION
MODELS 253
distinction between simultaneous and conditional spatial processes. Next, we turn

to a review of maximum likelihood and instrumental variables estimators for spatial
regression models. We close with a brief discussion of operational implementation
and software issues.
A. Spatial Stochastic Processes

As in the time domain, in order to carry out meaningful inference for a spatial pro-
cess, some degree of equilibrium must be assumed in the sense that the stochastic
generating mechanism is taken to work uniformly over space. In a strict sense, a
notion of “spatial stationarity” accomplishes this objective since it imposes the con-
dition that any joint distribution of the random variable under consideration over a
subset of the locations depends only on the relative position of these observations in
terms of their relative orientation (angle) and distance. Even stricter is a notion of
isotropy, for which only distance matters and orientation is irrelevant. For practical
purposes, the notions of stationarity and isotropy are too demanding and not veri-
fiable. Hence, weaker conditions are typically imposed in the form of stationarity
of the first (mean) and second moments (variance, covariance, or spatial autocor-
relation). Even weaker requirements follow from the so-called intrinsic hypothesis
in geostatistics, which requires only stationarity of the variance of the increments,
leading to the notion of a variogram (for technical details, see Ripley 1988, pp. 6-7;
Cressie 1993, pp. 52-68).
For stationary processes in the time domain, the careful inspection of auto-
covariance and autocorrelation functions is a powerful aid in the identification of
the model, e.g., following the familiar Box-Jenkins approach (Box et al. 1994). One
could transpose this notion to spatial processes and consider spatial autocorrela-
tion functions indexed by order of contiguity as the basis for model identification.
However, as Hooper and Hewings (1981)have shown, this is only appropriate for a
very restrictive class of spatial processes on regular lattice structures. For applied
work in empirical economics, such restrictions are impractical and the spatial de-
pendence in the model must be specified explicitly by means of the spatial lag and
spatial error structures reviewed in the previous section. Inference may be based
on the asymptotic properties (central limit theorems and laws of large numbers) of
so-called dependent and heterogeneous processes, as developed in White and Do-
mowitz (1984) and White (1984, 1994). Central to these notions is the concept of
mixing sequences, allowing for a trade-off between the range of dependence and the
extent of heterogeneity (see Anselin 1988a, pp. 45-46 for an intuitive extension of
this to spatial econometric models), While rigorous proofs of these properties have
not been derived for the explicit spatial case, the notion of a spatial weights matrix
based on a proper metric is general enough to meet the criteria imposed by mixing
conditions. In a spatial econometric approach then, a spatial lag model is considered
to be a special case of simultaneity or endogeneity with dependence, and a spatial
254 AND BERA
ANSELIN
error model is a special case of a nonspherical error term, both of which can be
tackled by means of generally established econometric theory, though not as direct
extensions of the time-series analog.
The emphasis on “simultaneity” in spatial econometrics differs somewhat from
the approach taken in spatial statistics, where conditional models are often consid-
ered to be more natural (Cressie 1993, p. 410). Again, the spatial case differs sub-
stantially from the time-series one since in space a conditional and simultaneous
approach are no longer equivalent (Brook 1964, Besag 1974, Cressie 1993, pp. 402-
410). More specifically, in the time domain a Markov chain stochastic process can be
expressed in terms of the joint density (ignoring a starting point to ease notation) as
where z refers to the vector of observations for all time points, and Qt is a function
that only contains the observation at t and at t - 1 (hence, a Markov chain). The
conditional density for this process is
illustrating the lack of memory of the process (i.e., the conditional density depends
only on the first-order lag). Due to the one-directional nature of dependence in time,
(23)and (24)are equivalent (Cressie 1993, p. 403).An extension of (23)to the spatial
domain may be formulated as
Prob[t] = n
i= 1
N
Q;[z;,z;; j E S;] (25)
where the z; only refer to those locations that are part of the neighborhood set S; of
i. A conditional specification would be
Prob[z;(z;, j # i] = Prob[zi(zj;j E S ; ] (26)
i.e., the conditional density of z;, given observations at all other locations only de-
pends on those locations in the neighborhood set of i. The fundamental result in this
respect goes back to Besag (1974),who showed that the conditional specification
only yields a proper joint distribution when the so-called Hammersley-Clifford the-
orem is satisfied, which imposes constraints on the type and range of dependencies
in (26). Also, while a joint density specification always yields a proper conditional
specification, in range of spatial interaction implied is not necessarily the same.
For example, Cressie (1993, p. 409) illustrates how a first-order symmetric spatial
autoregressive process corresponds with a conditional specification that includes
third-order neighbors (Haining 1990, pp. 89-90). Consequently, it does make a dif-
ference whether one approaches a spatially autocorrelated phenomenon by means of
SPATIAL DEPENDENCE
IN LINEAR REGRESSIONMODELS 255
(26) versus (25). This also has implications for the substantive interpretation of the
model results, as illustrated for an analysis of retail pricing of gasoline in Haining
(1984).
In practice, it is often easier to estimate a conditional model, especially for
nonnorrnal distributions (e.g., auto-Poisson, autologistic). Also, a conditional speci-
fication is more appropriate when the focus is on spatial prediction or interpolation.
For general estimation and inference, however, the constraints imposed on the type
and range of spatial interaction in order for the conditional density to be proper
are often highly impractical in empirical work. For example, an auto-Poisson model
(conditional model for spatially autocorrelated counts) only allows negative autocor-
relation and hence is inappropriate for any analysis of clustering in space.
In the remainder, our focus will be exclusively on simultaneously specified
models, which is a more natural approach from a spatial econometric perspective
(Anselin 1988a, Cressie 1993, p. 410).
B. Maximum Likelihood Estimation

The first comprehensive treatment of maximum likelihood estimation of regression
models that incorporate spatial autocorrelation in the form of a spatial lag or a spatial
error term was given by Ord (1975). The point of departure is a joint normal density
for the errors in the model, from which the likelihood function is derived. An impor-
tant aspect of this likelihood function is the Jacobian of the transformation, which
takes the form ( I - p W ( and ( I - hW( in respectively the spatial lag and spatial
autoregressive error models, with p and h as the autoregressive coefficient and W as
the spatial weights matrix. The need for this Jacobian can be seen from expression
(4) for the spatial lag model and (12) for the spatial autoregressive error model (for a
more extensive treatment, see Anselin 1988a, Chap. 6). In contrast to the time-series
case, the spatial Jacobian is not the determinant of a triangular matrix, but of a full
matrix. This would complicate computational matters considerably, were it not that
Ord (1975) showed how it can be expressed in function of the eigenvalues w ; of the
spatial weights matrix as
Using this simplification, under the normality assumption, the log-likelihood

function for the spatial lag model (3)follows in a straightforward manner as
256 ANSELIN
AND BERA
in the same notatiorl as used in Section 11. This expression clearly illustrates why,
in contrast to the time-series case, ordinary least squares (i.e., the minimization of
the last term in (28)) is not maximum likelihood, since it ignores the Jacobian term.
From the usual first-order conditions, the ML estimates for p and 0’ in a spatial lag
model are obtained as (for details, see Ord 1975, Anselin 1980, Chap. 4: Anselin
1988a, Chap. 6):
pML = (x’X)-lx’(I - pW)y (29)

and
Conditional upon p , these estimates are simply OLS applied to the spatially filtered
dependent variable and the explanatory variables in (6). Substitution of (29) and (30)
in the log-likelihood function yields a concentrated log-likelihood as a nonlinear
function of a single parameter p:
where eo and eL are residuals in a regression of y on X and Wy on X , respectively (for

technical details, see Anselin 1980, Chap. 4). A maximum likelihood estimate for p
is obtained from a numerical optimization of the concentrated log-likelihood function
(31). Based on the framework outlined in Heijmans and Magnus (1986a, 1986b),
it can be shown that the resulting estimates have the usual asymptotic properties,
including consistency, normality, and asymptotic efficiency. The asymptotic variance
matrix follows as the inverse of the information matrix
X’ w*xp X’X
-
02
0
0 2
where WA = W(1 - pW)-’ to simplify notation. Note that while the covariance
between p and the error variance is zero, as in the standard regression model, this
is not the case for p and the error variance. This lack of block diagonality in the
information matrix for the spatial lag model will lead to some interesting results on
SPATIAL DEPENDENCE
MODELS 257
the structure of specification tests, to which we turn in Section IV, It is yet another
distinguishing characteristic between the spatial case and its analog in time series.
Maximum likelihood estimation of the models with spatial error autocorrela-
tion that were covered in Section 1I.E can be approached by considering them as
special cases of general parametrized nonspherical error terms, for which E[&&’] =
a252(8), with 8 as a vector of parameters. For example, from (32) for a spatial au-
toregressive error term, it follows that
R(A) = [(I - AW)’(I - AV)]-’ (33)

As shown in Anselin (1980, Chap. 5), maximum likelihood estimation of such spec-
ifications can be carried out as an application of the general framework outlined in
Magnus (1978).Most spatial processes satisfy the necessary regularity conditions,
although this is not necessarily the case for direct representation models (Mardia
and Marshal1 1984, Warnes and Ripley 1987, Mardia and Watkins 1989).Under the
assumption of normality, the log-likelihood function takes on the usual form:
for example, with 52 (A) as in (33).First-order conditions yield the familiar general-
ized least-squares estimates for B, conditional upon A:
p M L = [x’52(A)- X I - X’R (A)- l y (35)

For a spatial autoregressive error process, 52(A)-’ = (I - AW)’(I - AW), so that
for known A, the maximum likelihood estimates are equivalent to OLS applied to the
spatially filtered variables in (12).Note that for other forms of error dependence, the
GLS expression (35)will involve the inverse of an N by N error covariance matrix.
For example, for the spatial moving average errors, as in (13),52 ( y ) - ’ = [I+ y ( W +
+
W’) y 2 W W ’ ] - ’ , which does not yield a direct expression in terms of spatially
transformed y and X.
Obtaining a consistent estimate for A is not as straightforward as in the time-
series case. As pointed out, OLS does not yield a consistent estimate in a spatial lag
model. It therefore cannot be used to obtain an estimate for A from a regression of
residuals e on We, as in the familiar Cochrane-Orcutt procedure for serially autore-
gressive errors in the time domain. Instead, an explicit optimization of the likelihood
function must be carried out. One approach is to use the iterative solution of the first-
order conditions in Magnus (1978, p. 283):
258 ANSELlN A N D BERA
where e = y - X/3 are GLS residuals. For a spatial autoregressive error process,
X2-'/8h = -W - W' +
AW'W. Solution of condition (36) can be obtained by
numerical means. Alternatively, the GLS expression for t9 and similar solution of
the first-order conditions for o2 can be substituted into the log-likelihood function
to yield a concentrated log-likelihood as a nonlinear function of the autoregressive
parameter h (for technical details, see Anselin 1980, Chap. 5):
with u'u = yLy~- ~ L X L [ X ~ ~ XXLi y] [-, ~, and y~ and X L as spatially filtered variables,
respectively y - hWy and X - hWX. The Jacobian term follows from In 1 f'2 (A)1 =
2 In 11 - h W ( and the Ord simplification in terms of eigenvalues of W .
The asymptotic variance for the ML estimates conforms to the Magnus (1978)
and Breusch (1980) general form and is block diagonal between the regression
( b )and error variance parameters o2 and 8. For example, for a spatial autoregres-
sive error, the asymptotic variance for the regression coefficients is AsyVar[B] =
a 2 [ X ; X ~ ] - ' .The variance block for the error parameters is
1
where, for ease of notation, Wjj = W(I - hW)-'. Due to the block-diagonal form
of the asymptotic variance matrix, knowledge of the precision of h does not affect
the precision of the /3 estimates. Consequently, if the latter is the primary inter-
est, the complex inverse and trace expressions in (38) need not be computed, as
in Benirschka and Binkley (1994). A significance test for the spatial error parame-
ter can be based on a likelihood ratio test, in a straightforward way (Anselin 1988a,
Chap. 8).
Higher-order spatial processes can be estimated using the same general prin-
ciples, although the resulting log-likelihood function will be highly nonlinear and the
use of a concentrated log-likelihood becomes less useful (Anselin 1980, Chap. 6).
The fit of spatial process models estimated by means of maximum likelihood
procedures should not be based on the traditional R 2 , which will be misleading in
the presence of spatial autocorrelation. Instead, the fit of the model may be assessed
by comparing the maximized log-likelihood or an adjusted form to take into account
the number of parameters in the models, such as the familiar AIC (Anselin 198813).
C. GMM/IV Estimation
The view of a spatially lagged dependent variable Wy in the spatial lag model as a
form of endogeneity or simultaneity suggests an instrumental variable (IV) approach
SPATIAL DEPENDENCE
to estimation (Anselin l980,1988a, Chap. 7; 1990b). Since the main problem is the
correlation between Wy and the error term in (3), the choice of proper instruments
for Wy will yield consistent estimates. However, as usual, the efficiency of these
estimates depends crucially on the choice of the instruments and may be poor in
small samples. On the other hand, in contrast to the maximum likelihood approach
just outlined, IV estimation does not require an assumption of normality.
Using the standard econometric results (for a review, see Bowden and Turk-
+
ington 1984),and with Q as a P by N matrix ( P >_ K 1) of instruments (including
K “exogenous” variables from X ) , the IV or 2SLS estimate follows as
with Z = [Wy XI, AsyVar(Blv) = a2[Z’Q(Q’Q)-] 0’21-’, and a2 = (y - ZBrv)’

(y - ZBlV>/N.
Clearly, this approach can also be applied to models where other endogenous
variables appear in addition to the spatially lagged dependent variable, as in a si-
multaneous equation context, provided that the instrument set is augmented to deal
with this additional endogeneity. It also forms the basis for a bootstrap approach to
the estimation of spatial lag models (Anselin 1990b). Moreover, it is easily extended
to deal with more complex error structures, e.g., reflecting forms of heteroskedastic-
ity or spatial error dependence (Anselin 1988a, pp. 86-88). The formal properties of
such an approach are derived in Kelejian and Robinson (1993)for a general methods
of moments estimator (GMM) in the model y = p Wy + +
Xg E with spatial error
components, E = We + +. The GMM estimator takes the form
BCMM= [z’Q(Q’hQ)-’ Q’Z]-’Z‘Q(Q’hQ)-’Q’y
where A is a consistent estimate for the error covariance matrix. The asymptotic vari-
ance for BGMMis [Z’Q(Q’hQ)-’Q’Z]-’. For the spatial error components model,
Kelejian and Robinson (1993, pp. 302-304) suggest an estimate for h = @11 +
$2 WW’, with $1 and $2 as the least-squares estimates in an auxilliary regression
of the squared IV residuals (y - ZBrv) on a constant and the diagonal elements
of WW‘.
A particularly attractive application of GLS-IV estimation in spatial lag mod-
els is a special case of the familiar White heteroskedasticity-consistent covariance
estimator (White 1984, Bowden and Turkington 1984, p. 91). The estimator is as in
(40), but Q’AQ is estimated by Q’QQ, where Q is a diagonal matrix of squared IV
residuals, in the usual fashion. This provides a way to obtain consistent estimates
for the spatial autoregressive parameter p in the presence of heteroskedasticity of
unknown form, often a needed feature in applied empirical work.
A crucial issue in instrumental variables estimation is the choice of the instru-
ments. In spatial econometrics, several suggestions have been made to guide the se-
lection of instruments for Wy (for a review, see Anselin 1988a, pp. 84-86; Land and
Deane 1992). Recently, Kelejian and Robinson (1993 p. 302) formally demonstrated
260 ANSELINAND BERA
the consistency of BGMM in the spatial lag model with instruments consisting of first-
order and higher-order spatially lagged explanatory variables ( W X , W 2 X , etc.).
An important feature of the instrumental variables approach is that estima-
tion can easily be carried out by means of standard econometric software, provided
that the spatial lags can be computed as the result of common matrix manipulations
(Anselin and Hudak 1992). In contrast, the maximum likelihood approach requires
specialized routines to implement the nonlinear optimization of the log-likelihood
(or concentrated log-likelihood). We next turn to some operational issues related to
this.
D. Operational Implementation and Illustration

To date, none of the widely available econometric software packages contain specific
routines to implement maximum likelihood estimation of spatial process models or
to carry out specification tests for spatial autocorrelation in regression models. This
lack of attention to the analysis of the lattice data structures that are most relevant
in empirical economics contrasts with a relatively large range of software for spatial
data analysis in the physical sciences, geared to point patterns and geostatistical
data. Examples of these are the GSLIB library (Deutsch and Journel 1992) and the
recent S+Spatialstats add-on to the S-PLUS statistical software (MathSoft 1996).
While the latter does include some analyses for lattice data, estimation is limited to
maximum likelihood of spatial error models with autoregressive or moving-average
structures. However, the spatial lag model is not covered and specification diagnos-
tics are totally absent.
The only self-contained software package specifically geared to spatial econo-
metric analysis in SpaceStat (Anselin 1992b, 1995). It contains both maximum like-
lihood and instrumental variables estimators for spatial lag and error models, as well
as ways to estimate heteroskedastic specifications and a wide range of diagnostics
for spatial effects. In addition, Spacestat also includes extensive features to carry out
exploratory spatial data analysis as well as utilities to create and manipulate spatial
weights matrices and interface with geographic information systems.
There are two major practical issues that must be resolved to implement the
estimation of spatial lag and spatial error models. The first is the need to construct
spatially lagged variables from observations on the dependent variable or residual
term. This is relevant for both instrumental variables (IV, 2SLS, GMM) as well as
maximum likelihood estimation, In principle, the lag can be computed as a simple
matrix multiplication of the spatial weights matrix W with the vector of observa-
tions, say Wy. This is straightforward to implement in most econometric software
packages that contain matrix algebra routines (specific examples for Gauss, Splus,
Limdep, Rats and Shazam are given in Anselin and Hudak 1992, Table 2, p. 514).
In practice, however, the size of the matrix that can be manipulated by economet-
ric software is severely limited and insufficient for most empirical applications, un-
IN LINEARREGRESSION
less sparse matrix routines can be exploited (avoiding the need to store a full N by
N matrix). This is increasingly the case in state-of-the-art matrix algebra packages
(e.g., Matlab, Gauss), but still fairly uncommon in application-oriented economet-
ric software; hence, the computation of spatial lags will typically necessitate some
programming effort on the part of the user (the construction of spatial lags based on
sparse spatial weights formats in SpaceStat is discussed in Anselin 1995).Once the
spatial lagged dependent variables are computed, IV estimation of the spatial lag
model can be carried out with any standard econometric package.
. The other major operational issue pertains only to maximum likelihood esti-
mation. It is the need to manipulate large matrices of dimension equal to the number
of observations in the asymptotic variance matrices (32) and (38) and in the Jaco-
bian term (27)of the log-likelihoods (31)and (37).In contrast to the time-series case,
the matrix W is not triangular and hence a host of computational simplifications are
not applicable. The problem is most serious in the computation of the asymptotic
variance matrix of the maximum likelihood estimates. The inverse matrices in both
WA = W(1- pW)-' of (32)and WB = W(1- hW)-' of (38)are full matrices which
do not lend themselves to the application of sparse matrix algorithms. For low values
of the autoregressive parameters, a power expansion of (I - pW)-' or (I - ATV)-'
may be a reasonable approximation to the inverse, e.g., (I - p W ) - ' = Ck pkWk+
error, with k = 0, 1 , . . . , K , such that p K < 6, where 6 is a sufficiently small value.
However, this will involve some computing effort in the construction of the powers of
the weights matrices and is increasingly burdensome for higher values of the autore-
gressive parameter. In general, for all practical purposes, the size of the problem for
which an asymptotic variance matrix can be computed is constrained by the largest
matrix inverse that can be carried out with acceptable numerical precision in a given
softwarehardware environment. In current desktop settings, this typically ranges
from a few hlindred to a few thousand observations. While this makes it impossible
to compute asymptotic t-tests for all the parameters in spatial models with very large
numbers of observations, it does not preclude asymptotic inference. In fact, as we ar-
gued in Section III.B, due to the block diagonality of the asymptotic variance matrix
in the spatial error case, asymptotic t-statistics can be constructed for the estimated
j? coefficients without knowledge of the precision of the autoregressive parameter h
(see also Benirschka and Binkley 1994, Pace and Barry 1996).Inference on the au-
toregressive parameter can be based on a likelihood ratio test (Anselin 1988a, Chap.
6). A similar approach can be taken in the spatial lag model. However, in contrast
to the error case, asymptotic t-tests can no longer be constructed for the estimated j?
coefficients, since the asymptotic variance matrix (32)is not block diagonal. Instead,
likelihood ratio tests must be considered explicitly for any subset of coefficients of
interest (requiring a separate optimization for each specification; see Pace and Barry
1997).
With the primary objective of obtaining consistent estimates for the parameters
in spatial regression models, a number of authors have suggested ways to manipu-
262 ANSELINAND BERA
late popular statistical and econometric software packages in order to maximize the
log-likelihoods (28) and (37). Examples of such efforts are routines for ML estima-
tion of the spatial lag and spatial autoregressive error model in Systat, SAS, Gauss,
Limdep, Shazam, Rats and S-PLUS (Bivand 1992, Griffith 1993, Anselin and Hu-
dak 1992, Anselin et al. 1993b). The common theme among these approaches is to
find a way to convert the log-likelihoods for the spatial models to a form amenable
for use with standard nonlinear optimization routines. Such routines proceed incre-
mentally, in the sense that the likelihood is built up from a sum of elements that
correspond to individual observations. At first sight, the Jacobian term in the spatial
models would preclude this. However, taking advantage of the Ord decomposition in
terms of eigenvalues, pseudo-observations can be constructed for the elements of the
Jacobian. Specifically, each term 1 - pwi is considered to correspond to a pseudo-
variable o;, and is summed over all “observations.” For example, for the spatial lag
model, the log-likelihood (ignoring constant terms) can be expressed as
which fits the format expected by most nonlinear optimization routines. Examples
of practical implementations are listed in Anselin and Hudak (1992, Table 10, p.
533) and extensive source code for various econometric software packages is given
in Anselin et al. (1993b).
One problem with this approach is that the asymptotic variance matrices com-
puted by the routines tend to be based on a numerical approximation and do not
necessarily correspond to the analytical expressions in (32) and (38). This may lead
to slight differences in inference depending on the software package that is used
(Anselin and Hudak 1992, Table 10, p. 533). An alternative approach that does not
require the computation of eigenvalues is based on sparse matrix algorithms to effi-
ciently compute the determinant of the Jacobian at each iteration of the optimization
routine. While this allows the estimation of models for very large data sets (tens of
thousands of observations), for example, by using the specialized routines in the Mat-
lab software, this does not solve the asymptotic variance matrix problem. Inference
therefore must be based on likelihood ratio statistics (for details and implementation,
see Pace and Barry 1996,1997).
To illustrate the various spatial models and their estimation, the results for
the parameters in a simple spatial model of crime estimated for 49 neighborhoods
in Columbus, Ohio, are presented in Table 1 . The model and results are based on
Anselin (1988a, pp. 187-196) and have been used in a number of papers to bench-
mark different estimators and specification tests (e.g., McMillen 1992, Getis 1995,
Anselin et al. 1996, LeSage 1997). The data are also available for downloading via
the internet from http://www.rri.wvu.edu/spacestat.htm.The estimates reported in
Table 1 include OLS in the standard regression model, OLS (inconsistent), ML, IV,
and heteroskedastic-robust IV for the spatial lag model, and ML for the spatial error
SPATIAL DEPENDENCE
IN LINEARREGRESSIONMODELS 263
Table I Estimates in a Spatial Model of Crimea
OLS Lag-OLS Lag-ML Lag-IV Lag-GIVE Err-ML
Constant 68.629 38.783 45.079 43.963 46.667 59.893

(4.73) (9.32) (7.18) (11.23) (7.61) (5.37)
P 0.549 0.431 0.453 0.419
(0.153) (0.118) (0.191) 0.139)
Income -1.597 -0.886 - 1.032 - 1.010 -1.185 -0.941
(0.334) (0.358) (0.305) (0.389) (0.434) (0.331)
Housing -0.274 -0.264 -0.266 -0.266 -0.234 -0.302
value (0.103) (0.092) (0.088) (0.092) (0.173) (0.090)
h 0.562
(0.134)
K2 0.552 0.652 0.620 0.633
Log-li k - 187.38 - 182.39 - 183.38
aData are for 49 neighborhoods in Columhus, Ohio, 1980. Dependent variable is per capita resitien-
tial burglaries and vehicle thefts. Income and housing values are in thousand dollars. A first-order
contiguity spatial weights matrix was used to construct the spatial lags.
model. The spatial lags for the exogenous variables ( W X ) were used as instruments
in the IV estimation. In addition to the estimates and their standard errors, the fit of
the different specifications estimated by ML is compared by means of the maximized
log-likelihood. For OLS and the IV estimates, the R2 is listed. However, this should
be interpreted with caution, since R2 is inappropriate as a measure of fit when spa-
tial dependence is present. All estimates were obtained by means of the Spacestat
software.
A detailed interpretation of the results in Table 1 is beyond the scope of this
chapter, but a few noteworthy features may be pointed out. The two spatial models
provide a superior fit relative to OLS, strongly suggesting the presence of spatial de-
pendence. Of the two, the spatial lag model fits better, indicating it is the preferred
alternative. Given the lack of an underlying behavioral model (unless one is willing to
make heroic assumptions to avoid the ecological fallacy problem), the results should
be interpreted as providing consistent estimates for the coefficients of income and
housing value after the spatial dependence in the crime variable is filtered out. The
most affected coefficient (besides the constant term) pertains to the income variable,
and is lowered by about a third while remaining highly significant. The estimates
for the autoregressive coefficient vary substantially between the inconsistent and
biased OLS and the consistent estimates, but the Lag-IV coefficient has a consider-
ably higher standard error. In some instances, OLS can thus yield “better” estimates
in an MSE sense relative to IV. Diagnostics in the Lag-ML model indicate strong re-
maining presence of heteroskedasticity (the spatial Breusch-Pagan test from Anselin
264 AND BERA
ANSELIN
1988a, p. 123, yields a highly significant value of 25.35, p < 0.00001). The robust
Lag-GIVE estimates support the importance of this effect: the estimate for the au-
toregressive parameter is quite close to the ML value while obtaining a significantly
smaller standard error relative to both OLS and the nonrobust IV. Moreover, the es-
timate for the Housing variable is no longer significant. This again illustrates the
complex interaction between heterogeneity and spatial dependence.
IV. TESTS FOR SPATIAL DEPENDENCE
As it happened in the mainstream econometrics literature, the initial stages of de-

velopment in spatial econometrics were characterized by an emphasis on estima-
tion. As discussed in the last section, Cliff and Ord (1973) and others formulated
the maximum likelihood approach which goes to back to work of Whittle (1954).In
mainstream econometrics, the test for serial correlation developed by Durbin and
Watson (1950, 1951)was the first explicit specification test for the regression model.
It has gained widespread acceptance since its inception. However, routine testing
for other specifications (such as homoskedasticity, normality, exogeneity, and func-
tional form) did not take prominence until the early eighties. A major breakthrough
was the rediscovery of the Rao (1947) score (RS) test (known as the Lagrange multi-
plier test in econometrics). The RS test became very popular due to its computational
ease compared to the other two asymptotically “equivalent” test procedures, namely
the likelihood ratio (LR) and Wald (W) tests (see Godfrey 1988 and Bera and Ullah
1991).
In a similar fashion, the origins of specification testing in spatial econometrics
can be traced back to Moran’s (1950a, 1950b) test for autocorrelation. This test laid
in obscurity until it was revived by Cliff and Ord (1972). It received further impetus
by Burridge (1980)as an RS test. However, the early spatial econometrics literature
on testing was dominated by the Wald and LR tests (for example, see Brandsma and
Ketellapper 1979a, 197933; Anselin 1980). Since the latter require the estimation of
the alternative model by means of nonlinear optimization (as discussed in Section
111), the advantages of basing a test on the least-squares regression of the null model,
offered by the RS test, were quickly realized. During the last 15 years, a number of
such tests were developed (see Anselin 1988a, 1988~).
Although mainstream econometrics and spatial econometrics literature went
through similar developments in terms of specification testing, the implementation
of the tests in spatial models turns out to be quite different from the standard case.
For example, most of the RS specification tests cannot be written in the familiar
“NR2” form (where R2 is a coefficient of determination) nor they can be computed
by running any artificial regression. In addition, the interaction between spatial lag
dependence and spatial error dependence in terms of specification testing is stronger
and more complex than its standard counterpart. There are, however, some common
SPATIAL DEPENDENCE
threads. As in the standard case, most of the tests for dependence in the spatial
model can be constructed based on the OLS residuals. In our discussion we will em-
phasize the similarities and the differences between specification testing in spatial
econometric models and the standard case.
We start the remainder of the section with a discussion of Moran’s I statistic
and stress its close connection to the familiar Durbin-Watson test. Moran’s I was
not developed with any specific kind of dependence as the alternative hypothesis,
although it has been found to have power against a wide range of forms of spatial
dependence. We next consider a test developed in the same spirit by Kelejian and
Robinson (1992). This is followed by a focus on tests for specific alternative hypoth-
esis in the form of either spatial lag or spatial error dependence. Tests for these two
kinds of autocorrelations are not independent even asymptotically, and their sepa-
rate applications when other or both kinds of autocorrelations are present will lead
to unreliable inference. Therefore, it is natural to discuss a test for joint lag and er-
ror autocorrelations. However, the problem with such a test is that we cannot make
any specific inference regarding the exact nature of dependence when the joint null
hypothesis is rejected. One approach to deal with this problem is to test for spa-
tial error autocorrelation after estimating a spatial lag model, and vice versa. This,
however, requires ML estimation, and the simplicity of tests based on OLS residuals
is lost. We therefore consider a recently developed set of diagnostics in which the
OLS-based RS test for error (lag) dependence is adjusted to take into account the
local presence of lag (error) dependence (Anselin et al. 1996). We then provide a
brief review of the small-sample properties of the various tests. Finally, the section
is closed into a discussion of implementation issues and our illustrative example of
the spatial model of crime.
A. Moran’s I Test
Moran’s (1950a, 1950b) I test was originally developed as a two-dimensional analog
of the test of significance of the serial correlation coefficient in univariate time series.
Cliff and Ord (1972, 1973) formally presented Moran’s I statistics as
1
where e = y - X g is a vector of OLS residuals, = ( X ’ X ) - ’ X ’ y , W is the spatial
weights matrix, N is the number of observations, and So is a standardization factor
equal to the sum of the spatial weights, xi xj
w ~For
. a row-standardized weights
matrix W , Sosimplifies to N (since each row sum equals 1)and the statistic becomes
e’ We
I=--- (43)
e’e
Moran did not derive the statistic from any basic principle; instead, it was suggested
as a simple test for correlation between nearest neighbors which generalized one
266 AND BERA
ANSELIN
of his earlier tests in Moran (1948). Consequently, the test could be given different
interpretations. The first striking characteristic is the similarity between Moran’s I
and the familiar Durbin-Watson (DW) statistic
e’Ae
DW=- (44)
e’e
where
-1 0 0 ... 0 0 0
2 -1 0 ... 0 0 0
-1 2 -1 ... 0 0 0
... .
... .
0 0 0 ... -1 2 -1
0 0 0 ... 0 -1 1
Therefore, both statistics equal the ratio of quadratic forms in OLS residuals and they
differ only in the specification of the interconnectedness between the observations
(neighboring locations). It is well known that the DW test is a uniformly most power-
ful (UMP) test for one sided alternatives with error distribution E~ = hct-l +cl (see,
e.g., King 1987). Similarly Moran’s I possesses some optimality properties. More
precisely, Cliff and Ord (1972) established a link between the LR and I tests. If we
take the alternative model as (8),i.e.,
&=hW&+t
then the LR statistic for testing Ho: h = 0 against the alternative H,: h = hi, when
E and o2are known, is proportional to
E‘ WE
&’(I+ ~ : G ) E
(45)
where G is a function of W . Therefore, I approaches the LR statistic as hl + 0, and

it can be shown to be consistent for Ho: h = 0 against H,: h # 0. As we discuss
later, Burridge (1980)also showed that I is equivalent to the RS test for h = 0 in (8)
(or y = 0 in the spatial moving average process (13))with an unscaled denominator.
Since we know that the LR and RS tests are asymptotically equivalent under the
null and local alternatives, Cliff and Ord’s result regarding asymptotic equivalence
of I and LR becomes very apparent. King and Hillier (1985) derived the locally
best invariant (LBI) test for the wider problem of testing Ho: h = 0 against Ha:
h > 0 when the covariance matrix of the regression disturbance is of the known form
0 2 Q(A) (as in our (lo)),and showed the test to be identical to the one-sided version
of the RS test. Combining this result with that of Burridge (1980), we can conclude
that Moran’s I must be an LBI test, which was demonstrated by King (1981).
IN LINEARREGRESSION
In practice the test is implemented on the basis of an asymptotically normal

standardized z-value, obtained by subtracting the expected value and dividing by
the standard deviation. One advantage of statistic like I is that under H0 : h = 0
and normality of E , e’e is distributed as central x2. Cliff and Ord (1972) exploited
this to derive the first two moments as
tr(MW)
E ( I ) = ___
N - K
and
V(1) =
tr(MWMW’) + tr(MWl2 + {tr(MW)I2 - CE(1)l2
(47)
(N -K)(N - K + 2)
where M = I - X ( X ’ X ) - ’ X ’ , and W is a row-standardized weights matrix.
It is possible to develop a finite-sample-bound test for I following Durbin and
Watson (1950,1951). However, for I , we need to make the bounds independent of not
only X but also of the weight matrix W . This poses some difficulties. Tiefelsdorf and
Boots (1995), using the results of Imhof (1961) and Koerts and Abrahamse (1968),
showed that exact critical values of I can be computed by numerical integration.
They first expressed I in terms of the eigenvalues y~ , y2, . . . , Y N - K of M W , other
than the K zeros, and N - K independent N ( 0 , 1) variables 6 1 , 6 2 , . . . , ~ N - K ;more
specifically,
N-K N-K
I = Yid’ / 6’
i= 1 i= 1
Then
(49)
Note that x!-7K(~i - Z O ) S ~is a weighted sum of ( N K) x; variables. Imhof’s

method simplifies the probability in (49)to
where
N-K
268 ANSELINAND BERA
The integral in (50)can be evaluated by numerical integration (for more on this, see
Tiefelsdorf and Boots 1995).
It is instructive to note that the computation of exact critical values of the DW
statistic involves the same calculations as for Moran’s I except that the yi is the
eigenvalues of MA, where A is the fixed matrix given by in (44).Even with the recent
dramatic advances in computer technology, it will take some time for practitioners
to use the above numerical integration technique to implement Moran’s I test.
B. Kelejian-Robinson Test
The test developed by Kelejian and Robinson (1992)is in the same spirit of Moran’s
I in the sense that it is not based on an explicit specification of the generating process
of the disturbance term. At the same time the test does not require the model to be
linear or the disturbance term to be normally distributed. Although the test does not
attempt to identify the cause of spatial dependence, Kelejian and Robinson (1992)
made the following assumption about spatial autocorrelation:
COV(&i,& j ) = Olj = zija (51)

where Zij is 1 by q vector which can be constructed from the independent variables
X , a is q by 1 vector of parameters, and i, j are contiguous in the sense that they are
neighbors in a general spatial “ordering” of the observations. The null hypothesis of
no spatial correlation can be tested by Ho : cy = 0 in (51).
For a given sample of size N , let C denote hN by 1 vector 0 i j 7 swhich are not
zero for i < j . Therefore, a test for a! = 0 can be achieved by running a regression of
C on the observation matrix Z which is of dimension hN by q consisting of Z;; values.
Since we do not observe the elements of C, they are replaced by the cross product of
e.
OLS residuals, eiej. The resulting hN by 1 vector is denoted by The test is based
on p = (Z’Z)-’Z’e and is given by
KR=-
p’z’zp
04
where 64 is a consistent estimator of 0’. For example, we can use [e’e/NI2 or

V
(e-
Z p ) ’ ( e - ZP)/hN for G4. Under Ho : a = 0, K R + x: (central chi-square
V
with q degrees of freedom), where + denotes convergence in distribution. Putting
p = (Z’Z)-*Z’e, K R can be expressed as
Pz ( Z’Z )-1Z’ e
KR = (53)
64
Since for the implementation of the test we need the distribution only under the null
hypothesis, it is legitimate to replace 0‘ by a consistent estimate under a = 0.
P P
Note.that under Ho, (?t/hN -+ 0 4 ,where + means convergence in probability.
Therefore, an asymptotically equivalent form of the test is
hN
elz(Z’Z)- 2’ e
(54)
*
e‘t
which has the familiar N R 2 form. Here R2 is the uncentered coefficient of determi-
c
nation of on 2 and h N is the sample size of this regression.
It is also not difficult to see an algebraic connection between K R and Moran’s
I . From (43)
12 =
’
(e We)
(e’e)2
N N
-
Using (53)’ we can write
where p;, are the elements of Z(Z’Z)-’Z’. Given that ti’s contain terms like ekel,
k < 1, it appears that the I 2 and K R statistics have similar algebraic structure.
C. Tests for Spatial Error Autocorrelation

In contrast to the earlier two tests, the alternative hypothesis is now stated explicitly
through the data generating process of E as in (8),i.e.,
and we test h = 0. All three general principles of testing, namely LR, W, and RS
can be applied. Out of the three, the RS test as described in Rao (1947)is the most
convenient one to use since it requires estimation only under the null hypothesis.
That is, the RS test can be based on the OLS estimation of the regression model (7).
Silvey (1959)derived the RS test using the Lagrange multiplier(s) of a constrained
optimization problem.
Burridge (1980) used Silvey’s form to test h = 0, although the Rao’s score
form, namely
RS = d ’ ( s ) z ( e ) - ’ d ( s ) (57)
270 ANSELINAND 6ERA
is more popular and much easier to use. In (57), d(8) = aL(8)/a8 is the score
vector, Z(8) = -E[a2L(8)/a(8)a(8)’] is the information matrix, L(8) is the log-
likelihood function, and 0 is the restricted (under the tested hypothesis) maximum
likelihood estimator of the parameter vector 8. For the spatial error autocorrelation
model 8 = (j?’,c?, A)’ and the log-likelihood function is given in (34). The test is
essentially based on the score with respect to A, i.e., on
We can immediately see the connection of this to Moran’s I statistic. After computing
Z(8) under Ho, from (36),we have the test statistic
where T = tr[(W’ +W ) W ] . Therefore, the test requires only OLS estimates, and
V
under Ho, RSA + xl.It is interesting to put W = W T (Section 1I.E) and obtain
T = N - 1 and RSA = ( N - 1)x2where = x c, e,e,-l/ c, ef-l in the time-series
context. Burridge (1980) derived the RS test (59)using the estimates of the Lagrange
multiplier following Silvey (1959). The Lagrangian function for this problem is
LR(e,p ) = L ( e ) - ,LA (60)

where p is the associated Lagrange multiplier. From the first-order conditions, we
have
PI
ah
i.e.,
and this results in the same statistic R S A .

A striking feature of the RS test is its invariance to different alternatives (for
-
details, see Bera and McKenzie 1986). The RS test uses the slope aL/a8 at 8 = 8,
and there may be many likelihood functions (models) which have the same slope at 0.
If we specify the alternative hypothesis as a spatial moving-average process (13)and
test Ho : y = 0, we obtain the same Rao’s score statistic RSA. Therefore, RSA will
be locally optimal for both autoregressive and moving-average alternatives. But this
also means that when the null hypothesis is rejected, the test does not provide any
guidance regarding the nature of the disturbance process, even when other aspects
IN LINEARREGRESSION
of the spatial model are resolved. This also raises the question whether RSA will
be inferior to other asymptotically equivalent tests such as LR and W, with respect
to power, since it does not use the precise information contained in the alternative
hypothesis. In the context of the standard regression model, Monte Carlo results of
Godfrey (1981) and Bera and McKenzie (1986)suggest that there is no setback in
the performance of RS test compared to the LR test. In Section IV.G, we discuss the
finite sample performance of R S A and other tests.
Computationally, the W and LR tests are more demanding since they require
ML estimation under the alternative, and the explicit forms of the tests are more
complicated. For instance, let us consider the W test which can be computed using
the ML estimate I by maximizing (34)with respect to /?,u2,and A. We can write the
W statistic as (Anselin 1988a, p. 104)
WSA =
x2
-c4
Asy Var (x)

where AsyVar (x) can be obtained from (38) as
tr( W j ) + tr( W iWB) - N
X
For implementation we need to replace h by in the above expression. In the stan-
dard time-series regression case the results are much simpler. For example, AsyVar
[d, A] is a diagonal matrix and AsyVar(k) is simply (1 - h 2 ) / ( N - 1). Therefore
the Wald test statistic can be simply written as
ws; = ( N1-- 1)X2

x2
Note that under h = 0, the asymptotic variance ( 1 - h " ) / ( N - 1) reduces to 1/(N -

l), the expression for AsyVar(X) used in the time series case to test the significance
of A.
The L R statistic can be easily obtained using the concentrated log-likelihood
function Lc in (37).We can write
LRk = 2[& - Lc] (65)

where the "hat" denotes that the quantities are evaluated at the unrestricted ML
x.
estimates 4,8*,and It is easy to see that LRA reduces to (Anselin 1988a, p. 104)
The appearance of the last term in (66)differentiate the spatial dependence situation
from the serial correlation case for time-series data.
272 ANSELINAND BERA
Finally, for higher-order spatial processes, it is easy to generalize the RS statis-

tic (59).For example, if we consider a qth-order spatial autoregressive model
E = A1 W-lE + A,W2E + + h,Wq& + 6
* * * (67)
and test Ho : A1 = A2 = - . . = A, = 0, the RS statistic will be given by
[e’%e/6”I2
RSA,...Aq =
1= 1 Tl
where Tl = tr[ Y’R+Y2],I = 1 , 2 , . . . , q. Under the null of no spatial dependence,

2)
RSA,...Aq ++ xi.Therefore, the test statistic for higher-order dependence is simply
the sum of corresponding individual tests. The same test statistic will result when
a moving average model as in (18) is taken as the alternative instead of (67). As
expected, the Wald and LR tests in this context will be more complicated as they
require ML estimation of A I ,h2, . . . , A,.
D. Tests for Spatial Lag Dependence

In this section, we consider tests on the null hypothesis Ho : p = 0 in (3) using the
log-likelihood function (26). Once again the RS test is the easiest one to use, and
Anselin ( 1 9 8 8 ~derived
) it explicitly (his equation (32)).The score with respect to p
is
=7 E‘ w y
ap p=o
The inverse of the information matrix is given in (30). The complicating feature of
this matrix is that even under p = 0, it is not block diagonal; the ( p , B) term is
equal to (X‘WX/3)/02, obtained by putting p = 0; i.e., WA = W . This absence
of block diagonality causes two problems. First, as we mentioned in Section 11, the
presence of spatial dependence implies that a sample contains less information than
an independent counterpart. This can now be easily demonstrated using (30).In the
absence of dependence ( p = 0 in (3)), the ML estimate of will have variance
02(X’X)-‘ which is the inverse of the information. But when p # 0, to compute the
variance of the ML estimate of B we need to add a positive-definite part to 0 2 ( X ’ X ) - ’
due to absence of block diagonality. Second, to obtain the asymptotic variance of d,,
even under p = 0 from (30), we cannot ignore one of the off-diagonal terms. This
was not the case for dA in Section 1V.C. Asymptotic variance of dA was obtained
just using the (2,Z) element of (36) (see (59)).For the spatial lag model, asymptotic
variance of d, is obtained from the reciprocal of the (1,l)element of
SPATIAL DEPENDENCE
Since under p = 0, WA = W and tr(W) = 0, the expression is Tl = [(WXP)'

+
M ( W X P ) Tu2]/u2,
where T is given in (59).Therefore, the RS statistic is given
by
where in T', , /3, and u2 are replaced by B and (?',respectively. Under Ho : p = 0,

V
RS, + xf, the Wald and LR tests will require maximization of the log-likelihood
function (26) or (29). Let 6 be the ML estimate of p. To get the asymptotic variance
of 6, we need the (1, 1) element of (30). Since the Wald test requires estimation
under the alternative hypothesis (i.e., p # O), the (1,3) element tr(WA)/a2 will also
be nonzero and the resulting expression will more complicated than 7'1 given above
(Anselin 1988a, p. 104). The LR statistic will have the same form as in (66) except
for the last term:
If ML estimation is already performed, LR, is much easier to compute than its Wald
counterpart. Under p = 0 both Wald and LR statistics will be asymptotically dis-
tributed as x:.
E. Testing in the Possible Presence of Both Spatial Error and
Lag Autocorrelation
The test described in the Sections 1V.C and 1V.D can be termed as one-directional
tests in the sense that they are designed to test a single specification assuming cor-
rect specification for the rest of the model. For example, we discussed RSA, WSh,
and LRA statistics for the null hypothesis Ho : h = 0 assuming that p = 0. Because
of the nature of the information matrix, these tests will not be valid even asymptot-
ically, when p # 0. For instance, we noted that under the null, Ho : h = 0 all
the three statistics are asymptotically distributed as central x2 with one degree of
freedom. This result is valid only when p = 0. To evaluate the effects of nonzero p
on RSA, WSA, and LRA, let u s write the model when both the spatial error and lag
autocorrelation are present:
where Wl and W2 are spatial weights matrices associated with the spatially lagged
dependent variable and the spatial autoregressive disturbances, respectively. Recall
from Section 1I.F that for model (73)to be identified, it is necessary that Wl # W2 or
274 ANSELINAND BERA
that the matrix X contain at least one exogenous variable in addition to the constant
term. An alternative specification of spatial moving-average error process for E as in
(1317
has no such problems and it also leads to identical results in terms of test statistics
discussed here. Using the results of Davidson and MacKinnon (1987) and Saikko-
nen (1989),we evaluate the impact of local presence of p on the asymptotic null
distribution of RSA, L R A , and WSA.Let p = 6/a 6 ,< 00, then it can be shown
that under Ho : h = 0, all three tests asymptotically converge to a noncentral x:,
with noncentrali ty parameter
+
where Gj = tr[KK$ y’K$], j = 1 , 2 (note that Tl2 = Tzl). Therefore, the tests will
reject the null of error autocorrelation even when h = 0 due to the local presence
of the lag dependence. In a similar way we can express the asymptotic distributions
of RS,, LR,, and WS,. Under p = 0 and local presence of error dependence, say,
h= t/a, t < 00. In this case the distributions remain x:, but with a noncentrality
parameter
where D = (Wl X P ) ’ M ( WI X P )+ TI .’01 Therefore, again we will have unwanted

“power” due to the presence of local error dependence. In the noncentrality param-
eters R , and RA, the crucial quantity is T 1 2 / n , which can be interpreted as the
covariance between the scores dA and d,. Note that if Tl2 = 0, then both R, and
RA vanish, and local presence of one kind of dependence cannot affect the test for
+
the other one. The trace term Tl2 = tr[ W1 V2 W,’W2],which will only be zero when
the nonzero elements in each row/column of the weights matrices Wl and W2 do not
overlap. In other words, this will be the case when the pattern of spatial dependence
in the lag term and in the error term pertain to a completely different set of neighbors
for each observation. However, in the typical case where Wl = W2 (or overlap to any
extent) then the noncentrality parameter will not vanish.
For valid statistical inference there is a need to take account of possible lag
dependence while we test for error dependence, and vice versa. In Anselin (1988~)
two different approaches are suggested. One is to test jointly for Ho : h = p = 0 in
(73)using the RS principle so that the test can be implemented with OLS residuals
. resulting joint test statistic is given by
(see Anselin 1 9 8 8 ~ )The
(77)
SPATIAL DEPENDENCE
IN LINEARREGRESSION
MODELS 275
where E = (D/a2)T22 - ( T I ~ )Note ~ . that this joint test not only depends on dk
and d, but also on their interaction factor with a coefficient 7’12. Expression (77) ap-
pears to be somewhat complicated but can be computed quite easily using only OLS
residuals. Also the expression simplifies greatly when the spatial weights matrices
W1 and W2 are assumed to be the same which is the case in most applications. Under
Wl = W2 = W , 2‘1 1 = T21 = Tl2 = T = tr[ ( W’ + W )W ] ,and (77) reduces to
Under Ho : h = p = 0, RSA, will converge to a central x2 with two degrees of

freedom. Because of this two degrees of freedom, the statistic will result in loss of
power compared to the proper one-directional test when only one of the two forms
of misspecification is present. To see this consider the presence of only h = t/n,
with p = 0. In this case the noncentrality parameter for both RSA and RSA, is
the same t 2 N T . Due to the higher degrees of freedom of the joint test RSA,, we
can expect some loss of power (Dasgupta and Perlman 1974). Another problem
with RSA, is that since it is an omnibus test, if the null hypothesis is rejected,
it is not possible to infer whether the misspecification is due to lag or error depen-
dence.
A second approach is to carry out an RS test for one form of misspecification in
a model where the other form is unconstrained. For example, this consists of testing
the null hypothesis Ho : h = 0 in the presence of p, i.e., based on the residuals
of a maximum likelihood estimation of the spatial lag model. The resulting statistic
RSA,, is given as
where the “hat” denotes quantities are evaluated at the maximum likelihood esti-
mates of the parameters of the model Y = pWiy + +
Xg 6 obtained by means
of nonlinear optimization. In (79) T 2 1 ~stands for tr[W2W1A-’ +
W,’WlA-’], with
A = I - PW]. Under Ho : h = 0, RSA,, will converge to a central x* with one
degree of freedom. Similarly, an RS test can be developed for Ho : p = 0 in the
presence of error dependence (Anselin et al. 1996).This test statistic can be writ-
ten as
where 2 is a vector of residuals in the ML estimation of the null model with spatial
+
AR errors, y = X g (I - AWz)-’< with Q = (g’, A ,a2)’,and B = I - xW2. The
terms in the denominator of (80)are
276 AND BERA
ANSELIN
- *
and Var(8 is the estimated variance-covariance matrix for the parameter vector 8.
It is also possible to obtain the W and LR statistics in the above three cases,
though these will involve the estimation of a spatial model with two parameters, re-
quiring considerably more complex nonlinear optimization. In contrast, RSLl, and
RS,lk are theoretically valid statistics that have the potential to identify the possible
source(s) of misspecification and can be derived from the results of the maximization
of the log-likelihood functions (32)and (26). However, this is clearly more computa-
tionally demanding than tests based on OLS residuals. We now turn to an approach
that accomplishes carrying out the tests without maximum likelihood estimation of
h and p.
F. Robust Test in the Presence of Local Misspecification
It is not possible to robustify tests in the presence of global misspecification (i.e.,

h and p taking values far away from zero). However, using the general approach
of Bera and Yoon (1993), Anselin et al. (1996)suggested tests which are robust to
local misspecifications, as defined in the previous subsection. The idea is to adjust
the one-directional score tests RSA and RS, by taking account of the noncentrality
parameters R, and RA, given in (75) and (76),so that under the null the resulting
test statistics have central x: distributions.
The modified test for Ho : h = 0 in the local presence of p is given by
When Wl = W2 = W , RS; becomes
Comparing RS; in (81)and RSA in (59), it is clear that the adjusted test modifies
RSA by correcting for the presence of p through 2, and 2'12, where the latter quantity
represents the covariance between dA and d,. Under Ho: h = 0 (and p = 6 / a ) ,
RS: converges to a central x: distribution; i.e., RS; has the same asymptotic distri-
bution as RSA based on the correct specification. This therefore produces asymptoti-
cally the correct size in the presence of local lag dependence. Also as noted for RS;,
SPATIAL IN LINEARREGRESSION
DEPENDENCE MODELS 277
we only need OLS estimation thus circumventing direct estimation of the nuisance
parameter p. However, there is a price to be paid for robustification and simplicity
in estimation. Consider the case when there is no lag dependence ( p = 0), but only
spatial error dependence through h = r / n . Under this setup, the noncentrality
parameters of RSA and RS; are respectively t2T22/N and r2(T22 - T,?p2D-')/N.
/ N 0, the asymptotic power of RS; will be less than that of
Since S ~ T ~ ~ C ? D - ' 2
RSA when there is no lag dependence. The above quantity can be regarded as a cost
of robustification. Once again, note its dependence on TQ. It is also instructive to
compare RS; with Anselin's RSkl, in (79). It is readly seen that RSkl, does not
have the mean correction factor. RSkl, uses the restricted ML estimator of p (under
h = 0) for which J, = 0. We may view RSA,, as the spatial version of Durbin's h
statistic, which can also be derived from the general RS principle. Unlike Durbin's
h, however, RSAI, cannot be computed using the OLS residuals.
In a similar way, the adjusted score test for Ho : p = 0, in the presence of local
misspecification involving spatial-dependent error process can be expressed as
Under Wl = W2 = W , the above expression simplifies to
[ J , - J*]2
RST, = (84)
3-26 -T
All our earlier discussion of RS; also applies to RS;.

Finally, consider the relationship among the five statistics RSA, RS,, RS;,
RS;, and RSA, given in (59),(71), (82), (84), and (78) respectively. RSA, is not the
sum of RSA and RS,; i.e., there is no additivity of the score tests along the lines
discussed in Bera and Jarque (1982)and Bera and McKenzie (1987). From (77), it
is clear that additivity follows only if T12 = 0 or T = 0 for the case of Wl = W2,
i.e., when d k and d, are asymptotically uncorrelated. In that case also RS; = RSA
and RS; = RS, (see (81), (59), (83), and (71)). Hence, for T = 0, the conventional
one-directional tests RSA and RS, are asymptotically valid in the presence of local
misspecification. However, as noted earlier T > 0 and Tl2 > 0 when W1 and W2
have some overlap in the neighbor structure. Under these circumstances (which are
the most common situation encountered in practice), the following very intriguing
result is obtained:
RSA, = RS; + RS, = RSA + RSZ (85)

i.e., the two-directional test for h and p can be decomposed into the sum of the ad-
justed one-directional test of one type of alternative and the unadjusted form for the
other. By construction, under h = p = 0, RS; and R S , are asymptotically inde-
pendently distributed, which cannot be said about RSA and RS,. By applying all the
278 ANSELINAND BERA
unadjusted and adjusted tests and exploiting the result (85),it is possible to identify
the exact nature of dependence in practice (Anselin et al. 1996). Finally, we should
mention that because of the complexity of the Wald and LR tests, it is not possible
to derive their adjusted versions that would be valid under local misspecification.
Of course, it is not computationally prohibitive to obtain these tests after the joint
estimation of both h and p.
G. Small Sample Properties

We have covered a number of procedures for testing spatial dependence. For ease of
implementation, we have emphasized Rao’s score test which in many cases can be
computed based on the OLS residuals. As we indicated, all these tests are of asymp-
totic nature; i.e., their justification derives from the presence of very large samples.
That is, however, not the case in most applications. The small sample performance
of the above tests both in terms of size and power is of major concern to practitioners.
There are only a few papers on the finite sample properties of tests on spatial
dependence compared to the vast literature on those for testing for serial correlation
for time-series data as summarized in King (1987). Bartels and Hordijk (1977) stud-
ied the behavior of Moran’s I . However, their focus was on the performance of dif-
ferent residuals, and they found that OLS residuals give the best results. Brandsma
and Ketellapper (1979b) included the LR test (LRh)in their study, but it performed
poorly compared to I . Both these studies were quite limited in terms of a small num-
ber of replications, few sample sizes, the use of only one type (irregular) weights
matrix and the narrow range of alternative values for the autocorrelation coefficient.
A first extensive set of Monte Carlo simulations was carried out by Anselin and Rey
(1991),who compared Moran’s 1 to RSh and R S , for different weights matrices and
error distributions. In terms of size, the small sample distributions of the statistics
corresponded close to their theoretical counterparts, except for the smallest size
( N = 25). In terms of power, Moran’s I had power against both kinds of depen-
dence, spatial lag and error autocorrelations. RSA and R S , had highest power for
their respective designated alternatives. These tests were found to possess superior
performance, but they fall short of providing a good strategy for identifying the exact
nature of dependence.
Anselin and Florax (1995b) provide the most comprehensive set of simulation
results to date. They carried out experiments for both regular (rook and queen) and
nonregular weight matrices, single- and multidirectional alternatives, and for differ-
ent error distributions, and included all the tests discussed earlier except the Wald
and LR tests. The results are too extensive to discuss in detail, and here we provide
only a brief summary of the main findings. First, the earlier results of Anselin and
Rey (1991) were confirmed on the power of I against any form of dependence and
the optimality of the RS tests against the alternatives for which they were designed.
Second, the specification of the spatial weights matrix impacted the performance of
SPATIAL DEPENDENCE
all tests, with a higher power obtained in the rook case. Third, as in Anselin and
Rey (1991), higher powers were achieved by lag tests relative to tests against error
dependence. This is important, since the consequences of ignoring lag dependence
are more serious. Fourth, the KR statistic did not perform well. For example, when
the errors were generated as lognormal, it significantly over rejected the true null hy-
pothesis in all configurations. There are two possible explanations. One is its higher
degrees of freedom. Another is that its power depends on the degree of autocorre-
lation in the explanatory variables which substitute for the weights matrix (compare
(55) and (56)).It is interesting to note that White’s (1980)test for heteroskedasticity
which is very similar to KR encounters problems of the same type. Fifth, the most
striking result is that the adjusted tests RST and RSp* performed remarkably well.
They had reasonable empirical sizes, remaining within the confidence interval in all
cases. In terms of power they performed exactly the way they were supposed to. For
instance, when the data were generated under p > 0, h = 0, although R S , had the
most power, the powers of RS; was very close to that of RS,. That is, the price paid
for adjustments that were not needed turned out to be small. The real superiority of
RS: was revealed when h > 0 and p = 0. It yielded low rejection frequencies even
for h = 0.9. The correction for error dependence in RS; worked in the right direc-
tion when no lagged dependence was present for all configurations. When p > 0,
the power function of RS; was seen to be almost unaffected by the values of A, even
for those far away from zero (global misspecification). For these alternatives RSA,
also had good power, but could not point to the correct alternative when only one
kind of dependence is present. RS; also performed well though not as spectacularly
as RS;. The adjusted tests thus seem more appropriate to test for lag dependence
in the presence of error correlation than for the reverse case. Again, this is impor-
tant since ignoring lag dependence has more severe consequences. Based on these
results Anselin and Florax (1995b) suggested a simple decision rule. When RS, is
more significant than R S k , and RS; is significant while RS; is not, a lag depen-
dence is the likely alternative. In a similar way presence of error dependence can
be identified through RST. Finally, the finite-sample performance of tests against
higher-order dependence R S A , A(see~ (68))and the joint test RSA, were satisfactory,
although these type of tests provide less insightful guidance for an effective specifi-
cation search. For joint and higher-order alternatives, these tests are optimal, and in
practice they should be used along with the unadjusted and adjusted one-directional
tests.
H. Operational Implementation and Illustration

As is the case for the estimation of spatial regression models, specification tests for
spatial dependence are notably absent from econometric software, with the exception
of SpaceStat (Anselin 199213,1995). Moreover, as pointed out, these tests cannot be
obtained in the usual N R 2 format, which lends itself to straightforward implemen-
280 ANSELINAND BERA
tation by means of auxiliary or augmented regressions. The closest to this situation

is the Kelejian-Robinson test (54), provided one has an easy way to select the pairs
of neighboring data points from the data. Typically, specification tests for spatial de-
pendence must be implemented explicitly either by writing special-purpose software
or by taking advantage of macros in econometric and statistical software. As in max-
imum likelihood estimation, the size of the weights matrix may be a constraint when
the number of observations is large. This is particularly the case for Moran’s I , where
several operations are involved in the computation of the expected value and vari-
ance (46)and (47). Examples of the implementation of this test for small data sets in
standard econometric software are given in Anselin and Hudak (1992) and Anselin
et al. (1993b), for Shazam, Rats and Limdep, among others.
Given their importance for applied work, we now briefly describe implemen-
tation strategies for the RS tests for spatial error and spatial lag autocorrelation, RSA
(59)and R S , (71). First, note that the squared expression in the numerator equals
N times a regression coefficient of an auxiliary regression of respectively We on e
(in (59)) and W y on e (in (71)). Once the lags are constructed, these coefficients can
be obtained using standard software. The denominator in the expressions is slightly
more complex. The trace elements T = tr(WW)+ tr(W’W) can easily be seen to
equal, respectively, xixj xi
wij,wji and C , ( W ~ , ) ~When
. the spatial weights ma-
trix consists of simple row-standardized contiguity weights, each element wii for a
given i equals l/ki, where k; is the number of neighbors for observation z. Hence,
xi C,(W;~)= ~ z i ( l / k ; ) , which can easily be computed. The other trace term
xiE,
is w;j.wji = C,(l/ki)[C, 8i,/k,], where 8;, is a binary variable indicating
whether or not wi, # 0. This requires only slighly more work to compute, similar
to the sorting needed to establish the neighbor pairs in the Kelejian-Robinson test.
Most importantly, the trace operations can be carried out without having to store a
full matrix in memory, taking advantage of the sparse nature of spatial weights (for
technical details, see Anselin 1995). Of course, for symmetric weights, the two traces
are equal. In practice, this may occur when all observations are considered to have
an equal number of neighbors, as in Pace and Barry (1996). The other term in the
denominator of (71) is the residual sum of squares in a regression with W X b (i.e., the
spatial lags for the predicted values from the OLS regression) on X , which can be
obtained in a straightforward way.
To illustrate the various specification tests, we list the results of Moran’s I , KR,
and the RS and LR tests for the spatial model of crime in Table 2 (using a slightly dif-
ferent notation, most of these results are reported in Table 2, p. 87 of Anselin et al.
1996). All results are part of the standard Spacestat regression diagnostic output.
They reflect a situation that is often encountered in practice: strong significance of
Moran’s I and KR, as well as of both one-directional RS and LR tests. Clearly, spa-
tial dependence is a problem, although without further investigation it is not obvious
which form of spatial dependence is the proper alternative. Convincing evidence is
provided by the robust tests RST and RSZ. While the former is not at all significant,
IN LINEAR REGRESSION MODELS
SPATIAL DEPENDENCE 28 I
Table 2 Specification Tests against Spatial Dependencen
Estimates Test (equation number) Value p-value
OLS 2.95 0.003

OLS 11.55 0.009
OLS 9.44 0.009
OLS 5.72 0.02
OLS 0.08 0.78
OLS 9.36 0.002
OLS 3.72 0.05
Lag-ML 9.97 0.002
Lag-ML 0.32 0.57
Err-ML 7.99 0.005
Err-ML 1.76 0.18
aSource: From Anselin (1988a, Chap. 12; 1992a, Chap. 26; 1995) and
Anselin et al. (1996).
the latter is significant at p slightly higher than 0.05. In other words, the impression
of spatial error autocorrelation that may be given by an uncritical interpretation of
Moran’s I is spurious, since no evidence of such autocorrelation remains after ro-
bustifying for spatial lag dependence. Instead, a spatial lag model is the suggested
alternative, consistent with the estimation results in Table 1.
V. CONCLUSIONS
In our review of methods to deal with spatial dependence in regression analysis, we

have emphasized the distinguishing characteristics of spatial econometrics relative
to time-series analysis. We highlighted the concept of spatial weights and the as-
sociated spatial lag operator which allow for the formal specification of neighbors
in space, a much more general concept than its counterpart in time. In the estima-
tion of spatial regression models, the maximum likelihood approach was shown to
be prevalent and requiring nonlinear optimization of the likelihood function. The
simplifying results from serial correlation in time series do not hold and estimation
necessitates the explicit manipulation of matrices of dimension equal to the number
of observations. Diagnostics for spatial effects in regression models may be based on
the powerful score principle, but they do not boil down to simple significance tests
of the coefficients in an auxiliary regression, as they do for time series.
The differences between the time domain and space are both puzzling and
challenging, in terms of theory as well as from an applied perspective. They are the
subject of active research efforts to develop diagnostics for multiple sources of mis-
282 ANSELINAND 6ERA
specification, to discriminate between heterogeneity and spatial dependence, and

to estimate models for complex forms of interaction in realistic data settings. Ex-
tensions to the space-time domain and to models for limited dependent variables
are particularly challenging. We hope that our review of the fundamental concepts
and basic methods will stimulate others to both apply these techniques as well as to
pursue solutions for the remaining research questions.
ACKNOWLEDGMENTS
We would like to thank Aman Ullah and an anonymous referee for helpful sugges-
tions, and Robert Rozovsky for very able research assistance. We also would like
to thank Naoko Miki for her help in preparing the manuscript. However, we retain
the responsibility for any remaining errors. The first author acknowledges ongoing
support for the development of spatial econometric methods by the U S . National Sci-
ence Foundation, notably through grants SES 87-21875, SES 89-21385, and SBR
94-10612 as well as grant SES 88-10917 to the National Center for Geographic In-
formation and Analysis (NCGIA). The second author acknowledges financial support
by the Bureau of Economic and Business Research of the University of Illinois.
REFERENCES
Albert, P. and L. M. McShane (1995), A Generalized Estimating Equations Approach for Spa-
tially Correlated Binary Data: Applications to the Analysis of Neuroimaging Data, Bio-
metrics, 51,627-638.
Amemiya, T. (1985), Advanced Econometrics, Harvard University Press, Cambridge, MA.
Ancot, J-P., J. Paelinck, and J. Prins (1986), Some New Estimators in Spatial Econometrics,
Economics Letters, 21,245-249.
Anselin, L. (1980), Estimation Methods for Spatial Autoregressive Structures, Regional Sci-
ence Dissertation and Monograph Series 8, Cornell University, Ithaca, NY.
Anselin, L. (1982), A Note on Small Sample Properties of Estimators in a First-Order Spatial
Autoregressive Model, Environment and Planning A , 14, 1023-1030.
Anselin, L. (1988a), Spatial Econometrics: Methods and Models, Kluwer, Dordrecht.
Anselin, L. (1988b), Model Validation in Spatial Econometrics: A Review and Evaluation of
Alternative Approaches, International Regional Science Review, 11, 279-316.
Anselin, L. (1988c), Lagrange Multiplier Test Diagnostics for Spatial Dependence and Spatial
Heterogeneity, Geographical Analysis, 20, 1-17.
Anselin, L. (1990a), Spatial Dependence and Spatial Structural Instability in Applied Re-
gression Analysis, Journal of Regional Science, 30, 185-207.
Anselin, L. (1990b), Some Robust Approaches to Testing and Estimation in Spatial Econo-
metrics, Regional Science and Urban Economics, 20, 141-163.
Anselin, L. (ed.) (1992a), Space and Applied Econometrics. Special Issue, Regional Science
and Urban Economics, 22.
IN LINEARREGRESSION
Anselin, L. (1992b), SpaceStat: A Programfor the Analysis of Spatial Data, National Center
for Geographic Information and Analysis, University of California, Santa Barbara, CA.
Anselin, L. (1995), SpaceStat Version 1.80 User’s Guide, Regional Research Institute, West
Virginia University, Morgantown, WV.
Anselin, L., A. K. Bera, R. Florax, and M. J. Yoon (1996), Simple Diagnostic Tests for Spatial
Dependence, Regional Science und Urban Economics, 26,77-104.
Anselin, L. and A. Can (19%), Spatial Effects in Models of Mortgage Origination, Paper pre-
sented at the Mid Year AREUEA Conference, Washington, DC, May 28-29.
Anselin, L., R. Dodson, and S. Hudak (1993a), Linking GIS and Spatial Data Analysis in
Practice, Geographical Systems, 1,3-23.
Anselin, L., R. Dodson, and S. Hudak (1993b), Spatial Data Analysis and GZS: Interfacing
G/S and Econometric Software, Technical Report 93-7, National Center for Geographic
Information and Analysis, University of California, Santa Barbara.
Anselin, L. and R. Florax (eds.) (199Sa), New Directions in Spcitial Econometrics, Springer-
Verlag, Berlin.
Anselin, L. and R. Florax (1995b), Small Sample Properties of Tests for Spatial Dependence
in Regression Models: Some Further Results, in L. Anselin and R. Florax (eds.), New
Directions in Spatial Econometrics, Springer-Verlag, Berlin, 21-74.
Anselin, L. and S. Hudak (1992), Spatial Econometrics in Practice, a Review of Software
Options, Regional Science and Urban Economics, 22,509-536.
Anselin. L. and S. Rey (1991), Properties of Tests for Spatial Dependence in Linear Regression
Models, Geogruphic Analysis, 23, 112-131.
Anselin, L. and S. Rey (eds.) (1997), Spatial Econometrics. Special Issue, International Re-
gional Science Review, 20.
Anselin, L. and 0. Smirnov (1996), Efficient Algorithms for Constructing Proper Higher Order
Spatial Lag Operators, Journal ofRegiona1 Science, 36, 67-89.
Arora, S. and M. Brown (1977), Alternative Approaches to Spatial Autocorrelation: An Im-
provement over Current Practice, Internationul Regional Science Review, 2,67-78.
Bartels, C. P. A. and L. Hordijk (1977), On the Power of the Generalized Moran Contigu-
ity Coefficient in Testing for Spatial Autocorrelation among Regression Disturbances,
Regional Science and Urban Economics, 7,83-101.
Bartels, C. and R. Ketellapper (eds.) (1979), Exploratory and Explanatory Arialysis of Spatial
Ilatn, Martinus Nijhoff, Boston.
Benirschka, M. and J. K. Binkley (1994), Land Price Volatility in a Geographically Dispersed
Market, American Journal of’Agriculturti1 Economics, 76, 185-195.
Bera, A. K. and C. M. Jarque (1982), Model Specification Tests: A Simultaneous Approach,
Bera, A. K. and C. R. McKenzie (1986), Alternative Forms and Properties of the Score Test,
Journal ofApplied Statistics, 13, 13-25.
Bera, A. K. and C. R. McKenzie (1987), Additivity and Separability of the Lagrange Multiplier,
Likelihood Ratio and Wald Tests, Joiimal of Quuntitntive Economics, 3, 53-63.
Beron, K. J., J. C. Murdoch, and W. P. M. Vijverberg (1996), Why Cooperate? An Interdepen-
dent Probit Model of Network Correlations, Working Paper, School of Social Sciences,
University of Texas at Dallas, Richardson, TX.
Bera, A. K. and A. Ullah (1991), Rao’s Score Test in Econometrics, Journal c$Quantitative
Economics, 7, 189-220.
284 ANSELINAND 6ERA
Bera, A. K. and M. J. Yoon (1993), Specification Testing with Misspecified Alternatives, Econo-
metric Theory, 9,649-658.'
Besag, J. (1974), Spatial Interaction and the Statistical Analysis of Lattice Systems, Journal
of the Royal Statistical Society, B, 36, 192-225.
Besley, T. and A. Case (1995), Incumbent Behavior: Vote-Seeking, Tax-Setting, and Yardstick
Competition, American Economic Review, 8 5 , 2 5 4 5 .
Bivand, R. (1992), Systat Compatible Software for Modeling Spatial Dependence among Ob-
servations, Computers and Geosciences, 18,951-963.
Blommestein, H. (1983), Specification and Estimation of Spatial Econometric Models: A Dis-
cussion of Alternative Strategies for Spatial Economic Modelling, Regional Science and
Urban Economics, 13,250-271.
Blommestein, H. (1985), Elimination of Circular Routes in Spatial Dynamic Regression Equa-
tions, Regional Science and Urban Economics, 15, 121-130.
Blommestein, H. J. and N. A. Koper (1992), Recursive Algorithms for the Elimination of
Redundant Paths in Spatial Lag Operators, Journal of Regional Science, 32,91-111.
Bolduc, D., M. G. Dagenais, and M. J. Gaudry (1989), Spatially Autocorrelated Errors in
Origin-Destination Models: A New Specification Applied to Urban Travel Demand in
Winnipeg, Transportation Research, B 23,361-372.
Bolduc, D., R. Laferrikre, and G. Santarossa (1992), Spatial Autoregressive Error Components
in Travel Flow Models, Regional Science and Urban Economics, 22,371-385.
Bolduc, D., R. Laferrihre, and G. Santarossa (1995), Spatial Autoregressive Error Components
in Travel Flow Models, an Application to Aggregate Mode Choice, in L. Anselin and R.
Florax (eds.), New Directions in Spatial Econometrics, Springer-Verlag, Berlin, 96-108.
Bowden, R. J. and D. A. Turkington (1984), Instrumental Variables, Cambridge University
Press, Cam bridge.
Box, G. E. P., G. M. Jenkins, and G. C. Reinsel(1994), Time Series Analysis, Forecasting and
Control, 3rd ed., Prentice Hall, Englewod Cliffs, NJ.
Brandsma, A. S. and R. H. Ketellapper (1979a), A Biparametric Approach to Spatial Auto-
correlation, Environment and Planning A , 11, 51-58.
Brandsma, A. S. and R. H. Ketellapper (1979b). Further Evidence on Alternative Procedures
for Testing of Spatial Autocorrelation among Regression Disturbances, in C. Bartels
and R. Ketellapper (eds.), Exploratory and Explanatory Analysis in Spatial Data, Mar-
tin Nijhoff, Boston, 111-136.
Brett, C. and C. A. P. Pinkse (1997), Those Taxes Are All over the Map! A Test for Spatial Inde-
pendence of Municipal Tax Rates in British Columbia, Znternational Regional Science
Review, 20, 131-151.
Breusch, T. (1980), useful Invariance Results for Generalized Regression Models, Journal of
Brook, D. (1964), On the Distinction between the Conditional Probability and Joint Probability
Approaches in the Specification of Nearest Neighbor Systems, Biometrika, 51, 481-
483.
Brueckner, J. K. (19%), Testing for Strategic Interaction among Local Governments: The Case
of Growth Controls, Discussion Paper, Department of Economics, University of Illinois,
Champaign.
Burridge, P. (1980), On the Cliff-Ord Test for Spatial Autocorrelation, Journal of the Royal
Statistical Society B, 42, 107-108.
IN LINEARREGRESSION
Can, A. (1992), Specification and Estimation of Hedonic Housing Price Models, Regional
Science and Urban Economics, 22, 453-474.
Can, A. (19%), Weight Matrices and Spatial Autocorrelation Statistics Using a Topological
Vector Data Model, Znternational Journul of Geographical Information Systems, 10,
1009- 1017.
Can, A. and I. F. Megbolugbe (1997),Spatial Dependence and House Price Index Construc-
tion, Journal of Real Estate Finance and Economics, 14,203-222.
Case, A. (1987),On the Use of Spatial Autoregressive Models in Demand Analysis, Discus-
sion Paper 135, Research Program in Development Studies, Woodrow Wilson School,
Princeton University.
Case, A. (1991),Spatial Patterns in Household Demand, Econometrica, 59,953-965.
Case, A. (1992). Neighborhood Influence and Technological Change, Regional Science and
llrban Economics, 22,491-508.
Case, A. C., H. S. Rosen, and J. R. Hines (1993), Budget Spillovers and Fiscal Policy In-
terdependence: Evidence from the States, Journal of Public Economics, 52, 285-
307.
Cliff, A. and J. K. Ord (1972),Testing for Spatial Autocorrelation among Regression Residu-
a l ~Geographic
, Analysis, 4, 267-284.
Cliff, A. and J. K. Ord (1973),Spatial Autocorrelation, Pion, London.
Cliff, A. and J. K. Ord (1981),Spatial Processes: Models and Applications, Pion, London.
Cressie, N. (1991), Geostatistical Analysis of Spatial Data, in National Research Council,
Spatial Statistics and Digital Zmage Analysis, National Academy Press, Washington,
DC, 87-108.
Cressie, N (1993),Statistics for Spatial Data, Wiley, New York.
Dasgupta, S. and M. D. Perlman (1974),Power of the Noncentral F-test: Effect of Additional
Variate on Hotelling’s T2-test,Journal of the American Statistical Association, 69,174-
180.
Davidson, R. and J. G. MacKinnon (1987), Implicit Alternatives and Local Power of Test
Statistics, Econometrica, 55, 1305-1329.
Davidson, R. and J. G. MacKinnon (1993),Estimation and Inference in Econometrics, Oxford
University Press, New York.
Deutsch, C. V. and A. G. Journel (1992), GSLZB: Geostatisticul Software Library and User’s
Guide, Oxford University Press, Oxford.
Doreian, P. (1980), Linear Models with Spatially Distributed Data, Spatial Disturbances or
Spatial Effects, Sociological Methods and Research, 9 , 2 9 4 0 .
Doreian, P., K. Teuter, and C-H. Wang (1984),Network Autocorrelation Models, Sociological
Methods and Research, 13, 155-200.
Dow, M. M., M. L. Burton, and D. R. White (1982), Network Autocorrelation: A Simulation
Study of a Foundational Problem in Regression and Survey Study Research, Social
Networks, 4, 169-200.
Dubin, R. (1988),Estimation of Regression Coefficients in the Presence of Spatially Autocor-
related Error Terms, Review of Economics and Statistics, 70,466-474.
Dubin, R. (1992), Spatial Autocorrelation and Neighborhood Quality, Regional Science and
Urban Economics, 22,433-452.
Durbin, J. and G. S. Watson (1950),Testing for Serial Correlation in Least Squares Regression
I, Biometrika, 37,409-428.
286 ANSELINAND BERA
Durbin, J. and G. S. Watson (1951), Testing for Serial Correlation in Least Squares Regression
11, Biometrika, 38, 159-179.
Florax, R. and S. Rey (1995), The Impacts of Misspecified Spatial Interaction in Linear Re-
gression Models, in L. Anselin and R. Florax (eds.), New Directions in Spatial Econo-
metrics, Springer-Verlag, Berlin, 111-135.
Fomby, T. B., R. C. Hill, and S. R. Johnson (1984), Advanced Econometric Methods, Springer-
Verlag, New York.
Getis, A. (1995), Spatial Filtering in a Regression Framework: Examples Using Data on Ur-
ban Crime, Regional Inequality, and Government Expenditures, in L. Anselin and
R. Florax (eds.), New Directions in Spatial Econometrics, Springer-Verlag, Berlin, 172-
185.
Godfrey, L. (1981), On the Invariance of the Lagrange Multiplier Test with Respect to Certain
Changes in the Alternative Hypothesis, Econometrica, 49, 1443-1455.
Godfrey, L. (1988), MisspeclJication Tests in Econometrics, Cambridge University Press, New
York.
Greene, W. H. (1993), Econometric Analysis, 2nd ed., Macmillan, New York.
Griffith, D. A. (1987), Spatial Autocorrelation, A Primer, Association of American Geogra-
phers, Washington, DC.
Griffith, D. A. (1993), Spatial Regression Analysis on the PC: Spatial Statistics Using SAS,
Association of American Geographers, Washington, DC.
Haining, R. (1984), Testing a Spatial Interacting-Markets Hypothesis, The Review of Eco-
nomics and Statistics, 66, 576-583.
Haining, R. (1988), Estimating Spatial Means with an Application to Remotely Sensed Data,
Communications in Statistics: Theory and Methods, 17, 573-597.
Haining, R. (1990), Spatial Data Analysis in the Social and Environmental Sciences, Cam-
bridge University Press, Cambridge.
Heijmans, R. and J. Magnus (1986a), On the First-Order Efficiency and Asymptotic Normality
of Maximum Likelihood Estimators Obtained from Dependent Observations, Statistica
Neerlandica, 40, 169-188.
Heijmans, R. and J. Magnus (1986b), Asymptotic Normality of Maximum Likelihood Estima-
tors Obtained from Normally Distributed but Dependent Observations, Econometric
Theory, 2,374-412.
Hepple, L. W. (1995a), Bayesian Techniques in Spatial and Network Econometrics. 1: Model
Comparison and Posterior Odds, Environment and Planning A , 27, 447-469.
Hepple, L. W. (1995b), Bayesian Techniques in Spatial and Network Econometrics. 2: Com-
putational Methods and Algorithms, Environment and Planning A , 27,615-644.
Holtz-Eakin, D. (1994), Public-Sector Capital and the Productivity Puzzle, Review of Eco-
nomics and Statistics, 76, 12-21.
Hooper, P. and G . Hewings (1981), Some Properties of Space-Time Processes, Geographical
Analysis, 13,203-223.
Hordijk, L. (1974), Spatial Correlation in the Disturbances of a Linear Interregional Model,
Regional and Urban Economics, 4, 117-140.
Hordijk, L. (1979), Problems in Estimating Econometric Relations in Space, Papers, Regional
Science Association, 42,99-11.5.
Hordijk, L. and J. Paelinck (1976), Some Principles and Results in Spatial Econometrics,
Recherches Economiques de Louvain, 42, 175-197.
IN LINEAR REGRESSION
Huang, J. S. (1984). The Autoregressive Moving Average Model for Spatial Analysis, Aus-
tralian Journal of Statistics, 26, 169-178.
Imhof, J. P. (1961), Computing the Distribution of Quadratic Forms in Normal Variables,
Biometrika, 48,419-426.
Isaaks, E. H. and R. M. Srivastava (1989), An Introduction to Applied Ceostatistics, Oxford
University Press, Oxford.
Johnson, N. L., and S. Kotz (1%9),Distributions in Statistics: Discrete Distributions, Houghton
Mifflin, Boston.
Johnston, J. (1984), Econometric Models, McGraw-Hill, New York.
Judge, G., R. C. Hill, W. E. Griffiths, H. Lutkepohl, and T-C. Lee (1982), Introduction to the
Theory and Practice of Econometrics, Wiley, New York.
Judge, G., W. E. Griffiths, R. C. Hill, H. Lutkepohl, and T.-C. Lee (1985), The Theory and
Practice of Econometrics, 2nd ed., Wiley, New York.
Kelejian, H. and D. Robinson (1992),Spatial Autocorrelation: A New Computationally Simple
Test with an Application to Per Capita Country Police Expenditures, Regional Science
and Urban Economics, 22,317-331.
Kelejian, H. H. and D. P. Robinson (1993), A Suggested Method of Estimation for Spatial
Interdependent Models with Autocorrelated Errors, and a n Application to a County
Expenditure Model, Papers in Regional Science, 72,297-312.
Kelejian, H. H. and D. P. Robinson (1995), Spatial Correlation: A Suggested Alternative to
the Autoregressive Model, in L. Anselin and R. Florax (eds.), New Directions in Spatial
Econometrics, Springer-Verlag, Berlin, 75-95.
King, M. L. (1981), A Small Sample Property of the Cliff-Ord Test for Spatial Correlation,
Journal of the Royal Statistical Society B , 43,263-264.
King, M. L. (1987), Testing for Autocorrelation in Linear Regression Models: A Survey, in
M. L. King and D. E. A. Giles (eds.), Specijication Analysis in the Linear Model, Rout-
ledge and Kegan Paul, London, 19-73.
King, M. L. and M. A. Evans (1986), Testing for Block Effects in Regression Models Based
on Survey Data, Journal ofthe American Statistical Association, 81,677-679.
King, M. L. and G. H. Hillier (1985), Locally Best Invariant Tests of the Error Covariance
Matrix of the Linear Regression Model, Journal ofthe Royal Statistical Society B , 47,
98-102.
Koerts, J. and A. P. I. Abrahamse (1968), On the Power of the BLUS Procedure, Journal of
the American Statistical Association, 63, 1227-1236.
Krugman, P. (1991), Increasing Returns and Economic Geography, Journal ofPolitica1 Econ-
omy, 99,483499.
Land, K. and G. Deane (1992), On the Large-Sample Estimation of Regression Models with
Spatial- or Network-Effects Terms: A Two Stage Least Squares Approach, in P. Marsden
(ed.), Sociological Methodology, Jossey Bass, San Francisco, 221-248.
Leenders, R. T. (1995), Structure and Injluence. Statistical Models for the Dynamics of Actor
Attributes, Network Structure and Their Interdependence, Thesis Publishers, Amster-
dam.
Legendre, P. (1993), Spatial Autocorrelation: Trouble or New Paradigm, Ecology, 74, 1659-
1673.
LeSage, J. P. (1993), Spatial Modeling of Agricultural Markets, American Journal of Agricul-
tural Economics, 75, 1211-1216.
288 ANSELINAND BERA
LeSage, J. P. (1997), Bayesian Estimation of Spatial Autoregressive Models, International

Regional Science Review, 20, 113-129.
Liang, K. Y. and S. L. Zeger (1986), Longitudinal Data Analysis Using Generalized Linear
Models, Biometrika, 73, 13-22.
Liitkepohl, H. (1991), Introduction to Multiple Time Series Analysis, Springer-Verlag, Berlin.
Magnus, J. (1978), Maximum Likelihood Estimation of the GLS Model with Unknown Pa-
rameters in the Disturbance Covariance Matrix, Journal of Econometrics, 7,281-312;
Corrigenda, Journal of Econometrics, 10,261.
Mardia, K. V. and R. J. Marshal1 (1984), Maximum Likelihood Estimation of Models for Resid-
ual Covariance in Spatial Regression, Biometrika, 71, 135-146.
Mardia, K. V, and A. J. Watkins (1989), On Multimodality of the Likelihood for the Spatial
Linear Model, Biometrika, 76,289-295.
MathSoft (1W6), S+Spatialstats User’s Manual, Version 1.U , MathSoft, Inc., Seattle.
McMillen, D. P. (1992), Probit with Spatial Autocorrelation, Journal of Regional Science, 32
335-348.
Moran, P. A. P. (1948), The Interpretation of Statistical Maps, Biometrika, 35,255-260.
Moran, P. A. P. (1950a), Notes on Continuous Stochastic Phenomena, Biometrika, 37, 17-23.
Moran, P. A. P. (1950b), A Test for the Serial Independence of Residuals, Biometrika, 37,
178- 181.
Murdoch, J. C., M. Rahmatian, and M. A. Thayer (1993), A Spatially Autoregressive Median
Voter Model of Recreation Expenditures, Public Finance Quarterly, 21, 334-350.
Murdoch, J. C., T. Sandler, and K. Sargent (1996), A Tale of Two Collectives: Sulfur versus
Nitrogen Oxides Emission Reduction in Europe, Working Paper, Department of Eco-
nomics, Iowa State University, Ames, IA.
National Research Council (1991), Spatial Statistics and Digital Image Analysis, National
Academy Press, Washington, DC.
Ord, J. K. (1975), Estimation Methods for Models of Spatial Interaction, Journal of the Amer-
ican Statistical Association, 70, 120-126.
Pace, R. K. and R. Barry (1996), Sparse Spatial Autoregressions, Statistics and Probability
Letters, 2158, 1-7.
Pace, R. K. and R. Barry (1997), Quick Computation of Spatial Autoregressive Estimators,
Geographical Analysis, 29 (forthcoming).
Paelinck, J. (1982), Operational Spatial Analysis, Papers, Regional Science Association, 50,
1-7.
Paelinck, J. and L. Klaassen (1979), Spatial Econometrics, Saxon House, Farnborough.
Pinkse, J. and M. E. Slade (1995), Contracting in Space, an Application of Spatial Statistics
to Discrete-Choice Models, Working Paper, Department of Economics, University of
British Columbia, Vancouver, BC.
Poirier, D. J. (1995), Intermediate Statistics and Econometrics. A Comparative Approach, The
MIT Press, Cambridge, MA.
Rao, C. R. (1947), Large Sample Tests of Statistical Hypotheses Concerning Several Parame-
ters with Applications to Problems of Estimation, Proceedings of the Cambridge Philo-
sophical Society, 44, 50-57.
Ripley, B. D. (1981), Spatial Statistics, Wiley, New York.
Ripley, B. D. (1988), Statistical Znfirence for Spatial Processes, Cambridge University Press,
Cambridge.
IN LINEAR REGRESSIONMODELS
SPATIAL DEPENDENCE 289
Romer, P. M. (1986),Increasing Returns and Long-Run Growth, Journal of Political Economy,

94,1002-1037.
Saikkonen, P. (1989), Asymptotic Relative Efficiency of the Classical Test Statistics Under
Misspecification, Journal of Econometrics, 42,351-369.
Silvey, S. D. (1959),The Lagrange Multiplier Test, Annals ofMathematicaZ Statistics, 30,389-
407.
Tiefelsdorf, M. and B. Boots (1995), The Exact Distribution of Moran’s I , Environment and
Planning A , 27,985-999.
Tobler, W. (1979), Cellular Geography, in S. Gale and G. Olsson (eds.), Philosophy in Geog-
raphy, Reidel, Dordrecht, 379-386.
Upton, G. J. and B. Fingleton (1985), Spatial Data Analysis by Example. Volume I : Point
Pattern and Quantitative Data, Wiley, New York.
Upton, G . J. and B. Fingleton (1989),Spatial Data Analysis by Example. Volume 2: Categorical
and Directional Data, Wiley, New York.
Warnes, J. J. and B. D. Ripley (1987), Problems with Likelihood Estimation of Covariance
Functions of Spatial Gaussian Processes, Biometrika, 74,640-642.
White, H. (1980), A Heteroskedastic-Consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity, Econometrica, 48,817-838.
White, €1. (1984),Asymptotic Theoryfor Econometricians, Academic Press, Orlando.
White, H. (1994), Estimation, Inference and Specijcntion Analysis, Cambridge University
Press, Cambridge.
White, €1. and I. Domowitz (1984),Nonlinear Regression with Dependent Observations, Econo-
metrica, 52, 143-161.
Whittle, P. (1954),On Stationary Processes in the Plane, Biometrika, 41,434-449.
Zeger, S. L. and K. Y. Liang (1986),Longitudinal Data Analysis for Discrete and Continuous
Outcomes, Biometrics, 42, 121-130.
Zeger, S. L., K. Y. Liang, and P. S. Albert (1988),Models for Longitudinal Data: A Generalized
Estimating Equations Approach, Biometrics, 44, 1049-1060.
Panel Data Methods
Badi H. Baitagi
Texas A&M University, College Station, Texas
1. INTRODUCTION
Panel data refers to data sets consisting of multiple observations on each sampling
unit. This could be generated by pooling time-series observations across a variety
of cross-sectional units including countries, states, regions, firms, or randomly sam-
pled individuals or households. Two well-known examples in the United States are
the Panel Study of Income Dynamics (PSID) and the National Longitudinal Survey
(NLS). The PSID began in 1968 with 4802 families, including an oversampling of
poor households. Annual interviews were conducted and socioeconomic character-
istics of each of the families and of roughly 31000 individuals who have been in
these or derivative families were recorded. The list of variables collected is over
5000. The NLS followed five distinct segments of the labor force. The original sam-
ples include 5020 older men, 5225 young men, 5083 mature women, 5159 young
women, and 12686 youths. There was an oversampling of blacks, hispanics, poor
whites, and military in the youths survey. The list of variables collected runs into
the thousands. Panel data sets have also been constructed from the U.S. Current
Population Survey (CPS), which is a monthly national household survey conducted
by the Census Bureau. The CPS generates the unemployment rate and other la-
bor force statistics. Compared with the NLS and PSID data sets, the CPS contains
fewer variables, spans a shorter period, and does not follow movers. However, it
covers a much larger sample and is representative of all demographic groups. Eu-
ropean panel data sets include the German Social Economic Panel, the Swedish
study of household market and nonmarket activities, and the Intomart Dutch panel
of households.
Some of the benefits and limitations of using panel data sets are listed in Hsiao
(1986). Obvious benefits are a much larger data set with more variability and less
29 I
292 BALTAGI
collinearity among the variables than is typical of cross-sectional or time-series data.

With additional, more informative data, one can get more reliable estimates and test
more sophisticated behavioral models with less restrictive assumptions. Another ad-
vantage of panel data sets are their ability to control for individual heterogeneity. Not
controlling for these unobserved individual specific effects leads to bias in the result-
ing estimates. Panel data sets are also better able to identify and estimate effects that
are simply not detectable in pure cross sections or pure time-series data. In particu-
lar, panel data sets are better able to study complex issues of dynamic behavior. For
example, with cross-sectional data set one can estimate the rate of unemployment
at a particular point in time. Repeated cross sections can show how this proportion
changes over time. Only panel data sets can estimate what proportion of those who
are unemployed in one period remain unemployed in another period.
Limitations of panel data sets include problems in the design, data collec-
tion, and data management of panel surveys (Kasprzyk et al. 1989). These include
the problems of coverage (incomplete account of the population of interest), nonre-
sponse (due to lack of cooperation of the respondent or because of interviewer error),
recall (respondent not remembering correctly), frequency of interviewing, interview
spacing, reference period, the use of bounding to prevent the shifting of events from
outside the recall period into the recall period and time in sample bias. Another lim-
itation of panel data sets is the distortions due to measurement errors. Measurement
errors may arise because of faulty response due to unclear questions, memory errors,
deliberate distortion of responses (e.g., prestige bias), inappropriate informants, mis-
recording of responses, and interviewer effects. Although these problems can occur
in cross-sectional studies, they are aggravated in panel data studies. Duncan and Hill
(1985)in a validation study of the PSID data set compare the records of a large firm
with the response of its employees and find the ratio of measurement error variance
to true variance to be of the order of 184%for average hourly earnings. These figures
are for a one-year recall (i.e., 1983 for 1982) and are more than doubled with two
years’ recall. Panel data sets may also exhibit bias due to sample selection problems.
For the initial wave of the panel, respondents may refuse to participate or the inter
viewer may not find anybody at home. This may cause some bias in the inference
drawn from this sample. While this nonresponse can also occur in cross-sectional
data sets, it is more serious with panels because subsequent waves of the panel are
still subject to nonresponse. Respondents may die, or move, or find that the cost of
responding is high. The rate of attrition differs across panels and usually increases
from one wave to the next, but the rate of increase declines over time. Becketti et al.
(1988) studied the representativeness of the PSID, 14 years after it started. They
find that only 40% of those originally in the sample in 1968 remained in the sample
in 1981. Typical panels involve annual data covering a short span of time for each
individual. This means that asymptotic arguments rely crucially on the number of
individuals in the panel tending to infinity. Increasing the time span of the panel
is not without cost either. In fact, this increases the chances of attrition with every
PANELDATAMETHODS 293
new wave and increases the degree of computational difficulty in the estimation of
qualitative limited dependent variable panel data models (Baltagi 1995b).
II. THE ERROR COMPONENTS REGRESSION MODEL
Although, random coefficient regressions can be used in the estimation and specifi-
cation of panel data models (Swamy 1971, Hsiao 1986, Dielman 1989),most panel
data applications have been limited to a simple regression with error components
disturbances
where i denotes individuals and t denotes time, Kit is a vector of observations of k

explanatory variables, B is a k-vector of unknown coefficients, FCL; is an unobserved
individual specific effect, A, is an unobserved time specific effect and v ; is ~ a zero-
mean random disturbance with variance of.The error components disturbances fol-
low a two-way analysis of variance (ANOVA). If Fi and At denote fixed parameters
to be estimated, this model is known as the fixed-effects (FE) model. The Xit’S are
assumed independent of the Vit’S for all i and t . Inference in this case is conditional
on the particular N individuals and over specific time-periods observed. Estimation
in this case amounts to including N - 1 individual dummies and T - 1 time dum-
mies to estimate these time invariant and individual invariant effects. This leads to
an enormous loss in degrees of freedom. In addition, this attenuates the problem of
multicollinearity among the regressors. Furthermore, this may not be computation-
ally feasible for large N and/or T . In this case, one can eliminate the p i ’ s and At’s
and estimate B by running least squares of yit = yil - -ri. r.t+r., on the Zil’s simi-
larly defined, where the dot indicates summation over that index and the bar denotes
averaging. This transformation is known as the within transformation and the corre-
sponding estimator of is called the within estimator or the FE estimator. Note that
the FE estimator cannot estimate the effect of any time-invariant variable like sex,
race, religion, or union participation. Nor can it estimate the effect of any individual
invariant like price, interest rate, etc., that vary only with time. These variables are
wiped out by the within transformation.
If pi and A, are random variables with zero means and constant variances c r i
and o:, this model is known as the random-effects model. The preceding moments
are conditional on the xil’s. In addition, pi, hl, and v ; are
~ assumed to be condition-
ally independent. The random-effects (RE) model can be estimated by GLS which
can be obtained using a least-squares regression of y: = yi, = 81yi, +
- 82y., 837..
on K:, similarly defined, where 81,& and 8s are simple functions of the variance
components a;, o f , and 0: (Fuller and Battese 1974). The corresponding GLS es-
timator of /3 is known as the RE estimator. Note that for this RE model one can
estimate the effects of time-invariant and individual-invariant variables. The best
294 BALTAGI
quadratic unbiased (BQU) estimators of the variance components are ANOVA type
estimators based on the true disturbances and these are minimum variance unbi-
ased (MVU) under normality of the disturbances. One can obtain feasible estimates
of the variance components by replacing the true disturbances by OLS residuals
(Wallace and Hussain 1969). Alternatively, one could substitute the fixed-effects
residuals as proposed by Amemiya (1971). In fact, Amemiya (1971) shows that the
Wallace and Hussain (1969)estimates of the variance components have a different
asymptotic distribution from that knowing the true disturbances, while the Amemiya
(1971)estimates of the variance components have the same asymptotic distribution
as that knowing the true disturbances. Other estimators of the variance components
were proposed by Swamy and Arora (1972) and Fuller and Battese (1974). Maximum
likelihood estimation (MLE) under the normality of the disturbances is derived by
Amemiya (1971). The first-order conditions are nonlinear, but can be solved using
an iterative GLS scheme (Breusch 1987). Finally one can apply Rao’s (1972) min-
imum norm quadratic unbiased estimation (MINQUE) methods. These methods are
surveyed in Baltagi (1995b).Wallace and Hussain (1969) compare the RE and FE
estimators of B in the case of nonstochastic (repetitive) %it’sand find that both are (i)
asymptotically normal, (ii) consistent and unbiased, and that (iii) BKE has a smaller
generalized variance (i.e., more efficient) in finite samples. In the case of nonstochas-
tic (nonrepetitive) xit’s they find that both BKE and f i ~ 1 . :are consistent, asymptotically
unbiased and have equivalent asymptotic variance-covariance matrices, as both N
and T + 00. Under the random effects model, GLS based on the true variance
components is BLUE, and all the feasible GLS estimators considered are asymptot-
ically efficient as N and T + 00. Maddala and Mount (1973)compared OLS, FE,
RE, and MLE methods using Monte Carlo experiments. They found little to choose
among the various feasible GLS estimators in small samples and argued in favor of
methods that were easier to compute. MINQUE was dismissed as more difficult to
compute, ancl tLle applied researcher given one shot at the data was warned to com-
pute at least two methods of estimation. If these methods give different results, the
authors diagnose misspecification. Taylor (1980)derived exact finite sample results
for the one-way error component model ignoring the time effects. He found the fol-
lowing important results. (1) Feasible GLS is more efficient that the FE estimator
for all but the fewest degrees of freedom. (2) The variance of feasible GLS is never
more than 17% above the Cramer-Rao lower bound. (3)More efficient estimators of
the variance components do not necessarily yield more efficient feasible GLS esti-
mators. These finite-sample results are confirmed by the Monte Carlo experiments
carried out by Baltagi (1981a).
One test for the usefulness of panel data models is their ability to predict. For
the FE model, the best linear unbiased predictor (BLUP) was derived by Wansbeek
and Kapteyn (1978) and Taub (1979). This derivation was generalized by Baltagi
and Li (1992) to the RE model with serially correlated remainder disturbances. More
recently, Baillie and Baltagi (1995) derived the asymptotic mean square prediction
error for the FE and RE predictors as well as two other misspecified predictors and
compared their performance using Monte Carlo experiments.
111. TEST OF HYPOTHESES
Fixed versus random effects has generated a lively debate in the biometrics liter-
ature. In econometrics, see Mundlak (1978). The random and fixed effects models
yield different estimation results, especially if T is small and N is large. A speci-
fication test based on the difference between these estimates is given by Hausman
(1978). The null hypothesis is that the individual and time effects are not correlated
with the xit’s. The basic idea behind this test is that the fixed effects estimator f i F E is
consistent whether the effects are or are not correlated with the xit’s. This is true be-
cause the fixed effects transformation described by y~~wipes out the pi and A, effects
from the model. However, if the null hypothesis is true, the fixed effects estimator is
not efficient under the RE specification, because it relies only on the within variation
in the data. On the other hand, the RE estimator &E is efficient under the null hy-
pothesis but is biased and inconsistent when the effects are correlated with the xLL’s.
BRE
The difference between these estimators ij = f i -~ ~ tend to zero in probability
limits under the null hypothesis and is nonzero under the alternative. The variance of
this difference is equal to the difference in variances, var(4) = var(&E)-var(fiRE),
since cov(4, BKE)= 0 under the null hypothesis. Hausman’s test statistic is based
upon m = ij’[var(Q)]-‘Q and is asymptotically distributed as x’
with k degrees of
freedom under the null hypothesis.* The Hausman test can also be computed as a
variable addition test by running y* on the regressor matrices X* and 8 testing that
I
the coefficients of are zero using the usual F-test. This test was generalized by
Arellano (1993)to make it robust to heteroskedasticity and autocorrelation of arbi-
trary forms. In fact, if either heteroskedasticity or serial correlation is present, the
variances of the FE and RE estimators are not valid and the corresponding Haus-
man test statistic is inappropriate. Ahn and Low (1996)show that the Hausman test
statistic can be obtained as (NT)R’ from the regression of GLS residuals on R and
X where the latter denotes the matrix of regressors averaged over time. Also, an al-
ternative generalized method of moments (GMM) test is recommended for testing the
joint null hypothesis of exogeneity of the regressors and the stability of regression
parameters over time. If the regression parameters are nonstationary over time, then
both BR~: BFE:
and are inconsistent even though the regressors may be exogenous.
Ahn and Low perform Monte Carlo experiments which show that both the Haus-
*For the one-way error components model with individual elfects only, Hausman and Taylor (1981)show
that Hausman’s specification test can also he based on two other contrasts that yield numerically iden-
tical results. Kang (1985) extends this analysis to the two-way error romponents model.
296 BALTAGI
man and the alternative GMM test have good power in detecting endogeneity of the
regressors. However, the alternative GMM test dominates if the coefficients of the
regressors are nonstationary. Li and Stengos (1992) propose a Hausman specifica-
tion test based on fl-consistent semiparametric estimators. They apply it in the
context of a dynamic panel data model of the form
yit = 6yi,t-1 + g(xit) + it, i = 1, . . . , N ; t = 1, . . . , T (2)

where the function g(.) is unknown, but satisfies certain moment and differentia-
bility conditions. The x;, observations are IID with finite fourth moments and the
disturbances uit are IID(0, a') under the null hypothesis. Under the alternative, the
disturbances uit are IID in the i subscript but are serially correlated in the t sub-
script. Li and Stengos base the Hausman test for &: E(u;,lyi,,-l) = 0 on the differ-
ence between two n - c o n s i s t e n t instrumental variables estimators for 6, under the
null and the alternative respectively.
For panels with large N and small T , testing for poolability of the data amounts
to testing the stability of the cross-sectional regression across time. In practice, the
Chow (1960) test for the equality of the regression coefficients is popular. This is
proper only under the spherical disturbances assumption. It leads to improper infer-
ence under the random-effects specification. In fact, Baltagi (1981a) shows that in
this case the Chow test leads to rejection of the null too often when in fact it is true.
However, if one accounts for the random-effects variance-covariance matrix, the F -
test for the equality of slopes performs well in Monte Carlo experiments. Recently,
Baltagi, Hidalgo, and Li (1996) derive a nonparametric test for poolability which is
robust to functional form misspecification. In particular, they consider the following
nonparametric panel data model:
yit = &(%it) + e;t (i = 1, . . . , N ; t = 1, . . . , T ) (3)

where gt(.) is an unspecified functional form that may vary over time, and xit is a
k x 1 column vector of predetermined explanatory variables with ( p 2 1) variables
being continuous and k - p (20). Poolability of the data over time is equivalent to
testing that gt(x)= gs(x)almost everywhere for all t and s = 1 , 2 , . . . , T versus
gt(x)# g,(x) for some t # s with probability greater than zero. The test statistic is
shown to be consistent and asymptotically normal and is applied to a panel data set
on earnings.
In choosing between pooled homogeneous parameter estimators versus non-
pooled heterogeneous parameter estimators, some mean-square error (MSE) criteria
can be used as described in Wallace (1972) to capture the trade-off between bias and
variance. Bias is introduced when the poolability restriction is not true. However, the
variance is reduced by imposing the poolability restriction. Hence, the MSE crite-
ria may choose the pooled estimator despite the fact that the poolability restriction
is not true. Ziemer and Wetzstein (1983) suggest comparing pooled and nonpooled
estimators according to their forecast risk performance. They show for a wilderness
recreation demand model that a Stein-rule estimator give better forecast risk per-
formance than the pooled or individual cross-sectional estimates. More recently, the
fundamental assumption underlying pooled homogeneous parameters models has
been called into question. For example, Robertson and Symons (1992) warned about
the bias from pooled estimators when the estimated model is dynamic and homoge-
neous when in fact the true model is static and heterogeneous. Pesaran and Smith
(1995)argued in favor of heterogeneous estimators rather than pooled estimators for
panels with large N and T . They showed that when the true model is dynamic and
heterogeneous, the pooled estimators are inconsistent, whereas an average estimator
of heterogeneous parameters can lead to consistent estimates as long as both N and
T tend to infinity. Using a different approach, Maddala, Srivastava, and Li (1994) ar-
gued that shrinkage estimators are superior to either heterogeneous or homogeneous
parameter estimates especially for prediction purposes. In this case, one shrinks
the individual heterogeneous estimates toward the pooled estimate using weights
depending on their corresponding variance-covariance matrices. Baltagi and Griffin
(1997) compare the short-run and long-run forecast performance of the pooled homo-
geneous, individual heterogeneous, and shrinkage estimators for a dynamic demand
for the gasoline across 18 OECD countries. Based on 1-, 5-, and 10-year forecasts,
the results support the case for pooling. Alternative tests for structural change in
panel data include Han and Park (1989), who used the cumulative sum and cusum
of squares to test for structural change based on recursive residuals. They find no
structural break over the period 1958-1976 in U.S. foreign trade of manufacturing
goods.
Testing for random individual effects is of utmost importance in panel data
applications. Ignoring these effects lead to huge bias in estimation (Moulton 1986).
A popular Lagrange multiplier (LM) test for the significance of the random effects
H l ; 0: = 0 was derived by Breusch and Pagan (1980).This test statistic can be
easily computed using least-squares residuals. This assumes that the alternative hy-
pothesis is two-sided when we know that the variance components are nonnegative.
A one-sided version of this test is given by Honda (1985).This is shown to be uni-
formly most powerful and robust to nonnormality. However, Moulton and Randolph
(1989) showed that the asymptotic N ( 0 , 1) approximation for this one-sided LM
statistic can be poor even in large samples. They suggest an alternative standard-
ized Lagrange multiplier (SLM) test whose asymptotic critical values are generally
closer to the exact critical values than those of the LM test. This SLM test statistic
centers and scales the one-sided LM statistic so that its mean is zero and its variance
IS one.
For Hob; 0; = a : = 0, the two-sided LM test is given by Breusch and Pa-
gan (1980) and is distributed as x; under the null. Honda (1985)does not derive a
uniformly most powerful one-sided test for Hk, but he suggests a “handy” one-sided
test which is distributed as N ( 0 , 1)under H i . Later Honda (1991)derives the SLM
version of this one-sided test. Baltagi, Chang, and Li (1992) derive a locally mean
298 BALTAGI
most powerful (LMMP) one-sided test for Hk and its SLM version is given by Baltagi
(1995b). Under H:, a; = cr: = 0, these standardized Lagrange multiplier statis-
tics are asymptotically N(0, 1)and their asymptotic critical values should be closer
to the exact critical values than those of the corresponding unstandardized tests. Al-
ternatively, one can perform a likelihood ratio test or an ANOVA-type F-test. Both
tests have the same asymptotic distribution as their LM counterparts. Moulton and
Randolph (1989) find that although the F-test is not locally most powerful, its power
function is close to the power function of the exact LM test and is therefore rec-
ommended. A comparison of these various testing procedures using Monte Carlo ex-
periments is given by Baltagi, Chang, and Li (1992). Recent developments include a
generalization by Li and Stengos (1994) of the Breusch-Pagan test to the case where
the remainder error is heteroskedastic of unknown form. Also, Baltagi and Chang
(1996) propose a simple ANOVA F-statistic based on recursive residuals to test for
random individual effects.
For incomplete (or unbalanced) panels, the Breusch-Pagan test can be eas-
ily extended; see Moulton and Randolph (1989)for the one-way error components
model and Baltagi and Li (1990)for the two-way error components model. For non-
linear models, Baltagi (1996) suggests a simple method for testing for zero random
individual and time effects using a Gauss-Newton regression. In case the regression
model is linear, this test amounts to a variable addition test, i.e., running the original
regression with two additional regressors. The first is the average of the least-squares
residuals over time, while the second is the average of the least-squares residuals
over individuals. The test statistic becomes the F-statistic for the significance of the
two additional regressors.
Baltagi and Li (1995)derive three LM test statistics that jointly test for serial
correlation and individual effects. The first LM statistic jointly tests for zero first-
order serial correlation and random individual effects, the second LM statistic tests
for zero first-order serial correlation assuming fixed individual effects, and the third
LM statistic tests for zero first-order serial correlation assuming random individual
effects. In all three cases, Baltagi and Li (1995)showed that the corresponding LM
statistic is the same whether the alternative is AR(1) or MA(1). In addition, Baltagi
and Li (1995)derive two simple tests for distinguishing between AR(1) and MA(1)
remainder disturbances in error components regressions and perform Monte Carlo
experiments to study the performance of these tests. For the fixed-effects model,
Bhargava, Franzini, and Narendranathan (1982) derived a modified Durbin-Watson
test statistic based on FE residuals to test for first-order serial correlation and a test
for random walk based on differenced OLS residuals. Chesher (1984) derived a score
test for neglected heterogeneity, which is viewed as causing parameter variation.
Also, Hamerle (1990) and Orme (1993) suggest a score test for neglected hetero-
geneity for qualitative limited dependent-variable panel data models.
Holtz-Eakin (1988) derives a simple test for the presence of individual ef-
fects in dynamic (autoregressive) panel data models, while Holtz-Eakin, Newey, and
Rosen (1988)formulate a coherent set of procedures for estimating and testing VAR
(vector autoregression) with panel data. Arellano and Bond (1991) consider tests
for serial correlation and overidentification restrictions in a dynamic random-effects
model, while Arellano (1990) considers testing covariance restrictions for error com-
ponents or first-difference structures with white noise, MA, or AR schemes.
Chamberlain (1982, 1984) finds that the fixed effects specification imposes
testable restrictions on coefficients from regressions of all leads and lags of the de-
pendent variable on all leads and lags of independent variables. These overidenti-
fication restrictions are testable using minimum chi-squared statistics. Angrist and
Newey (1991) show that, in the standard fixed effects model, this overidentifica-
tion test statistic is simply the degrees of freedom times the R2 from a regression
of within residuals on all leads and lags of the independent variables. They apply
this test to models of the union-wage effect using five years of data from the National
Longitudinal Survey of Youth and to a conventional human capital earnings function
estimating the return to schooling. They do not reject a fixed effect specification in
the union-wage example, but they do reject it in the return to schooling example.
Testing for unit roots using panel data has been recently reconsidered by Quah
(1994), Levin and Lin (1996), and Im, Pesaran, and Shin (1996). This has been ap-
plied by MacDonald (1996) to real exchange rates for 17 OECD countries based
on a wholesale price index, and 23 OECD countries based on a consumer price in-
dex, all over the period 1973-1992. The null hypothesis that real exchange rates
contain a unit root is rejected. Earlier applications include Bourmahdi and Thomas
(1991). who apply a likelihood ratio unit root panel data test to assess efficiency of
the French capital market. Using 140 French stock prices observed weekly from Jan-
uary 1973to February 1986 ( T = 671) on the Paris Stock Exchange, Boumahdi and
Thomas (1991)do not reject the null hypothesis of a unit root. Also, Breitung and
Meyer (1994) apply panel data unit roots test to contract wages negotiated on firm
and industry level in western Germany over the period 1972-1987. They find that
both firm and industry wages possess a unit root in the autoregressive representation.
However, there is weak evidence for a cointegration relationship.
IV. GENERALIZATIONS OF THE ERROR

COMPONENTS MODEL
The error components disturbances are homoskedastic across individuals. This may
be an unrealistic assumption and has been relaxed by Mazodier and Trognon (1978)
and Baltagi and Griffin (1988). A more general heteroskedastic model is given by
Randolph (1988)in the context of unbalanced panels. Also, Li and Stengos (1994)
proposed estimating a one-way error component model with heteroskedasticity of
unknown form using adaptive estimation techniques.
300 BALTAGI
The error components regression model has been also generalized to allow
for serial correlation in the remainder disturbances by Lillard and Willis (1978),
Revankar (1979), MaCurdy (1982),Baltagi and Li (1991, 1995),and Galbraith and
Zinde-Walsh (1995). Chamberlain (1982,1984) allows for arbitrary serial correlation
and heteroskedastic patterns by viewing each time period as an equation and treating
the panel as a multivariate setup. Also, Kiefer (1980), Schmidt (1983), Arellano
(1987), and Chowdhury (1994) extend the fixed-effects model to cover cases with an
arbitrary intertemporal covariance matrix.
The normality assumption on the error components disturbances may be un-
tenable. Horowitz and Markatou (1996)show how to carry out nonparametric estima-
tion of the densities of the error components. Using data from the Current Population
Survey, they estimate an earnings model and show that the probability that individ-
uals with low earnings will become high earners in the future are much lower than
that obtained under the assumption of normality. One drawback of this nonparamet-
ric estimator is its slow convergence at a rate of l/(log N ) , where N is the number .
of individuals. Monte Carlo results suggest that this estimator should be used for N
larger than 1000.
Micro panel data on households, individuals, and firms are highly likely to
exhibit measurement error; see Duncan and Hill (1985) who found serious measure-
ment error in average hourly earnings in the Panel of Income Dynamics. Using panel
data, Griliches and Hausman (1986) showed that one can identify and estimate a va-
riety of errors in variables models without the use of external instruments. Griliches
and Hausman suggest differencing the data j periods apart (yil - x,~-,),thus gener-
ating “different-lengths” difference estimators. These transformations wipe out the
individual effect, but they may aggravate the measurement error bias. One can cal-
culate the bias of the different-lengths differenced estimators and use this infor-
mation to obtain consistent estimators of the regression coefficients. Extensions of
this model include Kao and Schnell(1987a, 1987b), Wansbeek and Koning (1989),
Hsiao (1991), Wansbeek and Kapteyn (1992), and Biorn (1992). See also Baltagi and
Pinnoi (1995) for an application to the productivity of the public capital stock.
The error components model has been extended to the seemingly unrelated
regressions case by Avery (1977), Baltagi (1980), Magnus (1982), Prucha (1984),
and Kinal and Lahiri (1990). Some applications include Howrey and Varian (1984)
on the estimation of a system of demand equations for electricity by time of day, and
Sickles (1985) on the analysis of productivity growth in the U.S. airlines industry.
For the simultaneous equation with error components. Baltagi (1981b) derives
the error component two-stage (ECZSLS) and three-stage (EC3SLS) least-squares es-
timators, while Prucha (1985) derives the full-information MLE under the normality
assumption. These estimators are surveyed in Krishnakumar (1988). Monte Carlo
experiments are given by Baltagi (1984) and Mhtyhs and Lovrics (1990).Recent ap-
plications of EC2SLS and EC3SLS include (i) an econometric rational-expectations
macroeconomic model for developing countries with capital controls (Haque, Lahiri,
METHODS 30 I
PANELDATA
and Montiel 1993), and (ii) an econometric model measuring income and price elas-
ticities of foreign trade for developing countries (Kinal and Lahiri 1993).
Mundlak (1978) considered the case where the endogeneity is solely attributed
to the individual effects. In this case, Mundlak showed that if these individual ef-
fects are a linear function of the averages of all the explanatory variables across
time, then the GLS estimator of this model coincides with the FE estimator. Mund-
lak’s (1978) formulation assumes that all the explanatory variables are related to the
individual effects. The random-effects model, on the other hand, assumes no corre-
lation between the explanatory variables and the individual effects. Instead of this
“all or nothing” correlation among the xit’s and the p i ’ s , Hausman and Taylor (1981)
consider a model where some of the explanatory variables are related to the p i ’ s . In
particular, they consider
where the zi’s are cross-sectional time-invariant variables. Hausman and Taylor
(1981), hereafter HT, split the matrices X and 2 into two sets of variables: X =
[ X I ;X2]and 2 = [Zl; 221, where X I is n x k l , X2 is n x k 2 , Z l is n x gl, 2 2 is
n x g2, and n = N T . The terms X I and 21are assumed exogenous in that they
are not correlated with pi and u i t , while X 2 and Z2 are endogenous because they
are correlated with the pi’s but not the viL’s.The within transformation would sweep
the p i ’ s and remove the bias, but in the process it would also remove the Zi’s and
hence t h e within estimator will not give an estimate of the y’s. To get around that,
Hausman and Taylor (1981) suggest an instrumental variable estimator that uses
fl,f 2 , XI, and 21as instruments. Therefore, the matrix of regressors X I is used
twice, once as averages and another time as deviations from averages. This is an
advantage of panel data allowing instruments from within the model. The order con-
dition for identification gives the result that the number of X I ’ S (kl) must be at least
as large as the number of Z2k (g2). With stronger exogeneity assumptions between
X and the pi’s, Amemiya and MaCurdy (1986) and Breusch, Mizon, and Schmidt
(1989) suggest more efficient instrumental variable (IV) estimators. Cornwell and
Rupert (1988) apply these IV methods to a returns to schooling example based on a
panel of 595 individuals drawn from the PSID over the period 1976-1982. Recently,
Metcalf (1996)shows that for the Hausman-Taylor model given in (4), using less in-
struments may lead to a more powerful Hausman specification test. Asymptotically,
more instruments led to more efficient estimators. However, the asymptotic bias of
the inefficient estimator will also be greater as the null hypothesis of no correlation
is violated. The increase in bias more than offsets the increase in variance. Since the
test statistic is linear in variance but quadratic in bias, its power will increase.
Cornwell, Schmidt, and Wyhowski (1992) consider a simultaneous equation
model with error components that distinguishes between two types of exogenous vari-
ables, namely singly exogenous and doubly exogenous variables. A singly exogenous
variable is correlated with the individual effects but not with the remainder noise,
302 BALTAGI
while a doubly exogenous variable is uncorrelated with both the effects and the re-
mainder disturbance term. For this encompassing model with two types of exogene-
ity, Cornwell, Schmidt, and Wyhowski (1992) extend the three instrumental vari-
able estimators considered above and give them a GMM interpretation. Wyhowski
(1994) extend these results to the two-way error components model, while Revankar
(1992) establishes conditions for exact equivalence of instrumental variables in a
simultaneous-equation two-way error components model.
V. DYNAMIC PANEL DATA MODELS
The dynamic error components regression is characterized by the presence of a

lagged dependent variable among the regressors, i.e.,
where 6 is a scalar, XI,

is 1 x k , and /3 is k x 1. This model has been extensively stud-
ied by Anderson and Hsiao (1982) and Sevestre and Trognon (1985).*Since yi, is a
function of pi, yi,,-l is also a function of pi. Therefore, %,,-I, a right-hand regres-
sor in (5), is correlated with the error term. This renders the OLS estimator biased
and inconsistent even if the vit’s are not serially correlated. For the FE estimator,
the within transformation wipes out the p i ’ s , but yi9L-l will still be correlated with
Pi, even if the vit’s are not serially correlated. In fact, the within estimator will be
biased of U(l/T) and its consistency will depend upon T being large; see Nickel1
(1981) and Kiviet (1995),who shows that the bias of the FE estimator in a dynamic
panel data model has an O ( W ’ T3/’) approximation error. The same problem occurs
with the random effects GLS estimator. In order to apply GLS, quasi-demeaning is
performed, and yL:t-l will be correlated with U:,. An alternative transformation that
wipes out the individual effects yet does not create the above problem is the first-
difference (FD) transformation. In fact, Anderson and Hsiao (1982)suggested first-
differencing the model to get rid of the p i ’ s and then using Ayi,,-2 = ~ i , ~ --2 yi,,-3
or simplyyi,,-z as an instrument for Ayi,,-I = yi,,-l -%,,-2. These instruments will
not be correlated with Avi, = vi,, - V ~ , ~ - as I , long as the v;,’s themselves are not
serially correlated. This IV estimation method leads to consistent, but not necessar-
ily efficient, estimates of the parameters in the model because it does not make use
of all the available moment conditions (Ahn and Schmidt 1995) and it does not take
into account the differenced structure on the residual disturbances (AV;,).Arellano
*In particular, the assumptions made on the initial values are of utmost importance (Anderson and Hsiao
1982, Bhargava and Sargan 1983, Hsiao 1986). Hsiao (1986) summarizes the consistency properties of
the MLE and GLS under a RE dynamic model depending on the initial values assumption and the way
in which N and T tend to infinity.
(1989)finds that for simple dynamic error components models the estimator that uses
differences rather than levels yi,t-2 for instruments has a singularity point
and very large variances over a significant range of parameter values. In contrast, the
estimator that uses instruments in levels, i.e., ~ i , ~ - 2has
, no singularities and much
smaller variances and is therefore recommended. Additional instruments can be ob-
tained in a dynamic panel data model if one utilizes the orthogonality conditions
that exist between lagged values of yil and the disturbances vit (Holtz-Eakin 1988,
Holtz-Eakin, Newey, and Rosen 1988, Arellano and Bond 1991).Based on these ad-
ditional moments, Arellano and Bond (1991)suggest a GMM estimator and propose
a Sargan-type test for overidentifying restrictions.* Arellano and Bover (1995)de-
velop a unifying GMM framework for looking at efficient IV estimators for dynamic
panel data models. They do that in the context of the Hausman and Taylor (1981)
model given in (4). Ahn and Schmidt (1995)show that under the standard assump-
tions used in a dynamic panel data model, there are additional moment conditions
that are ignored by the IV estimators suggested by Arellano and Bond (1991).They
show how these additional restrictions can be utilized in a GMM framework. Ahn
and Schmidt (1995)also consider the dynamic version of the Hausman and Taylor
(1981) model and show how one can make efficient use of exogenous variables as
instruments. In particular, they show that the strong exogeneity assumption implies
more orthogonality conditions which lie in the deviations from mean space. These
are irrelevant in the static Hausman-Taylor model but are relevant for the dynamic
version of that model.
An alternative approach to estimating dynamic panel data models have been
suggested by Keane and Runkle (1992). Drawing upon the forward filtering idea
from the time-series literature, this method of estimation first transforms the model
to eliminate the general and arbitrary serial correlation pattern in the data. By doing
so, one can use the set of original predetermined instruments to obtain consistent
parameter estimates of the model. First differencing is also used in dynamic panel
data models to get rid of individual specific effect, and the resulting first-differenced
errors are serially correlated of an MA(1)type with unit root if the original uil’s are
classical errors. In this case, there will be gain in efficiency in performing the Keane
and Runkle filtering procedure on the FD model. Underlying this estimation proce-
dure are two important hypotheses that are testable. The first is H A ;the set of instru-
ments are strictly exogenous. In order to test H A , Keane and Runkle propose a test
based on the difference between fixed-effects 2SLS (FE-2SLS) and first-difference
2SLS (FD-2SLS). FE-2SLS is consistent only if H A is true. In fact if the matrix of in-
struments contains predetermined variables then FE-2SLS would not be consistent.
*Bhargava (1991) gives sufficient conditions for the identification of static and dynamic panel data models
with endogenous regressors.
304 BALTAGI
In contrast, FD-2SLS is consistent whether H A is true or not. If H A is not rejected,

one should check whether the individual effects are correlated with the set of instru-
ments. In this case, the usual Hausman and Taylor (1981) test applies. However, if
H A is rejected, the instruments are predetermined and the Hausman-Taylor test is
inappropriate. In this case, the test will be based upon the difference between FD-
2SLS and 2SLS. Under the null, both estimators are consistent, but if the null is not
true FD-2SLS remains consistent while 2SLS does not. These two tests are Hausman
(1978)-type tests except that the variances are complicated because Keane and Run-
kle do not use the efficient estimator under the null (Schmidt, Ahn, and Wyhowski
1992).Keane and Runkle (1992) apply their testing and estimation procedures to a
simple version of the rational expectations life-cycle consumption model. See also
Baltagi and Griffin (1995) for another application to liquor demand.
Alternative estimation methods of a static and dynamic panel data model with
arbitrary error structure are considered by Chamberlain (1982, 1984). Chamberlain
(1984) considered the panel data model as a multivariate regression of T equations
subject to restrictions and derives an efficient minimum distance estimator that is
robust to residual autocorrelation of arbitrary form. He also first-differenced these
equations to get rid of the individual effects and derived an asymptotically equiv-
alent estimator to his efficient minimum distance estimator based on 3SLS of the
T - 2 differenced equations. Building on Chamberlain’s work, Arellano (1990)de-
veloped minimum chi-square tests for various covariance restrictions. These tests are
based on 3SLS residuals of the dynamic error component model and can be calcu-
lated from a generalized linear regression involving the sample autocovariance and
dummy variables. The asymptotic distribution of the unrestricted autocovariance
estimates is derived without imposing the normality assumption. In particular, Arel-
lano (1990)considered testing covariance restrictions for error components or first-
difference structures with white noise, moving-average, or autoregressive schemes.
If these covariance restrictions are true, 3SLS is inefficient and Arellano (1990)pro-
posed a GLS estimator which achieves asymptotic efficiency in the sense that it has
the same limiting distribution as the optimal minimum distance estimator. More re-
cently, Li and Stengos (1996) derived a a - c o n s i s t e n t estimator for a semiparamet-
ric dynamic panel data model, while Li and Stengos (1995) proposed a nonnested
test for parametric versus semiparametric dynamic panel data models.
VI. INCOMPLETE PANEL DATA MODELS
Incomplete panels are more likely to be the norm in typical economic empirical
settings. For example, if one is collecting data on a set of countries over time, a re-
searcher may find some countries can be traced back longer than others. Similarly,
in collecting data on firms over time, a researcher may find that some firms have
dropped out of the market while new entrants emerged over the sample period ob-
served. For randomly missing observations, unbalanced panels have been dealt with
in Fuller and Battese (1974), Baltagi (1985), Wansbeek and Kapteyn (1989), and
Baltagi and Chang (1994).* For the unbalanced one-way error component model,
GLS can still be performed as a least-squares regression. However, BQU estimators
of the variance components are a function of the variance components themselves.
Still, unbalanced ANOVA methods are available (Searle 1987). Baltagi and Chang
(1994) performed extensive Monte Carlo experiments varying the degree of unbal-
ancedness in the panel as well as the variance components. Some of the main results
include the following: (i) As far as the estimation of regression coefficients are con-
cerned, the simple ANOVA-type feasible GLS estimators compare well with the more
complicated estimators such as MLE and MINQUE and are never more than 4 %
above the MSE of true GLS. (ii) For the estimation of the remainder variance compo-
nent a :, these methods show little difference in relative MSE performance. However,
for the individual specific variance component estimation, a :
, the ANOVA-type es-
timators perform poorly relative to MLE and MINQUE methods when the variance
component a : is large and the pattern is severely unbalanced. (iii) Better estimates
of the variance components, in the MSE sense, do not necessarily imply better es-
timates of the regression coefficients. This echoes similar findings in the balanced
panel data case. (iv) Extracting a balanced panel out of an unbalanced panel by ei-
ther maximizing the number of households observed or the total number of observa-
tions lead in both cases to an enormous loss in efficiency and is not recommended.?
For an empirical application, see Mendelsohn et al. (1992), who use panel data on
repeated single-family home sales in the harbor area surrounding New Bedford, Mas-
sachusetts, over the period 1969 to 1988 to study the damage associated with prox-
imity to a hazardous waste site. Mendelsohn et al. (1992) find a significant reduction
in housing values, between $7000 and $10,000 (1989 dollars), as a result of these
houses’ proximity to hazardous waste sites. The extension of the unbalanced error
components model to the two-way model including time effects is more involved.
Wansbeek and Kapteyn (1989) derive the FE, MLE, and a feasible GLS estimator
based on quadratic unbiased estimators of the variance components and compare
their performance using Monte Carlo experiments.
*Other methods of dealing with missing data include (i) inputting the missing values and analyzing the
filled-in data h y complete panel data methods, and (ii) discarding the nonrespondents and weighting
the respondents to compensate for the loss of cases; see Little (1988) and the section on nonresponse
adjustments in Kasprzyk et al. (1989).
t Chowdhury (1991) showed that for the fixed effects error cwnponent model, the within estimator based
on the entire unbalanced panel is efficient relative to any within estimator based on a sub-balanced
pattern. Also, Mgtyhs and Lovrirs (1991) performed some Monte Carlo experiments to compare the loss
in efficiency of FE and GLS based on the entire incomplete panel data and complete subpanel. They
find the loss in efficiency is negligible if N T > 250, but serious for N T < 150.
306 BALTAGI
Rotating panels attempt to keep the same number of households in the survey
by replacing the fraction of households that drop from the sample in each period by
an equal number of freshly surveyed households. This is a necessity in surveys where
a high rate of attrition is expected from one period to the next. For the estimation of
general rotation schemes as well as maximum likelihood estimation under normality
(Biorn 1981). Estimation of the consumer price index in the United States is based on
a complex rotating panel survey, with 20%of the sample being replaced by rotation
each year (Valliant 1991).With rotating panels, the fresh group of individuals that
are added to the panel with each wave provide a means of testing for time-in-sample
bias effects. This has been done for various labor force characteristics in the Current
Population Survey. For example, several studies have found that the first rotation
reported an unemployment rate 10% higher than that of the full sample (Bailar 1975).
While the findings indicate a pervasive effect of rotation group bias in panel surveys,
the survey conditions do not remain the same in practice, and hence it is hard to
disentangle the effects of time-in-sample bias from other effects.
For some countries, panel data may not exist. Instead the researcher may find
annual household surveys based on a large random sample of the populations. Ex-
amples of some of these cross-sectional consumer expenditure surveys include the
British Family Expenditure Survey, which surveys about 7000 households annually.
Examples of repeated surveys in the United States include the Current Population
Survey and the National Crime Survey. For these repeated cross-sectional surveys,
it may be impossible to track the same household over time as required in a genuine
panel. Instead, Deaton (1985) suggests tracking cohorts and estimating economic
relationships based on cohort means rather than individual observations. One co-
hort could be the set of all males born between 1945 and 1950. This age cohort is
well defined and can be easily identified from the data. Deaton (1985) argued that
these pseudo panels do not suffer the attrition problem that plagues genuine panels,
and may be available over longer time periods compared to genuine panels.* For
this psuedo panel with T observations on C cohorts, the fixed effects estimator BFE,
based on the within-"cohort" transformation, is a natural candidate for estimating
B. However, Deaton (1985)argued that these sample-based averages of the cohort
means can only estimate the unobserved population cohort means with measurement
error. Therefore, one has to correct the within estimator for measurement error using
estimates of the errors in measurement variance-covariance matrix obtained from the
individual data. Details are given in Deaton (1985). There is an obvious trade-off in
the construction of a pseudo panel. The larger the number of cohorts, the smaller is
*Blundell and Meghir (1990) also argue that pseudo panels allow the estimation of life-cycle models
which are free from aggregation bias. In addition, Moffitt (1993) explains that a lot of researchers in
the United States prefer to use pseudo panels like thr Current Population Survey because it has larger
more representative samples anti the questions asked are more consistently defined over time than the
available U S . panels.
the number of individuals per cohort. In this case, C is large and the pseudo panel is
based on a large number of observations. However, the fact that the average cohort
size n, = N / C is not large implies that the sample cohort averages are not precise
estimates of the population cohort means. In this case, we have a large number C of
imprecise observations. In contrast, a pseudo panel constructed with a smaller num-
ber of cohorts and therefore more individuals per cohort is trading a large pseudo
panel with imprecise observations for a smaller pseudo panel with more precise ob-
servations. Verbeek and Nijman (1992b) find that n, + 00 is a crucial condition for
the consistency of the within estimator and that the bias of the within estimator may
be substantial even for large n,. On the other hand, Deaton’s estimator is consistent
for /3, for finite n,, when either C or T tend to infinity.
Moffitt (1993)extends Deaton’s (1985) analysis to the estimation of dynamic
models with repeated cross sections. Moffitt illustrates his estimation method for the
linear fixed-effects life-cycle model of labor supply using repeated cross sections
from the U.S. Current Population Survey. The sample included white males, ages
20-59, drawn from 21 waves over the period 1968 to 1988. In order to keep the es-
timation problem manageable, the data was randomly subsampled to include a total
of 15,500 observations. Moffitt concludes that there is a considerable amount of par-
simony achieved in the specification of age and cohort effects. Also, individual char-
acteristics are considerably more important than either age, cohort, or year effects.
Blundell, Meghir, and Neves (1993) use the annual U.K. Family Expenditure Sur-
vey covering the period 1970-1984 to study the intertemporal labor supply and con-
sumption of married women. The total number of households considered was 43,671.
These were allocated to 10different cohorts depending on the year of birth. The aver-
age number of observations per cohort was 364. Their findings indicate reasonably
sized intertemporal labor supply elasticities. Collado (1995) proposed a GMM es-
timator corrected for measurement error to deal with a dynamic pseudopanel data
model. This estimator is consistent as C tends to infinity for a fixed T and n,.
VII. LIMITED DEPENDENT VARIABLES AND

PANEL DATA
In many economic studies, the dependent variable is discrete, indicating, for exam-
ple, that a household purchased a car or that an individual is unemployed or that he
or she joined the union. For example, let yil = 1 if the ith individual participates in
the labor force at time t . This occurs if y,: the difference between the ith individual’s
offered wage and his unobserved reservation wage is positive. This can be described
more formally as follows:
308 BALTAGI
where
That is, y; can be explained by a set of regressors xi, and error components distur-
bances. In this case,
The last equality holds as long as the density function describing the cumulative
distribution function F is symmetric around zero. For panel data, the presence of in-
dividual effects complicates matters significantly. For the one-way error component
model with random individual effects E(ui,ui,) = 02, for any t , s = 1,2, . . . , T ,
and the joint likelihood of (yil, . . . , y ~ , can
) no longer be written as the product of
the marginal likelihoods of the yit’s. This complicates the derivation of maximum
likelihood and will now involve bivariate numerical integration. On the other hand,
if there are no random individual effects, the joint likelihood will be the product of
the marginals and one can proceed as in the usual cross-sectional limited depen-
dent variable case. For the fixed effects model, with limited dependent variable, the
model is nonlinear and it is not possible to get rid of the p i ’ s by taking differences
or performing the FE transformation, as a result B and of cannot be estimated con-
sistently for T fixed, since the inconsistency in the p i ’ s is transmitted to /3 and 02,
(Hsiao 1986).The usual solution around this incidental parameters ( p i ’ s ) problem
is to find a minimal sufficient statistic for the p i ’ s which does not depend on the B’s.
Since the maximum likelihood estimates are in general functions of these minimum
sufficient statistics, one can obtain the latter by differentiating the log-likelihood
T
function with respect to pi. For the logit model, this yields the result that yit is
a minimum sufficient statistic for pi. Chamberlain (1980)suggests maximizing the
conditional likelihood function
rather than the unconditional likelihood function. For the fixed-effects logit model,
this approach results in a computationally convenient estimator. However, the com-
putations rise geometrically with T and are excessive for T > 10.
In order to test for fixed individual effects, one can perform a Hausman-type
test based on the difference between Chamberlain’s conditional maximum likelihood
estimator and the usual logit maximum likelihood estimator, ignoring the individual
effects. The latter estimator is consistent and efficient only under the null of no in-
dividual effects and inconsistent under the alternative. Chamberlain’s estimator is
consistent whether Ho is true or not, but it is inefficient under Ho because it may not
use all the data. Both estimators can be easily obtained from the usual logit maxi-
mum likelihood routines. The constant is dropped and estimates of the asymptotic
variances are used to form Hausman's 'x

statistic. This will be distributed as un- xi
der Ho. For an application studying the linkage between unemployment and mental
health problems in Sweden using the Swedish Level of Living Surveys (Bjorklund
1985).
In contrast to the fixed-effects logit model, the conditional likelihood approach
does not yield computational simplifications for the fixed-effects probit model. In
particular, the fixed effects cannot be swept away and maximizing the likelihood
over all the parameters including the fixed effects will in general lead to inconsistent
estimates for large N and fixed T.* Heckman (1981b)performed some limited Monte
Carlo experiments on a probit model with a single regressor. For N = 100, T = 8,
at = 1, and c r i = 0.5, 1, and 3, Heckman computed the bias of the fixed-effects
MLE of p using 25 replications. He found at most 10% bias for = 1, which was B
always toward zero.
Although the probit model does not lend itself to a fixed effects treatment, it has
been common to use it for the random-effects specification. For the random-effects
probit model, maximum likelihood estimation yields a consistent and efficient esti-
mator of p. However, MLE is computationally more involved. Essentially, one has
to compute the joint probabilities of a T variate normal distribution which involves
T-dimensional integrals (Hsiao 1986). This gets to be infeasible if T is big. How-
ever, by conditioning on the individual effects, this 7'-dimensional integral problem
reduces to a single integral involving the product of a standard normal density and
the difference of two normal cumulative density functions. This can be evaluated
using the Gaussian quadrature procedure suggested by Butler and Moffitt (1982).
This approach has the advantage of being computationally feasible even for fairly
large T . For an application, see Sickles and Taubman (1986),who estimate a two-
equation structural model of the health and retirement decisions of the elderly using
five biennial panels of males drawn from the Retirement History Survey. For a re-
cent Monte Carlo study on the random-effects probit model, see Guilkey and Murphy
(19933. Underlying the random-effects probit model is the equicorrelation assump-
tion between successive disturbances belonging to the same individual. In a study
of labor force participation of married women, Avery, Hansen, and Hotz (1983) re-
ject the hypothesis of equicorrelation across the disturbances and suggest a method
of moments estimator that allows for a general type of serial correlation among the
disturbances. Chamberlain (1984) apples both a fixed effects logit estimator and a
minimum-distance random-effects probit estimator to a study of the labor force par-
ticipation of 924 married women drawn from the Panel Study of Income Dynamics.
Lechner (1995) suggests several specification tests for the panel data probit model.
*In ca4es where the conditional likelihood function is not feasible as in the fixed-effects probit case,
Manski (1987)suggests a conditional version of his maximum score estimator which under fairly general
conditions provides a strongly consistent estimator of' p.
3 10 BALTAGI
These are generalized score and Wald tests employed to detect omitted variables,
neglected dynamics, heteroskedasticity, nonnormality, and random-coefficient vari-
ations. The performance of these tests in small samples is investigated using Monte
Carlo experiments. Also, an empirical example on the probability of self-employment
in West Germany is given which uses a random sample of 1926 working men selected
from the German Socio-Economic Panel and observed over the period 1984-1989.
Heckman and MaCurdy (1980) consider a fixed-effects tobit model to esti-
mate a life-cycle model of female labor supply. They argue that the individual ef-
fects have a specific meaning in a life-cycle model and therefore cannot be assumed
independent of the xit’s. Hence, a fixed effects rather than a random-effects speci-
fication is appropriate. For this fixed-effects tobit model, the model is given by (7),
-
with uit IIN(0, at) and
yit = yz if y: > 0
(9)
=0 otherwise
where yit could be the expenditures on a car. This will be zero at time t, if the ith
individual does not buy a car. In the latter case all we know is that yz 5 O.* As in
the fixed-effects probit model, the pi’s cannot be swept away and as a result /3 and
cr: cannot be estimated consistently for T fixed, since the inconsistency in the p i ’ s
is transmitted to /Iand 0:. Heckman and MaCurdy (1980) suggest estimating the
log-likelihood using iterative methods. Recently, Honor6 (1992) suggested trimmed
least absolute deviations and trimmed least-squares estimators for truncated and
censored regression models with fixed effects. These are semiparametric estimators
with no distributional assumptions necessary on the error term. The main assumption
is that the remainder error vit is independent and identically distributed conditional
on the xit’s and the p i ’ s , for t = 1, . . . , T . Honor6 (1992) exploits the symmetry in
the distribution of the latent variables and finds that when the true values of the pa-
rameters are known, trimming can transmit the same symmetry in distribution to the
observed variables. This generates orthogonality conditions which must hold at the
true value of the parameters. Therefore, the resulting GMM estimator is consistent
provided the orthogonality conditions are satisfied at a unique point in the parameter
space. Honor6 (1992)shows that these estimators are consistent and asymptotically
normal. Monte Carlo results show that as long as N 2 200, the asymptotic distri-
bution is a good approximation of the small-sample distribution. However, if N is
*Researchersmay also be interested in panel data economic relationships where the dependent variable
is a count of some individual actions or events, such as the number of patents filed, the number of drugs
introduced, or the number of jobs held. These models can he estimated by using Poisson panel data
regressions (Hausman, Hall, and Griliches 1984).
PANELDATAMETHODS 3 I I
small, the small-sample distribution of these estimators is skewed. Honor6 (1993)

extends his analysis to the dynamic Tobit model with fixed effects; i.e.,
with yil = max(0, y;} for i = 1, . . . , N , t = 1, . . . , T . The basic assumption is

that vit is IID(0, of)for t = 1, . . . , T , conditional on yio, x i t , and pi. Honor6 (1993)
shows how to trim the observations from a dynamic Tobit model so that the sym-
metry conditions are preserved for the observed variables at the true values of the
parameters. These symmetry restrictions are free of the individual effects and no
assumption is needed on the distribu5on of the hi’s or their relationship with the
explanatory variables. These restrictions generate orthogonality conditions which
are satisfied at the true value of the parameters. The orthogonality conditions can
be used in turn to construct method of moments estimators. Using Monte Carlo ex-
periments, Honor6 (1993)shows that MLE for a dynamic Tobit fixed-effects model
performs poorly, whereas the GMM estimator performs quite well, when S is the only
parameter of interest.
Recently, Keane (1994) derived a computationally practical simulation esti-
mator for the panel data probit model. Simulation estimation methods replace in-
tractable integrals by unbiased Monte Carlo probability simulators. This is ideal for
limited dependent variable models where for a multinominal probit model, the choice
probabilities involve multivariate integrals.* In fact, for cross-sectional data, McFad-
den’s method of simulated moments (MSM) involves an M - 1 integration problem,
where M is the number of possible choices facing the individual. For panel data,
things get more complicated, because there are M choices facing any individual at
each period. This means that there are M T possible choice sequences facing each
individual over the panel. Hence the MSM estimator becomes infeasible as T gets
large. Keane (1994) sidesteps this problem of having to simulate M T possible choice
sequences by factorizing the method of simulated moments first-order conditions into
transition probabilities. The latter are simulated using highly accurate importance
sampling techniques. This method of simulating probabilities is referred to as the
Geweke, Hajivassiliou, and Keane (GHK) simulator because it was independently
developed by these authors. Keane (1994) performs Monte Carlo experiments and
finds that even for large T and small simulation sizes, the bias in the MSM esti-
mator is negligible. When maximum likelihood methods are feasible, Keane (1994)
finds that the MSM estimator performs well relative to quadrature-based maximum
likelihood methods even where the latter are based on a large number of quadrature
*For good surveys of simulation methods, see Hajivassiliou and Ruud (1994)for limited dependent vari-
able models and Gourieroux and Monfort (1993) with special reference to panel data. The methods
surveyed include simulation of the likelihood, simulation of the moment functions, and simulation of the
score.
312 BALTAGI
points. When maximum likelihood methods are not feasible, the MSM estimator out-
performs the simulated maximum likelihood estimator even when the highly accurate
GHK probability simulator is used. Keane (1993)applies the MSM estimator to the
same data set used by Keane, Moffitt, and Runkle (1988) to study the cyclical be-
havior of real wages. He finds that the Keane, Moffitt, and Runkle conclusion of a
weakly procyclical movement in the real wage appears to be robust to relaxation of
the equicorrelation assumption.
Heckman (1981a, 1981b, 1981c) emphasizes the importance of distinguish-
ing between “true state dependence” and “spurious state dependence” in dynamic
models of individual behavior. In the true case, once an individual experiences an
event, his or her preferences change and he or she will behave differently in the fu-
ture as compared with an identical individual that has not experienced this event
in the past. In the spurious case, past experience has no effect on the probability of
experiencing the event in the future. However, one cannot properly control for all
the variables that distinguish one individual’s decision from another. In this case,
past experience which is a good proxy for these omitted variables shows up as a sig-
nificant determinant of the future probability of occurrence of this event. Testing for
true versus spurious state dependence is therefore important in these studies, but it
is complicated by the presence of the individual effects or heterogeneity. In fact, even
if there is no state dependence, Pr[yi,/x;,, y;,,-e] # Pr[yit/xit] as long as there are
random individual effects present in the model. If in addition to the absence of the
state dependence there is also no heterogeneity, then Pr[yi,/xi,, y;,,-e] = Pr[yi,/xi,].
A test for this equality can be based on a test for y = 0 in the model
wyi, = w i t 7 yit-11 =O : , B +Ui,t-I) (11)
by using standard maximum likelihood techniques. If y = 0 is not rejected, we
ignore the heterogeneity issue and proceed as in conventional limited dependent
variable models not worrying about the panel data nature of the data. However, re-
jecting the null does not necessarily imply that there is heterogeneity since y can
be different from zero due to serial correlation in the remainder error or due to state
dependence. In order to test for time dependence one has to condition on the individ-
ual effects, i.e., test Pr[Y,,/y;,,-e, x;,, p ; ] = Pr[yi,/x;,, pi]. This can be implemented
following the work of Lee (1987) and Maddala (1987). In fact, if y = 0 is rejected,
Hsiao (1996) suggests testing for time dependence against heterogeneity. If hetero-
geneity is rejected, the model is misspecified. If heterogeneity is not rejected then
one estimates the model correcting for heterogeneity. See Heckman (1981~) for an
application to married women’s employment decisions based on a three-year sample
from the Panel Study of Income Dynamics. One of the main findings of this study is
that neglecting heterogeneity in dynamic models overstate the effect of past experi-
ence on labor market participation.
In many surveys, nonrandomly missing data may occur due to a variety of
self-selection rules. One such self-selection rule is the problem of nonresponse of
PANELDATAMETHODS 3 I 3
the economic agent. Nonresponse occurs, for example, when the individual refuses
to participate in the survey or refuses to answer particular questions. This problem
occurs in cross-sectional studies, but it gets aggravated in panel surveys. After all,
panel surveys are repeated cross-sectional interviews. So, in addition to the above
kinds of nonresponse, one may encounter individuals that refuse to participate in
subsequent interviews or simply move or die. Individuals leaving the survey cause
attrition in the panel. This distorts the random design of the survey and questions the
representativeness of the observed sample in drawing inference about the popula-
tion we are studying. Inference based on the balanced subpanel is inefficient even in
randomly missing data since it is throwing away data. In nonrandomly missing data,
this inference is misleading because it is no longer representative of the population.
Verbeek and Nijman (1996)survey the reasons for nonresponse and distinguish be-
tween “ignorable” and “nonignorable” selection rules. This is important because, if
the selection rule is ignorable for the parameters of interest, one can use the standard
panel data methods for consistent estimation. If the selection rule is nonignorable,
then one has to take into account the mechanism that causes the missing observa-
tions in order to obtain consistent estimates of the parameters of interest.
We now consider a simple model of nonresponse in panel data. Following the
work of Hausman and Wise (1979), Ridder (1990), and Verbeek and Nijman (1996),
we assume that yit given by Eq. (1)is observed if a latent variable r: 2 0. This latent
variable is given by
where zit is a set of explanatory variables possibly including some of the xil’s. The
cess. The errors are assumed to be normally distributed ci

IIN(0,o;) with the only nonzero covariances being COV(E;,
-
one-way error components structure allows for heterogeneity in the selection pro-
IIN(0,op) and qLt
pi) = op6and cov
-
) oqV.
( q i t , u ; ~= In order to get a consistent estimator for #al,
generalization of Heck-
man’s selectivity bias correction procedure from the cross-sectional to the panel data
case can be employed. The conditional expectation of uit given selection now in-
volves two terms. Therefore, instead of one selectivity bias correction term, there
are now two terms corresponding to the covariances oWc and o q VHowever,
. unlike
the cross-sectional case, these correction terms cannot be computed from simple
probit regressions and require numerical integration. Fortunately, this is only a one-
dimensional integration problem because of the error component structure. Once the
correction terms are estimated, they are included in the regression equation as in the
cross-sectional case and OLS or GLS can be run on the resulting augmented model.
For details, see Verbeek and Nijman (1996), who also warn about heteroskedasticity
and serial correlation in the second step regression if the selection rule in nonignor-
able. Verbeek and Nijman (1996) also discuss MLE for this random-effects probit
model with selection bias.
314 BALTAGI
Before one embarks on these complicated estimation procedures one should

first test whether the selection rule is ignorable. Verbeek and Nijman (1992a) con-
sider a Lagrange multiplier test for Ho; oVq= ape= 0. The null hypothesis is
a sufficient condition for the selection rule to be ignorable for the random-effects
model. Unfortunately, this also requires numerical integration over a maximum of
two dimensions and is cumbersome to use in applied work. In addition, the LM test
is highly dependent on the specification of the selectivity equation and the distribu-
tional assumptions, Alternatively, Verbeek and Nijman (1992a) suggest some simple
Hausman-type tests based on GLS and within estimators for the unbalanced panel
and the balanced subpanel. All four estimators are consistent under the null hy-
pothesis that the selection rule is ignorable and all four estimators are inconsistent
under the alternative. In practice, Verbeek and Nijman (1992a) suggest including
three simple variables in the regression to check for the presence of selection bias.
These are (i) the number of waves the ith individual participates in the panel, (ii) a
binary variable taking the value 1 if and only if the ith individual is observed over
the entire sample, and (iii) a binary variable indicating whether the individual was
present in the last period. Testing the significance of these variables is recommended
as a check for selection bias. Intuitively, one is testing whether the pattern of miss-
ing observations affects the underlying regression. Wooldridge (1995)derives some
simple variable addition tests of selection bias as well as easy-to-apply estimation
techniques that correct for selection bias in linear fixed-effects panel data models.
The auxiliary regressors are either Tobit residuals or inverse Mill’s ratios, and the
disturbances are allowed to be arbitrarily serially correlated and unconditionally
heteroskedastic.
There are many empirical applications illustrating the effects of attrition bias;
see Hausman and Wise (1979)for a study of the Gary Income Maintenance exper-
iment. For this experimental panel study of labor supply response, the treatment
effect is an income guaranteehax rate combination. People who benefit from this
experiment are more likely to remain in the sample. Therefore, the selection rule
is nonignorable, and attrition can overestimate the treatment effect on labor sup-
ply. For the Gary Income Maintenance Experiment, Hausman and Wise (1979)
found little effect of attrition bias on the experimental labor supply response.
Similar results were obtained by Robins and West (1986) for the Seattle and
Denver Income Maintenance Experiments. For the latter sample, attrition was
modest (11% for married men and 7% for married women and single heads dur-
ing the period studied) and its effect was not serious enough to warrant extensive
correction procedures. More recently, Keane, Moffitt, and Runkle (1988) studied
the movement of real wages over the business cycle for a panel data drawn
from the National Longitudinal Survey of Young Men (NLS) from 1966 to 1981.
They showed that self-selection biased the behavior of real wage in a procyclical
direction.
PANELDATAMETHODS 3 I5
VIII. FURTHER READINGS
Supplementary readings on panel data include Hsiao’s (1986)Econometric Society

monograph, the standard reference on the subject, Maddala’s (1993) two volumes of
some of the classic papers in the field, a special issue of Empirical Economics edited
by Raj and Baltagi (1992), and a special issue of the Journal ofEconometrics edited
by Baltagi (1995a). Two recent books on panel data are Baltagi (1995b) and MAtyAs
and Sevestre (1996).
REFERENCES
Ahn, S. C. and S. Low (19%), A Reformulation of the Hausman Test for Regression Models
with Pooled Cross-Section Time-Series Data, Journal of Econometrics,71,309-319.
Ahn, S. C. and P. Schmidt (1995), Efficient Estimation of Models for Dynamic Panel Data,
Amemiya, T. (1971), The Estimation of the Variances in a Variance-Components Model, In-
ternational Economic Review, 12, 1-13.
Amemiya, T.and T. E. MaCurdy (1986), Instrumental-Variable Estimation of an Error Com-
ponents Model, Econometricu, 54,869-881.
Anderson, T. W. and C. Hsiao (1982), Formulation and Estimation of Dynamic Models Using
Panel Data, Journal of Econometrics, 18, 47-82.
Angrist, J. D. and W. K.Newey (1991), Over-identification Tests in Earnings Functions with
Fixed Effects, Journal of Business und Economic Statistics, 9,317-323.
Arellano, M. (1987), computing Robust Standard Errors for Within-Groups Estimators, Oxford
Bulletin of Economics and Statistics, 49, 431-434.
Arellano, M. (1989), A Note on the Anderson-Hsiao Estimator for Panel Data, Economics
Letters, 31,337-341.
Arellano, M. (1990), Some Testing for Autocorrelation in Dynamic Random Effects Models,
Review of Economic Studies, 57, 127-134.
Arellano, M.(1993), On the Testing of Correlated Effects with Panel Data, Journal of Econo-
metrics, 59, 87-97.
Arellano, M. and S. Bond (1991), Some Tests of Specification for Panel Data: Monte Carlo
Evidence and an Application to Employment Equations, Review of Economic Studies,
58,277-97.
Arellano, M. and 0. Bover (1995), Another Look at the Instrumental Variables Estimation of
Error-Component Models, Journal ofEconometrics, 68,2941.
Avery, R. B. (1977), Error Components and Seemingly Unrelated Regressions, Econometrica,
45,199-209.
Avery, R. B., L. P. Hansen, and V. J. Hotz (1983), Multiperiod Probit Models and Orthogonality
Condition Estimation, International Economic Review, 2 4 , 2 1-35.
Bailar, B. A. (1975), The Effects of Rotation Group Bias on Estimates from Panel Survey,
316 BALTAGI
Baillie, R. and B. H. Baltagi (1995), Prediction from the Regression Model with One-way
Error Components, Working Paper, Department of Economics, Texas A&M University,
College Station, Texas.
Baltagi, B. H. (1980), On Seemingly Unrelated Regressions with Error Components, Econo-
metrica, 48, 1547-1551.
Baltagi, B. H. (1981a), Pooling: An Experimental Study of Alternative Testing and Estimation
Procedures in a Two-way Error Components Model, Journal of Econometrics, 17,Zl-
49.
Baltagi, B. H. (1981b), Simultaneous Equations with Error Components, Journal of Econo-
rnetrics, 17, 189-200.
Baltagi, B. H. (1984), A Monte Carlo Study for Pooling Time-Series of Cross-Section Data in
the Simultaneous Equations Model, tnternational Economic Review, 25,603-624.
Baltagi, B. H. (1985), Pooling Cross-Sections with Unequal Time-Series Lengths, Economics
Letters, 18, 133-136.
Baltagi, B. H. (1995a), Editor’s Introduction: Panel Data, Journal of Econometrics, 68, 1-4.
Baltagi, B. H. (1995b), Econometric Analysis of Panel Data, Wiley, Chichester.
Baltagi, B. H. (1996), Testing for Random Individual and Time Effects Using a Gauss-Newton
Regression, Economics Letters, 50, 189-192.
Baltagi, B. H. and Y. J. Chang (1994), Incomplete Panels: A Comparative Study of Alternative
Estimators for the Unbalanced One- Way Error Component Regression Model, Journal
Baltagi, B. H. and Y. J. Chang (1996), Testing for Random Individual Effects Using Recursive
Residuals, Econometric Reviews, 15,331-338.
Baltagi, B. H., Y. J. Chang, and Q. Li (1992), Monte Carlo Evidence on Panel Data Regressions
with AR(1) Disturbances and an Arbitrary Variance on the Initial Observations, Journal
of Econometrics, 52, 371-380.
Baltagi, B. H. and J. M. Griffin (1988), A Generalized Error Component Model with Het-
eroscedastic Disturbances, International Economic Review, 29, 74.5-753.
Baltagi, B. H. and J. M. Griffin (1995), A Dynamic Demand Model for Liquor: The Case for
Pooling, Review of Economics and Statistics, 77, 545-553.
Baltagi, B. H. and J. M. Griffin (1997), Pooled Estimators vs. Their Heterogeneous Counter-
parts in the Context of Dynamic Demand for Gasoline, Journal of Econornetrics, 77,
303-327.
Baltagi, B. H., J. Hidalgo, and Q. Li (1996), A Non-parametric Test for Poolability Using Panel
Data, Journal of Econometrics, 75, 345-367.
Baltagi, B. H. and Q. Li (1990), A Lagrange Multiplier Test for the Error Components Model
with Incomplete Panels, Econometric Reviews, 9, 103-107.
Baltagi, B. H. and Q. Li (1991), A Transformation That Will Circumvent the Problem of Au-
tocorrelation in an Error Component Model, Journal of Econometrics, 48,385-393.
Baltagi, B. H. and Q. Li (1992), Prediction in the One-way Error Component Model with Serial
Correlation, Journal of Forecasting, 11, 561-567.
Baltagi, B. H. and Q. Li (1995), Testing AR(1) against MA(1) Disturbances in an Error Com-
ponent Model, Journal of Econometrics, 68, 133-151.
Baltagi, B. H. and N. Pinnoi (1995), Public Capital Stock and State Productivity Growth:,
Further Evidence from an Error Components Model, Empirical Economics, 20, 351-
359.
PANELDATAMETHODS 3 I7
Becketti, S., W. Gould, L. Lillard, and F. Welch (1988), The Panel Study of Income Dynamics
after Fourteen Years: An Evaluation, Journal of Labor Economics, 6,472-492.
Bhargava, A. (1991), Identification and Panel Data Models with Endogenous Regressors, Re-
view of Economic Studies, 58, 129-140.
Bhargava, A., L. Franzini, and W. Narendranathan (1982),Serial Correlation and Fixed Effects
Model, Review of Economic Studies, 49, 533-549.
Bhargava, A. and J. D. Sargan (1983),Estimating Dynamic Random Effects Models from Panel
Data Covering Short Time Periods, Econometrica, 51, 1635-1659.
Biorn, E. (1981),Estimating Economic Relations from Incomplete Cross-Section/Time-Series
Data, Journal of Econometrics, 16,221-236.
Biorn, E. (1992), The Bias of Some Estimators for Panel Data Models with Measurement Er-
rors, Empirical Economics, 15,221-236.
Bjorklund, A. (1985), Unemployment and Mental Health: Some Evidence from Panel Data,
The Journal of Human Resources, 20, 469-483.
Blundell, R. W. and C. H. Meghir (1990), Panel Data and Life-Cycle Models, in J. Hartog, G.
Ridder, and J. Theeuwes (eds.), Panel Data and Labor Market Studies, North-Holland,
Amsterdam, 231-252.
Blundell, R., C. Meghir, and P. Neves (1993), Labor Supply and Intertemporal Substitution,
Boumahdi, R. and A. Thomas (1991), Testing for Unit Roots Using Panel Data: Application
to the French Stock Market Efficiency, Economics Letters, 37, 77-79.
Breitung, J. and W. Meyer (1994),Testing for Unit Roots in Panel Data: Are Wages on Different
Bargaining Levels Cointegrated?, Applied Economics, 26,353-361.
Breusch, T. S. (1987), Maximum Likelihood Estimation of Random Effects Models, Journal
uf Econometrics, 36,383-389.
Breusch, T. S., G. E. Mizon, and P. Schmidt (1989), Efficient Estimation Using Panel Data,
Breusch, T. S. and A. R. Pagan (1980), The Lagrange Multiplier Test and Its Applications to
Model Specification in Econometrics, Review uf Econometric Studies, 47,239-253.
Butler, J. S. and R. Moffitt (1982), A Computationally Efficient Quadrature Procedure for the
One Factor Multinominal Probit Model, Econometrica, 50, 761-764.
Chamberlain, G. (1980), Analysis of Covariance with Qualitative Data, Review of Economic
Studies, 47,225-238.
Chamberlain, G. (1982), Multivariate Regression Models for Panel Data, Journal of Econo-
metrics, 18, 5-46.
Chamberlain, G. (1984), Panel Data, in Z. Griliches and M. Intrilligator (eds.), Handbook of
Econometrics, North-Holland, Amsterdam.
Chesher, A. (1984),Testing for Neglected Heterogeneity. Econometrica, 52,865-872.
Chow, G. C. (1960),Tests of Equality between Sets of Coefficients in Two Linear Regressions,
Chowdhury, G. (1991), A Comparison of Covariance Estimators for Complete and Incomplete
Panel Data Models, Oxford Bulletin of Economics and Statistics, 53,88-93.
Chowdhury, G. (1994), Fixed Effects with Interpersonal and Intertemporal Covariance, Em-
pirical Economics, 19,523-532.
Collado, M. D. (1995), Estimating Dynamic Models from Time-Series of Cross-Sections, Uni-
versity of Carlos 111 d e Madrid, Madrid, Spain.
322 BALTAGI
Ridder, G. (1990),Attrition in Multi-wave Panel Data, in J. Hartog, G. Ridder, and J. Theeuwes

(eds.), Panel Data and Labor Market Studies, North-Holland, Amsterdam, 45-67.
Robertson, D. and J. Symons (1992), Some Strange Properties of Panel Data Estimators, Jour-
nal of Applied Econometrics, 7, 175-189.
Robins, P. K. and R. W. West (1986), Sample Attrition and Labor Supply Response in Experi-
mental Panel Data: A Study of Alternative Correction Procedures, Journal of Business
and Economics Statistics, 4, 329-338.
Schmidt, P. (1983), A Note on a Fixed Effect Model with Arbitrary Interpersonal Covariance,
Journal of Econometrics, 22,391-393.
Schmidt, P., S.C. Ahn, and D. Wyhowski (1992), Comment, Journal ofBusiness and Economic
Searle, S. R., 1987, Linear Models for Unbalanced Data, Wiley, New York.
Sevestre, P. and A. Trognon (1985), A Note on Autoregressive Error Component Models, Jour-
nal of Econometrics, 28,231-245.
Sickles, R. C. (1985), A Nonlinear Multivariate Error Components Analysis of Technology
and Specific Factor Productivity Growth with an Application to U.S. Airlines, Journal
Sickles, R. C. and P. Taubman (1986),A Multivariate Error Components Analysis of the Health
and Retirement Study of the Elderly, Econometrica, 54, 1339-1356.
Swamy, P. A. V. B. (1971), Statistical Inference in Random Coeficient Regression Models,
Springer-Verlag, New York.
Swamy, P. A. V. B. and S. S. Arora (1972),The Exact Finite Sample Properties of the Estimators
of Coefficients in the Error Components Regression Models, Econometrica, 40, 261-
275.
Taub, A. J. (1979), Prediction in the Context of the Variance-Components Model, Journal of
Taylor, W. E. (1980), Small Sample Considerations in Estimation from Panel Data, Journal of
Valliant, R. (1991), Variance Estimation for Price Indexes from a Two-Stage Sample with Ro-
tating Panels, Journal of Business and Economic Statistics, 9,409-422.
Verbeek, M. and Th. E. Nijman (1992a),Testing for Selectivity Bias in Panel Data Models,
International Economic Review, 33,68 1-703.
Verbeek, M. and Th. E. Nijman (1992b),Can Cohort Data Be Treated as Genuine Panel Data?,
Empirical Economics, 17,9-23.
Verbeek, M. and Th. E. Nijman (1996), Incomplete Panels and Selection Bias, in L. Miityiis
and P. Sevestre (eds.), The Econometrics of Panel Data: A Handbook of Theory and
Applications, Kluwer, Dordrecht, 449-490.
Wallace, T. D. (1972), Weaker Criteria and Tests for Linear Restrictions in Regression, Econo-
netrica, 40,689-698.
Wallace, T. D. and A. Hussain (1969), The Use of Error Components Models in Combining
Cross-Section and Time-Series Data, Econometrica, 37, 55-72.
Wansbeek, T. J. and A. Kapteyn (1978),The Separation of Individual Variation and Systematic
Change in the Analysis of Panel Data, Annales de l’INSEE, 3631,659-680.
Wansbeek, T. J. and A. Kapteyn (1989), Estimation of the Error Components Model with In-
complete Panels, Journal of Econometrics, 41, 341-361.
Wansbeek, T. J. and A. Kapteyn (1992), Simple Estimators for Dynamic Panel Data Models
with Errors in Variables, in R. Bewley and T. Van Hoa (eds.), Contributions to Consumer
Demand and Econometrics: Essays in Honor of Henri Theil, St. Martin’s Press, New
York, 238-251.
Wansbeek, T. J. and R. H. Koning (1989), Measurement Error and Panel Data, Statistica Neer-
landica, 4 5 , 8 5 9 2 .
Wooldridge, J. M. (1995), Selection Corrections for Panel Data Models under Conditional
Mean Independence Assumptions, Journal of Econometrics, 68,115-132.
Wyhowski, D. J. (1994), Estimation of a Panel Data Model in the Presence of Correlation
between Regressors and a Two-way Error Component, Econometric Theory, 10, 130-
139.
Ziemer, R. F. and M. E. Wetzstein (1983), A Stein-Rule Method for Pooling Data, Economics
Letters, 11, 137-143.
Econometric Analysis in
Complex Surveys
Aman Ullah and Robert V Breunig

1. INTRODUCTION
In the last five decades there has been a significant growth of research in econo-
metric methods and their application in various areas of economics. Indeed, in the
last two decades, the tremendous growth in econometrics has dichotomized the sub-
ject into cross-sectional (micro) econometrics and time-series (macro) econometrics.
Whereas the new cross-sectional methodology was partly due to the nature of the data
and the empirical issues in microbased labor economics and industrial organiza-
tion, the new time-series methodology was an outcome of the challenging empirical
issues and data problems in macroeconomics and finance. Despite these develop-
ments, econometric inference methods (especially in cross-sectional econometrics)
have been confined to the assumptions that data is generated as a simple random
sample with replacement or that it is coming from an infinite population (Johnston
1991, Greene 1993).These assumptions are certainly not valid in the case of survey
data used in development and labor economics. Surveys usually have a well-defined
frame consisting of a finite population of individuals, households, or villages. Sample
data for analysis is generated from this finite population using a sampling design dif-
ferent from random sampling with replacement (RSWR). Sampling schemes such as
systematic sampling, stratified random sampling, and cluster sampling may be used
alone or in combination. These have been the subject of four decades of extensive
work in statistics literature (Kish 1965, Cochran 1953, Sukhatme 1984, Levy and
Lemeshow 1991, Thompson 1992).
The history of survey sampling can be traced back to the early eighteenth cen-
tury, and even earlier (see Hansen 1987 and Deaton 1994 for detailed references).
3 25
326 ULLAH AND BREUNIG
Engel (1895)used a nonrepresentative sample of 200 Belgian households to estab-

lish that the share of the budget allocated to food is higher for the poor. Kiaer (1897)
(see Kiaer 1976 for translation) was perhaps first to collect a large-scale, represen-
tative sample and discuss the principles, uses, and limitations of various sampling
designs. His paper was influential in the International Statistical Institute’s 1903
resolution supporting and promoting the use of the “representative method” of sam-
pling. A turning point in the history of survey sampling came in the 1920s with the
path-breaking work in statistical estimation theory and practice of R. A. Fisher at
the Rothamsted experimental station (Fisher 1925).Fisher emphasized randomiza-
tion, replication, and stratification in sampling design. His work led to the calcula-
tion of statistical estimates and their precision by Yates and Zacapony (1935) and
Cochran (1939).Indeed, Fisher’s work paved the way for Neyman’s (1934)classic
paper, which, for the first time, gave a systematic discussion of inference from ran-
dom samples drawn from a finite population, contained a comparison of purposive
sampling and random sampling, introduced the concept of the confidence interval,
established the asymptotic normality of the sample average, and provided the op-
timal sample sizes within strata independently of Tschuprow’s (1923)work. Later
Neyman (1938)developed the theory of what is known as “two-stage” sampling. In
the United States, around this period, important research work on sampling design
was done by the researchers at the Department of Agriculture and the statistical lab
at Iowa State University, and at the U.S. Bureau of the Census regarding labor force
surveys dealing with issues such as measuring employment and labor force partic-
ipation. These latter surveys, using systematic sampling designs, were necessitated
by the Great Depression; see Stephan (1948)for detailed references. Sampling with
probability proportional to measures of size at the successive stages of sampling was
introduced during the work on labor force studies.
A parallel development on the application of survey sampling took place under
the leadership of P. C. Mahalanobis at the Indian Statistical Institute in the 1930s
and P. V. Sukhatme at the Indian Council of Agriculture Research. Mahalanobis
introduced the concepts of developing sampling designs based on cost and variance
estimates and on methods of evaluating survey errors.
The work during the 1930s and the 1940s revolutionized the collection of
household survey data after the war. Major developments include the first national
sample survey data developed annually (1950-1970) and then every five years at
the Indian Statistical Institute, the household survey data now collected in United
States, United Kingdom, and Taiwan, the Living Standard Measurement Study
(LSMS) survey of Peru and the Ivory Coast by the World Bank, and Malaysian fam-
ily survey data by the Rand Corporation. These household data are now extensively
used in development economics to study poverty, income distribution, and economic
welfare.
A common feature of the survey data described above is that the ultimate
sample of households selected is rarely, if ever, based on RSWR. Yet, despite three
decades worth of developments in statistical inference based on survey data, econo-
ECONOMETRIC ANALYSIS
IN COMPLEX SURVEYS 327
metric analysis is carried on under the false assumption of RSWR; although see the
excellent works of Pudney (1989), Deaton (1994),and Howes and Lanjouw (1994)
for notable exceptions. This is especially a matter of concern in development eco-
nomics where measures of income inequality, poverty and elasticities are used in pol-
icymaking by governments and international agencies. We think there are perhaps
two reasons for this state of affairs. One is the statistical complexity of the various
sampling designs for an average development economist; the second is a complete
lack of exposure to the statistical literature on survey design in econometrics texts.
Given this deficiency a systematic development of the parametric and nonparametric
econometric inference (estimation and testing) of various econometric models, under
various practical sampling designs, is urgently needed. This is an ambitious project
and it is by no means attempted here. Instead, this chapter is a modest beginning in
this direction. Some new results are also presented. Essentially, our objectives are
as follows. The first is to provide a unified econometric framework of the five decades
of diverse statistical literature on estimating the population mean. We refer to this
as the mean model. The second is to explore the implications of results from the
mean model for the linear regression model and the nonparemetric kernel density
estimation. The third is to explore the implications of misspecifying the sampling
design on the properties of econometric estimators. It is hoped that this chapter will
contribute to further development of econometric inference results in other practical
econometric models and for other parameters of interest.
In Section I1 we present the estimation of the finite population mean, the den-
sity, and linear regression coefficients under RSWR and random sampling without
replacement (RSWOR). Section 111 deals with stratified sampling. Section IV con-
siders cluster sampling, systematic sampling, and two-stage sampling. In Section V
we give some limited simulation results. Finally, in the Appendix we provide some
technical details of the results in Section 11.
II. RANDOM SAMPLING

A. Random Sampling without and with Replacement
(RSWOR and RSWR)
Let us consider a finite population of size N for an economic variable Y , and write
the population mean model as
where Y; is the ith population observation, Ui is the ith error, and j3 and o2are the
population mean and variance, respectively, given by
N
N-1
S2
1= 1
N i= I
N
328 ULLAH AND 6REUNlG
where
N
i= 1
N-1
The errors U ; are nonsampling errors which sum to zero by the definition of B in
(2). Therefore, U ; and Y, are nonstochastic variables. However, if we treat the finite
population model (1) as generated from an infinite population or superpopulation
model, then U ; and Y, are stochastic. This case is not considered here.
A random sample without replacement of size n, often referred to in the liter-
ature as a simple random sample (SRS), is taken from the above finite population of
size N. We denote the sample observations as yi and write (1) for these observations
as
where U ; is now the ith sampling error.

Since the sampling is without replacement,
is the probability that the rth population unit is selected in the ith draw and
is the probability that the (T-,s)th unit is selected in the (i, j)th draw where i, j =
1 , . . . , n and r, s = 1 , . . . , N ( z # j , r # s). These probabilities provide
N
ITr = Cnr(i)
=-
n
N
i= 1
the probability of selection of the rth population unit in the sample of size n, and
ifj .
the probability of selection of the (r, s)th population unit in the sample.
In view of (4) to (7), we get
and, for i # j,
where p = -l/(N - 1) is the intraclass population correlation between y; and rj.

Further, we can verify that
Eu! = ( + 3)a4 = -
~ 2
1
N
E(Yi- /?)4
i= 1
1 N N
and for i # j # k # 1,
2y] (r3
____-
( N - 1) ( N - 2)
01234
2(Y2 +3) - N 04
(N - 1)(N - 2)
330 ULIAH AND BREUNIG
where y1 and y 2 are Pearson's measures of skewness and excess kurtosis. For nor-
mal distribution, y1 = y 2 = 0. The outcomes (8)to (11) indicate that RSWOR
represents a set of n identically distributed but correlated random variables yi.
In the case of RSWR the draws are independent, so
because a12 = 01112 = 0123 = 0 1 2 3 4 = 01123 = 0 and 0 1 1 2 2 = 04.This also

holds if we are sampling from an infinite population ( N + OO), the case usually
considered in econometrics. In what follows we analyze the effect of assuming RSWR
when the true design is RSWOR.
B. Estimation of Parameters
Let us write the model (3) in a more compact form as
where y is an n x 1 vector, 1 is an n x 1 vector of unit elements, is a scalar parameter

(finite population mean), and U is an n x 1 error vector such that
by using (8)and (9).

The n x n matrix Z can be rewritten as
c = a2(1- p ) I + _
[ P
_
1 - p"']
and its well-known inverse, using ( I + bb')-' =[I - bb'/(l + b; b)],where b is an

n x 1 vector, is
The least-squares (LS) estimators of B in (13) is obtained by minimizing u'u

with respect to /?.This gives
i= 1
IN COMPLEX SURVEYS 33 I
The estimator b is unbiased, Eb = /?,and its variance is
where the last equality gives the familiar expression of the variance of the sample
mean under RSWOR. The term 1 - n/N is known as the finite population correction
(fpc). For n + N , V ( b ) + 0.
The efficient generalized least-squares estimator of is
s = (l’x-ll)-+’x-ly = ( L ’ q ’ l ’ y = ji (19)
where the second equality follows by using (16).Thus, under (15),the two estimators
and their variances are the same.
When the sampling is RSWR or the population is infinite, C = 0’1 because
p = 0. In this case, Eb = B and
o2
V(6)= -
n
which also follows from the last equality of (18)where n/N + 0 as N + 00. From
(18)and (20)
The above results indicate that the LS estimator b is unbiased for both RSWOR
and RSWR. However, if the sampling is actually without replacement, the variance
formula in (20) is wrong and gives an overestimate of the correct variance (18).To
obtain the correct variance one needs to deflate (20) by ( N - n ) / (N - 1).For example
if n = 20 and N = 40 the correct variance will be approximately 50% smaller than
the wrong variance. The smallness of the variance of ~ R S V ( I O Ris due to the negative
correlation p.
In order to calculate the variance of b we look into an unbiased estimator of
S2. This is given by (using (13))
where M = I - u‘/n is an idempotent matrix. From the result ( 1 16) in the Appendix
it is straightforward to show that s2 is an unbaised estimator,
because M L = 0 and p = - ( N - l ) - * . For ratio estimators of S2 based on the

auxiliary information, see Prasad and Singh (1992).
Proposition I . When the sampling is without replacement we get
V ( S 2 )= 1
n
[{ y2 + 3 - n ?")( N - 1 J 0 4 (24)
Proof. V ( s 2 ) = Es2 - S4 = ( n - 1)-2E(u'Mu)2- S4. Now using (8)to (11)

and ( 118) in the Appendix we get (24).
Corollary I . When the sampling is with replacement we get
n+l N
N-1
This follows by substituting 01122 = o4 and 01112 = 01123 = 0 1 2 3 4 = 0

+
in (24). When the population is large V ( s 2 ) = a 4 ( y 2 / n 2 / ( n - 1 ) ) (Sukhatme
1984),which reduces to the well known result V ( s 2 ) = 2a4/(n- 1 ) under normality
( Y 2 = 0).
From the above results it is clear that, as in the case of the LS estimator b, the
estimator s2 is unbiased for both RSWOR and RSWR and the variance of s2 under
RSWOR differs from that under. RSWR. But, unlike b, the variance of s2 is affected
by nonnormality through y2. For example when the population is large the variance
of s2, compared to the normal case, is larger for all nonnormal distributions with yz >
0. Also the effect of nonnormality does not disappear for large n since the ratio of the
+
variance under nonnormality to the variance under normality is 1 y2(n - l ) / n ,
+
which converges to 1 y2 as n * 00.
Finally, we note that the estimator s2 is used to obtain an unbiased estimator
of V ( b ) :
N-n
P(b) = ___ s .
nN
This, however, does not guarantee that a(b ) be an unbiased estimator of

will ,/m;
see Ullah and Breunig (1996)for the magnitude of the bias. In the special
-
case of
sampling from an infinite, normal population, an unbiased estimator of J P ( b ) can
be obtained as s*/n, where
An application of knowing s2 and b is to obtain the estimator of the population

coefficient of variation (cv)
a
cv= B
This is given by
A S
cv = -
b
C. The Bias and MSE of the Coefficient of Variation

The sample c^v is extensively used in applied sciences for various purposes. In eco-
nomics c^v has been widely considered to compare inequalities in income of different
regions or groups of individuals or households (Kakwani 1980, Sen 1992). Beach et
al. (1994)consider the conditional cv for analyzing third-order stochastic dominance
when Lorenz curves cross. (We note that third-order stochastic dominance tests using
standardized income on income distributions with different mans can lead to test re-
sults which contradict the dominance ordering on the unstandardized distributions.)
Despite widespread use of the &, not much is known about its finite sampling prop-
erties under the nonnormal population; although see Bowman and Shenton (1981),
who have considered mean and variance of c^v under the Weibull distribution, Neuts
(1982), who has considered mixtures of distributions, and Singh (1993), who has an-
alyzed the sampling properties of the inverse of &-squared. Sampling of & under
normality has been well analyzed going back to McKay (1932). Warren (1982) con-
siders McKay’s results and provides other references. More recently, Gupta and Ma
(1996) analyze the & in k-normal populations. The results presented below could
be extended to the case of more than one variable.
Here we present the sampling properties of 8 = (c^v)2 of 0 = ( c v ) instead
~
of G. We analyze (&)2 because comparing 2 across regions/groups gives similar
results to comparing 8, and the approximate sampling properties of & require ex-
panding both s and l/b, whereas 8 require the expansion of l / b only.
In what follows, we present both small-a and Nagar-type large-n approxima-
tions for the mean and mean-squared error (MSE) of 8. For details on small-a ap-
proximation see Kadane (1971) and Srivastava et al. (1980), where they show a better
performance of the small-a approximation under certain situations.
Proposition 2. When the sampling is without replacement the bias of 8, up to
O(a4), and the M S E of 8, up to O(a4>, respectively, are given by
2 3r2 3e2
Bias(6) = -p8 - -0
n
(1 + 7(2 (29)
and
where
and 0's are as given in (9)to ( 1 1).

The proof of Proposition 2 is given in Appendix B.
Corollary 2. When the sampling is with replacement the bias of 8, up to O(04)
and the MSE of 8, up to O(04)),
respectively, are given by
The proof of Corollary 2 follows by substituting 01122 = o4 and 0123 =

01112 = 01123 = 0 1 2 3 4 = p = 0 in Proposition 2.
From (32)we can also obtain the bias, up to O(n-'), as
1
Bias(@ = -[302 - 2B3//"yl) (34)
n
This result compares with Breunig (1996),where the large-n expansion is used to
obtain (34). Breunig (1996)also provides the MSE, up to O(n-2),as
n
y2
+ 20+-
"1
+ 48 - 4y18'/' + 2-n - 1
n + l +20-) n
(35)
n-1 n-1
)
- 4e1l2 y 3 + 4 y l ___ 4-75@ - 10883/2y1
( n-1 1
From (34)we observe that the bias goes to zero as n + 00. Further, the asymp-
totic MSE (AMSE) is
Thus, for distributions with y2 > 0 (y2 < 0) the AMSE will be above (below) the
AMSE under normality. Further, we note that the bias, up to 0(a3)and O(n-’), is
positive for negatively skewed distributions, but the bias to O ( d ) is negative for pos-
itively skewed distributions. However, to O ( n - l ) , the bias is negative for positively
skewed distributions provided
2
cv = 01’2< ,)q (37)
Generally, 8 will give an underestimate of the true 0 for positively skewed distribu-
tions. Since income distributions are generally positively skewed, it is possible that,
in the past, the use of 8 in measuring income inequality gave an underestimation
of inequality. In view of this, Breunig (1996) hazuggested an estimator of 0 which
adjusts the bias of 8. This is given as = 8 - Bias(8), where BTs(8) is the bias of
8 in (34)with 8 replaced by 8 and y1 by p1 = C(Yi - T)’/ns^”.Although we do not
attempt it here, it will be interesting to analyze these results for sampling without
replacement. This could be done by using (29)and (30).Another possible extension
involves the use of the geometric mean, useful when the data are expressed in ra-
tio form. We could then formulate the sample cv as the ratio of the sample geometric
mean to the sample standard deviation. The authors are unaware of any development
of the finite-sample moments for such a statistic, even under random sampling with
replacement.
D. Sampling with Unequal Probabilities

When sampling is with unequal probabilities, nr(i) in (4), n, in (6),and 7 1 , s in (7)
are not constants. In this case, we first transform (13)by
and then obtain the LS estimator of /3 by minimizing the weighted squared error
U‘WU = (y - @ ) ’ W ( y - $I) ,
where W = Diag(w1, . . . , w,) is an n x n stochas-
tic diagonal weight matrix whose elements wi,known as the normalized expansion
factors, satisfy l‘wl = wi = 1.This gives the weighted LS estimator of /3 as
The stochastic weights w;are chosen such that (using (4) and (6))the sample is rep-
resentative of population in the sense that the sample mean, on average, is identical
to the population mean. That is,
/ n \ n N
\i=1 / i=l ;=1
This gives
1
w; = - (41)
Nni
and b , = N-’ C;(yJni), where ni is the probability of selection of the ith popu-
lation unit in the sample. When n; = n / N we get b , = b as given in (17).
An alternative way to obtain (41) is to write Eb, = E ( x ’ ; wiy;) = E ( x F
w;diY,)= Cr(wi.Ed;)U; = winiYi where di is a dummy random variable
which is 1 when Y, is in the sample, and 0 otherwise. Since the probability of selection
of the ith population unit in the sample is n;, we can verify that
cov(d,, d;) = n,;- n;n;
where Xi; is defined in (7). Using (42),
When ni = n / N and ni; = n ( n - l ) / N ( n - 1) we get (18).We point out here

again that both mean and variance of 6, contain U; since the error is coming from
the sampling, not from the errors in the superpopulation.
E. Estimation of Regression Model
Let us consider the mean of the conditional population of Y given a vector of k vari-
ables X I , . . . , X k as
Y = x*p +U (44)
where Y is an N x 1vector and X * is an N x k matrix. The model (44)is a conditional
mean model if the conditional mean of the nonsampling errors, E ( U l X ) , is zero or
/3 = (X *’ X*)-’X *’ Y is the finite population regression coefficient. As before, a

random sample without replacement of size n is taken from the population of size N.
The sample model is
y=xp+u (45)
where y is an n x 1 vector and X is an n x k matrix. Under RSWOR, E ( u ( X ) = 0,
and V ( u l x ) = E, where C is as given in (15).The LS and GLS estimators of /I and
their variances, respectively, are
b = (x’x)-’x’y, V ( b ) = (x’x)-’x’Ex(X’x)-’ (46)

bcrs = ( x ’ C - ’ x ) - l x ’ C - ’ y , V(bc,r,s) = (x’c-’x)-’
When sampling is with replacement, or as N + 00, p = 0 and we get b = bc1,s.

It is easy to verify that
V ( b ) = a2(1- p)(x’x)-’ + a2p(X’X)-’X’11’X(X’X)-~ (47)

+
= a2(x’X)-’ a2p(X’X)-1X’(ll’ - Z)X(X’X)
where 11’ - I is an indefinite matrix. Thus the sign of V ( b ) - a2(X’X)-’ is not

determined. This implies that if the sampling is without replacement and we consider
the variance of b as though we had RSWR then this variance will be an underestimate
or overestimate of the variance depending upon the design matrix X. Another point
is that .b(;ls is not identical to b numerically. These results are in contrast to the
mean model (13) where k = 1 and X = 1 . However, it follows from Zinde-Walsh
and Ullah (1987) and Ghosh and Sinha (1980) have shown that the F-test for testing
restrictions on /I using bcl-5, with C in (15),will be numerically the same as the F-test
based on b. This result also holds for the likelihood ratio, Lagrange multiplier, and
Wald tests.
F. Nonparametric Estimation of Density and Regression Functions

As in the earlier sections, let Y;, i = 1 , . . . , N , be the population of Y with proba-
bility density f,and yi = 1, . . . , n be the sample drawn without replacement from
f.Further, as in Section II.D, let d ; , i = 1 , . . . , N be the dummy random variables
such that d; = 1 if yi is selected in the sample and di = 0 otherwise. Note that the
di’s are not independent and Ed‘ = n/N and Edpd; = n(n - l)/N(N - 1) for any
p > 0 , q > 0.
where K ; = h,'K((Y, - Y ) / ~ N Rosenblatt's

). (1956)kernel estimator, based on y;,
is then
1 "
k; = -
Nh
l
cN
i= I
d;K; (49)
where k; = h,'K((y; -y)/h,) and h = n / N . Thus the finite population density esti-
mation problem reduces to the problem of estimating the population mean discussed
in (1) to (3).It therefore follows from the results in Section 1I.A that
where S i = ( N - l)-' cy(K; - K)2.However, if we treat the finite population Y,,

i = 1, . . . , N , itself as an i.i.d. random observation from an infinite (super) popula-
tion with density f *, then as ( n , N ) + 00 such that h + c > 0,
E f ( y ) = (Nh)-"(Ed,)
(E-K ('lb'))
N;
____ (51)
+f*
The details of (51) and (52) can be worked out by following Rosenblatt (1956) and
Pagan and Ullah (1995).For the asymptotic theory in the parametric, finite popula-
tion models, see Fuller (1984)and the references therein.
In the regression context we can write the nonparametric versions of the finite
population and sample models in (44)and (45), respectively as
y; = rn(X;) + ui (53)
and
where x, is a given point. More compactly,
y = ax,)&J +U (55)
where Z ( x , ) = [ l (x - LX,)], 6 ( x , ) = [ m ( x , ) rn'(x,)]', and V ( u ) = C as given
in (46).The local least-squares estimation of 6 ( x , ) can then be carried on by mini-
mizing U'C-'/'KC-'/~U.The properties of $(,) so obtained can be worked out by
using the procedures developed in Ruppert and Wand (1994).
111. STRATIFIED SAMPLING

A. Estimation of Mean Model
A large number of surveys use stratified random sampling, which involves dividing
the total population first into various strata (subpopulations), typically by geographi-
cal regions, such as rurahrban or states, or by certain characteristics such as blue-
collar and white-collar workers. Then RSWOR is used within each stratum. The
sampling units are usually households or individuals. There are several advantages
to this method as opposed to RSWOR or RSWR from the entire population. First,
the stratification provides a more representative sample overall and so reduces the
variance, especially when the variation within strata is small but the variation be-
tween strata is large. Stratification allows for different types of sampling schemes in
different strata, desirable perhaps because of cost considerations. For example, one
can perform SRS in urban areas, where households are closely concentrated, and
cluster sampling in rural areas, where households are widely dispersed (see Section
IV). Finally, stratification helps to obtain enough sample observations from small
subpopulations of special interest.
The population mean model for stratified sampling can be written as
X,=/?;+U,, i = l , ...,M, j = 1 , . . . , Ni (56)

where c, is the jth unit in the ith stratum, /?i is the mean of the ith stratum (subpop-
u lation),
and U,, is the error. The variance of the ith stratum is
In a more compact form
yi = pi + Ui (59)
where is an Ni x 1 vector of observations. The population size is N = Ni. E, M
The stratified sample observations, generated by RSWOR in each stratum,

follow
or, more compactly,

where ii is ni x 1 vector of unit elements, yi is an ni x 1 vector of sample observations

(yi;) in the ith stratum, and C; is C in (14) with n = n;. The total size of the sample
is n = ni. The model (60) for the ith subpopulation (stratum) is the same as
that of the population model in (13). Thus, the results for Section I1 go through for
the estimation of the ith stratum parameters. For example, the LS estimator of Bi is
and
Here, Mi is the matrix M in (22) with L = i i . The parameter of interest may, however,
be the overall mean of the population. That is,
M N, M
where 8i = N ; / N is the proportion of the total population in the ith stratum. An

estimator of B is then
which is unbiased. Further

M M
V(b,,) = C e : V ( b i ) = CO;
i= 1 i= 1
provided that the sample elements are chosen independently for each stratum.
To see the LS and weighted LS solutions of /3 we rewrite (60) as
yij =B + BT + uij (67)

where #?: = #?i - /I,Eui, = 0, Euz.LJ
= o;,Eu;juitjt = pia: for i = i’, j # j ’ , and 0
when i # i’. In the model (67), however, /3: and /3 are not identifiable. So we impose
the restriction Cr(n;/n)&+ = 0. Using this we get the LS estimators of pi and B by
minimizing Cy(yi; - /3 - /3,*)’. This gives bi* and hence bi as in (62), and
But n;/N;,the probability of selection of the jth population element in the sam-
ple of size ni in the ith stratum, may not be constant across strata. Therefore the
weighted LS described in Section 1I.D will be more useful here. For this we mini-
mize CYi wi,(y;j - j3 - s:)’with the restriction that w;jp; = 0 and
Cp;’ w;;= 1. This gives
M n;
and
This inflation or expansion factor w;;is chosen such that E b = #? and Eb; = pi.
This gives
If we substitute Tti; = n;/N;,we get bw = b,, and b;, = b;.

We observe that the LS estimator b will become unbiased if strata are homoge-
nous with regard to sampling fractions, ni/N; = n/N, in which case the combined
sample is SRS (this is known as proportional stratified sampling, a special case of
stratification in which the data will be self-weighting); or if @; = for all strata, in
which case the population is homogeneous with respect to means. Alternatively, if
we consider T
s to be random with mean zero and constant variance, then (67) can
be treated as a one-way error component model. In this case, the LS estimator b in
(68) is unbiased as well as efficient since b is also the GLS estimator in (67).
In contrast, if we assume #?i = #? and pool all the observations (ignoring strata
and treating the data as a random sample of size n from a population of size N ) , then
the variance of the pooled estimator b p = b, known as the SRS estimator, is
This is the same as V(b,,) or V ( b ) if n;/n = N ; / N and #?; = 6. In this case the
population is homogeneous and the combined sample is a simple random sample.
In general, V ( b s ~ s >
) V(bst),especially if within-strata heterogeneity is low and
between-strata heterogeneity is high. To see this, consider V ( b s L for
) case of propor-
tional stratified sampling, where n; = nN;/N. In this case
where we use S2 21 N-' ( Y i j - p)2 E +

0;s; @;(pi- /3)2. Thus,
if between-strata variation, V ( b ; ) ,is zero, V ( b s ~ s=
) V(b,,). But if V ( b i )is
. nonzero,
V(b,,) is smaller than V ( b s ~ s ) .
In order to minimize the variance of the weighted estimator, bst, one must
choose n; such that V(b,,) is a minimum subject to E, M
n; = n. The optimal n;
M
will be n; = nNis;/ x k = l nk.
In practice, the unbiased estimator of V ( b s l )can be calculated by substituting
S: with s? in (66). Further, if /?;= p, the unbiased estimate of the variance of b s ~ s
can be calculated as
Alternatively,
where s2 = ( n - M ) - ' C r ( n ;- 1)s; is the pooled estimator if S? = S2.If S; # S2

but ni/n = N ; / N , one may consider
However, if nil. # Ni/N and pi # p, then
where q is an unbiased estimator of ( N - l)-' Ni(pi - (Cochran 1953,

p. 99).
From the above analysis it is clear that if the sample observations y;i are gen-
erated by stratified random sampling then they should be reweighted to resemble
the population by replicating (inflating) sampling units, using the inflation or ex-
pansion factor, and treating the enlarged sample as if it were the population. The
inflation factor, O i j , for each sampling unit j in the ith stratum is the reciprocal of
its sampling probability; that is, Oi, = l/nij = N i / n ; . If we multiply each sample
observation by its inflation factor 8;;. we obtain an unbiased estimate of the popu-
lation total. Alternatively if we multiply the sample observations by their weights
wij = Oy/ Oi, = l/Nnij, the normalized inflation factor, we get an unbiased
estimate of population mean, as shown in (72) and Section 11. Exactly the same pro-
cedures can be used to obtain estimates of medians, variances, and other parameters.
We will examine the weighting for regression parameters.
B. Regression Model
Suppose now that the parameters of interest are no longer population means, but the
parameters of a linear regression model
where pi is a k x 1 vector and i = 1, . . . , M strata. As in the mean model the

parameters of interest are pi and p, where p is the overall parameter or the weighted
average of pi:
M
i= 1
By the analogy of the population mean case one may consider 0; = N i / N . If the
population is stratified on some economic grounds such as rural and urban, then the
estimates of will be useful in their own right. However, if the population is divided
into a large number of strata on administrative grounds, then studying /?will be more
meaningful. The estimates of are
bi = (xi'xL>-'xL!yi (80)
with V ( b ; ) = aT(Xl!Xi)-'. If we have RSWOR bi = ( X ~ C ~ ' X ; ) - ' X ( C ~ with
'yi
V ( b ; ) = ( X f C ; ' X i ) - ' , X i is C with 1 replaced by 1; and n by ni.
To estimate /3 we write
or, more compactly, for i = 1, . . . , M ,

where y is n x 1 , X is n x k, and X* is n x kM,n = n;. Then the pooled

(RSWR) estimator is
M -1 M
bp = bRsWR = ( x ' x ) - ' x ' y =
/ M \-' M M
provided 0: = 02, Q; = XL!X;/n;,and w; = (CE, piQ;)-'p;Qi. For RSWOR,

b, = bsRs = ( X ' S t - ' X ) - ' X ' S t - ' y , where St = Diag((aTZ;)). It is easy to see that
b i s a biased estimator. This bias is
However, if pi - p = 6; is random and independent of Xi then the bias is zero.

Also, the bias vanishes if pi = p. When X is stochastic, we are concerned with the
consistency of our estimator not its bias. We thus want to consider the probability
limit as n; + 00.
So far, the two situations under which the bias vanishes mimic the results from
the mean model. Also, as before, the bias of pooled, or unweighted, estimator bp
depends on the structure of the sample. This is so even when Q; = Q for all i.
However, unlike in the mean model case, even if n;/n = N i / N = 0; the bias still
depends on the sample. When Q; = Q and n;/n = N i / N , the bias vanishes provided
/3 is defined as in (79).
Now consider the weighted estimator of /3 by minimizing u'WU, where W =
( ( ~ i j ) ) wij
, = l/(nij N ) . This gives
*
6, = ( X ' W X ) - ' X ' W y (85)

When k = 1 and X = 1, we get the same result as in Section 11. Further, if ni; =
n;= n ; / N i , then
and
so that the bias does not depend on the sample values; wi is the same as in (83) with
pi replaced by 0i. Furthermore if Bi - p is random, Q; = Q or pi = p, then the bias
vanishes. For Q; = Q the bias vanishes when wi = 8;. Note that for = /?both
b, and b p are unbiased but b p will be more efficient by the Gauss Markov theorem
(DuMouchel and Duncan 1983). Kish and Frankel (1974), however, argue in favor
of b, and the parameter of interest
i= 1
where wi is as in (86). When Q; = Q, wi = 0; and (88)reduces to (79). Note that b,

is unbiased only if the parameter of interest is ,d in (88). If the parameter of interest
is given by (79), b, will not be unbiased.
In practice, one may be able to make a choice between bp and b, by comparing
bp and b, and performing a Hausman-type specification test. This amounts to doing
an artificial regression of the type
y= xp + m y + U (89)
and testing for y = 0 by the standard F-test. Alternatively, one can combine the two
by using Stein-type shrinkage estimators as
bs = (1 - +) bp + i b ,
The properties of bs are not known, but they can be developed by following Judge
and Bock (1978) or Vinod and Ullah (1981).
Magee et al. (1996)suggest an alternative estimator to weighted least squares
when the sampling probabilities are known but the form of the sample design is un-
known. They propose a conditional maximum likelihood estimator, which, under cer-
tain conditions, is superior (in the mean-squared error sense) to WLS or OLS. They
treat the weights (sampling probabilities) as having been generated by a stochas-
tic process and independently distributed throughout the population, an assumption
which is violated under either stratification or clustered sampling. They suggest us-
ing their ML estimator in any case, since information about the sample is usually
unknown. Magee (1996) and Magee et al. (1996) suggest a way to improve weighted
least-squares regression for survey data. Including weights often injects additional
346 ULlAH AND BREUNIG
heteroskedasticity into the model and the WLS estimator, though consistent, often
has a high variance. Magee (1996)suggests creating new weights by multiplying the
weights by a function parameterized so as to minimize variance. Again, indepen-
dence across the sample is assumed, necessitating some adjusting of the procedure
for use under stratification or clustered sampling.
For further discussion of regression, including inference on finite population
and superpopulation parameters, Kalton (1983) provides a clear and readable in-
troduction, Pfefferman (1993) reviews some of the recent work on regression mod-
els and weighting, Selden (1994) considers the case of weighted, generalized least
squares in the mean model case, and Godambe (1995,1997)provides a more general
model.
The estimators b p , b,, and bs are all weighted averages of 6;. Another weighted
average of bi follows by considering pi - /3 to be random with mean vector 0 and
diagonal covariance matrix A,. Thus, heterogeneity across strata is due to variance
+ +
only. In this case, X ; ( p ; - p ) ui has the variance a: X;AX,/ and we can get the
GLS estimator of j3 as
i= 1
+
where wi = ( [ C I ( A p &)-']-'(As + i$)-';& = 0f(XL!Xi)-').This is the well-
known Swamy (1971) estimator. The estimators b p and 6, will both be inefficient
under this scenario. Their standard errors need to be calculated by using a White
(1980) kind of adjustment.
The above procedures are useful when heterogeneity across strata is present.
This is important when the number of strata are large. If the number of strata are
small, one could do separate regressions, combine them by allowing stratum-specific
intercepts, or use strata dummies with the same variables.
Although we do not explore questions of prediction here, they will be important
in choosing the parameters of interest. The census parameter, /?in (79)would be of
interest when one would like to predict the change in the independent variable for a
small change in the dependent variable for every member of society. If a particular
policy would affect some strata differently than others, this might not be a parameter
which aids in prediction.
For the purpose of prediction, however, nonparametric analysis will be more
useful since it directly estimates the regression function and thus avoids the prob-
lem of defining the parameter of interest. This is not taken up here, but will be an
interesting subject of a future study.
Also not considered here are the usual diagnostic tests used in both parametric
and nonparametric ecomometric analysis. Obviously, these too must be adapted to
account for complex sampling procedures.
IV. CLUSTER AND SYSTEMATIC SAMPLING

A, Cluster Sampling
Let us consider Yij to be population observations of the ith group or cluster, i =

1, . . . , M and j = 1, . . . , N; elements in the ith cluster. Cluster sampling or single-
stage sampling involves drawing a sample of m clusters out of M and then sampling
all N;, i = 1, . . . , m, elements in the sample. The sampling is done, therefore, only
at the first stage. However, if we further take a random sample of n; elements out of
Ni at the second stage, then the overall sampling is called subsampling or two-stage
sampling. The first-stage units, clusters, could be villages or street blocks, and the
second-stage units could be households. The primary advantage of cluster sampling
is that it drastically reduces survey cost per second-stage unit. The disadvantage is
that it usually leads to higher variance due to correlation among the elements within
clusters. The standard assumption of uncorrelated observations in cross-sectional
data is certainly not true for clustered samples.
It is useful to begin by first considering the problem of estimating population
means b and pi under single-stage (cluster) sampling. For this we consider the pop-
ulation model as xj = /I !If, + + U , and the sample model as
y;j = j3 +j3: +U;,, i = 1 , . .. ,m , j = 1 , . . . , Ni (92)

where we assume RSWR such that
Euij = O , Eu:, = 02, Euijuijl = p~ 2 , j # j'
(93)
Eu..u.l
51 1 J
' I = 0, i # if
p > 0 is the intracluster correlation coefficient. Equation (93) implies that the ele-
ments within clusters are correlated but are uncorrelated across clusters. Thus,
Eu = 0 and Euu' = Diag(C1, . . . , E m ) = Q (94)

where S2 is an n x n block diagonal matrix with C; = a2[(1- p ) I + piii:]. The LS
estimators of /3; and /3, respectively, are
If /3; = /3, the estimator 6, is unbiased. Further, its variance is given by
where n = Cr;ZNi,and E = n-'Cr;ZNn? is the weighted mean of cluster sizes. As

shown in Section 11, the LS estimator b, is the same as the GLS estimator for the
structure of X i in (94). An alternative expression of (96) can be written by following

the results in Section 1I.A. This is V ( b c ) = o i / m where 0; = C;"(bi- b,)2/M.
For the equivalence of this expression with (96) see Kish (1965).
If we ignore the clustering and consider RSWR of size n, then
Thus, the usual formula of RSWR has to be inflated by d = 1 +

( E - 1)p) to
account for the intracluster correlation. For p 3 0, Eq. (97) may provide serious
underestimation of the true variance in (96).
It can be shown that
is an unbiased estimator of a:, where iiij = yij . b,. In practice, one could consider
i2'a/n, which is consistent. Further, for i = 1, . m,
* 7
gives
Deaton (1994) provides numerical examples of the effect of ignoring p in the calcu-
lation of standard errors by 6,. He showed that for estimated food price elasticities
in Pakistani villages, p is between .3 and .6, leading to underestimation of V ( b c ) by
a factor greater than 2 when the mean cluster size is 12. Now we turn to the case of
two-stage sampling where, within each selected cluster, we pick a sample of n; < Ni
units. The probability of selection of every element in the chosen cluster is mni/MNi.
It is easy to verify that the estimator of the mean /3 is 62s = E'; b;/m, where bi is
C:i yi;/ni. For the variance of b2s, see Kish (1965).
B. Systematic Sampling
Systematic sampling is one of the most common techniques used in development eco-
nomics. In systematic sampling, the sampling units are (usually) arranged in random
order with respect to the variable of interest. Of the first K units, one is selected at
random. Then every Kth unit is sampled in order. This sampling design is the easiest
to implement because it involves drawing only one sample. Systematic sampling can
be thought of as a kind of one-stage cluster sampling. The population is arranged into

K clusters, each with n elements. One of these clusters is chosen and every eIement
within that cluster is sampled. For simplicity, we assume that there are N elements
in the population and that N / K is an integer. (In other words, the N elements are
exhaustively and uniquely assigned to the K clusters, each of which has the same
number of elements n, an integer.)
Consider the finite population model Y = B +
U , where the data are ran-
domly ordered. The population mean and variance are defined as in Section I. If the
population is divided into K clusters, we can write the model as
Our sample would consist of one randomly chosen cluster k, written as
3 = /9 + uj, j = 1, . . . , n
The LS estimator of the mean, when cluster k is chosen, is
which will be unbiased when N / k is an integer, Further,
In general, we will not be able to estimate this variance. In the case where our data
consists of one systematic sample, the population mean, p, is unknown as are the
remaining K - 1 unsampled clusters. In some surveys, resampling is possible. In
this case, information can be gathered about the within and across-cluster hetero-
geneity and an approximation for V ( b s y s ) as a function of the intracluster correlation
coefficient, p :
In general, if the data is randomly arranged with respect to the variable of

interest, systematic sampling should give broad coverage of the population, the esti-
mate of the mean should be unbiased, and the approximation of V ( b s y s ) by assuming
simple random sampling will not be too unreasonable. If the data is clustered and the
clusters are ordered, then a systematic sample will perform better than either SRS or
clustered sampling. This follows because the systematic sample, picking every kth
elements, will cover most, if not all, clusters. This very broad coverage will give a
precise estimate of the mean.
350 ULLAH AND BREUNlG
C. Regression Model
In the regression model
where X is n x k, n = C’;n;, the LS and GLS estimators of /? and their variances,

respectively, are
6, = b u = (X’X)-’X’y, ) (X’X)-’X’StX(X’X)-’
V ( b ~ s= (106)
and
52 is as defined in Section 1II.B. Suppose X contains observations which vary only

across clusters but are constant within clusters. Further assume that the cluster sizes
are the same so that ii = n. Then it follows from Kloek (1981)that the LS estimator
is numerically the same as the GLS estimator. Further, the true variance of b, is
V(b,) = a 2 ( X ’ X ) - ’ d (108)
which reduces to the variance in the mean model where k = 1 and X = 1. Moulton
(1990)and Deaton (1994)provide examples of potential underestimation, using the
usual variance o~(x’x)-’.
The results for the case of observations in X changing with clusters or when
n; # n is not well developed in the literature, although see Pfefferman and Smith
(1985)who provide an upper bound of V(bc).Also, the efficiency of b ~ compared u
to b u needs to be analyzed.
An alternative to estimating St by first estimating p and cr2 is to consider $2 to
be of unknown form and then estimate X’QX consistently by X ’ h X , as suggested
by White (1980)and Arellano (1987).
D. Unequal Probabilityof Selection

Consider again the mean model
yij = p + ST + uij (109)

where cj?; = 0. Then the LS estimators b; and b of pi and j?, respectively, are
as given in Section 1V.A. In practice, two-stage sampling is carried out by using
the probability weight proportional to the size of the cluster. In this situation the
estimator b will be biased. An unbiased weighted estimator is
where ni is the probability of selection of the ith cluster in the sample. For RSWR,
7 t i = m/M. We note that wij = (Z'ij)-l, where Pi; is the probability of selecting
yi; = probability of selecting the ith cluster x probability of selecting the jth unit
of the population in the sample given the ith cluster is selected. This gives
If the cluster is chosen with probability proportional to size of the cluster (i.e., 7ri =
mNJN = mei,
1 "
b, = - bi
m
i=1
In practice 7ri is proportional to estimated size N;*. This gives
where 87 = N r / N * , N* = N:.
When we consider single-stage sampling or cluster sampling, the estimator bw
can be written as
For the RSWR xi = m / M .

In the regression case
yi = XiBi + ui
bi = (XfXi)-'Xfyi, and we can consider b, = C y ( B i b i / n i ) as before (Konijn
1962).For cluster sampling we can again consider 6, = m-l E';"bi/ni.
We can also estimate j3 by considering to be random so that yi = X i s ; +
+ +
= X;B Xi& U i , where 6; = B; - B. This will involve Swamy-type random
coefficient estimation of B described in Section I11 (also see Porter 1973).
Fuller (1975)estimates the one-way error component model
yi = Xi/3 + eili + ui
to capture the correlation across the elements of the ith cluster, where ei has mean
zero and constant variance.
352 ULIAH AND BREUNIG
V. SIMULATION
In this section, we present a summary of the results from a detailed simulation of the
mean model under complex sampling. The two primary objectives of the simulation
are (1) to illustrate the effect of ignoring sample design in data analysis, and (2)
to ascertain the properties of our estimators under various sample designs where
analytical results do not exist.
The first step in our simulation was the creation of several finite 44populations.”
The populations created ranged in size from 50 to 20,000, with means ranging from
1 to 2000. They were drawn from an “infinite” population of randomly distributed,
normal numbers. From these finite “populations” we then drew n observations, us-
ing the sampling design in question. For the questions under consideration in this
section, the shape of the distribution is irrelevant, so only normal random numbers
were considered. For analyzing other variables, such as the distribution of V ( s 2 )
considered in Section I, the shape of the distribution does matter, and conclusions
based on simulations using only normally distributed populations should be made
with caution.
To demonstrate the first point, a sample of size n was drawn using the sample
design of interest (stratified, clustered, etc.), then b and V ( b ) were estimated using
the information on how the sample was drawn. Then taking this same sample and
ignoring the sample design, we calculated ~ R S W Rand V(bRSWR)-i.e., treating the
sample as if it were a random sample drawn with replacement. These values are aver-
aged over 1000 repetitions. We then compare the average bias (b)and bias ( ~ R S W K )
and the ratio of the averages of the two variances, which can be interpreted as the de-
gree of over- or underestimation arising from ignoring (or misspecifying) the sample
design.
The second type of simulation we have undertaken, to answer the second ques-
tion raised above, involves drawing a separate sample for each sampling design of
interest. From these distinct samples, we calculate b and V ( b ) for each of the sample
designs. After r repetitions, we compute the average bias and the simulation variance
of the estimator for each design.
By way of example, let us consider a simulation of both kinds comparing sam-
pling under finite population and infinite population from Section I. Recall that
Following the first method, we take 1000 samples of size n from our population and
calculate V(bs~s),using the fact that the sample has been drawn without replace-
ment from a finite population. Then we calculate V ( ~ R S Wand
R ) compare the ratio of
the average of these two over the 1000 repetitions.
The ratio should be the inverse of the finite population correction. Indeed
this is confirmed in the results in Table 1. Results from the second type of simu-
Table I Effect of Ignoring Sample Design (Sampling with

Replacement)
Sample size
Population
size 5 10 20 50 100
50 1.11 1.25 1.67 - -

100 1 .OS 1.11 1.25 2.00 -
500 1.01 1.02 1.04 1.11 1.25
1000 1.00 1.01 1.02 1 .OS 1.11
Table entries show the degree of overestimationof the variance of the sample
mean: var(bRSIVR)
V ~ I ( ~ S R S )’
lation are presented in Table 2. Here, two separate samples are drawn from the same
population-one under SRS, the other under sampling with replacement. Results for
10,000 repetitions are reported. As expected, the results closely approximate those
in Table 1. One way to interpret these results is that for the same sample size, sam-
pling without replacement is more precise than sampling with replacement. (Since
the variance of the estimator b under SRS is, on average, smaller.) We can also think
of the ratio as representing the “cost” of assuming that sampling is from an infinite
population when in fact sampling is from a finite population.
Tables 3 through 6 present comparisons between RSWR and stratified sam-
pling without replacement, conducted under the first method described. In Table 3,
we consider two strata, each with a population of 1000, where a sample of size rt;
is drawn from each stratum. Since rtl # rt2 and both strata have a population of
1000, the sampling probability is unequal. As we saw in Section 111, the unweighted
estimator of /3 will be biased. As we can see from Table 3, the more unequal the
Table 2 Efficiency Gains from Sampling without Replacement vs.

Sampling with Replacement
Sample size
Population
size 5 10 20 50 100
50 1.08 1.22 1.63 - -

100 1.063 1.082 1.24 1.98 -
500 .993 1.009 1.054 1.12 1.253
1000 1.OO 1.oo 1.005 1.OS5 1.108
Entries in table are ” ~ $ ~ averaged

~ ) over 10,OOO repetitions of each design.
354 AND BREUNIG
ULLAH
Table 3 Sampling with Unequal Probabilities

~~~~~~ ~
Penalty of not considering sampling structure
Sample size Bias Ratio of variances Ratio of MSEs

(Stratum 1,2) Bias 6, ~RSWR var(6~s~~)/Vur(6,)MsE(b~sw~)/MsE(b,)
(5310) 0.94 17.97 1.46 1.81
(5920) -0.13 30.11 1.15 2.99
(5950) 0.30 41.62 0.60 4.75
(1095) 0.24 -16.62 1.23 1.37
(2075) 1.16 -29.68 0.72 1.97
(5095) -0.99 -41.37 0.27 3.07
2 Strata: j31 = 200; a1 = 59; j32 = 300; 02 = 74.8; j3 = 250.
sampling probabilities, the greater the bias in the unweighted estimator, ~ R S W R and
,
the greater the ratio of mean-squared errors. Most data in labor and development
economics is stratified, and the most common case is unequal sampling probabil-
ities, either by design or because of different rates of nonresponse across strata.
Thus, as the simulation shows, a potentially serious bias problem exists even in cal-
culating a simple mean. The general intuition behind these results extends to the
regression case.
In Tables 4 to 6 , we present results from the stratified case, but with equal
probabilities of selection in both strata. In Table 4, we see that even though ~ R S W R
Table 4 Sampling with Equal Probabilities
Population Ratio of variances Ratio of MSEs

Sample sizes means Bias bRSWR Var (bRS WR)/ M S E (6RS W R)/
(Stratum 1,2) (Stratum 1,2) = Bias 6, Var (b w ) MSE(6w )
(5, 5 ) (ZOO, 300) 0.33 1.59 1.30
(200, 400) 0.29 2.29 1.64
(200,600) 0.12 3.53 2.24
(200, 800) 2.22 5.46 3.19
(100, 100) (ZOO, 300) -0.20 1.74 1.36
(ZOO, 400) 0.15 2.37 1.67
(ZOO, 600) 0.05 3.68 2.35
(200, 800) 0.26 5.97 3.32
Table 5 Improved Efficiency: /31 # ,&

Population
Sample sizes means Ratio of variances Ratio of MSEs
(Stratum 1,2) (Stratum 1,2) Var(b~sw~)/var(b,) M~E(~Rs~R)/K~E(~
(57 5 ) (200,300) 2.20 1.56
(200,600) 20.22 10.44
(200,1000) 74.95 37.82
(20,20) (200, 300) 2.16 1.54
(200,600) 18.44 9.67
(200, 1000) 68.06 34.86
(50,50) (200,300) 2.20 1.61
(200, 600) 18.97 9.76
(200,1000) 69.58 34.51
(100,100) (200,300) 2.32 1.66
(200,600) 19.81 10.05
(200,1000) 72.75 35.55
(unweighted) is unbiased under equal sampling probability, it is not efficient com-

pared to 6,. In Tables 5 and 6, we consider across-strata heterogeneity in the mean
and the element variances separately, since both affect the ratio between ~ R ~ Wand R
bw. In Table 5 we first consider the case where a: = of,but /31 # /32. Table 6
presents the case where (T: # U ; , but = B 2 . We note that the increase in pre-
cision as measured by the ratio of variances is increasing as the distance between
the two strata means, /31 and P 2 , increases. It is not uncommon in development eco-
Table 6 Improved Efficiency: a: # a;

Ratio of variances Ratio of MSEs

Sample sizes Population U* Bias bRSwR I/ar(b~~w~)/ MsE(b~sp~)/
(Stratum 1,2) (Stratum 1,2) = Bias b , Var(b,) MSE( b W )
(57 5 ) (2500,3520) - 12.3 1.01 1.OO
(2500,9650) 13.4 1.01 1.OO
(50, 50) (2500,3520) -5.5 1.05 1.03
(2500,9650) -6.9 1.OS 1.03
Table 7 One-Stage Clustered Sampling

~ ~ ~~~~~~~~
# of Ratio of Expected
clusters Total True variances kish design
sampled sample size pop. Pop Var(bCLu)/ effect
(4 (n = c * = m) mean p P var(bRSWR) (4
5 100 1000 .12 .088 3.13 3.28
1000 .17 .13 5.00 4.23
1000 .26 20 6.48 5.94
1000 .44 .344 7.24 9.36
1000 .48 .38 10.02 10.12
10 200 1000 .12 .ll 2.39 3.28
1000 .17 .15 3.89 4.23
1000 .26 .24 5.78 5.94
1000 .44 .41 10.36 9.36
1000 .48 .45 11.78 10.12
25 500 1000 .12 .119 3.69 3.28
1000 .17 .169 4.29 4.23
1000 .26 .256 5.42 5.94
1000 .44 .439 9.69 9.36
1000 .48 .48 10.09 10.12
nomics to encounter stratified samples where the urban mean income is three times
that of rural mean income. In the case where = 200 and 82 = 600, we see that
this can lead to an overestimate of the variance of the population mean by a factor of
20. It increases for large sample sizes, because the finite population correction has a
proportionally larger effect. From Table 6, we see that stratification does not improve
efficiency when the strata have the same mean regardless of the difference in within-
stratum variance. The small increases in efficiency that we see are the result of the
increasing effect of the finite population correction as the sample size increases. In
other simulations, not reported here, we show that assuming a stratified structure
when the data does not have one leads to no gains in efficiency. This result follows
intuitively from the results in Table 6.
Table 7 presents the cost of ignoring the one-stage clustered sample design
and assuming that the sample is actually a RSWR. Here we have drawn a clustered
sample from one stratum with a population of 1000,which is divided into 50 clusters,
each of size 10. We present the ratio of variances, V ( b c ) / V ( b ~ s w where
~ ) , we have
calculated V ( b c ) using the estimated sample value of 8. We compare this to the
+
expected Kish design effect 1 (rn - 1 ) p given our knowledge of the true value of
p: 8 gives a slight underestimation, which disappears as n + N .
Table 8 Comparisons of SRS, Clustered, and Systematic Sampling Designs
Ratio of variances
# of Total True
clusters sample pop. Population SRSI SRS/ Clustered/
sampled size mean P clustered systematic systematic
5 100 10oO .12 .28 1.31 4.60

1000 .26 .15 1.88 12.4
1000 .48 .09 2.30 25.74
1000 .64 .06 2.92 47.37
10 200 1000 .12 .24 1.09 4.52
loo0 .26 .ll 1.69 14.7
1000 .48 .08 6.24 81.15
1000 .64 .07 1.04 14.78
Tables 8 and 9 present results comparing RSWOR (SRS),stratified sampling,

cluster sampling, and systematic sampling. We use the second method described
above for simulation.
Table 8 presents results from 5000 replications comparing SRS, clustered, and
systematic samples drawn from the same population. The last three columns com-
pare the variances of the different estimators. Systematic sampling performs best
in the simulation reported in Table 8. Since the data is ordered by cluster, the sys-
tematic sample gives the broadest coverage of the population, taking at least one
observation from each cluster. (See Section 1V.B) As expected, clustered sampling
gives the highest variance for given sample size. Table 9 compares stratified and sys-
Table 9 Systematic Sampling Compared with Stratified Systematic Sampling
Ratio of
Sample size variances Ratio of MSEs
(for each Strata Strata Bias Bias Vur(bsyS)/ MsE(bRswR)/
stratum) means pop- bw bsys VWbfU) M S E ( bW )
5 (2,900) 100 3.21 -0.09 1.06 1.03
(1000, 1000) 500 0.09 -0.54 1.05 1.02
25 (2,900) 100 4.11 -3.11 2.55 1.78
(1000, 1000) 500 -0.17 -0.04 1.46 1.23
50 (2,900) 100 1.89 0.30 1.89 1.44
(1000, 1000) 500 -0.16 0.77 1.74 1.37
tematic sampling in a stratified population. Results are for 10,000 repetitions. The
systematic sample does not perform as well as the stratified random sample since
stratification will more evenly cover the population over repeated sampling.
As the results of Sections I through IV demonstrate, problems of inference and
estimation arise when data is gathered under a complex sampling design. The simu-
lation helps to demonstrate that these are of more than trivial interest. Unequal sam-
pling probabilities are the rule, not the exception, and treating such data as having
been drawn under RSWR will lead to biased estimators. Even where the dispropor-
tion is 2 to 1, this leads to large bias as shown in Table 3.
As different strata will usually have different parameter means, ignoring strat-
ification will lead to large overestimates of the true variance of our estimate of ,!?. A
recent survey of income in Kenya showed that average rural income was one-third
that of average urban income. Ignoring stratification when calculating a population
mean in this case will lead to confidence intervals which are 20 times too wide. The
exact opposite problem occurs in clustering. Intraclass correlation coefficients of .5
are common in developing country studies. The simulation shows that ignoring the
sample design leads to an underestimate of the variance by a factor of 10, more if
the average cluster size is greater than 20.
Bias problems and misestimates of standard errors are exacerbated in more
complex sample designs which combine different aspects of stratification, cluster-
ing, and systematic sampling. Clearly, the same problems will arise in the regression
case. The simulation demonstrates that assuming away sample design effects as triv-
ial is unjustified. Instead, more careful attention should be paid to using available
methods of analysis and information on sampling to construct unbiased and more
precise estimates.
VI. APPENDIX
A. Some Useful Expectations
Suppose the elements of an n x 1 vector U satisfy (8)to (11). Let A and B be n x n
symmetric matrices of known constants, b an n x 1 vector of known constants, 1 an
*
n x 1 vector unit of elements, and A B the Hadamard product of A and B. Then
we can verify the following expectations:
E(u’Au) = E
[:+x ] [: + $
= a’tr A
uiujaij
pa’(1’ A1
=E
- tr A )
u?aii ~i~jaij] (116)
= a2(1- p)tr A + p&’ At

r n n n 1
E ( U’Au)(u’Bu)
When sampling is with replacement, 0 1 1 2 2 = o4 and p = 0 = 0112 =

01112 = 0 1 2 3 = 01123 = 01234. In this case (116) to (118) reduce to the standard
results in the literature:
E(u’Au) = o2tr A (119)

E(u’Au)(b’u) = YIO‘L’(Z* A)b (120)
E(u’Au)(u’Bu)= y2a4 tr(A * B ) + a4(tr A tr B + 2 tr A B ) (121)
B. Proof of Proposition 2
Let us use (2) and (17) in (28) and write the estimator 8 = ( G ) 2 of 8 = (cv)’ as
where = I - M = u’/n. Further we write the model (13)as
where U is the error vector where moments are determined by the moments of U in
+ +
(8)to (11).Then y’My = a2v’Mu and y’My = nB2 0(2/3u’i) 02u’Mv, and these
give, up to 0(04),
= a2a2 + 2 , s + a4a4
where
3 -
04a4 = -(u’Mu)(u‘Mu)
n284
Now, using (116)to (118)we can easily verify that

1
a2Ea2 = - O ( l - p)(n - 1)
n
ACKNOWLEDGMENTS
The research on this topic is an outcome of the first author’s visit to the Policy
Research Division of the World Bank and helpful discussions with C. Howes, E.
Jimenez, and M. Ravallion. The authors are thankful for comments by participants of
seminars at York University, University of Windsor, UCR, and University of Guelph
and participants in the Canadian Econometric Studies Group 19% meeting. Special
thanks to Gordon Anderson, David Giles, V. P. Godambe, and Lonnie Magee. Fi-
nancial support from the Academic Senate, UCR, is gratefully acknowledged by the
first author.
REFERENCES
Arellano, M. (1987),Computing Robust Standard Errors for Within-Group Estimators, Oxford

Bulletin of Economics and Statistics, 49, 431-434.
Beach, C. M., R. Davidson, and G. A. Slotsve (1994), Distribution Free Statistical Inference
for Inequality Dominance with Crossing Lorenz Curves, manuscript, Vanderbilt Uni-
versity, Nashville.
Bowman, K. 0. and L. R. Shenton (1981), Moment Series for the Coefficient of Variation
in Weibull Sampling, American Stntisticnl Association-Proceedings of the Statistical
Computing Section, 148-153.
Breunig, R. (1996), Almost Unbiased Estimation of an Inequality Measure, manuscript, Uni-
versity of California, Riverside.
Cochran, W. G. (1953), Sampling Techniques, Wiley, New York.
Cochran, W. G. (1939), The Use of the Analysis of Variance in Enumeration by Sampling,
Journul of the American StatisticaZ Association, 34, 492-510.
Deaton, A. (1994), The Analysis of Household Surveys: Microeconometric Analysis for De-
velopment Policy, manuscript, Princeton University.
DuMouchel, W. H. and G. J. Duncan (1983), Using Sample Survey Weights in Multiple Re-
gression Analysis of Stratified Samples, Journal Americun Statistical Association, 78,
535-543.
Engel, E. (1895), Die Productions-Und Consummations-Erhaltnisse des Konigreichs Sach-
sen, in Die Lehenkosten helgischer Arhielter FU (E. Engel, ed.), Dresden.
Fisher, R. A. (192S),Statistical Methodsfor Reseurch Workers, Oliver and Boyd, Edinburgh.
Fuller, W. A. (1975), Regression Analysis for Sample Survey, Sankhya: The Indian Journal of
.Statistics, Series C, 87, 117-132.
Fuller W. A. (1984), Least Squares and Related Analyses for Complex Survey Designs, Survey
Methodology, 10,97-I 18.
Ghosh, M. and B. K. Sinha (1980), On the Robustness of Least Squares Procedures in Re-
gression Models, Jortrrzal ojMultiuariate Analysis, 10,332-342.
Godambe, V. P. (1995), Estimation of Parameters in Survey Sampling: Optimality, Canadian
Journal OjStatistics, 23,227-243.
Godambe, V. P. (1997), Estimation of Parameters in Survey Sampling, Journal ofthe American
Statistical Association, forthcoming.
Greene, W. M. (1993), Econometric Analysis, 2nd ed., Macmillan, New York.
Gupta, R. and S. Ma (1996), Testing the Equality of Coefficients of Variation in K Normal
Populations, Communications in Stutistics-Theory and Methods, 25, 115-132.
Hansen, M. (1987), History of Survey Sampling, Statistical Science, 2, 180-190.
Howes, S. and J. 0. Lanjouw (1994),Making Poverty Comparisons Taking into Account Survey
Design: How and Why, manuscript, World Bank.
Johnston, J. (1991), Econometric Method.s, 3rd et]., McGraw-Hill, London.
Judge, G. G. and M. E. Bock (1978), The Statistical Implications of Pre-Test and Stein Rule
Estimators in Econometrics, North-Holland, Amsterdam.
Kadane, J. B. (1971), Comparisons of k-Class Estimators When the Disturbances are Small,
Kakwani, N. (1980), Income Inequality and Poverty: Methods of Estimation and Policy Appli-
cations, World Bank, Oxford University Press, Oxford.
Kalton, G. (1983), Introduction to Survey Sampling, Sage University Paper #35.
Kiaer, A. N. (1897), The Representative Method of Statistical Surveys (translation 1976, orig-
inal in Norwegian), Kristiania Videnskabsselskabets Skrifter, Historisk-jlosojske Klasse,
4: (34-56), Statistisk Sentralbyra, Oslo.
Kish, L. (1965), Survey Sampling, Wiley, New York.
Kish, L. and M. R. Frankel(l974) Inference from Complex Samples, Journal of Royal Statis-
tical Society, Series B, 1, 1-35.
Kloek, T. (1981), OLS Estimation in a Model Where a Microvariable Is Explained by Aggre-
gates and Contemporaneous Disturbances are Equi-correlated, Econometric, 49,205-
207.
Konjin, M. S. (1962), Regression Analysis in Sample Surveys, Journal of the American Sta-
tistical Association, 57, 590-606.
Levy, P. S. and S. Lemeshow (1991), Sampling of Populations: Methods and Applications, 2nd
ed., Wiley, New York.
Magee, L. (1W6), Improving Survey-Weighted Least Squares Regression, manuscript, Mc-
Master University.
Magee, L., A. L. Robb, and J. B. Burbidge (1996), On the Use of Sampling Weights
When Estimating Regression Models with Survey Data, manuscript, McMaster Univer-
sity.
McKay, A. T. (1932), Distributions of the Coefficient of Variation and the Extended t-Distri-
bution, Journal of the Royal Statistical Society, 95,695-698.
Moulton, B. R. (1990), An Illustration of a Pitfall in Estimating the Effects of Aggregate Vari-
ables on Microunits, Review of Economics and Statistics, 72,334-338.
Neuts, M. (1982), On the Coefficient of Variation of Mixtures of Probability Distributions,
Communications in Statistics-Simulation and Computation, 11,649-657.
Neyman, J. (1938), Contribution to the Theory of Sampling Human Populations, Journal of
the American Statistical Association, 33, 101-1 16.
Neyman, J. (1934), On the Two Different Aspects of the Representative Method: The Method of
Stratified Sampling and the Method of Purposive Selection, Journal of Royal Statistical
Society, 97, 558-606.
Pagan, A. and A. Ullah (199S), Nonparametric Econometrics, manuscript, Australian National
University, Australia.
Pfefferman, D. (1993), The Role of Sampling Weights When Modeling Survey Data, Interna-
tional Statistical Review, 61, 317-337.
Pfefferman, D. and T. M. F. Smith (1985), Regression Models for Grouped Populations in
Cross-Sectional Surveys, International Statistical Review. 53, 37-59.
Porter, R. D. (1973), On the Use of Survey Sample Weights in the Linear Model, Annals of
Economics and Social Measurement, 212, 141-158.
Prasad, B. and M.P. Singh (1992), Unbiased Estimators of Finite Population Variance Using
Auxiliary Information in Sample Surveys, Communications in Statistics-Theory and
Methods, 21,1367-1376.
Pudney, S.(1989),Modeling Individual Choice: The Econometrics of Corners, Kinks, and Holes,
Basil Blackwell, Oxford.
Rosenblatt, M. (1956), Remarks on Some Nonparametric Estimates of Density Function, An-
nals of Mathematical Statistics, 27, 832-837.
Ruppert, D. and M. P. Wand (1994), Multivariate Locally Weighted Least Squares Regression,
The Annals of Statistics, 22, 1346-1370.
Selden, T. (1994), Weighted Generalized Least Squares Estimation for Complex Survey Data,
Economics Letters, 46, 1 4 .
Sen, A. (1992), Inequality Reexamined, Russell Sage Foundation, Oxford University Press,
New York.
Singh, M. (1993), Behaviour of Sample Coefficient of Variation Drawn from Several Distribu-
tions, Sankhya: The Indian Journal of Statistics, 55,65-76.
Srivastava, V. K., T. D. Dwivedi, M. Beluisky, and R. Tiwari (1980), A Numerical Comparison
of Exact Large Sample and Small Disturbance Approximations of Properties of K-Class
Estimators, International Economic Review, 21,249-252.
Stephan, F. F. (1948),History of the Uses of Modern Sampling Procedures, Journal ofAmerican
Statistical Association, 43.
Sukhatme, P. V. (1984), Sampling Theory of Surveys with Applications, Iowa State University
Press, Ames, IA.
Swamy, P. A. V. B. (1971), Statistical Inference in Random Coeficient Regression Models,
Springer-Verlag, Berlin.
Thompson, S.(1992), Sampling, Wiley, New York.
Tschuprow, A. (1923), On the Mathematical Expectation of the Moments of Frequency Distri-
butions in the Case of Correlated Observation, Metron, 2,461493,646-680.
Ullah, A. and R. Breunig (19%), On the Bias of the Standard Errors of the LS Residual and
the Regression Coefficients under Normal and Nonnormal Errors, Econometric Theory,
Problems and Solutions, 12, 868.
Vinod, H.D. and A. Ullah (1981), Recent Advances in Regression Methods, Marcel Dekker,
New York.
Warren, W. G. (1982),On the Adequacy of the Chi-Squared Approximation for the Coefficient
of Variation, Communications in Statistics-Simulation and Computation, 11, 659-
666.
White, €I. (1980). A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity, Econometrica, 48, 817-838.
Yates, F. and I. Zacapony (193.5), The Estimation of the Efficiency of Sampling with Special
Reference to Sampling for Yield in Cereal Experiments, Journal of Agricultural Sci-
ence, 25,543-577.
Zinde-Walsh, V. and A. Ullah (1987),On Robustness ofTests of Linear Restrictions in Regres-
sion Models with Elliptical Error Distributions, in Time-Series Econometric Modeling
(I. B. Macneil and G . J. Umphrey, ttds.), Reidel, Holland.
10
Information Recovery
In Simultaneous-Equations Statistical Models
Amos Golan
University of California at Berkeley, Berkeley, California and
American University, Washington, D.C.
George judge
University of California at Berkeley, Berkeley, California
Douglas Miller
Iowa State University, Ames, Iowa
1. INTRODUCTION
For the last five decades a significant portion of econometric effort has been directed
to
1. Develop ways for analyzing multiple-equation statistical models that de-

scribe a range of underlying economic data generation processes
2. Develop a corresponding basis for estimation and inference to be used in
data reduction and information recovery
Despite the productive efforts of many, questions remain concerning the insecure
assumptions underlying the sampling theory, likelihood, and asymptotic approaches
and the usefulness of traditional multiple-equation estimation and inference proce-
dures in helping us find order when using the partial-incomplete underlying eco-
nomic data that is normally found in practice. Against this backdrop, we propose a
new method of estimation in multiple-equation statistical models that is widely ap-
plicable because it does not require the specification of a parametric family for the
likelihood function. The estimation rule is robust with respect to likelihood, is flexi-
3 65
366 GOLANETAL.
ble with respect to the dynamic, stochastic, and feedback nature of economic data as
well as to the introduction of prior information, and is computationally simple. Using
linear and quadratic risk measures, we compare the finite-sample performance of this
method to other widely used traditional estimation rules.
The organization of the chapter is as follows: In Section I1 the simultaneous-
equations statistical model is introduced, traditional estimation rules are identified,
and corresponding asymptotic and finite-sample performances are noted. In Sec-
tion 111 the maximum-entropy approach to recovering information in the case of
inverse problems with noise is formulated and corresponding asymptotic sampling
properties are developed. In Section IV sampling experiments are proposed as a
basis for comparing finite-sample performance of the alternative estimation rules,
and the resulting empirical sampling-risk results are evaluated. Section V contains
summary comments and recommendations.
II. THE SIMULTANEOUS-EQUATIONSSTATISTICAL

MODEL (SESM) AND TRADITIONAL ESTIMATORS
To provide a format for analyzing the SESM that reflects an instantaneous feedback
mechanism between some of the variables in the stochastic system of equations,
consider a statistical model consistent with the data generation process for a system
of G simultaneous equations:
where Y is a T x G matrix of observations on G endogenous variables, X is a T x K

matrix of observations on the K exogenous predetermined variables, I' and B are
comformable matrices of unknown parameters, and E is a T x G matrix of unobserv-
able equation errors. The rows of E are assumed to be independently distributed,
with joint density function that has a zero mean vector and a G x G covariance ma-
trix C.The reduced form counterpart of (1)is
where nr -B or
D has a joint density function f ( [ O ] , 52) and Y has some corresponding density func-
tion f ( Y l n , 52).
For single-equation-analysis purposes, it is conventional to assume that r con-
tains -1's on the diagonal and to rewrite the ith equation as
INFORMATIONRECOVERY
IN STATISTICAL MODELS 367
where yi and x represent the endogenous jointly determined variables in the ith
equation and X ; represents the exogenous predetermined variables in the ith equa-
tion, XT represents the exogenous predetermined variables appearing in the system
but not included in equation i, and X represents the K exogenous predetermined
+
variables in the system of equations. Let 2; be a T x (G; K ; ) matrix representing
the G; endogenous Y; and K ; exogenous predetermined X ; variables that appear in
the ith equation with nonzero coefficients. Further, 6; = (gl, bl)' is a (Gi + Ki)-
dimensional vector of unknown and unobservable parameters corresponding to the
endogenous and exogenous variables in the ith equation, and ei is a 3"-dimensional
random vector for the ith equation that is traditionally assumed to have mean 0 and
scale parameter oi;. The variables yi, Y;, Xi are observed, and 6; and e; are unob-
served and unobservable.
Given the ith equation (4), the complete system of G equations may be writ-
[:]
ten as
+ (5)
eCI
or compactly as
y=Zb+e
where, in the context of the traditional SESM, y and e are GT-dimensional random
+
vectors and 6 is an unknown and unobservable Ci(Gi K;)-dimensional vector.
Traditionally e is assumed to have mean 0 and cov(e) = C @I I T , where C is a
G x G unknown covariance matrix. Given (5) and (6),the corresponding system of
reduced form equations may be written as
where n is a GK-dimensional vector of unknown parameters and the random vector

d has mean vector 0 and cov(d) = !2 @I ZT.
From a sampling-theory point of view, besides direct least squares (DLS), the
following estimators for 6 that seek the consistency property are relevant: two-stage
368 GOWNET AL.
least squares (2SLS), k-class (KC), limited-information maximum likelihood (LIML),

three-stage least squares (3SLS), and full-information maximum likelihood (FIML).
These estimators are widely reported and reviewed in econometrics books such as
Judge et al. (1985) and Davidson and MacKinnon (1993). In terms of sampling prop-
erties, FIML estimation of the SESM pro’duces consistent and asymptotically effi-
cient structural parameter estimates given that the distribution function is correctly
speczjed. A nice review of finite-sample results for sampling-theory SESM estima-
tors is given by Zellner (1996), Tsurumi (1990), and Park (1982). These authors also
review Bayesian approaches to the SESM and report conditional and unconditional
performance results for the Bayesian estimators, and White (1994) develops prop-
erties for the quasi-ML approach. The experimental design underlying the finite-
sample results reported in these papers and the papers by Cragg (1966, 1967)form
the basis for the maximum-entropy estimator sampling performance results given in
Section IV.
111. A MAXIMUM-ENTROPY FORMULATION

AND ESTIMATOR
As an alternative to traditional frequentist and Bayes estimators for the unknown

structural parameters 6; and 6, we consider a variation of the generalized maximum
entropy (GME) estimation rule, proposed by Golan, Judge, and Miller (1996).This
estimation rule is based on the entropy measure of Shannon (l948), the maximum
entropy (ME) formalism of Jaynes (1957a, 1957b, 1984), Levine (1980), Shore and
Johnson (1980),Skilling (1989), and Csiszhr (1991),and the cross-entropy principle
of Gokhale and Kullback (1978), Good (1963), and Kullback (1959).
In formulating the GME estimator for the SESM, the possible outcomes for
the unknown and unobservable parameters and equation errors are viewed proba-
bilistically. To reflect this specification, the statistical model (7), involving a single
equation from the system of reduced-form equations, is reparameterized as
where Sq is a block-diagonal matrix reflecting transformation of the possible out-

comes of each n i k to the interval [0, 13 by defining a set of M 3 2 discrete support
points s z = [ s ; ~ , s z 2 , . . . , s Z M ] ’ and a vector of corresponding unknown weights
(probabilities) pz = [ p Z l , p z 2 , . . . , pfkM]’,such that
Similarly, V, is a matrix that provides a transformation of the possible outcomes

for d;, to the interval [0, 11 by defining a set of J 2 2 discrete support points
IN STATISTICAL MODELS
INFORMATIONRECOVERY 369
, . . . , vitj]’, distributed ungormly and evenly around zero, and a vec-

vit = [ v ; , ~vit2,
tor of corresponding unknown weights (probabilities) w;, = [will, wit27 . . . , witj]’
such that
i
When prior information is available, consistent with the set of discrete points
s$ in (9), this may be specified by corresponding prior probabilities q$ =
T I T
[qikl, q&2, . . . , q;,,,]’. Also, consistent with the set of discrete points V i l , we may
specify corresponding prior probabilities uit = [ u i t l ,u i ~. ., . , U ; ~ J ] ’ .For the com-
plete system of reduced-form equations, the statistical model (7) is reparameter-
ized as
where ST and V are block-diagonal matrices with Sn and Vi as the blocks.

Further, even though we do not make use of the structural equations directly,
r,
we reparameterize B and or similarly 6, in the following way. Let
rn
with rn = 1 , 2 , .. . , M and M 2 2.
To simplify notation, we also use the definitions
k = 1 , 2 , . . . , Gi, i = 1 , 2,...,G (13)
and
which are the two subsets of 6;.

Using these definitions and the corresponding reparameterizations, the gener-
alized maximum-entropy estimation problem for estimating a system of G equations
is formulated as
Max H(p*, w, p’) = -pX’ In pX - w’ In w - ps’In ps (15)
pK, W , P J
subject to the GT reparameterized reduced-form equations
y = ( I G @ x ) s X p n I/w+ (16)
the structural equation to reduced-form restrictions (3),reflecting (13) and (14),
m n n r n
370 GOLANET AL.
and the adding-up conditions

(16'; (8 1L)p: = l i G , (18)
( I K ;(8 1 d ) p q = l i K , (19)
(IT 8 1;)Wi = 1 i T (20)
where l i t is an l x 1 vector of unit values.
Carrying through the first-order conditions and solving the resulting system
yields the solution
where X[., i] is the ith column of X and Xi is a T x 1 vector of Lagrange multipliers

for equation 2's constraints in (16):
exp (- x
k @ik*nksynm) - exp (- xk @ik*nkSynm)
Pynm =
cm
exp (- x
k @ik*nksrnm) - Q,Y,( f i i )
(23)
and
where pi is the vector of Lagrange multipliers associated with (17) that refers to equa-
tion i, and Q(.) and @(.) are the partition (normalization) functions for the proba-
bilities.
Finally, using (9), (lO), and (12) yields
ei= S"pn 1 1 (25)

ai = Vtwi (26)
8; = s"p; (27)
or Ti =-Sy'py and A = SiBf piB .
To obtain the GME solution it is sufficient to have two points in the support
spaces that convert the elements from the real line into a [0, 13 space. The larger the
number of support points, the more moments can be captured in the optimization-
estimation process.
If prior information exists concerning the structural or reduced-form parame-
ters or the noise components, this information can be introduced through the prior
probabilities q:, q:, and ui. This information can be used in the following general-
ized cross-entropy (GCE) formulation:
INFORMATION RECOVERYIN STATISTICAL MODELS 37I
subject to (16)-(20). Carrying through the first-order conditions and solving the re-
sulting system yields the estimated probabilities
and
In a similar way, estimates are provided for the structural parameters pJ' and p B .
Finally, the point estimates 5,d, and 6 (or 'i/ and 8)
are recovered as in (25)-
(27).
A. Discussion and Large Sample Properties
As shown in Chapter 6 of Golan, Judge, and Miller (1996), for the traditional lin-
ear statistical (regression) model, the Hessian matrix of the GCE problem is posi-
tive definite for pi, wi >> 0, and thus satisfies the sufficient condition for a unique
global minimum. When prior information does not exist, both q i k and uit become
uniform (e.g., q i k m = 1/M for all rn and k ) and the GCE solution is equivalent
to the GME solution. Although the GME-GCE solutions do not have a closed
form, the dual unconstrained formulation proposed by Miller (1994) and Golan,
Judge, and Miller (1996) may be used to evaluate the sampling behavior of the
solutions within the extremum of M-estimation framework (Huber, 1981). In gen-
eral, the GME-GCE solutions may be viewed as discrete members of the ex-
ponential family of probability distributions, and these functional forms may be
used to relate the original parameter vector, pi, to the dual parameters, Ai
and pi. The large-sample properties of the GME-GCE estimators are based on
the asymptotic behavior of the dual parameters, and the relationship follows the
corresponding results in the exponential family literature (Brown 1986, Johansen
1979).
Following Golan, Judge, and Miller (1996, Chap. 6), we develop the dual un-
constrained GME-SESM. Given the Lagrangian for the optimization problem (15)-
(20), we substitute the maximum-entropy posteriors probabilities (21)-(24) into the
Lagrangian where, forsimplicity, we use pB and p)' instead of p'. Further, since these
372 GOWNETAL.
posteriors already satisfy the adding-up requirements, Eqs. (18)-(20) are omitted
from the Lagrangian. Using some simple algebra, one gets the dual unconstrained
problem. Specifically,
L ( X , p ) = -pn’ In pn - pDrIn pD - p ~In’ p~ - wrIn w
i k m
r 1
i k i t
i k i n
Minimizing the dual unconstrained GME model with respect to X and p yields fi
and p , which, in turn, yield fin, pJ’,p p , and*.Investigating the concentrated, or
dual, objective function (31) reveals the following properties. First, letting G = 1,
the system reduces to the simple (one equation) linear statistical model where the
last two terms disappear and the summation and indices involving i are deleted from
the first three, Thus, we have the GME estimator for the linear statistical model.
Second, the first three terms correspond to the reduced-form system of equations
and involve the data y’X and the sum of the partition functions for 7r and w, re-
lNFORMATlON RECOVERYIN STATISTICAL MODELS 373
spectively. This part can be viewed as an empirical likelihood function (Golan and
Judge, 1996a) for the reduced-form equations. Third, the last two terms correspond
to the definition (3) or its reparameterized form (17). There is no noise component
involved in these two terms, so they are related to the classical (pure) maximum-
entropy formulation.
What remains is to show that (i) .it is a statistically consistent estimator of m,
B
and (ii) i3/ and are consistent internally estimates of y and 8. Part (i) is a trivial
generalization of Golan, Judge, and Miller (1996, Proposition 6.3, p. 104). Part (ii)
follows the principle of classical (pure) maximum entropy. That is, given the esti-
mated 7r, which serves as the data for the (pure) ME problem, the entropies of y and
j3 are maximized. This ensures estimates that can be realized in the greatest number
of ways consistent with the data (Jaynes 1957a,b, Levine 1980, Golan, Judge, and
Miller 1996, Chap. 3). Furthermore, for those equations that are exactly identified,
the maximum-entropy approach yields the exact mathematical inversion.
An alternative consistency motivation may be based on the following heuristic
argument: Given the value of m, the problem (15)-(20) can be viewed as equivalent
to maximizing (15)subject to
y - ( l c 8 X ) m = vw (32)
7r = snpn (33)
and (17)-(20). The solution is then one of choosing pT with maximum entropy -pn’
In pT to satisfy (32), with the remainder of the problem (w and p6) being separable
from the choice of pn. Given the continuity of the constraint functions in the original
P
problem, and since .ir 7r, it is reasonable to view the problem of deriving w and
__+
p b through (32), (33),and (15)-(20) as leading to the informational consistency of

the separable maximum-entropy problem relating to /?and y.
B. Remarks
In applied work, many times emphasis is focused on one structural equation in the
system of equations. For this situation it is traditional to use the 2SLS method of
moments (Hansen 1982) or LIML estimators. Within a GME context several pos-
sibilities exist, and a range of these is discussed in Chapter 12 of Golan, Judge,
and Miller (1996). One straightforward GME possibility is just to make use of the
information, yi, zi, in the structural equation of interest. Although this formulation
ignores X:, the exogenous variables in the remainder of the system, the sampling
results provided by Golan, Judge, and Miller (1996) suggest it performs well relative
to traditional sampling-theory competitors.
A GME single-equation formulation and estimation rule consistent with the
objectives of LIML may be developed as a special case of the complete system of
374 GOLANET AL.
equations formulation (15)-(27). In this single-structural-equation formulation, the

objective function identifies the ith equation and it is maximized subject to (6),the
zero-restriction condition for the zth equation (17), and appropriate adding-up con-
ditions. Although this format is compatible with the LIML formulation, the distri-
butional (and other) specification requirements and the solution basis differ signifi-
cantly.
Finally, we note the relationship between the general method of moments
(GMM) estimator (Hansen 1982) and the GME estimation rule. To develop this rela-
tionship, we consider a single overidentified structural equation and premultiply by
the idempotent matrix X(X’X)-’X’, as proposed by Basmann (1957), to rewrite the
statistical model as
x’z;s;pi+ z,’x(x’x)-’x’v,wi
z,’x(x’X)-’x’yi= zL!x(x’x)-’ (34)
Letting Qj = 0 for all j yields the pure moment condition
ix (X’X)- ’X’ZlSiPi
zix (X’X)- I x’y; = z (35)
which, with the addition of a reparameterization for 6i,is identical to the usual first-
order conditions for the GMM estimator. If (35)replaces the relaxed moment condi-
tion (34) as the consistency relation in the GME formulation, then traditional GMM
estimates for Si result. If the relaxed moment relation (34) is used in the GME-GCE
formulation, and the bounded parameter space S contains the true parameter vec-
tor S;, then asymptotically the resulting estimates have, under standard regularity
conditions, the same large-sample properties as the GMM estimators (Judge et a1
1988, pp. 641-643). In finite samples, sampling results presented in Golan, Judge,
and Miller (1996, Chap. 12) suggest that, under a squared error measure, GME is a
superior performing estimator.
IV. SAMPLING EXPERIMENTS
Although analytic small-sample results are available in a few cases, much of the in-
formation that we have about the finite-sample properties of simultaneous-equations
estimation rules comes from sampling experiments conducted over a four-decade
period. Despite the usefulness of these studies, many questions are unresolved. For
example, when considered in a loss-risk context, the rankings of traditional estima-
tion rules remain somewhat in doubt. To obtain some experience with the MaxEnt
estimator specified in Section 111, and to gauge how they compare performance-wise
with traditional sampling-theory rules, we conducted a limited range of sampling
experiments. These experiments focused on some of the special characteristics of
nonexperimental economic data such as small samples, collinear relations among
INFORMATION RECOVERYIN STATISTICAL MODELS 375
variables, and the lack of independence between some of the right-hand-side vari-
ables and the equation errors. As a basis for judging estimator performance, we use
the quadratic loss measures 118 - 811’for some unknown 8.
A. Sampling Design
In the sampling experiments, we work with a linear simultaneous-equations model
involving three structural equations. The model follows Tsurumi (1990)and is a mod-
ification of the model structure employed by Cragg (1967).In the context of (l),
4.4
0 .74 0
r= ( -1.0
.222
0
.267
-1.0
.046 -y7)
-1.0
B=
.7
0
0 .53
0 .ll
.96 .13 0
0 0 .56
0
Using zero restrictions, all three equations are overidentified.

The exogenous variables x,2, . . . , xt7 are drawn from a normal ( 0 , l )distribu-
tion. To reflect the structure of the correlation among the exogenous variables, we use
the condition number K ( X ’ X ) , which is the ratio of the largest and smallest singular
values of X with columns scaled to unit length. We use K ( X ’ X ) = 1 for orthonormal
X’s and the condition number K ( X ’ X ) = 90 for moderately high collinearity. Each
experiment to be reported involved 1000 samples of size T = 20. The errors for
the structural equations were drawings from a multivariate normal distribution with
mean zero and covariance C 8 IT with
C= ( 1
-1
-.125
-1
4
0.625
-.125
.0625
8
(37)
For comparison purposes, the results for an alternative covariance and drawings from
a t ( 3 ) distribution are also reported. In addition to the zero and normalization restric-
tions, the support spaces specified for the structural and reduced-form parameters
and equation errors are s$ = s$ = [-5, -2.5,0, 2.5, 51’ fork = 2, 3, . . . , 7; s z =
s$ = [-20, - 1 O , O , 10,201’ for k = 1; and srn = [-2, - 1 , O , 1,2]’ for i, n =
1 , 2 , 3 .The errors’ support V is specified as v;, = [-3o,, 0, 3oM]’,where ofiis the
empirical standard deviation of yi. For all experiments, the sampling performances
(empirical risks) are reported for the whole system of structural equations and for a
single parameter in (1). For comparison purposes, 3SLS results that use the COT-
376 GOLANET AL.
rect covariance matrix (37) are reported. To provide results when the analysis focuses
only on one equation, a GME formulation using only the information in (yi, 2;)is
reported.
B. Sampling Results
The sampling results for a range of experiments are summarized in Table 1. The
results for the base experiment ( T = 20, K ( X ’ X ) = 1, and normal errors) are
given in the first row of Table 1. Focusing on the MSE results for the reduced-
form parameter vector T , the unrestricted LS empirical MSE is 125, and thus close
to its theoretical value. In contrast, the GME estimator of T that takes account of
zero restrictions in the system yields MSE(fi) = 4.11. This reduction in MSE
relative to the traditional unrestricted T estimator is impressive. In terms of the
structural parameters, note, relative to the 3SLS estimator with known error C O -
variance, the superior empirical risk performance of the GME estimator. The em-
pirical sampling variability of the GME estimator is given in parentheses for the
structural parameters r and B. These results reflect the relative stability of the
GME estimator, even under conditions of nonnormal errors or a high condition num-
ber. Intuitively speaking, significant improvement of the GME relative to the 3SLS
is due to (i) shrinkage possibilities for both the signal and noise components, (ii)
use of a dual loss function, and (iii) avoiding distributional assumptions (restric-
tions).
To reflect the sampling characteristics of the GME and 3SLS (with known CO-
variance) estimators, we follow Tsurumi (1990) and focus on the y12 parameter in
(1) and give in Figure 1 a frequency plot for the two estimators. Relative to the
3SLS known error covariance estimator, the high concentration of y l 2 and the re-
stricted range variability of the GME estimates are nicely reflected in the empirical
histogram. The empirical bias of the 3SLS (known covariance) estimator is slightly
smaller than that of the GME estimator.
In terms of results for (l),the GME results indicate the empirical-risk gain
when information from the whole system is used (column 4) relative to using only
information from (yi, Zi)in column 7. In contrast, note the empirical-risk superiority
of the GME ( y ; ,2;)estimation rule over the 3SLS (known covariance) estimation rule
for (1). Also note the GME (yi, Zi) estimation rule remains stable and is concentrated
near y12 as the condition number or error distribution changes. Finally, repeating
the experiment involving T = 20 and K ( X ’ X ) = l ) , but using relaxed moment
constraints
t t k m
yields estimates similar to those reported in the first row of Table 1.

Although we have chosen as a basis of comparison the unattainable traditional
3SLS estimator with known error covariance, the empirical-risk superiority under a
IN STATISTICAL MODES
INFORMATION RECOVERY 377
Table I Empirical-Risk Results from loo0 Experiments Using the SESM, GME, and 3SLs Estimators with MSE Performance Measures“
~
Results from an analysis of the complete system Using only y~ Y1, X1

Experiment - -
description MSE(*) MSE(@ MSE(P) MSE@ + f ) MSE& + p12 MSE@I +?1) P12
GME, T = 20 4.11 7.10 0.09 7.20 3.31 0.39 24.17 0.31

normal errors (2.95) (0.03) (1.30) (0.01) (0.06)
K(X’X) = 1
3SIS (known Cov) 699.70 0.16
normal errors
K(X’X) = 1 (0.13)
GME, T = 20 8.77 11.39 0.21 11.59 4.74 0.36 24.60 0.31
1 8 ) errors (7.81) (0.15) (3.09) (0.05) (0.11)
K(X’X) = 1
GME, T = 20 0.08 6.97 3.02 0.38 23.88 0.31
normal errors (0.03) (1.19) (0.01) (0.05)
K ( X ’ X ) = 90
GME, T = 206 4.21 6.41 0.08 6.50 2.50 0.33 24.05 0.31
normal errors (2.64) (0.02) (123) (0.01) (0.07)
K(X’X) = 1
“Numbers in parentheses are empirical variances.
bCov= (
1.4 -2.3 0.9
4.1 y)
378 GOLANET AL.
I I I 1
-2 0 2 4
Gamma12
Figure I Empirical histogram of the GME and 3SLS for Gamma 12.
MSE measure of the GME estimation rule appears to hold over a range of conditions
normally found in practice.
V. CONCLUDING REMARKS
We propose a new GME method for recovering the unknown parameters in a simul-
taneous-equations statistical model that is (i) robust in respect to likelihood; (ii) flex-
ible in respect to introducing sample and nonsample information; (iii) works well in
IN STATISTICAL MODELS
INFORMATION RECOVERY 379
both ill-posed (e.g., collinear X’s) and well-posed problems and with small sam-
ples of data; (iv) has the usual asymptotic sampling properties; (v) in finite samples,
under a squared error loss measure and relative to traditional estimators, is a high-
performing estimation rule; and (vi) is computationally simple.
In contrast to traditional estimation for simultaneous-equations models, it per-
mits the sample information to be introduced in either a data or moment form. It per-
mits information recovery in case of nonlinear and/or nonstationary expectational
models (Golan, Judge, and Karp 1996) and with discrete and/or limited endogenous
regressors (Golan, Judge, and Perloff 1995, 1996). Using the normalized entropy
concept provides a basis for selecting among alternative competing statistical mod-
els (Golan and Judge 199613). By employing the entropy measure for each of the
unknown endogenous and exogenous variables, when all the support spaces s are
defined to be symmetric about zero, it is possible to identify the extraneous variables
in each of the G equations. This problem will be further developed in future work.
The finite-sample results reported suggest, relative to the 3SLS rule with known
covariance, the superior performance of the GME rule under selected experimental
designs. What is needed at this point is extensive sampling experiments that make
a sharp comparison with traditional sampling theory and Bayes’ estimators for the
SESM.
ACKNOWLEDGMENTS
It is a pleasure to acknowledge the helpful suggestions of Ron Mittelhammer.
REFERENCES
Basmann, R. L. (1957), A General Classical Method of Linear Estimation of Coefficients in a

Structural Equation, Econometrica, 2.5, 77-83.
Brown, L. D. (1986), Fundamentals of Statistical Exponential Families, Institute of Mathe-
matical Statistics, Hayward, CA.
Cragg, J. G. (1966), On the Sensitivity of Simultaneous Equations Estimators to the Stochastic
Assumptions of the Models, Journal of the American Statistical Association, 61, 136-
151.
Cragg, J. G. (1967), On the Relative Small-Sample Properties of Several Structural-Equation
Estimators, Econometrica, 3.5,89-110.
Csiszar, I. (1991), Why Least Squares and Maximum Entropy? An Axiomatic Approach to
Inference for Linear Inverse Problems, Annals of Statistics, 19,2032-2066.
Davidson, R. and J. G. MacKinnon (1993), Estimation and Inference in Econornetrics, Oxford
University Press, New York.
Gokhale, D. V. and S.Kullback (1978), Thelnformation in Contingency Tables, Marcel Dekker,
New York.
380 GOLANETAL.
Golan, A. and G. Judge (1996a), A Maximum Entropy Approach to Empirical Likelihood

Estimation and Inference, University of California, Berkeley, unpublished paper.
Golan, A. and G. Judge (19%b), A Simultaneous Estimation and Variable Selection Rule,
University of California, Berkeley, unpublished paper.
Golan, A., G. G. Judge, and L. Karp (19%), A Maximum Entropy Approach of Estimation and
Inference in Dynamic Models or Counting Fish in the Sea Using Maximum Entropy,
Journal of Economic Dynamics and Control, 20,559-582.
Golan, A., G. G. Judge, and D.Miller (19%), Maximum Entropy Econometrics: Robust Esti-
mation with Limited Data, Wiley, New York.
Golan, A., G. Judge, and J. Perloff (1995), Estimation and Inference with Censored and Or-
dered Multinomial Response Data, University of California, Berkeley, unpublished
paper.
Golan, A., G. Judge, and J. Perloff (19%), A Maximum Entropy Approach to Recovering
Information from Multinomial Response Data, Journal of the American Statistical As-
sociation, 91, 841-853.
Good, I. J. (1%3), Maximum Entropy for Hypothesis Formulation, Especially for Multidimen-
sional Contingency Tables, Annals vf Mathematical Statistics, 34,911-934.
Hansen, L. P. (1982), Large Sample Properties of Generalized Method of Moments Estimators,
Econometrica,50, 1029-1054.
Huber, P. (1981), Robust Statistics, Wiley, New York.
Jaynes, E. T. (1957a), Information Theory and Statistical Mechanics, Physics Review, 106,
620430.
Jaynes, E. T. (1957b), Information Theory and Statistical Mechanics 11, Physics Review, 108,
171-190.
Jaynes, E. T. (1984), Prior Information and Ambiguity in Inverse Problems, in D. W. McLaugh-
lin (ed.), Inverse Problems, SIAM Proceedings, American Mathematical Society, Prov-
idence, RI, 151-166.
Johansen, S.(1979), An Introduction to the Theory of Regular Exponential Families, Institute
of Mathematical Statistics, Hayward, CA.
Judge, G. G., W. E. Griffiths, R. C. Hill, H. Lutkepohl, and T.-C. Lee (1985), The Theory and
Judge, G. G., R. C. Hill, W. E. Griffiths, H. Lutkepohl, and T.-C. Lee (1988), Introduction to
the Theory and Practice of Econometrics, 2nd ed., Wiley, New York.
Kullback, J. (1959), Information Theory and Statistics, Wiley, New York.
Levine, R. D. (1980), An Information Theoretical Approach to Inversion Problems, Journal
of Physics A, 13,91-108.
Miller, D. J. (1994), Solving Generalized Maximum Entropy Problems with Unconstrained
Numerical Techniques, University of California, unpublished paper.
Park, S.-B. (1982), Some Sampling Properties of Minimum Expected Loss (MELO) Estimators
of Structural Coefficients, Journal of Econometrics, 18,295-3 11.
Shannon, C. E. (1948), A Mathematical Theory of Communications, Bell System Technical
Journal, 2’7,379423.
Shore, J. E. and R. W. Johnson (1980), Axiomatic Derivation of the Principle of Maximum En-
tropy and the Principle of Minimum Cross-Entropy, ZEEE Transactionson Information
Theory, IT-26(1), 26-37.
Skilling, J. (1989), The Axioms of Maximum Entropy, in J. Skilling (ed.), Maximum Entropy
and Bayesian Methods in Science and Engineering, Kluwer, Dordrecht, 173-187.
INFORMATION RECOVERYI N STATISTICAL MODELS 38 I
Tsurumi, H. (1WO), Comparing Bayesian and Non-Bayesian Limited Information Estimators,

in S. Geisser, J. S. Hodges, S. J. Press, and A. Zellner (eds.), Bayesian and Likelihood
Methods in Statistics and Econometrics, North-Holland, Amsterdam, 179-207.
White, H. (1994), Estimation, Inference and Specijication Analysis, Cambridge University
Press, Cambridge.
Zellner, A. (1997), The Finite Sample Properties of Simultaneous Equations' Estimates and
Estimators: Bayesian and Non-Bayesian Approaches, Journal of Econometrics, Annals
Issue in Honor of Car1 F. Christ (L. R . Klein, ed.) (in press).
Diagnostic Testing in Econornetrics
Variable Addition, RESET, and Fourier
Approximations
Linda F. DeBenedictis
Ministry of Humon Resources, Victoria, British Columbia, Conado
David E. A. Giles
University of Victoria, Victoria, British Columbia, Canada
1. INTRODUCTION
The consequences of model misspecification in regression analysis can be severe in

terms of the adverse effects on the sampling properties of both estimators and tests.
There are also commensurate implications for forecasts and for other inferences that
may be drawn from the fitted model. Accordingly, the econometrics literature places
a good deal of emphasis on procedures for interrogating the quality of a model’s
specification. These procedures address the assumptions that may have been made
about the distribution of the model’s error term, and they also focus on the structural
specification of the model, in terms of its functional form, the choice of regressors,
and possible measurement errors.
Much has been written about “diagnostic tests” for model misspecification in
econometrics in recent years. The last two decades, in particular, have seen a surge
of interest in this topic which has, to a large degree, redressed what was previously
an imbalance between the intellectual effort directed toward pure estimation issues
and that directed toward testing issues of various sorts. There is no doubt that diag-
nostic testing is now firmly established as a central topic in both econometric the-
ory and practice, in sympathy with Hendry (1980, p. 403), urging that we should
“test, test, and test.” Some useful general references in this field include Kramer
and Sonnberger (1986), Godfrey (1988, 1996), and White (1994), among others.
383
384 AND GILES
D~ENEDICTIS
As discussed by Pagan (1984),the majority of the statistical procedures proposed

for measuring the inadequacy of econometric models can be allocated into one of
two categories-“variable-addition” methods and “variable-transformation” meth-
ods. More than a decade later, it remains the case that the first of these categories still
provides a useful basis for discussing and evaluating a wide range of the diagnos-
tic tests that econometricians use. The equivalence between variable-addition tests
and tests based on “Gauss-Newton regressions” is noted, for instance, by Davidson
and MacKinnon (1993, p. 194) and essentially exploited by MacKinnon and Magee
(1990).Indeed it is the case that many diagnostic tests can be viewed and categorized
in more than one way.
In this chapter we limit our attention to diagnostic tests in econometrics which
can be classified as “variable-addition” tests. This will serve to focus the discussion
in a manageable way. Pagan (1984)and Pagan and Hall (1983)provide an excellent
discussion of this topic. Our purpose here is to summarize some of the salient fea-
tures of that literature and then to use it as a vehicle for proposing a new variant of
what is perhaps the best-known variable addition test-Ramsey’s (1969)“regression
specification error (RESET) test.”
The layout of the chapter is as follows. In the next section we discuss some
general issues relating to the use of variable-addition tests for model misspecifica-
tion. Section I11 discusses the formulation of the standard RESET test and the extent
to which the distribution of its statistic can be evaluated analytically. In Section IV
we introduce a modification of this test, which we call the FRESET test (as it is based
on a Fourier approximation), and we consider some practical issues associated with
its implementation. A comparative Monte Carlo experiment, designed to explore the
power of the FRESET test under (otherwise) standard conditions, is described in Sec-
tion V. Section VI summarizes the associated results. The last section contains some
conclusions and recommendations which strongly favor the new FRESET test over
existing alternatives; and we note some work in progress which extends the present
study by considering the robustness of tests of this type to nonspherical errors in the
data-generating process.
II. VARIABLE-ADDITION TESTS IN ECONOMETRICS

A. Preliminaries
One important theme that underlies many specification tests in econometrics is the
idea that if a model is correctly specified, then (typically) there are many weakly con-
sistent estimators of the model’s parameters, and so the associated estimates should
differ very little if the sample size is sufficiently large. A substantial divergence of
these estimates may be taken as a signal of some sort of model misspecification (e.g.,
White 1994, Chap. 9). Depending on the estimates which are being compared, tests
€or various types of model misspecification may be constructed. Indeed, this basic
DIAGNOSTIC IN ECONOMETRICS 385
TESTING
idea underlies the well-known family of Hausman (1978) tests and the information
matrix tests of White (1982,1987). This approach to specification testing is based on
the stance that, in practice, there is generally little information about the precisefonn
of any misspecification in the model. Accordingly, no specific alternative specifica-
tion is postulated, and a pure significance test is used. This stands in contrast with
testing procedures in which an explicit alternative hypothesis is stated, and used in
the construction and implementation of the test (even though a rejection of the null
hypothesis need not lead one to accept the stated alternative). In the latter case, we
frequently “nest” the null within the alternative specification and then test whether
the associated parametric restrictions are consistent with the evidence in the data.
The use of likelihood ratio, Wald, and Lagrange multiplier tests, for example, in this
situation are common and well understood.
As noted, specification tests which do not involve the formulation of a specific
alternative hypothesis are pure significance tests. They require the construction of a
sample statistic whose null distribution is known, at least approximately or asymp-
totically. This statistic is then used to test the consistency of the null with the sample
evidence. In the following discussion we will encounter tests which involve a spe-
cific alternative hypothesis, although the latter may involve the use of proxy variables
to allow for uncertainties in the alternative specification. Our subsequent focus on
the RESET test involves a procedure which really falls somewhere between these
two categories, in that although a specific alternative hypothesis is formulated, it is
largely a device to facilitate a test of a null specification. Accordingly, it should be
kept in mind that the test is essentially a “destructive” one, rather than a “construc-
tive” one, in the sense that a rejection of the null hypothesis (and hence of the model’s
specification) generally will not suggest any specific way of reformulating the model
in a satisfactory form. This is certainly a limitation on its usefulness, so it is all the
more important that it should have good power properties. If the null specification
is to be rejected, with minimal direction as to how the model should be respecified,
then at least one would hope that we are rejecting for the right reason(s). Accord-
ingly, in our reconsideration of the RESET test in Sections I11 and IV we emphasize
power properties in a range of circumstances.
Variable-addition tests are based on the idea that if the model specification
is “complete,” then additions to the model should have an insignificant impact, in
some sense. As noted by Pagan and Hall (1983)and Pagan (1984), there are many
forms that such additions can take. For instance, consider a standard linear multiple
regression model, with k fixed regressors and T observations:
y=XBfu (1)
where it may be assumed that (y I X ) -
N [ X / 3 , 0‘1771. One could test this specifi-
cation in terms of the adequacy of the assumed conditional mean of y, namely X/?;
or one might test the adequacy of the assumed conditional covariance matrix, a 2 1 ~ .
The assumed normality could be tested with reference to higher-order moments, as
386 DEBENEDICTIS
AND GILES
in Jarque and Bera (1980).In most of these cases, tests can be constructed by fitting
auxiliary regressions which include suitable augmentation terms, and then testing
the significance of the latter.
B. Variable Addition and the Conditional Mean

For example, if it is suspected that the conditional mean of the model may be mis-
specified, then one could fit an “augmented” model
and test the hypothesis that y = 0. This assumes, of course, that W is known and
observable. In the event that it is not, a matrix of corresponding proxy variables, W * ,
may be substituted for W , and (2) may be written as
and we could again test if y = 0. As Pagan (1984, p. 106) notes, the effect of this
substitution will show up in terms of the power of the test that is being performed.
An alternative way of viewing (2) (or (3) if the appropriate substitution of the proxy
variables is made below) is by way of an auxiliary regression with residuals from (1)
as the dependent variable:
y -Xb = X ( B -6) + Wy+U (4)

where b = (X’X)-’X’y is the least-squares estimator of B in (1). This last model is
identical to (2), and the test of y = 0 will yield an identical answer in each case.
However, (4) emphasises the role of diagnostic tests in terms of explaining residual
values.
The choice of W (or W * )will be determined by the particular way in which the
researcher suspects that the conditional mean of the basic model may be misspeci-
fied. Obvious situations that will be of interest include a wrongly specified functional
form for the model or terms that have been omitted wrongly from the set of explana-
tory variables. There is, of course, a natural connection between these two types
of model misspecification, as we discuss further. In addition, tests of serial indepen-
dence of the errors, structural stability, exogeneity of the regressors, and those which
discriminate between nonnested models, can all be expressed as variable-addition
tests which focus on the conditional mean of the data-generating process. All of these
situations are discussed in some detail by Pagan (1984).Accordingly, we will simply
summarize some of the salient points here, and we will focus particularly on func-
tional form and omitted effects, as these are associated most directly with the RESET
test and hence with the primary focus of this chapter.
The way in which tests for serial independence can be cast in terms of variable-
addition tests is easily illustrated. Consider model (l),but take as the maintained
hypothesis an AR(1) representation for the disturbances. That is, assume that u t =
DIAGNOSTIC
TESTING
IN ECONOMETRICS 387
put-) + E ~ where
, lpl < 1. We wish to test the hypothesis that p = 0. The mean
of y in (1) is then conditional on the past history of y and X so it is conditional on
previous values of the errors. Accordingly, the natural variable-addition test would
involve setting W in (2) to be just the lagged value of U. Of course, the latter is unob-
servable, so the proxy variable approach of (3) would be used in practice, with W*
comprising just the lagged OLS residual series (U:,) from the basic specification,
(1). Of course, in the case of a higher-order AR process, extra lags of U* would be
used in the construction of W * , and we would again test if y = 0. It is important
to note that the same form of variable-addition tests would be used if the alternative
hypothesis is that the errors follow a moving-average process, and such tests are gen-
erally powerful against both alternatives. The standard Durbin-Watson test can be
linked to this approach to testing for model misspecification, and various other stan-
dard tests for serial independence in the context of dynamic models, such as those
of Godfrey (1978), Breusch (1978), and Durbin (1970), can all be derived in this
general manner. Tests for structural stability which can be given a variable-addition
interpretation include those of Salkever (1976), where the variables that are used to
augment the basic model are suitably defined “zero-one” dummy variables. Further,
the well-known tests for regressor exogeneity proposed by Durbin (1954),Wu (1973),
and Hausman (1978) can also be reexpressed as variable-addition tests which use
appropriate instrumental variables in the construction of the proxy matrix W * (e.g.,
Pagan 1984, pp. 114-1 15).
The problem of testing between (nonnested) models is one which has attracted
considerable attention during the last 20 years (e.g., McAleer 1987,1995). Such tests
frequently can be interpreted as variable-addition tests which focus on the specifica-
tion of the conditional mean of the model. By way of illustration, recall that in model
(1)the conditional mean of y (given X , and the past history of the regressors and of y)
is X g . Suppose that there is a competing model for explaining y, with a conditional
mean of X+p, where X and X + are nonnested, and p is a conformable vector of
unknown parameters. To test one specification of the model against the other, there
are various ways of applying the variable-addition principle. One obvious possibility
(assuming an adequate number of degrees of freedom) would be to assign W* = X +
in (3), and then apply a conventional F-test. This is the approach suggested by Pe-
saran (1974) in one of the earlier contributions to this aspect of the econometrics
literature. Another possibility, which is less demanding on degrees of freedom, is to
set W* = X+(X+’X+)-’X+’y (that is, using the ordinary least-squares (OLS) es-
timate of the conditional mean from the second model as the proxy variable), which
gives us the J-test of Davidson and MacKinnon (1981). There have been numer-
ous variants on the latter theme, as discussed by McAleer (1987), largely with the
intention of improving the small-sample powers of the associated variable-addition
tests.
Our main concern is with variable-addition tests which address possible mis-
specification of the functional form of the model or the omission of relevant explana-
388 DEBENEDICTIS
AND GILES
tory effects. The treatment of the latter issue fits naturally into the framework of
Eqs. (1)-(4).The key decision that has to be made in order to implement a variable-
addition test in this case is the choice of W (or, more likely, W*). If we have some
idea what effects may have been omitted wrongly, then this determines the choice of
the “additional” variables, and if we were to make a perfect choice then the usual
F-test of y = 0 would be exact and uniformly most powerful (UMP). Of course,
this really misses the entire point of our present discussion, which is based on the
premise that we have specified the model to the best of our knowledge and ability,
but are still concerned that there may be some further, unknown, omitted effects. In
this case, some ingenuity may be required in the construction of W or W * , which
is what makes the RESET test (and our modification of this procedure in this chap-
ter) of particular interest. We leave a more detailed discussion of the RESET test to
Section 111.
In many cases, testing the basic model for a possible misspecification of its
functional form can be considered in terms of testing for omitted effects in the con-
ditional mean. This is trivially clear if, for example, the fitted model includes simply
a regressor, xt,but the correct specification involves a polynomial in xt. Construct-
ing W* with columns made up of powers of xt would provide an optimal test in this
case. Similarly, if the fitted model included x, as a regressor, but the correct spec-
ification involved some (observable) transformation of xt,such as log(x,), then (2)
could be constructed so as to include both the regressor and its transformation, and
the significance of the latter could be tested in the usual way. Again, of course, this
would be feasible only if one had some prior information about the likely nature of the
misspecification of the functional form. (See also, Godfrey, McAleer, and McKenzie
1988).
More generally, suppose that model (1) is being considered, but in fact the
correct specification is
y= f(X9 P l y B2> + U (5)

where f is a nonlinear function which is continuously differentiable with respect to
the parameter subvector, 82. If (1)is nested within (5) by setting 8 2 = 3/,; then by
taking a Taylor series expansion off about the (vector) point &, we can (after some
rearrangement) represent (5) by a model which is of the form (3) and then proceed
in the usual way to test (1) against (5) by testing if y = 0 via a RESET test. More
specifically, this can be achieved, to a first-order approximation, by setting
where bl is the least-squares estimator of the subvector obtained subject to the

restriction that #?2 = @. As we will see again in Section 111, although the presence
of 61 makes W* random, the usual F-test of y = 0 is still valid, as a consequence of
the Milliken-Graybill (1970)theorem.
DIAGNOSTIC IN ECONOMETRICS
TESTING 389
As we have seen, many tests of model misspecification can be formulated as

variable-addition tests in which attention focuses on the conditional mean of the un-
derlying data-generating process. This provides a very useful basis for assessing the
sampling properties of such tests. Model (3)forms the basis of the particular speci-
fication test (the RESET test) that we will be considering in detail later. Essentially,
we will be concerned with obtaining a matrix of proxy variables, W * ,that better rep-
resents an arbitrary form of misspecification than do the usual choices of this matrix.
In this manner, we hope to improve on the ability of such tests to reject models which
are falsely specified in the form of Eq. (1).
C. Variable-Addition and Higher Moments

Although our primary interest is with variable-addition tests which focus on mis-
specification relating to the conditional mean of the model, some brief comments
are in order with respect to related tests which focus on the conditional variance
and on higher-order moments. Of the latter, only the third and fourth moments have
been considered traditionally in the context of variable-addition tests, as the basis of
testing the assumption of normally distributed disturbances (e.g., Jarque and Bera
1987). Tests which deal with the conditional variance of the underlying process have
been considered in a variable-addition format by a number of authors.
Essentially, the assumption that the variance, 02, of the error term in (1)is a
constant is addressed by considering alternative formulations, such as
of = o2 + ztb, (7)
where zt is an observation on a vector of r known variables, and 4 is r x 1. We
then test the hypothesis that 4 = 0. To make this test operational, (7) needs to be
reformulated as a “regression relationship” with an observable “dependent variable”
and a stochastic “error term.” The squared tth element of U in (1) gives us of, on
average, so it is natural to use the corresponding squared OLS residuals on the left
side of (7). Then
+
= o2 zt4 + (U;)2 - of = o2+ ztqb + vt (8)
where vt = ( u : ) ~- of. Equation (8) can be estimated by OLS to give estimates
of o2and 4 and to provide a natural test of 4 = 0. The (asymptotic) legitimacy of
the usual t -test (or F-test) in this capacity is established, for example, by Amemiya
(1977).
So the approach in (8) is essentially analogous to the variable-addition ap-
proach in the case of (2) for the conditional mean of the model. As was the situation
there, in practice we might not be able to measure the zl vector, and a replacement
proxy vector, z:, might be used instead. Then the counterpart to (3)would be
+ + (zt - z;)@ +
= o2 zt4 U, + z;(b +
= o2 U; (9)
390 DEBENEDICTIS
AND GILES
where we again test if 4 = 0, and the choice of z; determines the particular form of
heteroskedasticity against which we are testing.
For example, if z* is an appropriately defined scalar dummy variable then we
can test against a single break in the value of the error variance at a known point.
This same idea also relates to the more general homoskedasticity tests of Harrison
and McCabe (1979)and Breusch and Pagan (1979). Similarly, Garbade’s (1977) test
for systematic structural change can be expressed as the above type of variable-
addition test, with z; = txf; and Engle’s (1982) test against ARCH(1) errors amounts
to a variable addition test with z r = ( u , * - , ) ~Higher-order
. ARCH and GARCH
processes* can be accommodated by including additional lags of ( u : ) ~in the def-
inition of z*. Pagan (1984, pp. 115-118) provides further details as well as other
examples of specification tests which can be given a variable-addition interpretation
with respect to the conditional variance of the errors.
D. Multiple Testing
Variable-addition tests have an important distributional characteristic which we have
not yet discussed. To see this, first note that under the assumptions of model (l),the
UMP test of y = 0 will be a standard F-test if X and W (or W * ) are both non-
stochastic and of full column rank. In the event that either the original or “addi-
tional” regressors are stochastic (and correlated with the errors), and/or the errors
are nonnormal, the usual F-statistic for testing if y = 0 can be scaled to form a
statistic which will be asymptotically chi-square. More specifically, if there are T
observations and if rank(X) = Ic and rank(W) = p, then the usual F-statistic will
be FP,”,under the null (where U = T - k - p). Then p F will be asymptotically x:
under the null.? Now, suppose that we test the model’s specification by means of a
variable addition test based on (2) and denote the usual test statistic by F”. Then,
suppose we consider a second test for misspecification by fitting the “augmented”
model
where rank(2) = q, say. In the latter case, denote the statistic for testing if 6 = 0 by
F”. Asymptotically, p F ” is x:,and qF” is xi,
under the respective null hypotheses.
Now, from the usual properties of independent chi-square statistics, we know
that if the above two tests are independent, then pF” +
qF” is asymptotically x&,
under the null that y = 6 = 0. As discussed by Bera and McKenzie (1987) and
Eastwood and Godfrey (1992, p. 120), independence of the tests requires that plim
*Lee (1991)shows the equivalence of ARCH(p) and GARCH(p, Q ) tests under the null, for a constant
q , where p is any natural number.
?Its distribution under the alternative is discussed in Section 111.
DIAGNOSTIC
TESTING
IN ECONOMETRICS 39 I
(Z’V/Z‘) = 0 and that either plim(Z’X/T) = 0 or plim(W’X/T) = 0. The advan-

tages of this are the following. First, if these variable addition specification tests are
applied in the context of (2) and (10)sequentially, then the overall significance level
can be controlled. Specifically, if the (asymptotic) significance levels for these two
tests are set at a, and a, respectively, then* the overall joint significance level will
be 1 - (1 - a,) (1 - az). Second, the need to fit a “supermodel,” which includes all
of the (distinct) columns of X, V‘, and Z as regressors, can be avoided. The extension
of these ideas to testing subsets of the regressors is discussed in detail by Eastwood
and Godfrey (1992, pp. 120-122).
A somewhat related independence issue also arises in the context of certain
variable-addition tests, especially when the two tests in question focus on the differ-
ent moments of the underlying stochastic process. Consider, for example, tests of the
type discussed in Section 1I.B. These deal with the first moment of the distribution
and are simply tests of exact restrictions on a certain regression coefficient vector.
Now, suppose that we also wanted to test for a shift in the variance of the error term in
the model, perhaps at some known point(s) in the sample. It is well known? that many
tests of the latter type (e.g., those of Goldfeld and Quandt 1965, Harrison and Mc-
Cabe 1979, and Breusch and Pagan 1979) are independent (at least asymptotically
in some cases) of the usual test for exact restrictions on the regression coefficients.
Once again, this eases the task of controlling the overall significance level, if we
are testing for two types of model misspecification concurrently but do not wish to
construct “omnibus tests.”
The discussion assumes, implicitly, that the two (or more) variable-addition
tests in question are not only independent but that they are applied separately but
“concurrently.” By the latter, we mean that each test is applied regardless of the
outcome(s) of the other test(s). That is, we are abstracting from any “pretest testing”
issues. Of course, in practice, this may be too strong an assumption. A particular
variable-addition test (such as the RESET test) which relates to the specification of
the conditional mean of the model might be applied only if the model “passes” an-
other (separate and perhaps independent) test of the specification of the conditional
variance of the model. If it fails the latter test, then a dzerent variable addition test
for the specification of the mean may be appropriate. That is, the choice of thefonn of
one test may be contingent on the outcome of a prior test of a different feature of the
model’s specification. Even if the first-stage and second-stage tests are independent,
there remains a “pretesting problem” of substance.$ The true significance level of
*Essentially, this follows from Basu’s (19%)independenre theorem. For example, see Anderson (1971,
pp. 34-43, 116-134,270-276) and the asymptotic extensions discussed by Mizon (1977a, 1977b).
For some general discussion of this point, see Phillips and McCabe (1983), Pagan and Hall (1983), and
Pagan (1984, pp. 116-117,125-127). Phillips and McCabe (1989) also provide extensions to other tests
where the statistics can be expressed as ratios of quadratic forms in a normal random vector.
$For a more comprehensive discussion of this point, see Giles and Giles (1993, pp. 176-180).
392 DEBENEDICTIS
AND GILES
the two-part test for the specification of the conditional mean of the model will differ
from the sizes nominally assigned to either of its component parts, because of the
randomization of the choice of second-stage test which results from the application
of the first-stage test (for the specification of the conditional variance of the model).
More generally, specification tests of the variable-addition type may not be in-
dependent of each other. As noted by Pagan (1984, pp. 125-127), it is unusual for
tests which focus on the same conditional moment to be mutually independent. One
is more likely to encounter such independence between tests which relate to dzferent
moments of the underlying process (as in the discussion above). In such cases there
are essentially two options open. The first is to construct joint variable-addition tests
of the various forms of misspecification that are of interest. This may be a somewhat
daunting task, and although some progress along these lines has been made (e.g.,
Bera and Jarque 1982), there is still little allowance for this in the standard econo-
metrics computer packages. The second option is to apply separate variable-addition
tests for the individual types of model misspecification, and then adopt an “induced
testing strategy” by rejecting the model if at least one of the individual test statistics
is significant. Generally, in view of the associated nonindependence and the likely
complexity of the joint distribution of the individual test statistics, the best that one
can do is to compute bounds on the overall significance level for the “induced test.”
The standard approach in this case would be to use Bonferroni inequalities (e.g.,
David 1981, Schwager 1984), though generally such bounds may be quite wide and,
hence, relatively uninformative. A brief discussion of some related issues is given
by Kramer and Sonnberger (1986, pp. 147-155), and Savin (1984) deals specifically
with the relationship between multiple t-tests and the F-test. This, of course, is di-
rectly relevant to the case of certain variable-addition tests for the specification of
the model’s conditional mean.
E. Other Distributional Issues
There are several further distributional issues which are important in the context
of variable-addition tests. In view of our subsequent emphasis on the RESET test
in this chapter, it is convenient and appropriate to explore these issues briefly in
the context of tests which focus on the conditional mean of the model. However, it
should be recognized that the general points that are made in the rest of this section
also apply to variable-addition tests relating to other moments of the data-generating
process. Under our earlier assumptions, the basic form of the test in which we are
now interested is an F-test of y = 0 in the context of model (2). In that model, if
W is truly the precise representation of the omitted effects, then the F-test will be
UMP. Working, instead, with the matrix of proxy variables W”, as in (3), does not
affect the null distribution of the test statistic in general, but it does affect the power
of the test, of course. Indeed, the reduction in power associated with the use of the
proxy variables increases as the correlations between the columns of W and those
TESTINGIN ECONOMETRICS
DIAGNOSTIC 393
of W* decrease. Ohtani and Giles (1993)provide some exact results relating to this
phenomenon under very general distributional assumptions, and find the reduction
in power to be more pronounced as the error distribution departs from normality.
They also show that, regardless of the degree of nonnormality, the test can be biased*
as the hypothesis error grows, and they prove that the usual null distribution for the
F-statistic for testing if y = 0 still holds even under these more general conditions.
Of course, in practice, the whole point of the analysis is that the existence,
form, and degree of model misspecification are unknown. Although the general form
of W* will be chosen to reflect the type of misspecification against which one is
testing, the extent to which W* is a “good” proxy for W (and hence for the omitted
effect) will not be able to be determined exactly. This being the case, in general it is
difficult to make specific statements about the power of such variabIe-addition tests.
As long as W* is correlated with W (asymptotically), a variable-addition test based
on model (3) will be consistent. That is, for a given degree of specification error,
as the sample size grows the power of the test will approach unity. In view of the
immediately preceding comments, the convergence path will depend on the forms of
W and W * .
The essential consistency of a basic variable-addition test of y = 0 in (2)
is readily established. Following Eastwood and Godfrey (1992, pp. 123-125), and
assuming that X ’ X , W’W, and X’W are each O,(T), the consistency of a test based
on F” (as defined in Section II.D, and assuming independent and homoskedastic
disturbances) is ensured if plim(FW/T) # 0 under the alternative. Now, as is well
known, we can write
(RSS - USS)/p
F” =
USS/( T - k - p )
where RSS denotes the sum of the squared residuals when (2) is estimated by OLS
subject to the restriction that y = 0, and USS denotes the corresponding sum of
squares when (2) is estimated by unrestricted OLS. Under quite weak conditions the
denominator in (11)converges in probability to o2 (by Khintchine’s theorem). So,
by Slutsky’s theorem, in order to show that the test is consistent, it is sufficient to
establish that plim (RSS - USS)/T # 0 under the alternative.? Now, for our problem
we can write
* A biased test is one whose power can fall helow its significance level in some region of the relevant
parameter space.
?Clearly, this plim is zero if‘the null is true, because then both RSSIT and USSIT are consistent estimators
of (T2.
394 DEBENEDICTIS
AND GILES
where R = [Zk, O,], and X* = (X : W ) . Given our assumption about the orders in
probability of the data matrices, we can write plim(X*’X*/T) = Q*, say, where Q*
is finite and nonsingular. Then it follows immediately from (12) that plim[(RSS -
USS)/T] > 0, if y # 0, so the test is consistent. It is now clear why consistency is
retained if W* is substituted for W , as long as these two matrices are asymptotically
correlated. It is also clear that this result will still hold even if W is random or if
W* is random (as in the case of a RESET test involving some function of the OLS
prediction vector from (1) in the construction of W*).
Godfrey (1988, pp. 102-106) discusses another important issue that arises in
this context, and which is highly pertinent for our own analysis of the RESET test in
this chapter. In general, if we test one of the specifications in (2)-(4) against model
(l), by testing if y = 0, it is likely that in fact the true data-generating process differs
from both the null and maintained hypotheses. That is, we will generally be testing
against an incorrect alternative specification, In such cases, the determination of
even the asymptotic power of the usual F-test (or its large-sample chi-square coun-
terpart) is not straightforward, and is best approached by considering a sequence of
local alternatives. Not surprisingly, it turns out that the asymptotic power of the test
depends on the (unknown) extent to which the maintained model differs from the
true data-generating process.
Of course, in practice the errors in the model may be serially correlated and/or
heteroskedastic, in which case variable-addition tests of this type generally will be
inconsistent, and their power properties need to be considered afresh in either large
or small samples. Early work by Thursby (1979, 1982) suggested that the RESET
test might be robust to autocorrelated errors, but as noted by Pagan (1984, p. 127)
and explored by Porter and Kashyap (1984), this is clearly not the case. We abstract
from this situation in the development of a new version of the RESET test later in this
chapter, but it is a topic that is being dealt with in some detail in our current research.
Finally, we should keep in mind that a consistent test need not necessarily have high
power in small samples, so this remains an issue of substance when considering
specific variable-addition tests.
Finally, it is worth commenting on the problem of discriminating between two
or more variable-addition tests, each of which is consistent in the sense described
above. If the significance level for the tests is fixed (as opposed to being allowed to
decrease as the sample size increases), then there are at least two fairly standard
ways of dealing with this issue. These involve bounding the powers of the tests away
from unity as the sample size grows without limit. The first approach is to use the
“approximate slope” analysis of Bahadur (1960,1967). This amounts to determining
how the asymptotic significance level of the test must be reduced if the power of
the test is to be held constant under some fixed alternative. The test statistics are
O,(T), and they are compared against critical values which increase with T . The
approximate slope for a test statistic, S, which is asymptotically chi-square, is just
DIAGNOSTIC
TESTING
IN ECONOMETRICS 395
plim(S/T).* This is the same quantity that we considered in the determination of

the consistency of such a test. A choice between two consistent tests which test the
same null against the same alternative can be made by selecting so as to maximize
the approximate slope. Of course, once again this does not guarantee good power
properties in small samples.
A second approach which may be used to discriminate between consistent
tests is to consider a sequence of local alternatives. In this approach, the alternative
hypothesis is adjusted so that it approaches the null hypothesis in a manner that
ensures that the test statistics are Op(l). Then a fixed critical value can be used,
and the asymptotic powers of the tests that can be compared as they will each be
less than unity. In our discussion of the traditional RESET test in the next section
and of a new variant of this test in Section IV, neither of these approaches to dis-
criminating between consistent tests is particularly helpful. There, the tests that are
being compared have the same algebraic construction, but they differ in terms of the
number of columns in W* and the way in which these columns are constructed from
the OLS prediction vector associated with the basic model. In general, the “approxi-
mate slope” and “local alternatives” approaches do not provide tractable expressions
upon which to base clear comparisons in this case. However, in practical terms our
interest lies in the small-sample properties of the tests, and it is on this characteristic
that we focus.
111. T H E RESET TEST
Among the many “diagnostic tests” that econometricians routinely use, some variant
or other of the RESET test is widely employed to test for a nonzero mean of the error
term. That is, it tests implicitly whether a regression model is correctly specified in
terms of the regressors that have been included. Among the reasons for the popularity
of this test are that it is easily implemented and that it is an exact test whose statistic
follows an F-distribution under the null. The construction of the test does, however,
require a choice to be made over the nature of certain “augmenting regressors” that
are employed to model the misspecification, as we saw in Section 1I.B. Depending on
this choice, the RESET test statistic has a nonnull distribution which may be doubly
noncentral F or totally nonstandard. Although this has no bearing on the size of the
test, it has obvious implications for its power.
The most common construction of the RESET test involves augmenting the
regression of interest with powers of the prediction vector from a least-squares re-
gression of the original specification and testing their joint significance. As a result
*For example, see Geweke (1981), Magee (1987), and Eastwood and Godfrey (1992, p. 132).
396 DE~ENEDICTIS
AND GILES
of the Monte Carlo evidence provided by Ramsey and Gilbert (1972) and Thursby
(1989), for example, it is common for the second, third, and fourth powers of the
prediction vector to be used in this way.* Essentially, Ramsey’s original sugges-
tion, following earlier work by Anscombe (1961),involves approximating the un-
known nonzero mean of the errors, which reflects the extent of the model mis-
specification, by some analytic function of the conditional mean of the model. The
specific construction of the RESET test noted above then invokes a polynomial ap-
proximation, with the least-squares estimator of the conditional mean replacing its
true counterpart.
Other possibilities include using powers and/or cross products of the individ-
ual regressors, rather than powers of the prediction vector, to form the augmenting
terms. Thursby and Schmidt (1977) provide simulation results which appear to fa-
vor this approach. However, all of the variants of the RESET test that have been
proposed to date appear to rely on the use of local approximations, essentially of a
Taylor series type, of the conditional mean of the regression. Intuitively, there may be
gains in terms of the test’s performance if a global approximation were used instead.
This chapter pursues this intuition by suggesting the use of an (essentially unbiased)
Fourier flexible approximation. This suggestion captures the spirit of the develop-
ment of cost and production function modeling, and the associated transition from
polynomial functions (e.g., Johnston 1960) to Translog functions (e.g., Christensen
et al. 1971,1973) and then to Fourier functional forms (e.g., Gallant 1981, Mitchell
and Onvural 1995,1996).
Although Ramsey (1969) proposed a battery of specification tests for the lin-
ear regression model, with the passage of time and the associated development of the
testing literature, the RESET test is the one which has survived. Ramsey’s original
discussion was based on the use of Theil’s (1965, 1968) “BLUS” residuals, but the
analysis was subsequently recast in terms of the usual OLS residuals (e.g., Ramsey
and Schmidt 1976, Ramsey 1983), and we will follow the latter convention in this
chapter. As Godfrey (1988, p. 106)emphasizes, one of the principles which under-
lies the RESET test is that the researcher has only the same amount of information
available when testing the specification of a regression model as was available when
the model was originally formulated and estimated. Accordingly direct tests against
new theories, perhaps embodying additional variables, are ruled out.
A convenient way of discussing and implementing the standard RESET test is
as follows. Suppose that the regression model under consideration is (l), which we
reproduce here:
y=xp+u (13)
*For instance, the SHAZAM (1993)package adopts this approach. Clearly, the first power cannot be used
as an extra regressor in the “augmented” equation as the design matrix would then be perfectly collinear.
DIAGNOSTIC
TESTING
IN ECONOMETRICS 397
where X is T x k,of rank k, and nonstochastic; and it is assumed that U -

“0, 021r].
Such a model would be “correctly specified.” Now, suppose we allow for the possi-
bility that the model is misspecified through the omission of relevant regressors or
e
the wrong choice of functional form. In this case, E[u I XI = # 0. The basis of
the RESET test is to approximate 6 by 2 8 , which corresponds to W * y in (3), and fit
an augmented model,
y = xg+ ze f E (14)
We then test if 6 = 0 by testing if 13 = 0, using the usual F-test for restrictions
on a subset of the regression coefficients. Different choices of the T x p matrix Z
lead to different variants of the RESET test. As noted, the most common choice is to
construct 2 to have tth row vector
2, = [ ( X b ) T , (Xb)Q, . . . , ( X b ) f + ’ ]
where often p = 3; b = ( X ’ X ) - ’ X ’ y is the OLS estimator of # I (1);and ( X b ) ,

from
is therefore the tth element of the associated prediction vector, = X b .
If 2 is chosen in this way (or if it is nonrandom, or if it depends on y only
through b), then, as a consequence of the Milliken-Graybill(l970) theorem, the RE-
SET test statistic is F-distributed with p and T - k - p degrees of freedom under
the null that 8 = 0,* provided that the disturbances are NID(0, a2).If 2 is non-
stochastic (as would be the case if Thursby and Schmidt’s 1977 proposal of using
powers of the regressors to augment the model were followed) then the test statistic’s
nonnull distribution is doubly noncentral F with the same degrees of freedom? and
with numerator and denominator noncentrality parameters:
In this case, the power of the RESET test can be computed exactly for any given
degree of misspecification, 6, by recognizing that
*Note that is being approximated by ZO.

?For example, see Ramsey (1969)and Thursby and Schmidt (1977).
398 DEBENEDICTIS
AND G!LES
where F denotes a doubly noncentral F-statistic, and x2

denotes a noncentral chi-
square statistic, each with degrees of freedom and noncentrality parameters as shown.
The algorithm of Davies (1980), which is conveniently coded in the SHAZAM (1993)
package, provides an efficient and accurate means of computing such probabili-
ties, although we are not aware of any studies which do this in the context of the
RESET test.
If, as is usual in practice, 2 is constructed from powers of X b or is random
for any other reason, then A1 and A2 will be random and the nonnull distribution
of the RESET test statistic will no longer be doubly noncentral F . The power char-
acteristics associated with its nonstandard distribution will depend on the specific
choice of 2 (and hence on the number of powers of X b that may be used, if this
choice is adopted), and are then best explored via Monte Carlo simulation. This is
the approach that we adopt in this chapter.
IV. FOURIER APPROXIMATIONS A N D

T H E FRESET TEST
As noted, the essence of Ramsey’s RESET test is to approximate the unknown 6

by some analytic function of X/3 (or, more precisely, of the observable = Xb). A
power series approximation is one obvious possibility, but there are other approxi-
mations which may be more attractive. In particular, as Gallant (1981) has argued in
a different context, one weakness of Taylor series approximations (which include the
power series and Translog approximations, for example) is that they have only local
validity. Taylor’s theorem is valid only in a neighborhood of some unspecified size
containing a specific value of the argument of the function to be approximated. On
the other hand, a Fourier approximation has global validity. Such approximations
can take the form of a conventional sine/cosine expansion or the less conventional
Jacobi, Laguerre, or Hermite expansions. We consider using a sine/cosine expansion
of X b to approximate (.* Although Gallant (1981) has suggested that the precision
of a truncated Fourier series can generally be improved by adding a second-order
Taylor series approximation (see also Mitchell and Onvural 1995), we do not pursue
this refinement here.
In order to obtain a Fourier representation of 2 8 in (14), and hence a Fourier
approximation of the unknown t,by using X b , we need to first transform the elements
of this vector to lie in an interval of length 2 n , such as [-n,+XI. This is because
a Fourier representation is defined only if the domain of the function is in an inter-
*In view of the Monte Carlo evidence provided by Thurshy and Schmidt (1977), in principle it would also
be interesting to consider multivariate Fourier expansions in terms of the original regressors.
TESTING I N ECONOMETRICS
DIAGNOSTIC 399
val of length 2n.Mitchell and Onvural (1995, 1996) and other authors use a linear
transformation.* In our case, this amounts to constructing
where (Xb),;, and (Xb),,, are respectively the smallest and largest elements of the
prediction vector. We also consider an alternative sinusoidal transformation, based
on Box (1966):
Then the 2 matrix for (2) is constructed to have tth row vector
2, = [sin(w,), cos(wt), sin(2wt), cos(2wt), sin(3wt),

cos(3wl), . . . , sin(p’w,), cos(p’w,)]
for some arbitrary truncation level p’. This recognizes that the Fourier approxima-
tion, g ( x ) , of a continuously differentiable function, f(x), is
where?
ug =
uj =
Equation (22) gives an exact representation off (x) by g ( x ) , except near x = -n

and x = +n.An approximation is obtained by truncating the range of summation
in (22) to a finite number of terms, p ’ , and this approximation will be globully valid.
The failure of the representation at the exact endpoints of the [-n,+ n ] interval can
generate an approximation error which is often referred to as “Gibbs’ phenomenon.”
Accordingly, some authors modify transformations such as (19)or (20) to achieve a
range which is just inside this interval. We have experimented with this refinement
to the above transformations and have found that it makes negligible difference in
the case considered. The results reported in Section VI do not incorporate this re-
finement.
+
*If the data have to be positive, as in the case of cost functions, then a positive interval such as [c, c 2x1
would be appropriate.
tThe range of summation in (22) is from j = 1 to j = 00; the ranges of integration in (23) are each from
--II to +R.
400 AND GILES
DE~ENEDICTIS
In our case, the u;’s and v;’s in (23) correspond to elements of 8 in(14), so they
are estimated as part of the testing procedure.* Note that 2 is random here, but only
through b. Our FRESET test involves constructing 2 in this way and then testing if
8 = 0 in (14). Under this null, the FRESET test statistic is still central F with 2p’
and t - k - 2p‘ degrees of freedom. Its nonnull distribution will depend upon the
form of the model misspecification, the nature of the regressors, and the choice of p’.
This distribution will differ from that of the RESET test statistic based on a (trun-
cated) Taylor series of X b . In the following discussion we use the titles FRESETL
and FRESETS to refer to the FRESET tests based on the linear transformation (19)
and the sinusoidal transformation (20) respectively.
V. A M O N T E CARLO EXPERIMENT
We have undertaken a Monte Carlo experiment to compare the properties of the new
FRESET tests with those of the conventional RESET test for some different types
of model misspecification. Table 1 shows the different formulations of the tests that
we have considered in all parts of the experiment. Effectively, we have considered
choicest of p = 1 , 2 , or 3 and p’ = 2 or 3. In the case of the RESET test, the vari-
ables whose significance is tested (i.e., the “extra” 2 variables which are added to
the basic model) comprise powers of the prediction vector from the origind model
under test, as in (15). For the FRESET test they comprise sines and cosines of mul-
tiples of this vector, as in (21), once the linear transformation (19)or the sinusoidal
transformation (20) has been used for the FRESETL and FRESETS variants respec-
t i vel y.
Three models form the basis of our experiment, and these are summarized in
Table 2. In each case we show a particular data-generating process (DGP), or “true”
model specification, together with the model that is actually fitted to the data. The
latter “null” model is the one whose specificatiori is being tested. Our model 1 allows
for misspecification through static variable omission, and corresponds to models 6-8
(depending on the value of y ) of Thursby and Schmidt (1977, p. 638). Our model 2
allows for a static misspecification of the functional form, and our model 3 involves
the omission of a dynamic effect. In each case, x p , x3, and x4 are as in Ramsey and
*The parameter uo gets “absorbed” into the coefficient of the intercept in the model (14).
We found that setting p’ > 3 resulted in a singular matrix when constructing the FRESET tests. East-
wood and Gallant (1991)suggest that setting the numher of parameters equal to the sample size raised
to the two-thirds power will ensure consistency anti asymptotic normality when estimating a Fourier
function. Setting p’ = 2 or p’ = 3 is broadly in keeping with this for our sample sizes. As Mitchell
and Onvural(l996) note, increasing p’ will incwase the variance of test statistics. In the context of the
FRESET test it seems wise to limit the value of p’.
TESTING 40 I
Table I Test Variables

Test p Test variables
RESET
Pt
FRESETL 2 1)sin(w,), cos(wt), sin(2wl), cos(2wt)
(linear transformation) 3 2) sin(w,), cos(wI), sin(2wt), cos(2wL),
sin(3wl), cos(3w,)
FRESETS 2 3 ) sin(w,), cos(w,), sin(2wl), cos(2wl)
(sinusoidal transformation) 3 4) sin(w,), cos(w,), sin(2wl), cos(2wL),
sin (3w t) , cos (3wt)
Table 2 Models
~~
Model Specification Problem
1 + + +
DGP: y, = 1.0 - 0 . 4 ~ 3 , ~ 4 , y ~ g l U ( Omitted variable
+
Null: y, = Bo B P 3 l B 4 W + + U, (omitted static effect)
2 + + +
DGP: y, = 1.0 - 0 . 4 ~ 3 , xal( 1 y x z r ) ut Incorrect functional fon..
+
Null: yt = Bo B3X.3, + +
64x4, U , (omitted multiplicative effect)
3 + + +
DGP: yt = 1.0 - 0 . 4 ~ 3 , ~ 4 1 ~ ~ l - 1 U , Incorrect functional form
+
Null: YL = P O P 3 X 3 , P@4,+ + U, (omitted dynamic effect)
Gilbert (1972) and Thursby and Schmidt (1977),* and sample sizes of T = 20 and
T = 50 have been considered.
Various values of y were considered in the range [-8.0, +8.0] in models 1
and 2, though in the latter the graphs and tables reported relate to a “narrower”
range as the results “stabilize” quite quickly. Values of y in the (stationary) range
[ -0.9, +0.9] were considered in model 3. If y = 0 the fitted (null) model is correctly
specified. Other values of y generate varying degrees of model misspecification, and
we are interested in the probability that each test rejects the null model (by rejecting
the null hypothesis that 8 = 0 in (14)) when y # 0. For convenience, we will term
these rejection rates “powers” in the ensuing discussion. However, care should be
*Ramsey and Gilhert (1972, p. 185) provide data for two series, xi and x g , for a sample size of T = 10.
(We follow them in “repeating” these values to generate regressors which are “fixed in repeated samples”
+
when considering samples of size T = 20, SO.) As in Thurshy and Schmidt (1977), xg = xi x 2 ; and
x4 = x;/10.
402 DEBENEDI~TISAND GILES
1.2
0.8
1 0.6
RESET
FRESETL
0.4
0.2
Figure 1 Model 1, p = 3, p' = 3, T = 20, 10%significance level.
1.2
0.0
FRESETL
]0.6
c FRESETS L
0.4
0.2
0
~ ~ ~ Q T Q ~ ~ O - W F
ormv

DIAGNOSTIC
TESTING
IN ECONOMETRICS 403
1.2
0.8
1 0.8
0.4
0.2
Figure 3 Model 2, p = 3 , p' = 3 , T = 20, 10% significance level.
1.2
0.8
1 0.6 FRESETL
FR ESET6
0.4
0.2
0
c c
Figure 4 Model 2, p = 3, p' = 3, T = SO, 10%significance level.

404 DEBENEDICTIS
AND GILES
1.2
0.8
RESET
1 0.B
FRESETL
0.4
0.2
1.2
0.8
1 0.6
0.4
0.2
a 1

TESTING 405
taken over their interpretation in the present context. Strictly, of course, the power
of the RESET or FRESET test is the rejection probability when 0 # 0, As noted
in Section 11, this power can be determined (in principle), as the test statistics are
typically doubly noncentral F when 8 # 0. Only in the very special case where 2 8
in (14) exactly coincides with the specification error, t, would “powers” of the sort
that we are computing and reporting actually correspond to theformal powers of the
tests. (In model 1, for example, this would require that the RESET test be applied by
using just xp,fortuitously, as the only “augmenting” variable rather than augmenting
the model with powers of the prediction vector.) Accordingly, it is not surprising that,
in general, the shapes of the various graphs reported in the next section do not accord
with that for the true power curve for an F-test.
The error term, u, was generated to be standard normal, though of course the
tests are scale invariant and so the results are invariant to the value of the true error
variance. The tests were conducted at the 5% and 10%significance levels. As the
RESET and FRESET test statistics are exactly F-distributed if y = 0, there is no
“size distortion” if the appropriate F critical values are used-the nominal and true
significance levels coincide. Precise critical values were generated using the Davies
(1980)algorithm as coded in the DISTRIB command in the SHAZAM (1993)pack-
age. Each component of the experiment is based on 5000 replications. Accordingly,
from the properties of a binomial proportion, the standard error associated with a
rejection probability, ~t ( in Tables 3 4 , takes the value [n(1 - ~t)/5000]’/~, which
takes its maximum value of 0.0071 when y = 0.5. The simulations were undertaken
using SHAZAM code on both a PC and a DEC Alpha 3000/400.
VI. M O N T E CARLO RESULTS
In this section, we present the results of Monte Carlo experiments designed to gather
evidence on the power of the various RESET and the above FRESET tests in Table 1
for each model in Table 2. The experimental results and the graphs of the rejection
rates are given below. For convenience we will refer to the graphs below as “power”
curves. As discussed, it is important to note that these are not conventional power
curves. Only results for case 3 for the RESET and case 2 for the FRESET tests in
Table 1 are presented in detail, for reasons which will become evident. The entries
in Tables 3-5 are the proportions of rejections of the null hypothesis. Not surpris-
ingly, for all models considered, the RESET, FRESETL, and FRESETS tests exhibit
higher rejection rates at the 10% significance level than at the 5% level. However,
the “pattern” of the power curves is insensitive to the choice of significance level,
hence we will focus on the 10%significance level in the remainder of the discussion.
Generally, the patterns of the power curves differ only marginally when the
sample size is increased from 20 to 50. The results for a sample size of 50 display
higher power than the comparable sample-size-20 results, reflecting the consistency
Table 3 Model 1, p = 3, p’ = 3
T = 20 T = 50
5% 10% 5% 10%
y RESET FRESETS FRESETL RESET FRESETS FRESETL RESET FRESETS FRESETL RESET FRESETS FRESETL
-8.0 0.188 0.893 I .WO 0.710 1.000 0.942 1.ooo 0.998 1.000 1.Ooo 1.OOO 0.998
-7.5 0.209 0.909 1.ooo 0.709 1.000 0.946 1.ooo 0.999 1.000 1.000 1.Ooo 0.999
-7.0 0.233 0.913 1.000 0.708 1.ooo 0.944 1.ooo 0.999 1.ooo 1.ooo 1.OOO 1.WO
-6.5 0.258 0.874 1.000 0.705 1.000 0.918 1.000 0.998 1.000 1.000 1.OOO 0.999
-6.0 0.284 0.868 1.ooo 0.700 1.ooo 0.915 1.ooo 0.996 1.ooo 1.ooo 1.Ooo 0.999
-5.5 0.309 0.870 1.000 0.695 1.ooo 0.911 1.000 0.995 1.000 1.000 1.oOO 0.998
-5.0 0.331 0.879 1.ooo 0.687 1.ooo 0.925 1.ooo 0.998 1.ooo 1.000 1.000 0.999
-4.5 0.350 0.801 1.ooo 0.674 1.ooo 0.869 1.000 0.992 1.ooo 1.000 1.Ooo 0.996
-4.0 0.366 0.8% 1.000 0.658 1.000 0.937 1.ooo 0.998 1.000 1.000 1.OOO 0.998
-3.5 0.378 0.872 1.ooo 0.639 1.000 0.918 1.ooo 0.994 1.ooo 1.000 1.000 0.996
-3.0 0.384 0.795 0.997 0.609 1.ooo 0.870 0.999 0.998 1.000 1.000 LOO0 0.998
-2.5 0.369 0.767 0.965 0.576 0.992 0.854 0.996 0.995 1.ooo 0.999 1.OOO 0.998
-2.0 0.337 0.646 0.850 0.516 0.930 0.766 0.975 0.991 1.000 0.992 1.000 0.995
-1.5 0.266 0.414 0.580 0.419 0.738 0.565 0.866 0.941 0.997 0.941 0.998 0.%5
-1.0 0.137 0.216 0.254 0.235 0.392 0.334 0.479 0.706 0.791 0.617 0.882 0.806
-0.5 0.064 0.092 0.091 0.133 0.168 0.161 0.152 0.207 0.233 0.247 0.352 0.315
0.0 0.046 0.050 0.054 0.101 0.106 0.105 0.052 0.054 0.053 0.104 0.102 0.101
0.5 0.141 0.095 0.096 0.233 0.173 0.173 0.339 0.237 0.249 0.471 0.366 0.355
1.o 0.117 0.204 0.176 0.206 0.289 0.332 0.409 0.713 0.643 0.556 0.766 0.817
1.5 0.061 0.381 0.267 0.126 0.413 0.528 0.365 0.916 0.907 0.524 0.954 0.953
2.0 0.037 0.617 0.608 0.091 0.745 0.749 0.406 0.987 0.9% 0.593 0.999 0.992
2.5 0.027 0.720 0.679 0.083 0.771 0.807 0.559 0.980 0.997 0.748 1.000 0.989
3.0 0.02 1 0.778 0.983 0.085 0.994 0.866 0.727 0.998 1.000 0.881 1.Ooo 0.999
3.5 0.019 0.960 0.992 0.087 0.999 0.979 0.854 1.000 1.000 0.956 1.000 1.000
4.0 0.017 0.829 0.999 0.089 1.ooo 0.887 0.935 0.995 1.000 0.987 1.000 0.997
4.5 0.014 0.853 1.000 0.091 1.ooo 0.910 0.976 0.998 1.000 0.998 1.OOO 0.999
5.0 0.011' 0.857 1.000 0.093 1.000 0.908 0.993 0.995 1.000 0.999 1.OOO 0.996
5.5 0.009 0.879 1.000 0.094 1.000 0.929 0.998 0.996 1. O N 1.000 1.OOO 0.997
6.0 0.007 0.915 1.000 0.095 1.ooo 0.956 0.999 1.000 1.000 1.000 1.000 1.000
6.5 0.005 0.920 1.000 0.096 1.ooo 0.949 1.000 0.999 1.000 1.000 1.000 0.999
7.O 0.005 0.846 1.000 0.097 1.000 0.907 1.ooo 0.999 1.000 1.000 LOO0 1.000
7.5 0.003 0.811 1.000 0.098 1.ooo 0.903 1.ooo 0.994 1.000 1.000 1.Ooo 0.997
8.0 0.002 0.913 1.ooo 0.096 1.ooo 0.946 1.000 0.999 1.000 1.000 1.OOO 0.999
408 DEBENEDICTIS
AND GILES
of the tests. This is also in accord with the fact that the larger sample size yields
“smoother” power curves in the FRESETL and FRESETS cases for models 1 and 2,
as in Figs. 1 to 4.
The probability of rejecting a true null hypothesis when specification error is
present depends, in part, on the number of variables included in Z t . In general, our
results indicate, regardless of the type of misspecification, that the use of p = 3 in
the construction of Z , as in (15)yields the most powerful RESET test. However, this
does not always hold, such as in model 3 where the RESET test with only the term jf
included in the auxiliary regression yields higher power for relatively large positive
misspecification ( y > 0.3) and large negative misspecification ( y < -0.7).
The FRESETL and FRESETS tests with p’ = 3 terms are generally the most
powerful of these tests. The pattern of the power curves tends to fluctuate less and
the results indicate higher rejection rates than in the comparable p’ = 2 case. This
is not surprising, as we would expect a better degree of approximation to the omitted
effect as more terms are included. However, the ability to increase the number of
test variables included in the auxiliary regression is constrained by the degrees of
freedom. We focus primarily on the RESET test with p = 3 and the FRESET tests
with p f = 3.
In all cases, the FRESET tests perform equally as well as, and in many cases
yield higher powers than, the comparable RESET tests. A comparison of the rejec-
tion rates of the various tests for the three models considered indicates FRESETL is
the most powerful test for models 1 and 2. The FRESETS test yields higher power
for model 3 than the FRESETL and RESET tests, with the exception of high levels
of misspecification, where FRESETL exhibits higher rejection rates. The FRESETS
test yields higher rejection rates than the comparable RESET test for models 1 and 2,
with two exceptions. First, model 1 in the presence of a high degree of misspecifica-
tion; second, model 2 in the presence of positive levels of misspecification ( y > 0).
However, FRESETL yields higher rejection rates than the RESET test for the two
exceptions. The FRESETL test dominates the RESET test for model 3 , as in Figs. 5
and 6.
The power of the RESET test is excellent for models 1 and 2, and p = 3, with
sample size 50. Then for larger coefficients of the omitted variable, the proportion of
rejections increases to 100%.For model 1, the use of the squares, cubes, and fourth
powers of the predicted values as the test variables for the RESET test results in
power which generally increases as the coefficient of the omitted variable becomes
increasingly negative. In the presence of positive coefficients of the omitted vari-
able, the rejection rate generally increases initially as the level of misspecification
increases but decreases as the coefficient of the omitted variable continues to in-
crease. However, power begins to marginally increase again at moderate levels of
misspecification ( y = 3 ) .
Our results for model 2 indicate the power of the RESET test increases as the
coefficient of the omitted variable increases for lower and higher levels of misspec-
T = 20 T = 50
5% 10% 5% 10%
-1.0 1.0oO 0.891 1.Ooo 1.Ooo 1.Ooo 0.922 1.Ooo 0.995 1.Ooo 1.Ooo 1.Ooo 0.998
-0.9 1.Ooo 0.865 1.Ooo 1.Ooo 1.Ooo 0.902 1.Ooo 0.989 1.Ooo 1.Ooo 1.Ooo 0.992
-0.8 1.Ooo 0.924 0.994 1.0oO 0.999 0.950 1.Ooo 0.998 1.Ooo 1.Ooo 1.Ooo 0.999
-0.7 0.996 0.910 0.832 1.Ooo 0.954 0.937 1.Ooo 0.994 1.Ooo 1.Ooo LOO0 0.996
-0.6 0.883 0.938 0.843 0.980 0.942 0.958 1.Ooo 0.998 1.Ooo 1.Ooo 1.Ooo 0.998
-0.5 0.339 0.931 0.996 0.604 0.999 0.953 1.Ooo 0.996 1.m 1.Ooo 1.Ooo 0.997
-0.4 0.013 0.926 1.o00 0.037 1.Ooo 0.950 0.309 0.998 1.Ooo 0.481 1.Ooo 0.999
-0.3 0.243 0.866 0.999 0.391 1.Ooo 0.909 0.870 0.989 1.Ooo 0.926 1.Ooo 0.994
-0.2 0.880 0.735 0.874 0.945 0.947 0.847 1.Ooo 0.994 1.Ooo 1.Ooo 1.Ooo 0.997
-0.1 0.404 0.221 0.23 1 0.552 0.363 0.340 0.899 0.697 0.776 0.952 0.862 0.792
0.0 0.046 0.050 0.054 0.101 0.105 0.106 0.043 0.M9 0.048 0.094 0.101 0.105
0.1 0.387 0.216 0.290 0.539 0.428 0.339 0.898 0.694 0.847 0.947 0.917 0.785
0.2 0.014 0.647 0.809 0.032 0.902 0.766 0.148 0.975 1.Ooo 0.290 1.0oO 0.985
0.3 0.022 0.851 0.980 0.076 0.995 0.899 0.729 0.991 1.Ooo 0.888 1.Ooo 0.9%
0.4 0.015 0.892 0.945 0.090 0.973 0.930 0.955 0.993 1.Ooo 0.994 1.Ooo 0.995
0.5 0.005 0.825 0.940 0.047 0.969 0.873 0.984 0.98 1 1.Ooo 0.998 1.Ooo 0.988
0.6 0.001 0.912 1.Ooo 0.021 1.000 0.938 0.994 0.993 1.Ooo 1.Ooo 1.Ooo 0.995
0.7 0.001 0.831 1.Ooo 0.014 1.Ooo 0.873 0.999 0.988 1.Ooo 1.Ooo 1.Ooo 0.992
0.8 0.Ooo 0.840 1.Ooo 0.020 1.Ooo 0.882 1.000 0.991 1.Ooo 1.Ooo 1.Ooo 0.994
0.9 0.001 0.910 1.Ooo 0.041 1.Ooo 0.939 1.Ooo 0.998 1.Ooo 1.Ooo 1.Ooo 0.999
1.o 0.001 0.936 1.000 0.099 1.000 0.960 1.Ooo 0.997 1.Ooo 1.Ooo 1.Ooo 0.999
T = 20 T = 50
5% lW0 5% 1WO
~
-0.9 O.OO0 0.858 0.989 0.Ooo 0.995 0.899 0.Ooo 0.997 1.Ooo 0.099 1.Ooo 0.998
-0.8 0.001 0.849 0.987 0.009 0.998 0.891 0.065 0.990 1.Ooo 0.278 1.Ooo 0.993
-0.7 0.020 0.851 0.913 0.069 0.970 0.909 0.586 0.995 1.Ooo 0.815 1.0oO 0.998
-0.6 0.055 0.709 0.672 0.147 0.825 0.819 0.816 0.994 1.Ooo 0.924 1.Ooo 0.996
-0.5 0.073 0.573 0.402 0.173 0.572 0.708 0.801 0.986 0.998 0.908 1.OO0 0.993
-0.4 0.068 0.482 0.252 0.160 0.376 0.609 0.674 0.964 0.934 0.805 0.967 0.982
-0.3 0.060 0.337 0.159 0.129 0.254 0.474 0.449 0.896 0.631 0.603 0.734 0.936
-0.2 0.052 0.174 0.094 0.104 0.167 0.286 0.222 0.617 0.293 0.338 0.410 0.740
-0.1 0.046 0.077 0.064 0.095 0.120 0.150 0.090 0.200 0.108 0.160 0.191 0.309
0.0 0.045 0.053 0.052 0.096 0.102 0.101 0.059 0.055 0.051 0.101 0.103 0.109
0.1 0.042 0.074 0.057 0.094 0.110 0.136 0.069 0.179 0.110 0.130 0.192 0.285
0.2 0.031 0.156 0.085 0.069 0.150 0.249 0.099 0.591 0.329 0.172 0.427 0.712
0.3 0.020 0.255 0.132 0.047 0.211 0.373 0.142 0.840 0.4% 0.241 0.575 0.900
0.4 0.014 0.327 0.181 0.037 0.272 0.462 0.233 0.836 0.527 0.372 0.m 0.882
0.5 0.011 0.360 0.197 0.039 0.287 0.478 0.382 0.839 0.518 0.555 0.618 0.883
0.6 0.011 0.357 0.172 0.037 0.251 0.469 0.524 0.909 0.588 0.695 0.700 0.941
0.7 0.007 0.285 0.122 0.029 0.184 0.388 0.550 0.883 0.722 0.741 0.820 0.928
0.8 0.003 0.162 0.076 0.013 0.115 0.242 0.371 0.786 0.659 0.593 0.781 0.861
0.9 O.OO0 0.037 0.019 0.002 0.037 0.059 0.034 0.277 0.157 0.100 0.250 0.381
DIAGNOSTIC
TESTING
IN ECONOMETRICS 4I I
1.2
0.8
1 0.6
0.4
0.2
. I
‘.
’.
0 . I
c
amrr
Figure 7 Model 2, RESET test ( p = l), 10%significance level.
ification. This result generally holds, but as can be seen by Figure 7, it is possible
for power to decrease as the level of misspecification increases. The test yields low
power at positive levels of misspecification for a sample size of 20 when there is an
omitted multiplicative variable. For model 3 and both sample sizes, the rejection
rate initially increases as the coefficient of the omitted variable increases and then
falls as the degree of misspecification continues to increase.
The powers of the FRESETL and FRESETS tests are excellent for models 1
and 2, when p’ = 3. The proportion of rejections increases to 100%as the coeffi-
cient of the omitted variable increases, with the exception of the FRESETS test* for
model 1. The inclusion of three sine and three cosine terms of the predicted values
as the test variables for the FRESETL and FRESETS tests results in power gener-
ally increasing as the coefficient of the omitted variable increases for models 1 and
2 with both sample sizes. However, as can be seen by Figure 8, it is possible for the
rejection rate to decrease as the coefficient of the omitted variable increases in the
p‘ = 2 case. For models 1 and 2 the “power” curve increases at a faster rate and is
“smoother” for sample size 50.
*For FRESETS, the number of rejections is greater than 90% for higher levels of misspecification.
4 12 DEBENEDICTIS
AND GILES
1.2
0.8
1 0.6
0.4
0.2
v
0 I
9 3 9 9 y q C y y o - N " ' m ~ t - 0
QumU
Figure 8 Model 1, FRESETL test (p' = Z ) , 10%significance level.
For model 3, our results indicate that the rejection rate increases initially as
the coefficient of the omitted variable increases and then decreases as the level of
misspecification continues to increase for positive omitted coefficients of the lagged
variable. However, the rejection rate increases as misspecification increases for neg-
ative coefficients of the omitted lagged variable.
Finally, we have also considered the inclusion of an irrelevant explanatory
variable in order to examine the robustness of the RESET and FRESET tests to an
overspecification of the model. We consider model 1 where the DGP now becomes
yt = Bo + B3x3t + B4x4t + vt
and the "null" becomes
yt 1.0 - 0.4~3, + + yxzt +
~ 4 t (25)
In this case, the coefficient ( y ) of the redundant regressor is freely estimated and
therefore we cannot consider a range of preassigned y values. Our results indicate
that the power results differ negligibly from the true significance levels, as the re-
jection rates fall within two maximum standard deviations of the size.* That is, the
tests appear to be quite robust to a simple overspecification of the model.
*As calculated by binomial proportions.

TESTING 4 I3
VII. CONCLUSIONS
We have considered the problem of testing a regression relationship for possible

misspecification in terms of its functional form and/or selection of regressors, in the
spirit of the family of variable-addition tests. We have proposed a new variant of
Ramsey’s (1969) RESET test, which uses a global (rather than local) approximation
to the misspecified part of the conditional mean of the model. Rather than basing
the variable-addition procedure on polynomial terms of a prediction vector from the
basic model, we suggest the use of a Fourier series approximation. Two ways of trans-
forming the predicted values are considered so that they lie in an admissible range
for such an approximation-a linear transformation gives rise to what we have termed
the FRESETL test, while a sinusoidal transformation results in what we have called
the FRESETS test.
These two new test statistics share the property of the usual RESET statistic of
being exactly F-distributed under the null hypothesis of no conditional mean mis-
specification, at least in the context of a model with nonstochastic regressors and
spherical disturbances. We have undertaken a Monte Carlo experiment to determine
the power properties of the two variants of the FRESET test, and to compare these
with the power of the RESET test for different forms of model misspecification, and
under varying choices of the number of “augmentation” terms in all cases. Our sim-
ulation results suggest that using the global Fourier approximation may have advan-
tages over using the more traditional (local) Taylor’s series approximation in terms
of the tests’ abilities to detect misspecification of the model’s conditional mean.
Our results also suggest that using a Fourier approximation with three sine and
cosine terms results in a test which performs well in terms of “power.” The empirical
rates of rejection for false model specifications exhibited by the FRESET tests are
at least as great as (and generally greater than) those shown by the RESET test. The
FRESETL test is generally the best overall when the misspecified model is static,
while the FRESETS test is best overall when the model is misspecified through the
omission of a dynamic effect. In practical terms, this favors a recommendation for
using the latter variant of the test.
Although the proposals and results in this chapter seem to offer some improve-
ments on what is, arguably, the most commonly applied variable-addition test in em-
pirical econometric analysis, there is still a good deal of research to be undertaken
in order to explore the features of both the RESET and the FRESET tests in more
general situations. Contrary to earlier apparent “evidence,” such as that of Thursby
(1979, 1982), it is now well recognized (e.g., Pagan 1984, p. 127; Godfrey 1988, p.
107; Porter and Kashyap 1984) that the RESET test is not robust to the presence of
autocorrelation and/or heteroskedasticity in the model’s errors. The same is clearly
true of the FRESET tests. Work in progress by the authors investigates the degree
of size distortion of these tests and their ability to reject falsely specified models in
these more general situations. As well as considering the same forms of the RESET
and FRESET tests that are investigated here, we are also exploring the properties of
4 14 DE~ENEDICTIS
AND GILES
the corresponding tests when the robust error covariance matrix estimators of White
(1980) and Newey and West (1987)are used in their construction. In these cases,
as well as in the case of models in which the null specification is dynamic and/or
nonlinear in the parameters, asymptotically valid (chi-square) counterparts to the
RESET and FRESET tests are readily constructed, as in Section 1I.D. The finite-
sample qualities of these variants of the tests are also under investigation by the
authors.
ACKNOWLEDGMENTS
We are grateful to Karlyn Mitchell for helpful correspondence in relation to the use of
Fourier functional forms and for supplying unpublished material, to Lindsay Tedds
for her excellent research assistance, and to Peter Kennedy, John Small, and Michael
Veal1 for their helpful comments. This research was supported by University of Vic-
toria Internal Research Grant 1-41566.
REFERENCES
Amemiya, T. (1977), A Note on a Heteroscedastic Model, Journal of Econometrics, 6, 365-

370.
Anderson, T. W. (1971), The Statistical Analysis of Time Series, Wiley, New York.
Anscombe, F. J. (1%1), Examination of Residuals, Proceedings of the Fourth Berkeley Sym-
posium on Mathematical Statistics and Probability, 1, 1-36.
Bahadur (196), Stochastic Comparison of Tests, Annals of Mathematical Statistics, 31, 276-
295.
Bahadur (1967), Rates of Convergence of Estimates and Tests Statistics, Annals of Mathemat-
ical Statistics, 38, 303-324.
Basu, D. (1955), On Statistics Independent of a Complete Sufficient Statistic, Sankhyb, 15,
377-380.
Bera, A. K. and C. M. Jarque (1982), Model Specification Tests: A Simultaneous Approach,
Bera, A. K. and C. R. McKenzie (1987), Additivity and Separability of the Lagrange Multiplier,
Likelihood Ratio and Wald Tests, Journal of Quantitative Economics, 3 , 5 3 4 3 .
Box, M. J. (1966), A Comparison of Several Current Optimization Methods, and the Use of
Transformations in Constrained Problems, Computer Journal, 9,67-77.
Breusch, T. S. (1978), Testing for Autocorrelation in Dynamic Linear Models, Australian Eco-
nomic Papers, 17,334-355.
Breusch, T. S. and A. R. Pagan (1979), A Simple Test for Heteroscedasticity and Random
Coefficient Variation, Econometrica, 47, 1287-1294.
Christensen, L. R., D. W. Jorgensen, and L. J. Lau (1971), Conjugate Duality and the Tran-
scendental Logarithmic Production Function (abstract), Econometrica,39,255-256.
TESTING 4 15
Christensen, L. R., D. W. Jorgensen, and L. J. Lau (1973), Transcendental Logarithmic Pro-

duction Frontiers, Review of Economics and Statistics, 55,28-45.
David, H. A. (1981), Order Statistics, Wiley, New York.
Davidson, R. and J. G. MacKinnon (1981), Several Tests for Model Specification in the Pres-
ence of Alternative Hypotheses, Econometrica, 49,781-793.
Davidson, R. and J. G. MacKinnon (1993). Estimation and Inference in Econometrics, Oxford
Davies, R. B. (1980), The Distribution of a Linear Combination of ,y2 Random Variables:
Algorithm AS 155, Applied Statistics, 29,323-333.
Durbin, J. (1954), Errors in Variables, Review of the International Statistical Institute, 22,
23-32.
Durbin, J. (1970), Testing for Serial Correlation in Least Squares Regression When Some of
the Regressors Are Lagged Dependent Variables, Econometrica, 38,41042 1.
Eastwood, A. and L. G. Godfrey (1992),The Properties and Constructive Use of Misspecifica-
tion Tests for Multiple Regression Models, in L. G. Godfrey (ed.), The Implementation
and Constructive Use of Mis-specijcation Tests in Econometrics, Manchester University
Press, Manchester, 109-175.
Eastwood, B. J. and A. R. Gallant (1991), Adaptive Rules for Semiparametric Estimators that
Achieve Asymptotic Normality, Econometric Theory, 7,307-340.
Engle, R. F. (1982),Autoregressive Conditional Heteroscedasticity with Estimates of the Vari-
ance of U.K. Inflation, Econometrica, 50,987-1007.
Gallant, A. R. (1981), On the Bias in Functional Forms and an Essentially Unbiased Form:
The Fourier Flexible Form, Journal of Econometrics, 15,211-245.
Garbade, K. (1977),Two Methods of Examining the Stability of Regression Coefficients, Jour-
nal of the American Statistical Association, 72, 54-63.
Geweke, J. (1981), The Approximate Slopes of Econometric Tests, Econometrica, 49, 1427-
1442.
Giles, J. A. and D. E. A. Giles (1993),Pre-test Estimation and Testing in Econometrics: Recent
Developments, Journal of Economic Surveys, 7, 145-197.
Godfrey, L. G. (1978), Testing against General Autoregressive and Moving Average Error
Models When the Regressors Include Lagged Dependent Variables, Econometrica, 46,
1293-1 302.
Godfrey, L. G. (1988), Misspecijication Tests in Econometrics: The Lagrange Multiplier Princi-
ple and Other Approaches, Cambridge University Press, Cambridge.
Godfrey, L. G. (1996), Misspecification Tests and Their Uses in Econometrics, Journal of
Statistical Planning and Inference, 49,24 1-260.
Godfrey, L. G., M. McAleer, and C. R. McKenzie (1988), Variable Addition and Lagrange
Multiplier Tests for Linear and Logarithmic Regression Models, Review of Economic
Studies, 70,492-503.
Goldfeld, S. M. and R. E. Quandt (1965), Some Tests for Homoscedasticity, Journal of the
American Statistical Association, 60, 539-547.
Harrison, M. J. and B. P. M. McCabe (1979), A Test for Heteroscedasticity Based on Ordi-
nary Least Squares Residuals, Journal of the American Statistical Association, 74,
494499.
Hausman, J. A. (1978), Specification Tests in Econometrics, Econometrica, 46, 1251-1272.
Hendry, D. F. (1980), Econometrics: Alchemy or Science?, Economica, 47,387-406.
4 16 AND GILES
DEBENEDICTIS
Jarque, C. J. and A. K. Bera (1980), Efficient Tests for Normality, Homoscedasticity and Serial
Independence of Regression Residuals, Economics Letters, 6,255-259.
Jarque, C. J. and A. K. Bera (1987), A Test for Normality of Observations and Regression
Residuals, International Statistical Review, 55, 163-172.
Johnston, J. (1960), Statistical Cost Analysis, McGraw-Hill, New York.
Kramer, W. and H. Sonnberger (1986), The Linear Regression Model under Test, Physica-
Verlag, Heidelberg.
Lee, J. H. H. (1991), A Lagrange Multiplier Test for GARCH Models, Economics Letters, 37,
265-27 1.
MacKinnon, J.G. and L. Magee (1990),Transforming the Dependent Variable in Regression
Models, International Economic Review, 3 1 , 315-339.
Magee, L. (1987), Approximating the Approximate Slopes of LR, W and LM Test Statistics,
Econometric Theory, 3,247-271.
McAleer, M. (1987), Specification Tests for Separate Models: A Survey, in M. L. King and
D. E. A. Giles (eds.), Specijcation Analysis in the Linear Model, Routledge and Kegan
Paul, London, 146-196.
McAleer, M. (1995), Sherlock Holmes and the Search for Truth: A Diagnostic Tale, reprinted
in L. Oxley et al. (eds.), Surveys in Econometrics, Blackwell, Oxford, 91-138.
Milliken, G. A. and F. A Graybill (1970), Extensions of the General Linear Hypothesis Model,
Journal ojthe American Statistical Association, 65, 797-807.
Mitchell, K. and N. M. Onvural (1995), Fourier Flexible Cost Functions: An Exposition and
Illustration Using North Carolina S & L’s, mimeo., College of Management, North Car-
olina State University.
Mitchell, K. and N. M. Onvural(19%), Economies of Scale and Scope at Large Commercial
Banks: Evidence from the Fourier Flexible Functional Form, Journal ofMoney, Credit
and Banking, 28,178-199.
Mizon, G. E. (1997a), Inferential Procedures in Nonlinear Models; An Application in a U.K.
Cross Section Study of Factor Substitution and Returns to Scale, Econometrica, 45,
1221-1242.
Mizon, G. E. (1977b), Model Selection Procedures, in M. J. Artis and A. R. Nobay (eds.),
Studies in Modern Economic Analysis, Blackwell, Oxford, 97-120.
Newey, W. and K. West (1987), A Simple, Positive-Definite, Heteroskedasticity and Autocor-
relation Consistent Covariance Matrix, Econometrica, 55, 703-708.
Ohtani, K. and J. A. Giles (1993),Testing Linear Restrictions on Coefficients in a Linear Model
with Proxy Variables and Spherically Symmetric Disturbances, Journal of Economet-
rics, 57,393-406.
Pagan, A. R. (1984), Model Evaluation by Variable Addition, in D. F. Hendry and K. F. Wallis
(eds.), Econometrics and Quantitative Economics, Blackwell, Oxford, 103-133.
Pagan, A. R. and A. D. Hall (1983), Diagnostic Tests as Residual Analysis, Econometric Re-
views, 2, 159-218.
Pesaran, M. H. (1974), On the Problem of Model Selection, Review of Economic Studies, 41,
153-17 1.
Phillips, G. D. A. and B. P. M. McCabe (1983), The Independence of Tests for Structural
Change in Econometric Models, Economics Letters, 12,283-287.
Phillips, G. D. A. and B. P. M. McCabe (1989), A Sequential Approach to Testing for Structural
Change in Econometric Models, in W. Kramer (ed.), Econometrics ojStructural Change,
Physica-Verlag, Heidelberg, 87-101.
DIAGNOSTIC
TESTING
IN ECONOMETRICS 4 17
Porter, R. D. and A. K. Kashyap (1984),Autocorrelation and the Sensitivity of RESET, Eco-

nomics Letters, 14,229-233.
Ramsey, J. B. (1%9),Tests for Specification Errors in Classical Linear Least-Squares Regres-
sion Analysis, Journal of the Royal Statistical Society B , 31, 350-371.
Ramsey, J. B. (1983), Comment on “Diagnostic Tests a s Residual Analysis,” Econometric
Reviews, 2,241-248.
Ramsey, J. B. and R. Gilbert (1972),A Monte Carlo Study of Some Small Sample Properties of
Tests for Specification Error, Journal of the American Statistical Association, 67, 180-
186.
Ramsey, J. B. and P. Schmidt (1976), Some Further Results in the Use of OLS and BLUS
Residuals in Specification Error Tests, Journal of the American Statisticctl Association,
71,389-390.
Salkever, D. S. (1976),The Use of Dummy Variables to Compute Prediction Errors and Con-
fidence Intervals, Journal of Econometrics, 4, 393-397.
Savin, N. E. (1984),Multiple Testing, in Z. Griliches and M. D. Intriligator (eds.), Handbook
of Econometrics, Vol. 2, North-Holland, Amsterdam, 827-879.
Schwager, S. J. (1984),Bonferroni Sometimes Loses, American Statistician, 38, 192-197.
SHAZAM (1993), SHAZAM Econometrics Progrum: User’s Reference Manual, Version 7.0,
McGraw-Hill, New York.
Theil, H. (1965),The Analysis of Disturbances in Regression Analysis, Journal ofthe Amer-
Theil, H. (1968),A Simplification of the BLUS Procedure for Analyzing Regression Distur-
bances, Journal of the American Statistical Association, 63,242-251.
Thursby, J. G. (1979),Alternative Specification Error Tests: A Comparative Study, Journal of
the American Statistical Association, 74,222-223.
Thursby, J. G. (1982),Misspecification, Heteroskedasticity, and the Chow and Goldfeld-Quandt
Tests, Review of Economics and Statistics, 64, 314-321.
Thursby, J. G. (1989),A Comparison of Several Specification Error Tests for a General Alter-
native, International Economic Review, 30,217-230.
Thursby, J. G. and P. Schmidt (1977),Some Properties of Tests for Specification Error in a
Linear Regression Model, Journal oj”the American Statistical Association, 72, 635-
641.
White, H. (1980),A Heteroskedastic-Consistent Covariance Matrix and a Direct Test for Het-
eroskedasticity, Econometrica, 48, 421-448.
White, H. (1982), Maximum Likelihood Estimation of Misspecified Models, Econometrica,
50, 1-25.
White, H. (1987), Specification Tests in Dynamic Models, in T. Bewley (ed.), Advances in
Econometrics-F$h World Congress, Vol. 1, Cambridge University Press, New York,
1-58.
White, H. (1994), Estimation, Injerence and Speci&cation Analysis, Cambridge University
Press, Cambridge.
Wu, D. M. (1973),Alternative Tests of Independence between Stochastic Regressors and Dis-
turbances, Econometrica, 41,733-750.
I2
Applications of the Bootstrap in
Econometrics and Economic Statistics
Michael R. Veall
McMaster University Hamilton, Ontario, Canada
1. INTRODUCTION
Empirical economists and econometricians have long looked to simulation studies

for guidance concerning the choice and performance of estimators whose theoret-
ical justification is often only asymptotic. For example, in any econometrics class
that describes a method based on an asymptotic distribution, an instructor will al-
most always encounter the question, “How large a sample is large enough?” It is a
good question, particularly when in most applications the character of the data (e.g.,
trending vs. nontrending) matters as much as the actual number of observations.
Most instructors will give an answer based on simulation studies, but in an applica-
tion there is often a considerable gap between cases that have been studied by Monte
Carlo methods and the case at hand. The bootstrap method may be regarded as a sim-
ulation study that is tailored to the actual data being studied, with the results used
either to fill in statistical gaps that do not yield easily to analytic methods (such as
providing standard errors or confidence intervals when they are otherwise unavail-
able) or to adjust the original statistical estimates in an attempt to improve finite-
sample accuracy. It is therefore not surprising that the bootstrap has proven useful
to many empirical researchers in economics, especially as the approach replaces
difficult or intractable theoretical calculations with computer calculations that are
becoming cheaper and cheaper over time. While bootstrap-like notions had existed
previously, even within econometrics, the seminal work for these developments is
Efron (1979),the classic paper in statistics that named the bootstrap, developed it
as a unified technique, and demonstrated how computer power could widen the scope
of its implement ati on.
419
420 VEALL
We shall suggest that there are two principal stages in the development of the
bootstrap. In the first stage, the bootstrap was used as a replacement for analytic
methods of calculating standard errors, confidence intervals, etc. This makes obvi-
ous sense when there are no available analytic methods (the “better than nothing”
case). This first-stage bootstrap has also sometimes been used by researchers under
the impression that its finite-sample properties would prove superior to the analytic
alternatives. While this is sometimes true, there is little general theoretical support
for this position. Section I1 focuses on this first stage of development. In Section
111 a number of applications are described, emphasizing cases such as confidence
intervals for forecasts and inference after specification search, where the bootstrap
may be used because there is no good alternative. Section IV emphasizes the second
stage of the bootstrap literature concerning cases where asymptotic, analytic tools
are available but in which bootstrap refinements are used to improve finite-sample
performance. These are most valuable when the original estimation is plagued by
substantial bias. Section V considers more applications, sometimes comparing the
results of different approaches. The cases of system estimation and nonlinearities are
emphasized. Section VI summarizes briefly and concludes, A principal conclusion
is that the bootstrap, while no panacea, may be an important step toward a style of
econometric practice that routinely checks the applicability of inferential tools that
are not exact in the statistical sense and hence depend on some form of potentially
unreliable asymptotic approximation. This argument for simulation is in addition to
that of McCloskey and Ziliak (1996),who suggest simulation as a tool for elucidating
the economic meaning of econometric results.
Finally a caveat: no paper could now cover all the published applications of
the bootstrap in empirical economics, let alone the theoretical developments which
may prove relevant to future practice. This chapter instead tries to describe the
development of the bootstrap along the lines indicated, using the vocabulary most
economists will know from their experience with regression methods, with emphasis
on the standard techniques involving confidence intervals or hypothesis tests. The
focus is on what an economic statistician or econometrician can reasonably expect
from the bootstrap at its current stage of development. Hence, only a few areas of
bootstrap application are considered, including inference problems when a system of
equations has been estimated or when the estimate of interest is a nonlinear function
of the parameter estimates. These are chosen to illustrate the strengths and weak-
nesses of the bootstrap approach. There are a number of surveys that cover many
applications, including Veal1 (1989),Jeong and Maddala (1993),Vinod (1993),and
Li and Maddala (1996).The last survey is particularly recommended for researchers
interested in bootstrapping in time-series contexts, with potential problems due to
unit roots, lagged dependent variables, and serial correlation. There are also books
that treat the statistical background of the bootstrap in more detail, with Efron’s early
monograph (1982) still a useful reference. Efron and Tibshirani (1993) is probably
the most straightforward reference for econometric practitioners, while some of the
APPLICATIONS
OF THE BOOTSTRAP 42 I
discussion in Hall (1992) is at a more difficult theoretical level and includes more
emphasis on the nature of bootstrap approximation. (LePage and Billard 1992 also
include theoretical papers on the limits of the bootstrap: See, for example, Arcones
and Gin6 1992, who prove that the bootstrap distribution of any M-estimator con-
verges weakly to the true limiting distribution in probability. This provides the basis
for a similar result in Hahn 1996 for generalized method of moments estimators.)
It should be noted that the Hall and Efronmibshirani books represent different ap-
proaches on a key aspect of bootstrap practice, as will be discussed further in Sec-
tions IV, V, and VI.
II. THE SIMPLE BOOTSTRAP

A. The Basic Method
Because it is so familiar to economists, we use the linear regression model as our
example, and, where possible, we describe the bootstrap using the language of a
Monte Carlo experiment. The linear regression model is
=X;S+e;, i = l , ...,n
and the row k-vector X ; are observations on the dependent and independent vari-
B
ables respectively with the first element of X i a 1, ,6 and are column vectors of
parameters and their estimates respectively, U ; are random errors which will be as-
sumed to be independently and identically distributed (i. i. d) with mean zero and
variance 0 2 ,and e; are regression residuals. The bootstrap estimate of the variance
matrix in this case is calculated by the following simulation experiment.
Step 1: Create an artificial sample:
(2)
The elements er' are created by resampling from e;, that is drawing from e; randomly
but with replacement. (We have assumed there is an intercept so that the e; have
mean zero; otherwise we would have to work with deviations in e; by subtracting off
the mean.) The artificial sample can be thought of as the result of one trial of a Monte
Carlo experiment where the independent variables are set as the actual values of X;,
fi
the parameters are set as the actual data estimates and its disturbances are the
draws from the empirical distribution function of the residuals e ; , that is the discrete
probability distribution function that attaches probability l/n to each value e;.
Step 2: Estimate B*' on artificial sample 1.
Step 3: Repeat steps 1 and 2 B times.
Step 4: The sampling distribution of the B*J a
- over the B Monte Carlo
bootstrap trials is an estimate of the sampling distribution of fi - @. In particular,
422 VEAU
B
the variance of is estimated by fi*(B*J), the sample variance matrix of the over B*J
the B bootstrap samples.
An alternative, known as the parametric bootstrap, is identical to the above
except the disturbances are drawn from a particular distribution with its parame-
ters set as estimates from the residuals, most commonly a normal random number
generator with mean zero and variance equal to the residual variance. It is also pos-
sible to generate the disturbances as weighted averages of draws from the empirical
distribution function of the residuals and the parametric distribution; this method
may be called the smoothed bootstrap with the weights determining the degree of
smoothing. Other techniques in Monte Carlo analysis, such as importance sampling
and antithetic variates, can also be applied to the bootstrap. (See Efron and Tib-
shirani 1 9 9 3 for some discussion and references.) These alternatives have the same
essential properties as the ordinary bootstrap, which will remain our focus as it is
the method of resampling that has almost exclusively been employed in econometric
applications.
B. Why the Bootstrap Works

To consider why this approach may work, consider the sampling distribution of a B * j ,
where j denotes the jth bootstrap sample drawn from a population based on the real
sample in the manner described in the previous section. Call v ( B * j ) the variance ma-
trix of ) * j based on this distribution function. Just as in any simulation experiment,
as B grows large a*@*’) approaches v(B*j). Note that B * j = ( X T X ) - ’ X T Y * J =
B+ ( X T X ) - ’ X T e * J , where X is the n x k matrix of the row vectors X i , and Y*J
and e*i are the N-vectors corresponding to the jth artificial sample. Treating e*J as
a random variable drawn from the empirical distribution function, it is straightfor-
ward to show that, because the e;’s sum to zero and have equal probability weights
l / n , E ( e * j ) = 0 and v(e*J) = ( e r e / n ) Z n , where e is the vector of ei’s and In is the
n x n identity matrix. Therefore, E ( B * j ) = and v(B*J) = v ( ( X T X ) - ’ X T e * j ) =
( X T ~ ) - ’ X T v ( e ; j ) X ( X T X ) - ’= ( e T e / n ) ( X T X ) - ’ . But note this last expression is
just the usual estimate of v(B) as calculated by a regression package without boot-
strapping, with the exception that the division is by n rather than n - k . (This is
sometimes corrected in practice by scaling the original residuals by ( n / ( n - k))1’2.)
We have therefore shown that we can compute the ordinary least-squares (OLS) esti-
mate of the covariance matrix of the coefficients if we use as our population a proba-
bility model based on the sample. As it is sometimes put for bootstrapping statistics
more generally, as the sample is drawn from an underlying population, a resample
from that sample can be thought of as being drawn from something similar to the
underlying population. Hence we may infer the population variance of a statistic by
observing its behavior in resampling. A similar approach can be used to calculate
other statistics such as the limit points of confidence intervals.
APPLICATIONS
OF THE BOOTSTRAP 423
Why have we bothered to show this simple result, particularly as it seems the
bootstrap is redundant in this case, in that it is not essentially different from the
analytic estimate? First, the fact that the bootstrap is so close to the right answer
in a case where we know the right answer is reassuring given that the bootstrap is
also proposed in other contexts in which analytic formulas are not known. Second,
that the bootstrap estimate has divisor n when we know that the exact test requires
(among other things) that the divisor be n - k emphasizes that the justification for
the use of the simple bootstrap is typically only asymptotic, although in this case
a simple rescaling eliminates the discrepancy. (We shall refer to this issue later.)
Third, it shows that the bootstrap estimate of the variance, just like the standard an-
alytic variance estimate, will only be “correct” for inference if the initial model is
correctly specified. For example, the disturbances must be identically and indepen-
dently distributed; hence practitioners should check these assumptions by suitable
diagnostics before bootstrapping. Finally, the exercise emphasizes that the bootstrap
does not create any additional information. It is simply a computational device to uti-
lize information already in the original sample.
The type of bootstrap just described in the regression context was called “resid-
ual resampling” by Kiviet (1984), because it kept the fixed structure of the indepen-
dent variables and only resampled on the residuals. Alternatively we could use what
Kiviet called “complete resampling,” where resampling is from the row (k 1)- +
vectors (x, X i ) and then the bootstrap algorithm proceeds as usual. As discussed
in Jeong and Maddala (1993, pp. 577-578), this method should give a consistent
estimate of the variance-covariance matrix of the estimated coefficients even in the
presence of random X’s and/or heteroskedasticity. The discussion in Efron and Tib-
shirani (1993, Chap. 9) goes in a different direction, pointing out that the bootstrap
will yield an estimate of the sampling distribution of b provided only that the origi-
+
nal observation vectors were chosen from the underlying (k 1)-variate probability
distribution. This does not even depend on the existence of a “true” linear model
as in the first line of (1).Naturally, however, if there is no true model, interpretation
will be difficult: we shall have an estimate of the sampling distribution of with-
out knowing its relationship, if any, to the parameter being estimated, if any. Even
granting knowledge of the true model, the assumption of random draws in the com-
plete resample precludes most time-series estimation; it, however, may be a natural
approach to use in cross-sectional estimation.
C. An Example
As an example consider the Theil(l971) textile data where Yt,the log of Netherlands
annual textile consumption for 1923-1939, is regressed on X,1, the log of real income
per capita, and X,2, the log of the relative price of textiles:
Pt = /I0 + B l X t i + / I Z X t 2 = 1.37 + 1.14Xll - .83X,;! (3)

Table I Various Standard Error Estimates for the Theil Data
~~ -- ~
Analytic n.a. .3061 .lS60 .0361

Bootstrap (unscaled) 50 .2893 .1460 .0321
100 .2616 .1352 .0310
1000 .2824 .1449 .0332
Bootstrap (scaled) SO .3188 .1609 .0353
100 .2883 .1489 .0342
1000 .3120 .1608 .0362
Bootstrap (parametric) 50 .3272 .1627 .0348
100 3029 .1526 .0366
1000 .3003 .1S31 .0347
Bootstrap (complete resampling) 50 .2926 .1641 .0379
100 .2617 .1476 .0430
1000 .2613 .1375 .0391
=s.e. = standard error, and R = number of bootstraps.
Residual bootstrapping is based on the assumption that the underlying disturbance

terms are independently and identically distributed, so these assumptions should be
checked (as has been done for this example extensively in the numerous exercises
in the manual for the computer package SHAZAM 1993). Table 1 reports a variety
of standard error estimates, where the bootstrap standard errors are calculated in the
usual manner as the square root of the diagonal elements of the bootstrap estimate
of the variance matrix.
The first thing to notice from the table is that all the residual bootstrap meth-
ods yield answers close to the analytic case; the complete resampling bootstrap re-
sults are not as close, nor should they be, for as described above they are more like
estimates of the heteroskedasticity-consistent standard errors rather than the ordi-
nary standard errors. It is also clear that the number of bootstraps is not that impor-
tant, although more seems to be better. Scaling the residuals (by ( n / ( n- IL))'/~=
(17/14)li2 in this case) increases the bootstrap standard errors somewhat as would
be expected, from being a bit smaller than their analytic counterparts to being a bit
larger. The parametric bootstrap, which uses the normal distribution with mean zero
and variance e T e / ( n- k) instead of the empirical distribution function of the residu-
als in order to generate the artificial samples, yields very similar answers to the other
residual bootstrap techniques.
D. Bootstrapping Confidence Intervals

The main reason to calculate standard errors is for inference, by means of hypothesis
tests or confidence regions. Hypothesis tests involving the coefficients are execut-
APPLICATIONS
able by determining whether corresponding confidence regions cover the null hy-
pothesis parameter values. Noting this equivalence, we shall concentrate our dis-
cussion on bootstrapping confidence regions. Moreover for simplicity and because
the theory is better developed in this area, we focus on confidence regions involv-
ing one parameter, which of course correspond to t-statistics in a hypothesis-testing
framework.
One way to use the bootstrap in inference would be to use the bootstrap stan-
dard errors instead of analytic standard errors in the basic confidence interval for-
mula. We call these bootstrap standard error confidence intervals, and they are of the
si
form ft:2k -s.e.*(bi).As the bootstrapped standard errors were similar to the an-
alytic standard errors in the linear regression context, therefore bootstrap confidence
intervals of this form will not be much different from their analytic counterparts. An-
other more popular bootstrap confidence interval has been the percentile method.
In this, the artificial sample values of the statistic of interest, such as a regression
coefficient, are sorted and the 1 - 2a confidence intervals are set as the a t h and
(1 - a)th percentiles. While some find it intuitively more pleasing, in terms of re-
gression coefficient estimates these intervals tend to be similar to analytic or boot-
strap standard error confidence intervals, even when the disturbances are nonnormal
(although the combination of severe nonnormality, few degrees of freedom, and very
small a can lead to substantial differences). Regression estimates are weighted av-
erages, and very commonly this ensures (by way of a central limit theorem), that
the distribution function of 6" is approximately normal, as it is for 6. Hence for
constructing confidence intervals of linear regression coefficients, there is relatively
little to be gained by bootstrap methods.
More generally, Hall (1992) demonstrates that in most cases (essentially for
any root-n-consistent statistic that may be expanded in Edgeworth form), the end-
points of the bootstrap percentile confidence intervals and of the bootstrap standard
error confidence intervals are accurate only to 0 ( n - ' j 2 ) , which in general is the
accuracy that any analytic asymptotic method can be expected to achieve. Similar
results hold for the accuracy of the tail coverage-that is, the degree of approxima-
tion of the area in each tail outside the confidence interval to its putative value of
a. Given this, the best case for using these types of bootstrap methods is when there
is really no alternative. We shall discuss two cases: (1) forecasting from the linear
regression model and (2) inference after a specification search.
111. APPLICATIONS OF FIRSTSTAGE BOOTSTRAPPING

METHODS
A. Confidence Intervals for Forecasts
One context where the bootstrap may be very useful is in the confidence intervals for
8,
forecasts. If we stay with the linear regression context, the forecast is xn+l where
426 VEALL
xn+l is the row vector of observations on the k right-hand-side variables for period
+ f is
n 1. The forecast error e,+l
While the variance of this expression can be estimated analytically, note that in con-
structing confidence intervals, a central limit theorem does not ensure normality:
the second term of (4) may tend to the normal distribution as n grows large, but the
first will only be normal if the disturbances themselves are normal. To deal with this
nonnormality is difficult analytically but is straightforward using the bootstrap. Fol-
lowing Freedman and Peters (1984a, 1984b), we bootstrap just as above only we
focus on the forecast x,+lD*’ as the statistic of interest. We then calculate a “sim-
ulated actual” by adding an additional single bootstrapped residual to the actual
forecast %,+lb.The difference between the forecast and the simulated actual is the
simulated forecast error; we obtain an estimate of the probability distribution of the
forecast error by repeating the process B times.
Moreover, unlike the standard OLS case, in bootstrap simulation it is easy to
incorporate uncertainty in the x,+1 in the forecast confidence intervals, the impor-
tance of which is stressed by Feldstein (1971). Early contributions to this approach
are the stochastic simulation methods of, for example, Brown and Mariano (1984)
and especially Fair (1979, 19801, who, independently of the bootstrap literature,
proposed the same bootstrap method of evaluating forecast uncertainty. In addition,
Fair proposes a modification to the basic bootstrap uncertainty measures to make
allowance for specification error and applies the method to macroeconomic forecast-
ing in the United States. Freedman and Peters (1984a) use the bootstrap technique
to develop forecast standard errors in a generalized least-squares application in-
volving United States electricity consumption by region. Veall (1987a) applies the
method, with emphasis on the percentile method and the uncertainty in the inde-
pendent variable forecasts, to forecasting the demand for electricity in Ontario in a
time-series context; Veall (1987b) is a Monte Carlo study that confirms the reliability
of the approach for this problem. Bernard and Veall (1987) extend the same exercise
for Quebec emphasizing the dynamics still more. Prescott and Stengos (1987) use
the same approach for studying the United States supply of pork.
B. The Bootstrap and Specification Search

Much of the focus in econometric theory is on sampling error, yet, in practice, speci-
fication error is the more vexing question in many econometric applications. We typ-
ically cannot create more data by laboratory-type experiments, and it is relatively
rare to have such a large data set that we can select the specification on a training
set and then estimate on different data. Hence, specification search and estimation
are typically on the same data. While the process of changing a specification in re-
APPLICATIONS
sponse to initial results may be an important part of empirical modeling, the result-
ing “pretesting” may lead to the coefficients and standard errors estimated from the
eventual model being seriously biased.
There have been a few attempts to use simulation methods as a way of treat-
ing the specification error problem. The work of Fair (1979, 1980), which began to
deal with specification uncertainty, has already been cited in the prediction con-
text. Efron and Gong (1983)attempted to grapple with the problem of specification
search by studying the sampling distribution for estimates from an entire data-mining
procedure by bootstrap simulation and provide an example relating to hepatitis di-
agnosis. The entire decision tree of the investigator is laid out (e.g., step 1: esti-
mate whole model; step 2: drop all variables whose coefficients have t-statistics less
than 1, etc.) and then applied on the data. Then the same entire decision tree is
applied to each of the bootstrap samples, generated either by the complete resam-
pling method or the residual resampling method using either the first-stage or the
final model.
Brownstone (1990)and Veal1 (1992) apply this technique to econometric ex-
amples, the former also estimating the standard errors of Stein-James shrinkage
estimators in this context. (See Vinod and Raj 1988 for an application involving
ridge regression.) Freedman, Navidi, and Peters (1988) and Dijkstra and Veldkamp
(1988)study this kind of method for stylized data-mining procedures and, in general,
find that the simulation method does not yield accurate standard errors for the data-
mining estimator. The results of Freedman, Navidi, and Peters are particularly dis-
couraging, although it should be noted that their example is based on an initial stage
of estimating 75 coefficients from 100 observations. It must also be remembered that
the method is only valid if the estimation procedure can be modeled as if it were a
prespecified decision tree: if new hypotheses and approaches were entertained only
after seeing the first set of results, then, strictly speaking, even bootstrapping the
entire procedure as run does not solve the fundamental problem. Nonetheless this
method does seem to be the only feasible possibility at this time for dealing with the
pretest issue in any real estimation context. It does seem a minimal requirement for
any estimated econometric model that if Monte Carlo samples are generated from it
and the entire data-mining procedure is applied to those samples, the results over the
Monte Carlo samples should be consistent in all important respects with the results
from the original data.
Finally, in a Bayesian context, Geweke (1986)proposes a useful method of im-
plementing strong priors on coefficient signs in regression methods, essentially by
bootstrapping an unrestricted regression model but basing all estimates on the arti-
ficial samples for which the estimates meet the sign restrictions. Hence, if the only
prior is that the income coefficient is positive, the estimate of the income coefficient
will be the average of all the positive estimates over the bootstrap samples and the
standard error will be the standard deviation of these estimates. Some researchers
may wonder why they should ignore the negative estimates in this context, but this is
the consequence of the prior they purport to hold. Chalfant, Gray, and White (1991)
is an application of the same technique.
IV. BOOTSTRAP REFINEMENTS

A. Simple Bias Correction
As we have discussed, application of bootstrap confidence intervals or test proce-
dures based upon bootstrap standard errors or on the percentile method does not
lead to a more accurate (Edgeworth) approximation than asymptotic methods, and
we know of no other theoretical method of comparison. Hence there seems to be little
justification for using the bootstrap when asymptotic methods are available. There
is, however, an informal argument that if one can regard the bootstrap as a Monte
Carlo experiment, the bootstrap distribution of an estimator should in some sense
agree with the analytic distribution. If, for example, the analytic standard errors are
smaller than the bootstrap standard errors, it suggests that there is a downward bias
in the analytic standard errors, and hence perhaps the “true” standard errors are
bigger than both the analytic and the bootstrap standard errors, with the bootstrap
standard errors therefore a better choice as they are closer to the truth. Such a view
seems to be behind the early bootstrap application of Korajczyk (1985),for exam-
ple, who found for a system of econometric equations modeling the foreign exchange
market, that bootstrap standard errors exceeded analytic standard errors and that
switching to bootstrap standard errors meant that a null hypothesis associated with
rational behavior could no longer be rejected. Given a common tendency for boot-
strap standard errors to be larger than their analytic counterparts, we hope it will not
become the practice to invoke the bootstrap approach only when the economist in
one thinks that the null hypothesis should be protected.
We should differentiate at least two possible bias problems that come up in
standard regression modeling and its extensions. There can be a bias in coefficient
estimation, perhaps best captured by the well-known bias toward zero of estimated
serial correlation coefficients. There can also be bias in standard error estimation.
Both of these can affect inference.
A simple method of calculating bias for coefficient estimates (Efron, 1982, p.
33) is to subtract the actual value from the average of the bootstrap values over the B
bootstrap samples. Hence, if the actual data coefficient estimate is .5 and the average
of the bootstrap values is .7, the bias estimate is .2 and one could even propose a bias-
adjusted estimate of .5 - .2 = .3. The bias-adjusted result, which is very much like
a jackknife estimate, may, however, be subject to very high sampling variability.
Freedman and Peters (1984a, 198413) follow an instructive approach along
these same lines for the standard error, which is to calculate for each coefficient
a bias factor as the ratio of the bootstrap standard error to some average of the an-
alytic standard errors as calculated on each of the bootstrap samples. This gives a
OF THE BOOTSTRAP
APPLICATIONS 429
better idea of how the analytic standard errors are biased, and it therefore seems
natural to come up with a “bias-adjusted” standard error by multiplying the original
analytic standard error estimate by the bias factor. We note that, in the linear regres-
sion context, if standard errors have been based on a variance estimate in which n
rather than n - k has been used as the divisor, the Freedman and Peters approach
will “automatically” adjust the standard error for degrees of freedom. Marais (1986)
refines this approach and finds the method can be quite accurate in the context of
estimating systems of regression equation. However, because the purpose of stan-
dard errors is for inference, the Freedman and Peters approach has been eclipsed
by approaches that directly adjust the potential bias in the test procedures or in the
confidence intervals.
B. The Bias-Corrected Percentile Method

Efron (1987) proposes one attempt to solve the problem of finite-sample bias. As-
sume we have a scalar /?,which is not necessarily a single-equation regression model
coefficient and is more likely a coefficient from a system of equations or perhaps a
nonlinear function of coefficients that can be estimated by linear methods. Consider
the interval
where B*(a) is the l o a t h percentile of the B*’s over the B bootstrap runs. If we set
a1 = a and a2 = 1 - a , we have the ordinary bootstrap percentile confidence inter-
vals with intended coverage 1 - 2a but with no adjustment for bias. Efron proposes
the BCa confidence intervals where instead we implement (5) with
where
and
where <p is the standard normal cumulative distribution function, P* is the empirical
cumulative distribution function of the b*’s, W O ( 1 - a ) is defined analogously to
wo(a),and B is called the acceleration constant and will be discussed.
Consider first the case where a = 0, which yields what have been called
the bias-corrected (BC) percentile confidence intervals. This method primarily deals
with bias in the coefficient estimates. If there were no tendency toward bias in the
coefficient estimate of the b, we should expect that in the empirical distribution
430 VEALL
B
function of the B*'s, half should be above and half below; i.e., that P*(B) = .5 and
hence that 20 = 0-'(.5) = 0. This in turn implies that a1 = a and a2 = 1 - a.
Hence, with no coefficient bias, the BC percentile confidence intervals are just the
ordinary bootstrap percentile confidence intervals. Now suppose P* (b)
exceeded .5,
suggesting a negative bias in the coefficient estimates. Therefore 20 will be positive
+ +
and a1 = a(2.20 z(OL))and a2 = 0(220 z('-OL))will exceed a and 1 - a re-
spectively. Hence, the BC method adjusts to the downward bias in the coefficient
estimate by shifting the entire confidence interval up. (One option to adjust the co-
efficient estimate itself is to use a = .5 in the above formula.)
The role of the acceleration constant 6 is less obvious, but it is partially related
to bias in the estimation of the standard error. For example, if 20 = 0 and given that
z(OL)is negative and z ~ is- positive,
~ it can be shown that changing 6 from zero to
a small positive value will widen the BCa bootstrap percentile confidence intervals.
More generally, we could argue that the usual normal approximation
(where se(B) is the standard error of S) may be generalized to
for some increasing transformation rn, where seo(rn(b>) denotes the standard error
of m(B) when the true value B equals any conveniently chosen PO:recall the point
of the exercise is that in finite samples, se(rn(6)) will depend on the value of /3 and
the approximation in the denominator in (10)attempts to capture this. Efron (1987)
or Efron and Tibshirani (1993, pp. 326-327) show that if we use the normalizing
transform rn, calculate confidence intervals based on the normal distribution, and
then transform back using r n - l , we obtain the BCa intervals except that a and t o
need to be estimated. Note that rn does not need to be known. These papers also
argue that in one-parameter families, a good approximation for d is one-sixth of the
skewness coefficient of the score function of /3, evaluated at B;
for multiparameter
families, they offer a formula based on the infinitesimal jackknife. However, most
econometricians will prefer, at least computationally, the simpler jackknife formula
(Efron and Tibshirani, 1993, p. 186):
B
where the summations run from 1 to n, B ( i ) is calculated on a sample with the ith
observation deleted, and B j , the jackknife estimator of /3, is the average of the B ( i ) .
OF THE BOOTSTRAP
APPLICATIONS 43 I
C. Percentile-t Methods
An alternative refinement to BCa methods is the percentile-t bootstrap confidence

B*
intervals. Suppose after estimating and se(B) on the original sample, and se(fi*)
are estimated on each bootstrap sample. (A key requirement for this method is that
some form of standard error estimate is available for both the original and boot-
strapped data.) If one thinks of the bootstrap process as a Monte Carlo experiment,
it is natural to think ofB* B
= as the null hypothesis to be tested in each trial and
hence natural to calculate the t-ratio t* = (fi*
- B)/se(B*) on each trial. The boot-
strap procedure therefore essentially generates a distribution for this t-ratio under a
particular null hypothesis and the 1 - 2a percentile-t confidence intervals become
where t*'Y is the ath percentile of the t*'s. Essentially this technique uses the boot-
strap to create its own critical values instead of using those supplied by the usual
t-distribution.
D. Comparison of the BCa and Percentile-t Methods

Both the BCa and the percentile-t corrections are a refinement of the standard asymp-
totic confidence intervals. As discussed, if we consider either the endpoints them-
selves or the tail-area coverage, analytic confidence intervals, bootstrap standard
error confidence intervals and percentile confidence intervals are all only first-order
accurate, meaning that the approximation error is a term of O(n-'/'). The BCa and
percentile-t confidence intervals are both second-order accurate in that the approx-
imation error is a term of O ( n - ' ) ; that is, the approximation error goes to zero sig-
nificantly faster as n increases (Hall 1988, 1992; Efron and Tibshirani 1993).
There has been considerable discussion of the relative merits of the BCa
method versus the percentile-t method. Hall (1988, 1992), one of the leading pro-
ponents of the percentile-t method, notes that it avoids the calculations discussed
above. Although the jackknife method using (1 1) is not excessively difficult in many
contexts, certainly the concept of the acceleration constant is not a comfortable or
familiar one to many practitioners. There is indeed some reason to be uncomfortable.
While possible, the generalization in BCa to the multivariate confidence regions case
is somewhat difficult. Moreover, even in the single-coefficient case above, it can be
seen that the bounds (5) based on (6)-(8)are not necessarily monotonic in a: i.e., it
is possible, because of the jump discontinuity in (7) when the denominator is zero,
for increases in a to widen rather than narrow the confidence bounds. The investiga-
tor needs to be careful if a has sufficient magnitude for this to occur in the relevant
range of a.
432 VEALL
The percentile-t method is not without its flaws. It is not transformation invari-
ant: we cannot obtain the confidence interval endpoints for h ( 8 )by simply plugging
the endpoints of the 8 confidence intervals into the function h. This is a familiar prob-
lem in statistics (described in an econometrics context by Gregory and Veal1 1985),
but leaves open the possibility that different ways of specifying restrictions or pa-
rameters can lead to markedly different substantive results. A solution proposed by
Tibshirani (1988)is to implement the percentile-t method with a variance-stabilizing
transform.
While the BCa method is transformation invariant in the single-parameter
case, it is not in the multiparameter case if the transformation h involves the pa-
rameters: the confidence interval for 8182 - l is not the same as the confidence
interval for 81 - 1/82. While this seems quite natural, if the idea is to perform a
test for the null hypothesis that 81/32 - 1 = 0 (equivalent to seeing whether the
confidence interval of 81/32- 1 covers zero), it is disconcerting that it matters if we
instead test the algebraically equivalent null hypothesis 81 - 1/82 = 0.
E. The Barnard Method

A final method we shall discuss was developed before the bootstrap but can be
thought of as a type of percentile-t bootstrap. Barnard (1963) proposed simulat-
ing the distribution of a test statistic under the null hypothesis. In the context of
confidence intervals, this means that rather than calculating the actual confidence
interval around the estimated parameter and determining whether it covers the pa-
rameter value corresponding to the null hypothesis, we simulate the confidence in-
terval under the null hypothesis and see whether it covers the estimated value. These
confidence intervals will be
where 80 is the value of under the null hypothesis and t** are the t-values gener-
ated by a bootstrap simulation using the estimates that are generated with the null
hypothesis. In the case, for example, when there are no other parameters to estimate,
a confidence interval calculated in this manner (or the corresponding test) is exact in
finite samples. This result extends if there are parameters to estimate even under the
null hypothesis, provided those other parameters do not enter the null distribution
of the parameter under test. (An example would be in the (static) linear regression
case, where the null distribution of each coefficient does not depend on the other
coefficients.) However, if this property does not hold, such as in many dynamic or
nonlinear regression contexts, the property of exactness is lost due to the estimation
of the other "nuisance" parameters: see Dufour (1995),who incidentally attributes
the initial idea behind this procedure to Dwass (1957)rather than to Barnard. Theil
and his associates provide examples of the Barnard method (Rosalsky, Finke, and
Theill984, Theil, Shonkwiler, and Taylor 1985, Taylor, Shonkwiler, and Theil 1986,
APPLICATIONS
Shonkwiler and Theil 1986, Theil and Shonkwiler 1986), some of which will be dis-
cussed subsequently. Van Giersbergen and Kiviet (1 993) promote the approach in
the context of a dynamic regression model. The method of “calibration” now com-
monly used in macroeconomics, in which the distribution of statistics is generated
under a null hypothesis imposed under an elaborate computer model called an “ar-
tificial economy,” can be seen as an extension of this basic idea. See Gregory and
Smith (1993)for a survey.
V. APPLICATIONS O F BOOTSTRAP REFINEMENTS

A. Size Correction
While it is sometimes called a bootstrap correction, what we have called here the
Barnard approach can be used as a size correction for almost any test. Horowitz
(19940, for example, uses this approach to correct the size of the information matrix
test for heteroskedasticity, Fan and Li (1995) use it to correct the J-test for testing
a nonnested hypothesis, Theil and Shonkwiler (1986) use it to study tests for se-
rial correlation, and Davidson and MacKinnon (1996)use both the J-test and serial
correlation test examples in a more general discussion of this type of bootstrap test-
ing. Horowitz argues that 100 bootstraps are sufficient for many cases where a is
not too small (so that t*.” would be the fifth smallest value of t* and t*.9s would be
the 96th smallest, that is the fifth largest) (see also Marriot 1979, Hall 1986).While
Horowitz’s results are supported by his Monte Carlo simulation results which check
for empirical size, size is not the only criterion of interest, and it is possible that the
use of such a small number of bootstraps may affect power.
B. Estimation of Systems of Equations
Various bootstrap methods have been applied to tests on systems of equations, in

what is often called the SUR model. Laitinen (1978) illustrated that finite-sample
inference for this model could be a problem in the context of an expenditure system
for rn goods:
where Wit is the expenditure share of good i at time t , x: is a row vector consisting of
a 1 (corresponding to a constant), total expenditure, and all rn prices, and U;,repre-
sents a disturbance term which has no correlation with any disturbance term at any
other time but which may have a contemporaneous correlation across commodities.
The homogeneity property from consumer theory suggests that the last rn + 1 ele-
ments of each should sum to zero, implying that if all prices and total expenditure
were changed in the same proportion there would be no change in expenditure shares
and, hence, in physical quantities purchased. Because the right-hand-side variables
434 VEALL
are the same in each equation, if the disturbances are normally distributed, OLS is
maximum likelihood.
Laitinen noted that before his paper this kind of proposition was commonly
tested with a Wald test based on the OLS estimates of each share equation and a
cross-equation variance-covariance matrix estimated from the OLS residuals. The
resulting test is asymptotically distributed as chi-square, with n - 1 degrees of free-
dom. Checking the asymptotic distribution using a simulation experiment based on
data from Theil (1975), when the number of commodities is 14, a true (by construc-
tion) null hypothesis is rejected by the Wald test based on a nominal 5% level, 87
times out of 100 rather than the expected 5. Laitinen argued that this is one reason
there are typically so many rejections of homogeneity in actual applications. One
intuition is that the variance-covariance matrix is badly misestimated by maximum
likelihood methods because there is no adjustment for degrees of freedom lost due to
parameter estimation, particularly as the number of estimated parameters increases
directly with the number of equations.
As it turns out, for the special case of the homogeneity test, Laitinen finds an
exact test using the Hotelling T 2 distribution. But his exact solution does not apply
in other contexts, for example the closely related problem of testing for the property
of symmetry in demand systems. Meisner (1979) and Bera, Byron, and Jarque (1981)
use simulation methods to examine the test for symmetry and find very inaccurate
test sizes and poor power as well.
Theil, Shonkwiler, and Taylor (1985)and Taylor, Shonkwiler, and Theil(l986)
apply the Barnard method directly to the demand homogeneity and symmetry Wald
tests; Shonkwiler and Theil (1986) use the same method to develop critical values
for alternative, non-Wald tests that they show can have superior power. Raj and Tay-
lor (1989)apply a Barnard-type bootstrap to testing within-equation restrictions in
demand systems.
Other researchers generate ordinary bootstrap standard errors and base tests
on these. Korajczyk (1985)and Freedman and Peters (1984a, 198433)have been dis-
cussed. Atkinson and Wilson (1992) have a different reading of Freedman and Peters
than ours and believe Freedman and Peters are arguing for direct application of the
bootstrap standard errors with no adjustment for bias. Given our discussion above
that ordinary bootstrap quantities are not theoretically more accurate than their ana-
lytic counterparts, it is not surprising that Atkinson and Wilson find in Monte Carlo
analysis based on relatively small systems that SUR standard errors (calculated en-
tirely using OLS residuals) may be no more accurate than ordinary bootstrap stan-
dard errors. Rilstone and Veal1 (1996a), emphasizing that the purpose of estimat-
ing the variance-covariance matrix and standard errors is for inference, find that
percentile-t confidence intervals are considerably more accurate than those based
on OLSISUR, although the BCa confidence intervals performance is only fair. Ril-
stone (1993) has similar findings for percentile-t (see also Rayner 1991) and BCa
intervals in a single-equation regression context with AR(1) errors.
APPLICATIONS
C. Nonlinearities
While discussing the lack of invariance of percentile-t methods, it was pointed out
that sometimes test results and confidence intervals are sensitive to nonlinearities.
In demand systems of the kind just described, for example, even if we ignore the
SUR problem there are potential difficulties in inferences because the desired esti-
mates of the price and income elasticities are nonlinear in the estimated parameters.
Green, Rocke, and Hahn (1987) find that the bootstrap estimates of the standard er-
rors of the estimated price elasticities are not much different from their asymptotic
counterparts, but there are substantial differences for the income elasticity. Krinsky
and Robb (1986, 1990, 1991) find no large differences between analytic and boot-
strap alternatives for a different data set and a translog system of equations, nor do
McManus and Rosalsky (1985) in a nonlinear earnings equation. George, Oksanen,
and Veall (1995)find some differences in a context where desired stock is estimated
as a nonlinear quantity, as do Veall and Zimmermann (1993) in another dynamic,
nonlinear context where they also use simulation to estimate power.
The point of these examples is that sometimes nonlinearities seem to matter
and sometimes they do not. If we return to the simple example of Gregory and Veall
(1985), it is easy to see what can cause the problem. If we compare the quantities
8182 - 1and 81 - 1/82,for example, it is immediately clear that the second is much
more nonlinear as 82 approaches zero. Rilstone and Veall (199613)examine the use
of percentile-t methods and BCa methods and find that neither works well at all in
this simple nonlinear example. While it may be simply that any approximation will
break down with enough nonlinearity (e.g., if small enough 82 above), nonetheless
we must conclude that the bootstrap is not yet a complete answer to the problems
associated with finite-sample inference involving nonlinear quantities.
VI. SUMMARY AND CONCLUSIONS
An important role of the bootstrap is to provide standard errors and other tools of
inference when there are no other available methods. A s discussed, such methods
are no more accurate, at least in a theoretical sense, than analytic methods based
on asymptotic approximations, but in the typical case where the use of the bootstrap
can be justified asymptotically, bootstrap standard errors cautiously used may be
valuable as “better than nothing” when analytic alternatives are not available. (This
is particularly true when the appropriateness of the bootstrap is itself checked by
simulation as discussed in the final paragraph.)
However, this survey has raised some questions with respect to whether boot-
strap methods are necessarily more accurate than analytic, asymptotic methods.
Only bootstrap refinements, such as the BCa method with its analytic bias correc-
tion based on an estimated “acceleration” coefficient or the percentile-t method, are
436 VEALL
more accurate in the Edgeworth sense. Yet even these have their flaws: it is possible
for BCa (1 - 2a) confidence intervals to narrow as a increases; percentile-t confi-
dence intervals are not transformation invariant. These flaws seem to be reflected in
actual performance problems, for example the severe shortcomings of the bootstrap
in the simple nonlinear case just discussed. Only some applications of the Barnard
method, involving simulation under a null hypothesis which does not depend on nui-
sance parameters, yield bootstrap-based tests which are exact in finite samples.
Fortunately, while simulation reveals the problem it may also provide the an-
swer. The first point to emphasize is that analytic, asymptotic methods often have
bad finite-sample performance. However, the quality of finite-sample performance
of such methods is usually based on speculation unless a simulation experiment is
done. If a simulation is done which tends to verify the accuracy of asymptotic meth-
ods, further simulations are not a priority. But based on current understanding of the
bootstrap, if the bootstrap and analytic, asymptotic methods differ it may be that the
bootstrap results are slightly to be preferred (especially refined bootstrap methods),
but what really needs to be done is further simulation study of the bootstrap itself.
(See Beran 1988, 1990 or Beran and Ducharme 1991 for approaches along these
lines, in which, for example, the percentile-t bootstrap is performed using bootstrap
standard errors.) While this “bootstrapping the bootstrap” approach seems compu-
tationally tedious, computer time is the one thing that is getting cheaper and,. except
in the very rare case of exact tests, we now see that many kinds of results cannot be
relied upon in finite samples unless they can be confirmed in a simulation experi-
ment. In some sense, our answer to the student’s question in the introduction as to
how big a sample is big enough to use asymptotics reliably is, “It depends, but I can
tell you a way that may help to find out for any given problem.” Hence, while adding
layers of simulation to our standard econometric practice may seem difficult, it is
comforting to know that there is at least one feasible method to check the asymptotic
approximations that are so widespread in econometrics.
ACKNOWLEDGMENTS
The author acknowledges the research assistance of Deb Fretz, the useful comments
of a referee, and the financial support of the Social Sciences and Humanities Re-
search Council of Canada.
REFERENCES
Arcones, M. and E. Gin6 (1992),On the Bootstrap of M-estimators and Other Statistical Func-
tionals, in P. E. LePage and L. Billard (eds.), Exploring the Limits ofthe Bootstrap,
Wiley, New York, 13-47.
APPLICATIONS OF THE BOOTSTRAP 437
Atkinson, S. E. and P. W. Wilson (1992),The Bias of Bootstrapped versus Conventional Stan-

dard Errors in the General Linear and SUR Models, Econometric Theory, 8,258-275.
Barnard, G. A. (1963), Contribution to Discussion, Journal of the Royal Statistical Society B ,
25,294.
Bera, A., R. P. Byron, and C. M. Jarque (1981), Further Evidence on Asymptotic Tests for
Homogeneity and Symmetry in Large Demand Systems, Economics Letters, 8,101-105.
Beran, R. (1988), Prepivoting Test Statistics: A Bootstrap View of Asymptotic Refinements,
Beran, R. (1990), Refining Bootstrap Simultaneous Confidence Sets, Journal of the American
Statistical Association, 85,417-426.
Beran, R. and G. Ducharme (1991), Asymptotic Theory for Bootstrap Methods in Statistics,
Centre de Recherches Mathematiques, Montreal.
Bernard, J.-T. and M. R. Veal1 (1987),The Probability Distribution of Future Demand, Journal
of Business and Economics Statistics, 5,417-424.
Brown, B. and R. Mariano (1984), Residual-Based Stochastic Prediction and Estimation in a
Nonlinear Simultaneous System, Econometrica, 52, 32 1-343.
Brownstone, D. (1990), Bootstrapping Improved Estimators for Linear Regression Models,
Chalfant, J., R. S. Gray, and K. J. White (1991), Evaluating Prior Beliefs in a Demand System:
The Case of Meat Demand in Canada, American Journal ofAgriculturul Economics, 73,
476-490.
Davidson, R. and J. G. MacKinnon (1996),The Size Distortion of Bootstrap Tests, paper pre-
sented at the meetings of the Canadian Econometric Study Group, Waterloo, September.
Dijkstra, T, K. and J. H. Veldkamp (1988), Data-Driven Selection of Regressors and the Boot-
strap, in T.K. Dijkstra (ed.), On Model Uncertainty and Its Statistical Implications,
Springer, Berlin, 17-38.
Dufour, J.-M. (1995), Monte Carlo Tests with Nuisance Parameters: A General Approach to
Finite-Sample Inference and Nonstandard Asymptotics in Econometrics, paper pre-
sented at the meetings of the Canadian Econometric Study Group, Montreal, Septem-
ber.
Dwass, M. (1957), Modified Randomization Tests for Nonparametric Hypotheses, Annals of
Mathematics and Statistics, 28, 181-187.
Efron, B. (1979), Bootstrap Methods: Another Look at the Jackknife, Annals of Statistics, 7,
1-26.
Efron, B. (1982), The Jackkncfe, the Bootstrap and Other Resampling Plans, Society for Indus-
trial and Applied Mathematics, Philadelphia, Pa.
Efron, B. (1987),Better Bootstrap Confidence Intervals (with discussion), Journal of the Amer-
Efron, B. and G. Gong (1983), A Leisurely Look at the Bootstrap, the Jackknife and Cross-
Validation, American Statistician, 3 7 , 3 6 4 8 .
Efron, B. and R. J. Tibshirani (1993), An Introduction to the Bootstrap, Chapman and Hall,
New York.
Fair, R. C. (1979), An Analysis of the Accuracy of Four Macroeconometric Models, Journal
ofPolitical Economy, 87, 701-718.
Fair, R. C. (1980), Estimating the Expected Predictive Accuracy of Econometric Models, In-
ternational Economic Review, 21,355-378.
438 VEALL
Fan, Y. and Q. Li (1995), Bootstrapping J-type Tests for Non-nested Regression Models, Eco-
nomics Letters, 48, 107-1 12.
Feldstein, M. (1971), The Error and Forecast in Econometric Models When the Forecast-
Period Exogenous Variables Are Stochastic, Econometrica, 39, 55-60.
Freedman, D. A., W. Navidi, and S. C. Peters (1988), On the Impact of Variable Selection
in Fitting Regression Equations, in T. K. Dijkstra (ed.), On Model Uncertainty and Its
Statistical Implications, Springer, Berlin, 1- 16.
Freedman, D. A. and S. C. Peters (1984a), Bootstrapping a Regression Equation: Some Em-
pirical Results, Journal of the American Statistical Association, 79,97-106.
Freedman, D. A. and S. C. Peters (1984b), Bootstrapping an Econometric Model: Some Em-
pirical Results, Journal of Business and Economics Statistics, 2, 150-158.
George, P. J., E. H. Oksanen, and M. R. Veall (1995), Analytic and Bootstrap Approaches to
Testing a Market Saturation Hypothesis, Mathematics and Computers in Statistics, 29,
3 11-3 15.
Geweke, J. (1986), Exact Inference in the Inequality Constrained Normal Linear Regression
Model, Journal of Applied Econometrics, 1, 127-141.
Green, R., D. Rocke, and W. Hahn (1987), Standard Errors for Elasticities: A Comparison of
Bootstrap and Asymptotic Standard Errors, Journal of Business and Economics Statis-
tics, 4, 145-149.
Gregory, A. W. and G. W. Smith (1993),Statistical Aspects of Calibration in Macroeconomics,
in G. S. Maddala, C. R. Rao, and H. D. Vinod (eds.), Handbook ofStatistics, Vol. 11,
Elsevier, Amsterdam, 703-719.
Gregory, A. W. and M. R. Veall (1985), On Formulating Wald Tests of Nonlinear Restrictions,
Hahn, J. (1996),A Note on Bootstrapping Generalized Method of Moments Estimators, Econo-
metric Theory, 12, 187-197.
Hall, P. (1986), On the Number of Bootstrap Simulations Required to Construct a Confidence
Interval, Annals of Statistics, 14, 1453-1462.
Hall, P. (1988), Theoretical Comparison of Bootstrap Confidence Intervals (with discussion),
Annals of Statistics, 16, 927-952.
Hall, P. (1992),TheBootstrap and Edgeworth Expansions, Springer, Berlin.
Horowitz, J. L. (1994), Bootstrap-Based Critical Values for the Information Matrix Test, Jour-
nal of Econometrics, 61, 395-411.
Jeong, J. and G. S. Maddala (1993), A Perspective on Application of Bootstrap Methods, in
G. S. Maddala, C. R. Rao, and H. D. Vinod (eds.), Handbook of Statistics, Vol. 11,
Elsevier, Amsterdam, 573-610.
Kiviet, J. F. (1984), Bootstrap Inference in Lagged Dependent Variable Models, University of
Amsterdam Working Paper.
Korajczyk, R. A. (1985), The Pricing of Forward Contracts for Foreign Exchange, Journal of
Political Economy, 93,346-368.
Krinsky, I. and A. L. Robb (1986),On Approximating the Statistical Properties of Elasticities,
Review of Economics and Statistics, 68, 715-719.
Krinsky, I. and A. L. Robb (1990), On Approximating the Statistical Properties of Elasticities:
A Correction, Review of Economics and Statistics, 72, 189-190.
Krinsky, I. and A. L. Robb (1991), Three Methods for Calculating the Statistical Properties of
Elasticities: A Comparison, Empirical Economics, 16, 199-209.
APPLICATIONS
Laitinen, K. (1978), Why Is Demand Homogeneity So Often Rejected?, Economics Letters, 1,

187-191.
LePage, P. E. and L. Billard (eds.) (1992), Exploring the Limits ofthe Bootstrap, Wiley, New
York.
Li, H. and G. S.Maddala (1996), Bootstrapping Time Series Models (with discussion), Econo-
metric Reviews, 15, 115-195.
Marais, M. L. (1986), On the Finite Sample Performance of Estimated Generalized Least
Squares in Seemingly Unrelated Regression: Nonnormal Disturbances and Alterna-
tive Standard Error Estimators, Graduate School of Business, University of Chicago
Working Paper.
Marriot, F. H. C. (1979), Barnard’s Monte Carlo Test, How Many Simulations?, Applied Statis-
tics, 28, 75-77.
McCloskey, D. N. and S. T. Ziliak (1996),The Standard Error of Regressions, Journal of Eco-
nomic Literature, 34,97-114.
McManus, W. S.and M. C. Rosalsky (1985), Are All Asymptotic Standard Errors Awful?,
Meisner, J. F. (1979), The Sad Fate of the Asymptotic Slutsky Symmetry Test for Large Sys-
tems, Economics Letters, 2,231-233.
Prescott, D. M. and T. Stengos (1987), Bootstrapping Confidence Intervals: An Application
to Forecasting the Supply of Pork, American Journal of Agricultural Economics, 69,
266-273.
Raj, B. and T. C. Taylor (1989), Do “Bootstrap Tests” Provide Significance Levels Equivalent
to the Exact Test? Empirical Evidence from Testing Linear Within-Equation Restric-
tions in Large Demand Systems, Journal of Quantitative Economics, 5, 73-89.
Rayner, R. K. (1991), Resampling Methods for Tests in Regression Models with Autocorre-
lated Errors, Economics Letters, 36,281-284.
Rilstone, P. (1993), Some Improvements for Bootstrapping Regression Estimators under First-
Order Serial Correlation, Economics Letters, 42,335-339.
Rilstone, P. and M. R. Veall (1996a), Using Bootstrapped Confidence Intervals for Improved
Inferences with Seemingly Unrelated Regression Equations, Econometric Theory, 12,
569-580.
Rilstone, P. and M. R. Veall (19%b), Comment on Bootstrapping Time Series Models by H.
Li and G . S.Maddala, Econometric Reviews, 15, 177-181.
Rosalsky, M. C., R. Finke, and H. Theil(1984), The Downward Bias of Asymptotic Standard
Errors of Maximum Likelihood Estimates of Nonlinear Systems, Economics Letters, 14,
207-2 11.
SHAZAM User’s Reference Manual Version 7.0 (1993), McGraw-Hill, New York.
Shonkwiler, J. S. and H. Theil (1986), Some Evidence on the Power of Monte Carlo Tests in
Systems of Equations, Economics Letters, 20, 53-54.
Taylor, R. G., J. S. Shonkwiler, and H. Theil (1986), Monte Carlo and Bootstrap Testing of
Demand Homogeneity, Economics Letters, 2 0 , 5 5 5 7 .
Theil, H. (1971), Principles of Econometrics, Wiley, New York.
Theil, H. (1975), Theory and Measurement of Consumer Demand, Vol. 1, North-Holland, Am-
sterdam.
Theil, H. and J. S.Shonkwiler (1986), Monte Carlo Tests of Autocorrelation, Economics Letters,
20,157-160.
440 VEALL
Theil, H., J. S. Shonkwiler, and T. G. Taylor (1985), A Monte Carlo Test of Slutsky Symmetry,
Tibshirani, R. (1988), Variance Stabilization and the Bootstrap, Biometrika, 75,433-444.
van Giersbergen, N. P. A. and J. F. Kiviet (1993), How to Implement Bootstrap Hypothesis
Testing in Regression Models, University of Amsterdam Working Paper.
Veall, M. R. (1987a),Bootstrapping the Probability Distribution of Peak Electricity Demand,
International Economic Review, 28,203-2 12.
Veall, M. R . (1987b), Bootstrapping and Forecast Uncertainty: A Monte Carlo Analysis, in
I. B. MacNeill and G. J. Umphry (eds.), Time Series and Econometric Modelling, Reidel,
Dordrecht, 373-384.
Veall, M. R. (1989), Applications of Computationally-Intensive Methods to Econometrics,
Bulletin of the International Statistical Institute, Proceedings of the 4 7th Session, 3,
75-78.
Veall, M. R. (1992), Bootstrapping the Process of Model Selection: An Econometric Example,
Journal of Applied Econometrics, 7,93-99.
Veall, M. R. and K. F. Zimmermann (1993), The Size and Power of Integrability Tests in Dy-
namic Demand Systems, Computational Statistics, 8, 127-139.
Vinod, H. D. (1993), Bootstrap Methods: Applications in Econometrics, in G. S. Maddala,
C. R. Rao, and H. D. Vinod (eds.), Handbook ofstatistics, Vol. 11, Elsevier, Amsterdam,
629-66 1.
Vinod, H. D. and B. Raj (1988), Economic Issues in Bell System Divestiture: A Bootstrap
Application, Applied Statistics, 37,251-261.
Detection of Unusual Observations
in Regression and Multivariate Data
Ali S. Hadi
Cornell Universiv, Ithaca, New York
Mun S. Son
University of Vermont Burlington, Vermont
1. INTRODUCTIONAND MOTIVATING EXAMPLES
Regression and multivariate analysis techniques are commonly used to analyze data
from many fields of study including economic and social sciences. These data often
contain unusual observations. Unusual observations, usually referred to as outliers,
are observations that do not conform to the pattern (model) suggested by the major-
ity of the observations in a data set. If they exist in the data, outliers can distort the
analysis of data,and the conclusions based on the analysis. For example, outliers can
distort parameter estimation, invalidate test statistics, and lead to incorrect statisti-
cal inference. We illustrate this point and the methods presented in this chapter by
the following data set.
Example I : Financial Data. In this chapter we make a repeated use of the fol-
lowing data set, which we refer to as the financial data. The data set was collected
and thoroughly analyzed by Jeff M. Semanscin (a student in one of the authors’ ap-
plied regression methods class) using Standard & Poor’s Compustat PC Plus. The
purpose of using the data here is only to illustrate the methods presented in this
chapter. There are several variables in the data set, but for illustrative purposes we
consider only a subset consisting of the following three variables:
44 I
442 HADIAND SON
X I : Book value in dollars per share at the end of 1992

X2 : Net sales in millions of dollars in 1992
X3:Sales-to-assets ratio in 1992
This subset of the data is shown in Table 1. It consists of 26 observations (financial

companies).
Let us first think of this data set as a trivariate data. Figure 1 shows the trivari-
ate scatter plot after it has been rotated to show the outliers in the data. The four
outlying points marked on the graph are detected by the multivariate outliers detec-
tion method presented here. The mean and covariance matrix of the data are
22.06 168.14 23.44 -1.51
( ::if) and ( 23.44 35.24

-1.51 -0.62
-0.62
0.18
respectively. When the outliers are deleted, mean and covariance matrix become
(1)
1
21.65 125.24 32.59 -0.45
( E) and ( 32.59 18.47 -0.11

-0.45 -0.11 0.00
respectively. Note the dramatic effects of outliers on the estimated variances and co-
(2)
variances. To illustrate how the confidence regions can change substantially because
of outliers, let us examine the bivariate scatter plot of X Iversus X 2 . The scatter plot,
Table I Financial Data for 29 Financial Companies
Number X1 x2 Xn Number X1 x2 x3
~~
1 14.58 26.961 0.17 14 2.46 0.247 0.26

2 21.15 4.816 0.15 15 15.75 2.213 0.09
3 19.26 3.394 0.08 16 25.19 2.825 0.09
4 39.93 5.455 0.08 17 34.30 7.281 0.07
5 6.12 1.495 1.52 18 39.26 7.382 0.10
6 32.25 9.112 0.09 19 30.80 9.228 0.08
7 32.43 11.078 0.08 20 13.51 3.364 0.14
8 8.30 0.806 0.28 21 15.96 3.840 0.08
9 16.68 4.461 0.08 22 6.75 0.196 0.16
10 24.79 14.559 0.09 23 31.84 15.372 0.12
11 19.26 1.114 1.82 24 1.52 0.909 0.17
12 19.42 4.190 0.09 25 17.70 5.099 0.23
13 27.02 2.009 0.08 26 57.44 5.067 0.10
DETECTION
OF UNUSUAL OBSERVATIONS 443
Figure I Financial data: Trivariate Scatter plot of X I , Xz,and Xs with the outliers indicated
by their numbers.
together with two ellipses expected to contain 95% of the observations, are shown
in Figure 2. The larger ellipse is based on the mean and covariance matrix of the
full data (all 26 observations) and the smaller ellipse is based on the mean and CO-
variance matrix of the data without the outliers (indicated by their numbers on the
scatter plot). Observe the huge difference between the two ellipses in terms of their
I
69m73
44 m98
XI 20.23
-4 m52
/)!
.....
-29m27 I I I I I
-15.13 -4.61 5m92 16.44 26.96
x2
Figure 2 Financial data: Bivariate scatter plot of XI versus X , with two ellipses (expected
to contain 95%of the observations). The larger ellipse is based on the mean and covariance
matrix of the full data (all 26 observations), and the smaller ellipse is based on the mean and
covariance matrix of the data without the outliers (indicated by their numbers).
444 HADIAND SON
57.44- 26 0
0
0
#
0
#
#
#
0
#
. . .I
1.524 %
I I I f
0.20 6.89 13.58 20.27 26.96
*2
Figure 3 Financial data: Scatter plot of X1 versus X2 with two superimposed regression lines.
The solid line is the least-squares regression line obtained using the full data. The dotted line
is obtained when the outliers (1 and 26) are deleted.
sizes, orientations, and shapes. Note also how the larger ellipse is affected by the
outliers. The larger ellipse detects only one observation as an outlier, whereas the
smaller ellipse declares four observations as outliers.
Let us now think of the data as regression data, where distinction between
dependent and independent variables has to be made. Consider, for example, the
simple regression of X I on X 2 . The scatter plot of X1 versus X z , with two super-
imposed least-squares regression lines, is shown in Figure 3. The solid line is the
least-squares regression line obtained using the full data. The dotted line is obtained
when the two outlying observations 1 and 26 are deleted. Again, we obtain two sub-
stantially different lines.
The above example shows that outliers can lead to misleading or erroneous
conclusions. It is therefore important for data analysts to first identify and examine
outliers if they exist in the data, before making conclusions based on data.
Before we proceed any further, we wish to make an important point. After read-
ing the literature on outlier detection, some people are left with the incorrect impres-
sion that once outliers are identified, they should be deleted from the data and the
analysis continues. We do not advocate automatic deletion (or even automatic down-
weighing) of outliers because outliers are not necessarily bad observations. On the
contrary, if they are correct, they may be the most informative points in the data. For
example, they may indicate that the data did not come from a normally distributed
population, as commonly assumed by almost all multivariate analysis techniques; or
they may indicate that the model is not linear. For this reason the outliers should be
DETECTION
2030
1523
y 1016
5 09
3
0 7..5 15 22.5 30
X
Figure 4 A scatter plot of population size, Y , versus time, X. The curve is obtained by fit-
ting an exponential function to the full data. The straight line is the least-squares line when
observations 22 and 23 are deleted.
called the unusual or even the interesting observations. To emphasize that outliers
can be the most informative points in the data, we use the exponential growth data
described in the following example.
Example 2: Exponential Growth Data. Figure 4 is the scatter plot of two vari-
ables, the size of a certain population, Y , and time, X . As can be seen from the
scatter of points, the majority of the points resemble a linear relationship between
population size and time as indicated by the straight line in Figure 4. According to
this model, points 22 and 23 in the upper right corner are outliers. If these points,
however, are correct, they are the only observations in the data set that indicate that
the data follow a nonlinear (e.g., exponential) model, such as the one shown in the
graph. Think of this as a population of bacteria which increases very slowly over a
period of time, then once somebody sneezes the population size explodes.
What do we do with outliers once they are identified? Because outliers can
be the most informative observations in the data set, they should not be automati-
cally discarded without justification. Instead, they should be examined to determine
why they are outlying. Based on this examination, appropriate corrective actions can
then be taken. These corrective actions include correction of errors in the data, dele-
tion or downweighing outliers, transforming the data, considering a different model,
redesigning the experiment or the sample survey, collecting more data, etc.
Outliers in multivariate data are intrinsically more difficult to detect than out-
liers in univariate and bivariate data. For example, in simple regression and in uni-
variate, bivariate, and trivariate data, outliers can be detected easily by graphing the
data. In higher than three dimensions and in the presence of multiple outliers, it is
difficult to detect outliers because
446 HADIAND SON
188
181
168
162
163 169 175 181 187
Height
Figure 5 A scatter plot of weight versus height with the box plots on the margin of each
variable. The two bivariate outliers cannot be detected by examining only univariate graphs
such as box plots.
1. It is difficult to graph high-dimensional data. Furthermore, examining

lower-dimensional plots of the data does not necessarily detect multivari-
ate outliers. This point is illustrated using the following example.
Example 3: The Weight-Height Data. The two bivariate outliers which appear
in the scatter plot of weight versus height in Figure 5 cannot be detected by exam-
ining only univariate graphs (e.g., box plots) of weight and height separately. Note,
however, that methods that work for higher-dimensional data will continue to work
for lower-dimensional data, but the converse is not generally true.
2. If the data contain a single outlier, the problem of identifying the out-
lier is simple, but if the data contain more than one outlier the problem
of identifying them becomes difficult due to the masking and swamping
problems. Masking occurs when a method fails to detect outlying observa-
tions (false negative decisions). Swamping occurs when a method declares
good points as outliers (false positive decisions). Masking and swamping
are serious problems and are the cause of the failure of many outlier de-
tection methods. Note that methods that work in the presence of multiple
outliers will continue to work if the data contain a single outlier or no
outliers at all, but the converse is not generally true.
In this chapter we discuss some methods for the detection of outliers in re-
gression and multivariate data. We concentrate our attention on methods that have
been recently developed and that require reasonable computational effort. Rele-
vant outlier detection methods are found, for example, in Rohlf (1975), Hawkins
DETECTION
(1980), Schwager and Margolin (1982), Beckman and Cook (1983), Barnett and
Lewis (1984), Hampel et al. (1986), Bacon-Shone and Fung (1987), Rousseeuw and
Leroy (1987), Fung (1988), Rasmussen (1988), and Caroni and Prescott (1992). The
rest of the chapter is organized as follows. Section I1 describes a unified framework
for the detection of outliers in both multivariate and regression data. Sections I11
and IV discuss the specifics of this unified framework for the detection of outliers in
multivariate and regression data, respectively. Section V deals with the problem of
outliers detection in very large data sets.
II. A UNIFIED FRAMEWORK FOR OUTLIER DETECTION
In regression analysis situations the data set contains a response variable y, con-
sisting of n observations, and a matrix X,consisting of n rows (observations) and p
columns (covariates). In multivariate analysis situations there is no y. The data set
contains only X,but we think of X as a random sample generated from a multivari-
ate elliptically symmetric distribution such as a multivariate normal or a multivariate
t-distribution. Our objective is to identify outliers if they exist in the data set in each
of these two situations.
It has been long recognized that classical methods, such as least-squares resid-
uals or Mahalanobis distances, are not effective in the detection of outliers because
they are not robust; that is, they are affected by the outliers that they are supposed to
detect. One way out of this problem is to replace classical methods by robust meth-
ods, which produce estimates that are resistant to the presence of outliers and/or to
violations of distributional assumptions. Indeed, several books have been devoted
either entirely or in large part to robust methods and/or outlier detection techniques;
see, for example, Barnett and Lewis (1984), Hawkins (1980), Huber (1981), Hampel
et al. (1986),Rousseeuw and Leroy (1987), and Chatterjee and Hadi (1988). Other
relevant articles include Maronna (1976), Campbell (1980),Rousseeuw and Yohai
(1984),and Lopuha (1989).
Robust methods have been suggested for many years now, but they have not
yet been widely used in practice because they involve extensive computations. The
most widely known of the robust methods are the least median of squares (for regres-
sion data) and minimum-volume ellipsoids (for multivariate data) estimators. These
methods are highly effective because they are not affected by outliers, but they are
computationally prohibitive, like other robust estimation methods.
Another way out of the problem has been recently developed by Hadi (1992a,
1994) and Hadi and Simonoff (1993, 1994, 1997). The main idea of these methods
is to first form a basic subset of about half of the data which is presumably free of
outliers, then add observations that are consistent with the basic subset. If all the
observations are added to the basic subset, the data set is declared to be free of out-
liers; otherwise the observations that are not consistent with the basic subset are
448 HADIAND SON
declared to be outliers. To determine whether the observations are consistent with

the basic subset, a suitable metric is chosen to measure the distance between each
observation in the data and the center of the observations in the basic subset (multi-
variate data) or the least-squares line based on the observations in the basic subset
(regression data). Thus, the method is carried out in two stages; the first is to find the
basic subset, and the second is to test for the outliers. To implement this method the
following questions have to be addressed:
1. How do we divide the data set into basic and nonbasic subsets?
2. What distance measure should we use?
3. How large must the distances be for the corresponding observations to be
declared as outliers?
The answers to these questions depend on whether we are dealing with multivariate
or regression data.
111. OUTLIERS IN MULTIVARIATE DATA
In multivariate analysis situations the data matrix X is viewed as a sample from a

multivariate elliptically symmetric distribution. Thus, an elliptically symmetric dis-
tribution is assumed to be the underlying model. To detect the outliers in X we need
to measure the distance between the ith observation x; and the fitted model. If a good
distance measure is chosen, then observations with large values would indicate that
they are outliers. Since the model is elliptically symmetric, an appropriate measure
of distance would be the elliptical distance
The elliptical distance d ( c , V) measures the distance between the ith observation,
x;,and a location (center) estimator c, relative to a measure of dispersion, V. The
classical choices of c and V are c = x (the sample mean) and V = S (the sample
covariance matrix), respectively. This choice of c and V gives
which is known as the Mahalanobis distance. If the data come from a p-variate nor-
mal distribution, the dV(Z,S) follows approximately a X2-distribution with p de-
grees of freedom. Thus, using an a-level of significance, values of di(%, S) larger
than Jm are declared to be outliers.
Unfortunately, d ; ( X , S) is affected by outliers. For example, some of the out-
liers may still have small values of d; (%, S) (masking) and some of the observations
which are not outliers may have large di(x,S) (swamping). This is illustrated by the
following example.
DETECTION
Table 2 Financial Data: The Mahalanobis Distances for the Trivariate and
Bivariate Data Sets Graphed in Figures 1 and 2
Data set Data set
Number Trivariate Bivariate Number Trivariate Bivariate
1 3.97 3.% 14 1.67 1.60

2 0.33 0.18 15 0.92 0.69
3 0.68 0.43 16 0.78 0.66
4 1.47 1.47 17 0.96 0.95
5 3.03 1.29 18 1.34 1.34
6 0.85 0.85 19 0.79 0.77
7 1.04 1.04 20 0.85 0.70
8 1.24 1.20 21 0.77 0.51
9 0.69 0.43 22 1.49 1.34
10 1.49 1.49 23 1.63 1.63
11 3.79 0.80 24 1.77 1.63
12 0.57 0.31 25 0.36 0.34
13 0.99 0.89 26 2.92 2.91
Example 4:Mahalanobis Distance. Consider the financial data described in

Example 1. The Mahalanobis distances for the trivariate and bivariate data sets
graphed in Figures 1 and 2 are given in Table 2. The corresponding index plots
are shown in Figure 6a and b, respectively. In both cases, the Mahalanobis distance
declares only observation 1 as an outlier (the distances for observation 1 of 3.97
and 3.96 are slightly larger than the cutoff point of 3.86). By comparison with the
3.971.
0
2.07
eo
.
e ee
eee eee
e e
0.334
1 1 1 1 1 1 1 1 1 1 1
1 5 9 13 17 21 25 1 5 9 13 17 21 25
(a) Trivariate data (b) Bivariate data
Figure 6 Financial data: Index plot of the Mahalanobis distance for (a) the trivariate data
{ X l , X2,Xs}and (b) the bivariate data {XIand Xz}.
450 HADIAND SON
trivariate and bivariate scatter plots in Figures 1 and 2, we see that the Mahalanobis
distance failed to identify observations 5, 11, and 26 in the trivariate case and ob-
servations 10,23 and 26 in the bivariate case.
The Mahalanobis distance fails to detect the outliers because it depends on the
sample mean and covariance matrix, which are known to be sensitive to the presence
of outliers. One way to solve this problem is to replace the mean and the covariance
matrix by more robust estimators of the location and scale. There are many robust
estimators for multivariate data. One problem with robust methods, however, is that
they are computationally extensive and at times practically infeasible. This may ex-
plain why robust statistics have not been widely implemented in statistical packages,
although they have appeared in the literature for so many years. Alternatively, using
the mean and the covariance matrix of the basic subset, we obtain
d;(xb,sb) = J(xi - X~)TS;'(X~ - ~ b ) , i = 1, . . . , n (5)

where X b and S b are the mean and the covariance matrix of the basic subset, respec-
tively. Thus, the distance di(Xb,S b ) depends on the observations contained in the
basic subset. If the data come from a p-variate normal distribution, then d?(xb,S b )
follows approximately a X2-distribution with p degrees of freedom. If the basic sub-
set is free from outliers, then %b and S b will not be affected by the outliers and the
d;(Xb,Sg) would be effective in the detection of the outliers. Thus, an important task
remaining here is how to obtain a basic subset which is likely to be outlier-free. Let
+ +
h be the integer part of ( n p 1)/2, m the vector containing the coordinatewise
medians, and
n
i= 1
The two stages of the method in the multivariate case are given in Algorithms 1 and
2 (Hadi 1994). In these algorithms Xb and S b are the mean and covariance matrix of
the observations in the current basic subset xb.
Algorithm I : Finding the Basic Subset
Input: An n x p matrix of multivariate data.
Output: A basic subset of size h observations that are likely to be free from outliers.
Step 0: Compute d;(m, A). Let X b and s b be the mean and the covariance
matrix of the observations with the h smallest values of d;(iii,A).
Compute di (Xg, Sb). Rearrange the observations in ascending order
according to di(Xb,S b ) . Divide the observations into two initial sub-
+
sets: a basic subset containing the first p 1 observations and a non-
basic subset containing the remaining n - p - 1 observations.
DETECTION
OF UNUSUAL OBSERVATIONS 45 I
Step 1: If the basic subset is of full column rank, compute di(%b, S b ) , where
-
Xb and S b are the mean and covariance matrix of the observations
in the current basic subset. If the basic subset is not of full rank,
increase the basic subsets by as many observations as needed for
the basic subset to become full rank. If needed, the observations are
added according to their ranked order.
Step 2: Rearrange the n observations in ascending order according to di (xg,
S b ) . Let s be the number of observations in the current basic subset.
Divide the observations into two initial subsets: a basic subset con-
+
taining the first s 1 observations and a nonbusic subset containing
the remaining n - s - 1 observations.
Step 3: Repeat Steps 1 and 2 until the basic subset contains h observations.
Algorithm 2: Testing for Outliers
Input: An n x p matrix of multivariate data and a basic subset Xb of size h obser-

vations obtained by Algorithm 1.
Output: The set of observations identified as outliers.

Step 0: Let s = h and %b and s b be the mean and the covariance matrix of
the observations in the current basic subset.
Step 1: Compute di(%b, S b ) . Rearrange the observations in ascending order
according to di (Xb, S b ) .
Step 2: Let
2
cff = ( 1 +
n-1-3p n-p
be the Bonferroni-adjusted critical value based on a X2-distribution

with p degrees of freedom multiplied by a correction factor. If min
{di(%b,S b ) ; i E nonbasicsubset} 2 c,, stop and declare all observa-
tions with di(%b, S b ) to be outliers. Otherwise, go to Step 3.
+
Step 3: For a new basic subset by taking the first s 1 observations ordered
+
according to di(%b, S b ) . If n = s 1, then declare no outliers in the
data set and stop; otherwise go to Step 1.
If desired, the final basic subset obtained in Algorithm 2 can be used to com-
pute the final distance for each of the observations in the data set. The method pro-
posed here is easy to compute. It is also effective in identifying the outliers when
tried on real as well as simulated data. This method has been implemented in some
commercially available statistical packages (e.g., Stata; see Gould and Hadi 1993).
Example 5: financial Data. Consider again the financial data described in Ex-
ample 1. The di(Z6,S b ) for the trivariate and bivariate data sets graphed in Figures 1
452 HADIAND SON
Table 3 Financial Data: The di(%b, S b ) for the Trivariate and Bivariate Data
Sets Graphed in Figures 1and 2
Data set Data set
Number Trivariate Bivariate Number Trivariate Bivariate
1 5.73 10.27 14 1.73 1.27

2 0.44 0.22 1s 1.03 0.51
3 0.90 0.23 16 0.92 1.07
4 1.75 1.87 17 0.94 1.01
5 21.93 1.02 18 1.53 1.40
6 0.77 1.34 19 0.74 1.43
7 1.03 2.05 20 0.59 0.56
8 2.04 0.91 21 1.10 0.41
9 1.05 0.54 22 1.10 1.06
10 2.10 4.05 23 1.99 3.83
11 27.63 1.19 24 1.46 1.38
12 0.69 0.12 25 1.55 0.68
13 1.26 1.60 26 3.86 4.01
and 2 are given Table 3. The corresponding index plots are shown in Figure 7a and
b, respectively. In both cases, the Mahalanobis distance declares only observation 1
as an outlier (the distances for observation 1 of 3.97 and 3.96 are slightly larger than
the cutoff point of 3.86). By comparison with the trivariate and bivariate scatter plots
in Figures 1 and 2, we see that di(Zb,St,) identifies all outliers in both data sets.
27.63 - 10.27 -
20.83 - 7.74 -
14.04- 5.20-
* *
7.24- 2.66 -
..
e
0 ** * 0 ** *.**.** **
.Oo ~ o ' o * ~~**..*~e***** *e
0.44- 0.12-
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
(a) Trivariate data (b) Bivariate data

Figure 7 Financial data: Index plot of d ;(Xb,S b ) for (a) the trivariate data {XI,
X2,X3} and
(b) the bivariate data ( X I , X z } .
DETECTION
IV. OUTLIERS IN REGRESSION DATA
In regression analysis situations the data consist of an n-response vector y and an

n x p covariate matrix X assumed to be of full-column rank. A model that relates
y and X is usually assumed to be of a linear form, y = Xg + E , where /? is the
vector of regression coefficients, E is a vector of random errors assumed to have a
multivariate normal distribution with mean 0 and covariance matrix 021n,o2is an
unknown scalar, and I, is an identity matrix of order n.
Unlike the case of multivariate analysis data, in which all unusual observa-
tions are labeled as outliers, in the regression analysis case unusual observations
are classified into three classes: outliers, high-leverage points, and inpuential obser-
vations (see, e.g., Chatterjee and Hadi 1986). We deal with the detection of each of
these types of observations separately.
A. Detection of Outliers
As we mentioned, outliers are observations that do not conform to the pattern (model)
suggested by the majority of the observations in a data set. To detect the outliers in X
we need to measure the distance between the ith observation xi and the fitted model.
The classical choice of a distance here is the least-squares standardized residual
D .
where e; is the ith element of the residual vector e = (I, - P)y and pii is the ith
diagonal element of the projection matrix
P = x(xTx)-'xT (9)
Replacing o by 6 = d m ,we obtain
i = 1, . . . , n (10)
which is known as the internally Studentized residual. Replacing o or & ( i ) , we obtain

the externally Studentized residual,
where 6(i)is the estimate of o when the ith observation is deleted. For simplicity
of notation, we write r; and r: instead of ri(6) and r;(6(il),respectively. Note that r;
and r: are related by
ri = r;
n-p-r:
hence ri and r; are equivalent.

454 HADIAND SON
43-461
xi
j j
29.48
15.50
e d 14
1-52
0.07 0.51 0.95 1.38 1.82
Figure 8 Financial data: Scatter plot of X I versus X3. The solid line is the regression line
based on all 26 observations and the dotted line is obtained when the observations 5,11, and
26 are deleted.
Outliers tend to have large absolute values of r; or r:, but unfortunately this
is not always the case. Outliers can have small residuals (masking), and observa-
tions which are not outliers can have large residuals (swamping). The reason for
this, again, is that ri and r;* are affected by outliers. For example, the two points in
the lower right corner of the scatter plot of X I versus X 3 in Figure 8 have very small
residuals because they are close to the solid regression line, which is based on all 26
observations, yet they are very far from the dotted regression line, which is obtained
when the outliers 5, 11, and 26 are deleted. Thus, using the ri or r:, observations
5 and 11 will be masked. On the other hand, observation 14, which is close to the
dotted regression line, is far from the solid regression line (swamping).
To deal with masking and swamping problems, we replace r; and r;* by a more
robust residual. One alternative is to use the least median of squares (LMS) residual.
However, the LMS is computationally intensive. Another alternative is to define the
residual with respect to the fitted model based only on a basic subset which is likely
to be free from the outliers. Let yb and Xb be the observations in the basic subset.
Let b b and 6;be the least-squares estimate of /3 and 0’ based on the observations in
the basic subset, respectively. A robust version of the residual (which can actually
be thought of as the scaled prediction error) can be defined as
DETECTION
OF UNUSUAL 455
OBSERVATIONS
The two stages of the method in the regression case are given in Algorithms 3
and 4 (Hadi and Simonoff 1993).In these algorithms B b and 3; are the least-squares
estimates of B and 0
' based on the current basic subset yb and xb.
Algorithm 3: finding the Bosic Subset

Input: An n-vector y and an n x p matrix X .
Output: A basic subset of size h observations that is likely to be free from outliers.
Step 0: Compute ri. Rearrange the observations in ascending order accord-
ing to Iril. Divide the observations into two initial subsets: a basic
+
subset containing the first p 1 observations and a nonbasic subset
containing the remaining n - p - 1 observations.
Step 1: If X b is of full-column rank, compute di in (13). If X b is not of full
rank, increase the basic subsets by as many observations as needed
for Xb to become full rank. If needed, the observations are added
according to their ranked order.
Step 2: Rearrange the n observations in ascending order according to d;. Let
s be the number of observations in the current basic subset. Divide
the observations into two initial subsets: a basic subset containing
+
the first s 1 observations and a nonbasic subset containing the re-
maining n - s - 1 observations.
Step 3: Repeat Steps 1 and 2 until the basic subset contains h observations.
Algorithm 4: Testing for Outliers
Input: An n-vector y, an n x p matrix X,and a basic subset Y b and x6 of size h
observations obtained by Algorithm 3.
Output: The set of observations identified as outliers.
Step 0: Let s = h and let B b and 8; be the least-squares estimates of B b and
0;based on the current basic subset yb and Xb.
Step 1: Compute di as in (13) but use the current yb and xb. Rearrange the
observations in ascending order according to d;.
Step 2: Let
be the Bonferroni-adjusted critical value based on a t-distribution

with s - p degrees of freedom. If min{di; i E nonbasic subset} >_
c,, stop and declare all observations with d ; 2 c, to be outliers.
Otherwise, go to Step 3.
Step 3: Form a new basic subset by taking the first s+ 1 observations ordered
+
according to di. if n = s 1, then declare no outliers in the data set
and stop; otherwise go to Step 1.
456 HADIAND SON
Table 4 Financial Data: The di(Kb,s b ) for the Trivariate

and Bivariate Data Sets Graphed in Figures 1 and 2
Number di Number di
1 -3.70 14 - 1.25
2 0.07 15 -0.02
3 0.16 16 1.01
4 2.24 17 1.15
5 -1.06 18 1.75
6 0.50 19 0.30
7 0.09 20 -0.55
8 -0.64 21 -0.35
9 -0.40 22 -0.70
10 -1.84 23 - 1.09
11 0.67 24 -1.51
12 0.00 25 -0.41
13 1.43 26 4.29
If desired, the final basic subset obtained in Algorithm 4 can be used to com-
pute the final residual for each of the observations in the data set.
Notice the similarity between Algorithms 1 and 2 for the detection of outliers
in multivariate data and Algorithms 3 and 4 for the detection of outliers in regression
data. They are special cases of the unified framework discussed in Section 2.
Example 6:financial Data. Consider again the financial data described in
Example 1. The final di obtained by Algorithm 4 for the simple regression of X I on
X 2 is given in Table 4. The corresponding index plot is shown in Figure 9. Comparing
these results with the scatter plot in Figure 3, we see that the method identified the
two outliers marked on the graph.
B. Detection of High-Leverage Points

Observations in the X-space that exert undue leverage in determining the fitted line
are called high-leverage points. This can be seen by writing the ith fitted value as
where pi; is the ijth element of the matrix P in (9), and observing the pii is the
weight or leverage attached to yi in determining the ith fitted values. Two important
and interesting properties of pi; are
DETECTION
4 -29-
2.29-
e*
0 - 3 0 - .a*. . *. * . . . .* . . e.. . 0 . . . . . . . . .
** **.
-1.70 -
-3.70 7
Figure 9 Financial data: Index plot of the final di obtained by Algorithm 4 for the simple
regression of X1 on X p . Two observations (1 and 26) are declared outliers.
0 F pii i 1
and
e2
pi; + 1< 1
eTe -
For proofs of these and many other properties of the matrix P and its elements, see
Chatterjee and Hadi (1988, Chapter 2). Consequently, the larger the p;; the higher
the leverage of the yi in determining fi. For example, in the extreme case where
pii = 1, we have y; = yi and ei = 0; that is yi completely determines fi. Thus,
observations with large values of pi; are called high-leverage points and the pii’s are
called leverage values.
The presence of high-leverage points, individually or in groups, makes it very
difficult to identify the outliers. Therefore, the X data should be examined for the
presence of high-leverage points. High-leverage points tend to have large values of
pii. Unfortunately, high-leverage points may not always have large leverage values
because a group of points can collaborate together and collectively induce high lever-
age, although their individual leverage values are not high. In other words, while all
points with large leverage values are high-leverage points, some observations with
small leverage values may be collectively a high-leverage group. Such a group of
high-leverage points can be identified by exploiting the relationship between the
concept of high-leverage and outliers in the multivariate X-space. One can think
of high-leverage points simply as outliers in the X-space. Thus, to identify high-
leverage points, one can think of X as multivariate data and apply Algorithms 1 and
2 to identify the outliers in X (see also Rousseeuw and van Zomeren 1990,1991).In
the context of regression, these outliers are called high-leverage points.
458 HADIAND SON
C. Detection of Influential Observations

Observations, individually or collectively, that excessively influence the fitted re-
gression equation as compared to the other observations are called influential obser-
vations. A bewilderingly large number of statistical quantities have been proposed
to study outliers and influential observations in regression analysis (Chatterjee and
Hadi 1986). One of the commonly used measures of influence is known as Cook’s
distance (Cook 1977), which is defined as
where b(i)
is the least-squares estimate of B when the ith observation is deleted. A
comparison with (3) shows that Ci is the squared elliptical distance between and B
b(i).Thus, a large value of Ci indicates that the ith observation is influential on B.
After some algebraic manipulations, one can show that
from which it follows that C; is a multiplicative function of the residual and leverage
values. Although a large value of Ci indicates that the ith observation is influential
S,
on a small value of C; does not necessarily indicate that the ith observation is not
influential. This can be seen from (17) because a high-leverage point tends to have a
small residual, hence a small value of Ci. From (19)it can be seen that an observation
B
will be influential on if it is an outlier (large value of Iril), a high-leverage point
(large value of pii), or both. Hadi (1992b) utilizes this idea and develops the additive
influence measure
where df = ef/eTe is the square of the ith normalized residual. The first term is a
leverage term which measures outlyingness in the X-space. The function pii/( 1-pii)
is known as the potential function. The second term in Hi is a residual term which
measures outlyingness of the observation in the y-direction. Since Hi is an additive
function of the residual and potential functions, it will be large if the observation
is an outlier in either the X-space, the y-space, or both. To determine which is the
case, Hadi (1992b) suggests plotting the potential versus the residual function, that
is, the scatter plot of
This plot is referred to as the potential-residual (P-R) plot. In the P-R plot, high-
leverage points are located in the upper area of the plot and observations with large
DETECTION
Table 5 Financial Data: Measures of Outlyingness, Leverage, and Influence Obtained

When XI is Regressed on X2
~ ~
1 -2.89 0.54 3.80 1.80 14 - 1.33 0.07 0.07 0.23

2 -0.02 0.04 0.00 0.04 15 -0.31 0.05 0.00 0.07
3 -0.09 0.05 0.00 0.05 16 0.41 0.05 0.00 0.07
4 1.51 0.04 0.04 0.24 17 0.91 0.04 0.02 0.11
5 -1.07 0.06 0.04 0.16 18 1.33 0.04 0.04 0.20
6 0.65 0.05 0.01 0.09 19 0.52 0.05 0.01 0.08
7 0.56 0.07 0.01 0.10 20 -0.55 0.05 0.01 0.07
8 -0.85 0.07 0.03 0.13 21 -0.38 0.04 0.00 0.06
9 -0.35 0.04 0.00 0.05 22 -0.95 0.07 0.04 0.16
10 -0.25 0.12 0.00 0.15 23 0.29 0.14 0.01 0.17
11 0.03 0.06 0.00 0.07 24 - 1.45 0.07 0.07 0.25
12 -0.12 0.04 0.00 0.04 25 -0.31 0.04 0.00 0.05
13 0.61 0.06 0.01 0.09 26 3.53 0.04 0.17 1.10
prediction error are located in the area to the right. Both Hi and the P-R plot have
been implemented in commercially available statistics packages such as Data Desk
and Stata.
Example 7: Financial Data. Consider again the financial data in Example 1.

The r f , pii, Ci, and Hi obtained when Xi is regressed on X2 are given in Table 5.
The corresponding index plots are shown in Figure 10. The P-R plot is given in
Figure 11. Observation 26 has the largest value of r:, which is the only observation
that exceeds the cutoff value of t ( n - p , a!/2n) = 3.50. Observation 1 is the only
high-leverage point in the data. Observation 1 is also identified by Ci as the only
influential observation in the data. Two observations (1 and 26) are identified by H;
as influential. These two observations are separated from the rest of other points in
the P-R plot. Observation 26 is an outlier because it is located in the lower-right
corner of the plot. Observation 1 is both an outlier and a leverage point.
V. DEALING WITH VERY LARGE DATA SETS
The methods presented in Sections 111 and IV have been shown to perform well in
many real-life and simulated data. They produce results in a reasonable amount of
time for small to medium data sets. But for large data sets, increasing the basic subset
one observation at a time can be time consuming. Hadi and Velleman (1997) adapt
these methods to large data sets as follows:
460 HADIAND SON
3.53 . 1
. .
0.54
........ 0=42i
.
1.93
0.32 ........0.
. ..
Oi.,. ...............
-1.28
0 .
.
..................... . 0.I6
-2.89 .
1
1
1
5
1 1 1
9
1
13
1 1
17
1 1
21
1 1 1
25
1 1
1
1 1
5
1 1
9
1 1
13
1 1
17
1 1
21
0 .
1 1
25
1
3.80-’. 1.80-1.
2.85- 1.36 -
1.90- 0.92 -
-
0.95
0.00 ..........................
0.48
0.04-
-
.
.................. 0..
I 1 I l l I l l 1 I l l 1 1 - 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Figure I 0 Financial data: Index plot of (a) r:, (b) p i ; , (c) Ci,and (d) Hi obtained when XI is
regressed on X2.
1. In some applications the data analyst may have some reasons to believe
that a certain subset of the data is free from outliers. In these applications,
this subset can be used as the first basic subset instead of the one obtained
in Step 0 of Algorithms 1 and 3 . Actually, this suggestion is applicable
to data sets of all sizes, but the computational savings that result from
eliminating Step 0 of Algorithms 1 and 3 increase as the size of the data
set increases. Additional computational savings can also be achieved if
the size of the chosen basic subset is larger than the initial size of p 1. +
2. In Step 2 of Algorithms 1 and 3, the basic subset size is increased by
one observation at a time. Computational savings can be realized here by
adding to the basic subset all observations that are below a certain cutoff
point. In this way the subset size grows more rapidly than in the original
algorithms.
DETECTION
OF UNUSUAL OBSERVATIONS 46 I
1.19 1 '1
-.- 0-90 -
m
Y
0.62-
22
E
0.33 -
0
26
0.04- -8S 0
Figure I I Financial data: The Potential-Residual plot obtained when X Iis regressed on Xz.
3. In Algorithms 2 and 4, testing starts only after the size of the basic subset
reaches h. Computational time can be saved by starting testing as soon as
the basic subset stabilizes or includes a certain prespecified number of
observations per parameter (e.g., six or more observations per parameter).
Note that in Algorithms 1 4 ,the nonbasic subset size can be at most n - h observa-
tions. These observations, which constitute less than 50%of the data, are interpreted
as outliers. The above modifications imply that the nonbasic subset can be as large
as n - s, where s is the size of the initial basic subset chosen by the data analyst. In
this case, the basic and nonbasic subsets are regarded as two distinct subgroups in
the data set. Each of these subgroups can be further divided into two smaller subsets
by applying the above modified algorithms as many times as desired. In this way, the
modified algorithms can be thought of as methods for finding homogeneous groups,
rather than finding outliers, in data sets.
REFERENCES
Bacon-Shone, J. and W. K. Fung (1987),A New Graphical Method for Detecting Single and
Multiple Outliers in Univariate and Multivariate Data, Journal ofthe Royal Statistical
Society ( C ) 36, 153-162.
Barnett, V. and T. Lewis (1984),Outliers in Statistical Data, 2nd ed., Wiley, New York.
Beckman, R. and R. D. Cook (1983),Outlier.. . s, Technometrics, 25, 119-149.
Campbell, N. A. (1980),Robust Procedures in Multivariate Analysis. I: Robust Covariance
Estimation, Applied Statistics, 29, 231-237.
Caroni, C. and P. Prescott (1992),Sequential Application of Wilks's Multivariate Outlier Test,
Applied Statistics, 41, 355-364.
462 HADIAND SON
Chatterjee, S.and A. S.Hadi (1986), Influential Observations, High Leverage Points, and
Outliers in Linear Regression (with discussions), Statistical Science, 1,379-416.
Chatterjee, S.and A. S.Hadi (1988),Sensitivity Analysis in Linear Regression, Wiley, New
York.
Cook. R. D. (1977),Detection of Influential Observations in Linear Regression, Technometrics,
19,lS-18.
Fung, W. K. (1988),Critical Values for Testing in Multivariate Statistical Outliers, Journal of
Statistical Computation and Simulation, 30, 195-212.
Gould, W. and A. S. Hadi (1993),Identifying Multivariate Outliers, Stata Technical Bulletin,
11,z-5.
Hadi, A. S. (1992a), Identifying Multiple Outliers in Multivariate Data, Journal ofthe Royal
Statistical Society, Series ( B ) ,54, 761-771.
Hadi, A. S.(1992b), A New Measure of Overall Potential Influence in Linear Regression,
Computational Statistics and Data Analysis, 14, 1-27.
Hadi, A. S. (1994),A Modification of a Method for the Detection of Outliers in Multivariate
Samples, Journal of the Royal statistical Society, Series ( B ) , 56, No. 2,393-396.
Hadi, A. S. and J. S. Simonoff (1993),Procedures for the Identification of Multiple Outliers
in Linear Models, Journal of the American Statistical Association, 88, 424, 1264-
1272.
Hadi, A. S.and J. S.Simonoff (1994),Improving the Estimation and Outlier Identification
Properties of the Least Median of Squares and Minimum Volume Ellipsoid Estimators,
Parisankhyan Sammikkha, 1,61-70.
Hadi, A. S. and J. S.Simonoff (1997),A More Robust Outlier Identifier for Regression Data,
Bulletin of the International Statistical Institute, 281-282.
Hadi, A. S. and P. F. Velleman (1997),Computationally Efficient Adaptive Methods for the
Identification of Outliers and Homogeneous Groups in Large Data Sets, Proceedings of
the Statistical Computing Section, American Statistical Association (in press).
Hampel, F. R., E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel (1986),Robust Statistics:
The Approach Based on Influence Functions, Wiley, New York.
Hawkins, D. M. (1980),Identijlcicntion of Outliers, Chapman and Hall, London.
Huber, P. (1981),Robust Statistics, Wiley, New York.
Lopuha, H. P. (1989),On the Relation between S-Estimators and M-Estimators of Multivariate
Location and Covariance, Annals ofstatistics, 17, 1662-1683.
Maronna, R. A. (1976),Robust M-Estimators of Multivariate Location and Scatter, Annals of
Rasmussen, J. L. (1988),Evaluating Outlier Identification Test: Mahalanobis D Squared and
Comrey Dk, Multivariate Behavioral Research, 23, 189-202.
Rohlf, F. J. (1975),Generalization of the Gap Test for the Detection of Multivariate Outliers,
Biometrics, 31, 93-101.
Rousseeuw, P. J. and A. M. Leroy (1987),Robust Regression and Outlier Detection, Wiley, New
York.
Rousseeuw, P. J. and B. C. van Zomeren (1990),Unmasking Multivariate Outliers and Lever-
age Points (with discussion), Journcd ofthe American Statistical Association, 85,633-
651.
DETECTION
Rousseeuw, P. J. and B. C. van Zomeren (19911, Robust Distances: Simulations and Cutoff
Values, in Directions in Robust Statistics and Diagnostics: Part II (W. Stahel and S.
Weisberg, eds.), Springer-Verlag, New York, 195-203.
Rousseeuw, P. J. and V. J. Yohai (1984), Robust Regression by Means of S Estimators, in
Robust and Nonlinear Time Series Analysis, Lecture Notes in Statistics, Springer-Verlag,
New York,' Vol. 26,256-272.
Schwager, S. J. and B. H. Margolin (1982), Detection of Multivariate Normal Outliers, Annals
of Statistics, 10,943-954.
14
Union-Intersection and Sample-Split
Methods in Econometrics with
Applications to MA and SURE Models
Jean-Marie Dufour
University ofMontreal, Montreal, Quebec, Canada
Olivier Torres
Universitc! de M e , Villeneuve d’Ascq, France
1. INTRODUCTION
Straightforward application of usual inference procedures (tests and confidence re-

gions) in econometrics is often impossible. The problem usually comes from an in-
sufficient specification of the probability distribution generating the data, as occurs
for example when one makes assumptions only on the first few moments of an error
distribution. However, the problem also arises in parametric models that specify the
data generating process up to a finite vector of parameters. This is typically the case
when the assumptions made on the distribution depart from those made in the stan-
dard linear regression framework, such as the absence of serial independence and
homoskedastici ty.
This chapter treats in a unified way two apparently distinct categories of prob-
lems where distributional results are difficult to establish. The first one consists of
comparing and pooling information about parameter estimates from samples whose
stochastic relationship is totally unspecified. In such cases, it is not possible to write
a usable joint likelihood function and standard finite-sample or asymptotic meth-
ods are not applicable. The second one consists of making inferences in models for
which the distributions of standard test and confidence set procedures are difficult
to establish, e.g., because of the presence of nuisance parameters, but for which rel-
465
466 DUFOUR
AND T O R R ~ S
evant test statistics based on appropriately selected subsamples are distributionally

more tractable.
To illustrate the problems we will study, consider the case where we have m 2
1 regression equations of the form
where pi is an unknown k; x 1 vector (ki < Ni), is a k; x 1 vector of fixed (or

strictly exogenous) regressors, and Q i is an unknown positive-definite nonsingular
Ni x Ni matrix, i E I = { 1 , 2 , . . . , m } . This setup describes situations frequently
met in econometrics. Special cases of interest include the following.
1. Models in which each equation expresses a similar relationship between

analogous variables (i.e., the coefficients pi have the same economic in-
terpretation), but each one corresponds to a different sample and the dif-
ferent samples may be dependent in a way that is difficult to specify (e.g.,
this includes many panel data models).
2. Models with structural change: this situation is a special case of the previ-
ous one, where the different samples correspond to different subperiods.
3. Stacked regressions where each equation represents a different economic
relation, possibly with different numbers of regressors which have differ-
ent economic interpretations.
4. Time-series models where the dependence between the m equations in
(1) is induced by serial dependence.
For example, take m = 2. A model of type 1 could express the relation be-
tween the log of the wage and a variable measuring the level of education for two
individuals. The coefficient p is then interpreted as the return to education (Ashen-
felter and Krueger 1992), and we may wish to test whether this return is the same for
individuals 1 and 2. In models of type 2 we may wish to know whether the parameter
linking variable y to variable x is the same over the whole period of observation. An
example of a type 3 model could be two equations where ~ 1 and , ~~ 2 represent
, ~ the
consumption of two different goods and x 1,I and x2, are different vectors of explana-
tory variables. Model 3 is composed of two distinct relationships, but for some reason
we want to test the equality of the two coefficients. An important example of a type
4 model is a linear regression model with errors that follow a moving-average (MA)
process of order 1, where the first equation contains the odd-numbered observations
and the second equation the even-numbered observations.
The most common practice in such situations is to rely on asymptotic infer-
ence procedures. The lack of reliability of such methods is well documented in the
literature. This feature of asymptotic tests has been established by Park and Mitchell
(1980), Miyazaki and Griffiths (1984), Nankervis and Savin (1987), and DeJong et al.
UNION-INTERSECTION AND SAMPLE-SPLIT METHODS 467
(1992) in the context of AR(1) models. Burnside and Eichenbaum (1994)provide ev-
idence on the poor performance of GMM-based Wald test statistics. For more general
theoretical results on the inaccuracy of asymptotic methods, the reader may consult
Dufour (1997); see also Nelson, Startz, and Zivot (1996),Savin and Wiirtz (1996),
and Wang and Zivot (1996). Furthermore, there are situations where usual asymp-
totic procedures do not apply. For instance, consider a model for panel data with
time-dependent errors: if no assumption is made on the dependence structure, it is
not at all clear what should be done.
The main characteristic of model (1) is that the vector of dependent variables
y = (’yi,. . . ,y’,)’ is in some way divided into rn subsamples (different individu-
als and/or different subperiods), whose relationship is unknown. Because the joint
distribution of the vector of errors U = (U;,U;, . . . , u i ) ’ is not specified, usual
inference methods based on the whole sample y are not applicable. This chapter
develops inference procedures which are valid in such contexts.
The general issues we consider can be described as follows. Given several
data sets whose stochastic relationship is not specified (or difficult to model), but on
which we can make inferences separately, we study the following problems: (I) how to
combine separate tests for an hypothesis of interest bearing on the different data sets
(more precisely, how to test the intersection of several related hypotheses pertaining
to different data sets); for example, in model (l), we may wish to test whether the
linear restrictions Cip; = yio, i = 1, 2, . . . , rn, holdjointly; (11) how to test cross-
restrictions between the separated models (such as = pz = - . = Bm, when k; =
k, i = 1 , 2 , . . . , m), which involves testing the union of a large (possibly infinite)
numberof hypotheses of the preceding type (e.g., pi = PO,i = 1 , 2 , . . . , n,for some
PO); (111) how to combine confidence sets (e.g., confidence intervals or confidence
ellipsoids) for a common parameter of interest and based on different data sets in
order to obtain more accurate confidence sets. All these problems require procedures
for pooling information obtained from separate, possibly nonindependent, samples
and for making comparisons between them.
Besides being applicable to situations where the stochastic relationship be-
tween the different samples is completely unknown, the methods proposed will also
be useful for inference on various models in which the distribution of a standard
statistic based on the complete sample is quite difficult to establish (e.g., because
of nuisance parameters), while the distributional properties of test statistics can be
considerably simplified by looking at properly chosen subsamples. This is the case,
for example, in seemingly unrelated regressions (SURE) and linear regressions with
MA errors.
The methods proposed here rely on a systematic exploitation of Boole-Bonfer-
roni inequalities (Alt 1982) which allow one to bound the probability of the union (or
intersection) of a finite set of events from their marginal probabilities, without any
knowledge of their dependence structure. Although such techniques have been used
in the simultaneous inference literature to build simultaneous confidence intervals,
468 DUFOUR
AND TORR~S
especially in standard linear regressions (Miller 1981, Savin 1984), it does not ap-
pear they have been exploited for the class of problems studied here. In particular,
for general problems of type I, we discuss the use of induced tests based on rejecting
the null hypothesis when at least one of the several separate hypotheses is rejected
by one of several separate tests, with the overall level of the procedure being con-
trolled by Boole-Bonferroni inequalities. For problems of type 11, we propose using
empty intersection tests which reject the null hypothesis when the intersection of a
number of separate confidence sets (or intervals) is empty. In the case of confidence
intervals, this leads to simple rules that reject the null hypothesis when the distance
between two parameter estimates based on separate samples is greater than the sum
of the corresponding critical points. We also discuss how one can perform empty in-
tersection tests based on confidence ellipsoids and confidence boxes. For problems
of type 111, we propose using the intersection of several separate confidence sets as
a way of pooling the information in the different samples to gain efficiency. These
common characteristics have led us to use the terminology union-intersection (UI)
methods.
The techniques discussed in this chapter for type I problems are akin to pro-
cedures proposed for combining test statistics (Folks 1984) and for meta-analysis
(Hedges and Olkin 1985). Meta-analysis tries to combine the evidence reported in
different studies and articles on particular scientific questions; it has often been used
to synthesize medical studies. However, these studies have concentrated on situa-
tions where the separate samples can be treated as independent and do not deal
with econometric problems. Conversely, these methods are practically ignored in
the econometric literature. Note also that the techniques we propose for problems
of types I1 and 111 can be viewed as extensions of the union-intersection method
proposed by Roy (1953) (Arnold 1981, pp. 363-364) for testing linear hypotheses
in multivariate linear regressions, in the sense that an infinity of relatively simple
hypothesis tests are explicitly considered and combined. A central difference here
comes from the fact that the “simple” null hypotheses we consider are themselves
tested via induced tests (because we study quite distinct setups) and from the differ-
ent nature of the models studied.
As pointed out, our methods have the advantage of being versatile and straight-
forward to implement, even when important pieces of information are missing. These
also turn out to be easily applicable in various problems where the distributional
properties of test statistics can be considerably simplified by looking at appropri-
ately selected subsamples. We show in particular that this is the case for several
inference problems in SURE models and linear regressions with MA errors. This
provides original and rather striking examples of “sample-split techniques” for sim-
plifying distributional properties. For other recent illustrations of this general idea,
the reader may consult Angrist and Krueger (1994),Dufour and Jasiak (1995), and
Staiger and Stock (1997).In the first reference, the authors propose a sample-split
technique to obtain IV estimators with improved properties, while the two other pa-
pers suggest similar methods to obtain more reliable tests in structural models.
The paper is organized as follows. Section I1 presents the general theory: in the
context of a general statistical model, we derive procedures for testing null hypothe-
ses of types I and 11. In Section 111, we consider the problem of pooling confidence
sets obtained from different data sets (type I11 problems). In Section IV, we apply our
results to test the equality of linear combinations of parameters of different equa-
tions in a SURE model, an interesting setup where standard tests and confidence
sets only have an asymptotic justification (for a review, see Srivastava and Giles
1987). In particular, we impose no restrictions on the contemporaneous covariance
matrix, allowing for different variances and instantaneous cross-correlation. In Sec-
tion V, we study inference for linear regression models with MA(q) errors. We show
that our inference technique is very well suited for testing hypotheses on regression
coefficients in the presence of MA errors. We study in detail the case of an MA(1)
process and consider the problem of testing an hypothesis about the mean. We com-
pare our procedure with some alternative tests. It appears much easier to implement
than other commonly used procedures, since it does not require estimation of MA
parameters. We also study the performance of our method by simulation. The re-
sults show that sample-split combined test procedures are reliable from the point of
view of level control and enjoy surprisingly good power properties. We conclude in
Section VI.
II. HYPOTHESIS TESTING: GENERAL THEORY
In this section, we consider a general statistical model characterized by a sample

space y and a family L = {PO: 8 E 0 )of probability distributions parameterized
by 8, where 0 is the set of admissible values of 8. Let Lo be a subset of L and suppose
we wish to test Ho : POE LO against H1 : POE L\&. If the model is identified,
which will be assumed, this amounts to testing HO : 8 E 00against H1 : 8 E 0 1 ,
+
where8 E 00 POE ,Co.
We consider here three classes of inference problems concerning 8. First, we
study situations where 00can be expressed as a finite intersection of subsets of 0;
i.e., 00 = nyEr Goy,where r is an index set of the form I' = { 1 , 2 , . . . , r } and
0 0 ,c~0, y E r. Second, we examine null hypotheses which restrict 8 to a subset
00of 0, where 00can be written as 0 0 = UyEr 0 o ( y ) , 0 0 ( y ) c 0,y Er.
In this case, r is not constrained to be a finite set. Third, we consider situations
where the information about 8 is available from different subsamples whose joint
distribution is unknown. We then try to pool these pieces of information by combining
inferences based on each subsample.
470 DUFOUR
AND TORR~S
A. Ho as the Finite Intersection of Subhypotheses

The test procedure in this section is based on the fact that although Ho may not be
easily testable it can be expressed as the intersection of subhypotheses, HO, : 0 E
Oo,, each one of which can be tested by usual procedures. The decision rule is built
from the logical equivalence that Ho is wrong if and only if any of its components
Ho, is wrong.
Assume that we can test Ho, using a statistic T, such that, for any 0 E Go,,
Pe(Cy E Y : T,@) 2 x}) is known, for all x E Iw, y E r = { 1 , 2 , .. . , r } . The
relation between these statistics is unknown or difficult to establish [as it is the case
in model (l)]. We wish to combine the information on the true probability distribution
of the model, brought by each of the r statistics. Since Ho is true if and only if all
the H0,’s are individually true, a natural way of testing Ho is to proceed as follows.
Using the r statistics T,, we build r critical regions W,(a,) = T,-’([t,(a,), OO)),
where t,(ay) is chosen so that PQ[W,(~,)] = a , under Ho,. We reject the null
hypothesis Ho when the vector of observations y lies in at least one of the W,(a,)
regions, or equivalently if T,@) 2 t,(a,) for at least one y . The rejection region
corresponding to this decision rule is UYEr , W (a ,) . Such a test is called an induced
test of Ho (Savin 1984).Its size is impossible or difficult to determine since the joint
distribution of the statistics T, is generally unknown or intractable. It is, however,
possible to choose the individual levels a y so that the induced test has level a E
(0, l ) , for by subadditivity
for any 0 E 0 0 , = 00. Therefore, if we want the induced test to have level
a , we only need to choose the ay’sso that they sum to a (or less).
To our knowledge, there is no criterion for choosing the ay’s in an optimal
manner. Without such a rule, in most of our applications we will give the null hy-
potheses Ho, the same degree of protection against an erroneous rejection by taking
a , = a0 = a / r , V y E r. However, there may exist situations where we wish to weigh
the H0,’s in a different way. In particular, if for some reason we know that one of the
decisions d,l (say, accepting or rejecting Ho,~) is less reliable than the other deci-
sions, we are naturally led to give d,l less impact on the final decision concerning
the acceptance or rejection of Ho. In other words, we will choose a,/ c a,, V y # y’.
In the case where we choose a , = a0 = a / r , V y E r, we reject Ho, at level
a0 when y is in W,(ao).Assuming F,,e(x) is continuous in x, this region of Y can
be reexpressed as
UNION-INTERSECTION AND SAMPLE-SPLIT METHODS 47 I
for all 0 E 0 0 , where F,,Q(z) = Pe[T,-’((-oo, z ] ) ]is the probability distribution

function of T,(Y). Then an equivalent alternative rejection criterion is to reject Ho,
when the p-value A,@) = 1 - F;,e[ty@)]is less than ayg. So according to this rule,
a rejection of Ho occurs whenever at least one of the r p-values A, is not greater
than a,, and the critical region of this test procedure is
If we assume that the statistics T, are identically distributed under the null hypothe-
sis, then F,,e = Fe, VO E 00and t,(ayg) = t(ao),V y E r, hence (with probability
1)
W ( a ) = Cy E Y : max T,@) 2 t(ao>}
This criterion is derived heuristically from the logical equivalence that Ho is

true if and only if all the Hyg,’s are true. It is similar to Tippett’s (1931) procedure
for combining inferences obtained from independent studies: using the fact that the
r p-values are i.i.d. U[O,J]under Ho, Tippett (1931)suggested the rule:
reject HO at level a if min{A, : y E r}5 1 - (1 -

Such procedures were also proposed for meta-analysis (Hedges and Olkin 1985),
but have seldom been used in econometrics. We show that an extension of Tippett’s
procedure to the case where the p-values are not independent can be fruitfully ap-
plied to several econometric problems. The analogy with meta-analysis comes from
the fact that inference on the ith equation in (1) is made “independently” from in-
ference on any other equation, although the test statistics may not be stochastically
independent. Since dependence of the test statistics is a common situation in econo-
metrics, we do not assume the independence of the p-values, which leads one to
use a/r instead of 1 - (1 - When a is small, the difference between a/r
and 1 - (1 - a)’’, is typically quite small. For some applications of this approach
to independent test statistics, see Dufour (1990),McCabe (1988),and Phillips and
McCabe (1988, 1989).
It is possible to demonstrate optimality properties for induced tests. For ex-
ample, consider a test procedure which combines r p-values A 1, A2, . . . , A, so
that it rejects Hyg when S(A1, A2, . . . , A,) 5 s, where S is some function from
R‘ into R and s is a critical value such that P[S(Al, A2, . . . , A,) 5 s] 5 a un-
der Ho. Bimbaum (1954) showed that every monotone combined test procedure is
admissible in the class of all combined test procedures.* A combined test proce-
dure S is monotone if S is a nondecreasing function, i.e., if x’ 5 xi, i = 1,2, . . . ,
*On admissibility of decision rules, see Lehmann (1986, Section 1.8, p. 17).
472 DUFOUR
AND TORR~S
r 3 S(xT, xa, . . . , x:) 5 S ( x l , x2, . . . , xr). In our case, S(A1, Az, . . . , Ar) =
min(A1, Az, . . . , Ar}, is clearly nondecreasing. For further discussion of admissi-
bility issues in such contexts, see Folks (1984).
B. Ho as the Union of Subhypotheses

Let us consider a null hypothesis of the form Ho : 8 E 0 0 , where 00= U y E r0 o ( y ) .
The solution to this testing problem is very similar to the one suggested above in the
case of an intersection of subhypotheses. Once again it is based on the fact that Ho
is wrong if and only if each of its components is wrong. If each hypothesis H o ( y ) :
8 E 0 o ( y ) can be tested using the rejection region W y ( a y )= T-’([t,(ay), m)),
satisfying Pe[Wy(ay)]= a y ,V8 E 0 0 ( y ) , it would appear natural to consider the
overall rejection region W(a,, y E r)
= n, W y ( a y )for a test of Ho.
However, difficult problems arise when one wants to implement this procedure
as described above. First, if r contains a finite number p of elements, we have
P
Pe[W(ay, Y E r>121 - C[l-P ~ [ w ~ ( ~ ~ ) I I
y= 1
which provides a lower bound for the probability of making a type I error. Of course,
this type of bound is of no use since we try to bound from above the probability of
an erroneous rejection of Ho. Appropriate upper bounds for the probability of an
intersection are difficult to obtain. Second, when r is infinite, it is impossible to
build W y ( a y )for every y Er.
It is however interesting to note that some null hypotheses can be written as
the union of several hypotheses (possibly an infinite number of such hypotheses).
It is then natural to construct an overall rejection region which is equivalent to the
infinite intersectionnYEr . example, consider the hypothesis Ho : 81 =
W y ( a y ) For
82 = . . = 8,, where 8; is a q x 1 subvector of the initial parameter vector 8. We
note that HO is true if and only if 380 E RQsuch that 81 = 8 2 = . . - = 8, = 80,
where 80is the unknown true value of 8; under the null. Defining Oo(80) (8 E 0 :
81 = 82 = . . . = 8, = O O } , we have 0 0 = U B O E R 4 0 ~ ( HO 8 0 )can
. be expressed
as an infinite union of subhypotheses Ho(80) : 8 E Oo(80).Therefore Ho is true if
and only if anyone of the Ho(8o)’s is true.
Obviously, it is impossible to test every Ho(80). Instead, we propose the follow-
ing procedure. For each i E { 1 , 2 , . . . , m } , we build a confidence region C; (yi,ai)
for 8i with level 1 - ai using the sample yi, where the ayi’s are chosen so that
ELl ai = a. This region satisfies
Pe[A;(8;,ai)] 2 1 - ai, V8 E 0
where Ai(8i, ai) = Cy E y : C;(y;, ai) 3 8 i } , i = 1,2,. . . , rn, and G 3 x means
that the set “G contains x.” In particular, if 80 is the true value of 8;, we have
Pe[A;(80,ai)]2 1 - a ; , V8 E 00
AND SAMPLE-SPLIT METHODS 473
UNION-INTERSECTION
Proposition I . A (conservative) a-level test of Ho : 01 = 02 = . - . = 0, is

given by the rejection region
where ai,i = 1,2, . . . , m satigy ai I a.

Proof: We need to show that P o [ W ( a , m ) ]5 a, V0 E 0 0 ,Va E (0, 1). Note
that V00 E Rq,
Hence, using the Boole-Bonferroni inequality
Pe[Wa,4 1IPO E Y : c i o i , ail 9 001

i= 1
rn m
i=l i= 1
We shall call a critical region of the form of W(a,m) an empty intersection test.
In our notation, W ( a ,m ) does not depend directly upon a, but on how the a;’s are
chosen to satisfy the constraint ELla; 5 a. For this procedure to be applicable,
we need to have confidence regions C i O i , a;)with levels 1 - ai.This is of course
possible in model (1) as long as 52i = i E { 1 , 2 , . . . , m } . We describe three
interesting special cases for which the procedure takes a simple and appealing form.
I. Intersection of Confidence Intervals: The Sum of Critical Points Rule
Consider a situation where q = 1.Typically, C;@;, a;)has the form
where 6, is some estimator of 0;, ELl

ai 5 a E (0, l), and C i L O i , ai) > 0,
c;uCyi, ai) > 0 for all possible values ofy;. Furthermore, it is usually the case that
c i ~ o ia;)
, = ciu@i, a ; )but we shall not need this restriction here. It is easy to see
that the following lemma holds.
Lemma I . The intersection of afinite number m of intervals Ii = [Li , U i ] c

R, i = 1, 2 , . . . , m , with nonempty interiors is empty ifand only if
min(U; : i = 1,2, . . . , m } < max{li : i = 1 , 2 , . . . , m } .

474 DUFOUR
AND TORR~S
Proof: Define U M = min{U; : i = 1 , 2 , . . . , m } ,LM = max{L; : i =

1 , 2 , . . . , m } ,and I = {x E R : LM 5 x 5 U M } Then
.
f # 0 ~ 3 x s u c h t h a t L i5x 5 U i f o r i = 1 , 2 , . . . , rn
@ 3x such that LM 5 x 5 U M
or equivalently,
From Lemma 1 and Proposition 1, we reject Ho if and only if
rninI8; + c i u b ; , ai) : i = 1 , 2 , . . . , m }
n
c max(8; - c ; ~ , ( y ai)
i, : i = 1,2, . ..,m}
But this condition is equivalent to
3j, k E {1,2, . . . , m } such that 8; -t c;u(yj, aj) < 8 k - ckL(yk, ak)
or
Finally, we reject Ho if and only if
In the case where m = 2, with c;ucyi, a;) = c j ~ ( y ;a,)

, = cjcyi, a;),j = 1 , 2 , this
criterion takes the simple form: reject the null hypothesis when the distance between
the two estimates is larger than the sum of the two corresponding “critical point. ” The
rejection region is then
W ( a ,2) = cy E Y : 18, - 821 > c1 (yl 7 a1) + c202,

For m > 2, we reject the null hypothesis when at least one of the distances l 8 k - 8, I
+
is larger than the sum cj (y; , a;) C k ( y k , ak). We will now extend this procedure to
multidimensional parameters and consider confidence ellipsoids.
2. Intersection of Two Confidence Ellipsoids

Consider the null hypothesis HO : 81 = 8 2 , where 0i is a q x 1 vector. As before, HO
can be restated as HO : 8 E (8 E 0 : 300 E Rq : 191 = 02 = 80). Suppose that for
i = 1 , 2 , we have a confidence ellipsoid Ci(yi,a ; )for 8;, such that
C;(y;, a ; ) 3 80 (8i - 0,)’A;(8i - 00) 5 c i ( a i )

where A; is a q x q positive definite matrix whose elements depend on y ; , Si is an

estimator of Oi, and c ; ( a ; )is a constant such that
PO[@E y : C;&, ai) 3 O } ] >_ 1 - a ; , vo E 0
the rejection criterion C10,1, a1) n C202, a2) = 0 of Proposition 1 is seen to be

equivalent to E l ( a 1 ) n Ez(a2) = 0.
To determine whether the intersection of the two ellipsoids is empty, it is suf-
ficient to find the set E,*(a2)of solutions of the problem
and check whether there is at least one element of E,*(a2)lying in El (al),in which
case the two confidence ellipsoids have a non empty intersection and Ho is accepted
at level al+ a2. This is justified by the following lemma.
Lemma 2. Let E,* (az) c E2 (a2)be the set of the solutions of (2). Then
Proof: (+) Let E l ( a l ) n E,*(a2)# 0. Since Ez(a2) C E ~ ( a 2 it

) , follows
trivially that El ( a l )n E2(a2) # 0.
(+) Let E1 ( a l )n Ez(a2) # 0. Then we can find p such that p E El ( a l )and
p E Ea(a2). In other words, p is an element of E2(az) that satisfies the condition
- P11I2 5 c ~ ( a ~which
) , entails that minyEE2(cr2) - P l / I 2 5 cl(a1). Now
suppose El(a1) n E,*(az) = 0. This means that the following implication must
hold:
Since E;(a2) is not empty, it follows that minyEE2(a2))Iy - PI 11’ > c ~ ( a l )a, con-
tradiction. Thus we must have El ( a l )n E,*(a2) # 0.
476 DUFOUR
AND TORR~S
Although any numerical calculus computer package is able to solve (2), we

propose a simple two-step procedure for deciding whether the intersection of two
ellipsoids is empty. This method can prove useful for high-dimensional problems.
The two steps are the following:
1. Check whether 81 E Cz(y2, a2) or 8 2 E Cl(y1, a 1 ) . If one of these events

is realized, then C l ( y 1 , a 1 ) nC2 @2, a2) is not empty, and Ho is accepted.
Otherwise, go to the second stage of the procedure.
2. Since 81 4‘ C2(yy2, a 2 ) , it follows (by convexity) that E g ( a 2 ) is a subset of
the boundary a E z ( a 2 ) and
so that we can check whether E l ( a 1 ) n E2(a2) # 0 by checking if

E l (al)na E z ( a 2 ) # 0. If the latter condition is satisfied, Ho is accepted;
otherwise Ho is rejected.
To be more specific, step 2 simply requires one to find the vector 9 which minimizes
IIy - 91 subject to the restriction ( y - p2)’D(y - p2) = c 2 ( a 2 ) , and then to
reject Ho when 119 - 91 (I2 > c l ( a 1 ) .
3. intersection of Two Confidence Boxes

When the covariance matrices of the estimator 8, are unknown, one cannot typically
build confidence ellipsoids. To illustrate such situations, consider the following ex-
ample. Two published papers investigate econometric relationships of the form
-
where ui N ( 0 , aTIN,),and is a k x 1 vector of unknown parameters.
We wish to compare the two parameter vectors and 82. However, only the
standard errors of the coefficients are known (or reported), not their covariance ma-
trices. Then it is not possible to use the previous procedure. But it is possible to use
simultaneous inference techniques and build simultaneous confidence boxes (hy-
perrectangles). Various methods for building such confidence sets are described in
Miller (1981) and Savin (1984). More precisely, let us build for each of the two re-
gressions in (4) k simultaneous confidence intervals, denoted by C { ( y i , a{)for the
component /3i of /I;j,= 1 , 2 , . . . , k, i = 1 , 2 , such that
+
Then choosing the a i ’ s so that a1 a2 = a,and applying the results of Proposition
1, we reject Ho : = 8 2 at level a when the intersection of the two hyperrectangles
is empty.
Checking whether the intersection of the two boxes is empty is especially
simple because one simply needs to see whether the confidence intervals for each
component of @; have an empty intersection (as in Section II.B.l). Furthermore it
is straightforward to extend this technique in order to compare more than two re-
gressions. Similarly, although we proposed a test for the null hypothesis that all
the parameter vectors /I;are equal (then imposing that in each equation has the
same number of parameters), it is easy to extend this procedure in order to test the
equality of linear transformations of Pi,i = 1,2, . . . , rn. Indeed, the method re-
lies only on the ability to derive confidence regions for parameters which are re-
stricted to be equal under the null. This is clearly possible whenever the param-
eters of interest are of the form R;&. The procedure is actually applicable to any
function h(0) of the parameter, provided we are able to build a confidence region
for h(8).
111. CONFIDENCE SET ESTIMATION
In the previous section, we described a general method for testing hypotheses in

several contexts. The main feature of the procedure is that a single final decision
concerning a family of probability distributions is taken by combining several indi-
vidual (partial) decisions on that family.
In many situations, we may wish to go a step further. For instance, consider
again model (1) and the null hypothesis Ho : B1 = p 2 = . = pn. The results
of Section I1 show how to test such an hypothesis. Suppose Ho is taken for granted.
It is then natural to ask what could be a valid confidence region for j3, the unknown
common value of @;, i = 1 , 2 , . . . , rn. The main difficulty here comes from the fact
that only the marginal distributions of the separate samples are specified, not their
joint distribution. Suppose each of these marginal distributions can be used to build
a confidence region for /3. The problem is then to find a way of pooling these pieces
of information on the true value of B and derive a single confidence region which
is based on the whole sample. This can be done as follows. Suppose each one of
the separate observations vectorsyl ,y2, . . . , y n has a distribution which depends
on 0 E 0. Although the joint distribution of y = @’, ,yL, . . . ,yh)’ is unknown,
we assume it is possible to build rn separate confidence regions Ci(cu;,y;) for 8
such that
478 DUFOUR
AND TORR~S
Then a natural way to exploit simultaneously these different pieces of information

consists of taking the intersection npl
C;(a;,y;) of the different confidence regions.
It is easy to see that
i= 1
Thus selecting a1, a2, , . . , am so that E:=,

a; = a,we can get any desired con-
fidence level. This procedure can be especially useful when one of the subsamples
yields a particularly accurate confidence set for 8.
In the next sections, we show how the procedures described in Sections I1 and
I11 can be used to make inference in SURE and regressions with MA errors.
IV. EXACT INFERENCE IN SURE MODELS

A. The Model and the Procedures
In this section we consider the SURE-type model:
where Xi is a Ni x k; fixed matrix of rank k; < Ni,P; is a k; x 1 vector of unknown

and E(u;,uis) = 0, V t # s. Note we do
parameters, U ; = ( u i l , ui2, . . . , u;~;)’,
not impose any restriction on the relationship between the rn equations, so that the
above model is more general than the standard SURE model. The null hypotheses
of interest are H;’) : A; = AO;, i = 1 , 2 , . . . , m, and Hf’ : A1 = A2 = = Am,
where A; = Rib;, Ri is a known qi x k; matrix with rank 4; 5 k;,i = 1 , 2 , . . . , m,
and Ao; is a known qi x 1 vector, i = 1 , 2 , . . . , m.* An interesting special case of
H i 1 ) is = 8 2 = . - - = Bm = P O , which is obtained by choosing k; = k, R; =
Zk, Ao; = P O , a known FE x 1 vector, in the above setup.
sumption ( A l ) : U = ( u { , u $ ., . . ,U;)’ -
We will consider two versions of (5), depending on whether we make the as-
N ( 0 , a 2 Z ~ where
), N =
Under A l , there exists an optimal test of Hi1) given be the critical region associ-
N;. EL1
*For H f ’ we must have q = qi, Vi = 1,2, . . . , rn and y 5 min{k; : i = 1 , 2 , . . . , m}.

ated with the Fisher F-statistic, based on the stacked modely = Xg + U , where
y = (Y’,,Y;, . . . ,yL)’, B = ( P i , . . . Pi)’, X = diag(X;);=i,z ,..., and
(A - Ao>’[s2R(X’X)-’R’]-’(A - A*)
F=
Q
I t
withA = (A’,, A;, . . . , A;)’, A = ( A l l A,, . . . , & ) I l A0 = (A;’, Ahz, . . . , &)I,
A; = R;, pi,& = (XiXi)-‘Xiyi, i = 1 , 2 , .. . , rn, R = diag(R;);,l,z ,...,m , s2 =
A A
[ [ ( I N- X(X’X)-’X’)Y~~~/(N - K ) , Q = CL1 qi and K = k;.

When we introduce heteroskedasticity in the model by allowing the variances
to differ across equations, our procedure is still valid, but the Fisher procedure is
not. As an alternative, one would typically use an asymptotic method based on a
generalized least-squares estimation and a critical region defined by a Wald, a La-
grange multiplier or a likelihood ratio statistic. But, as we already mentioned in the
introduction, it is well known that these approximations are not reliable.
An induced test of HA’) consists in testing HhL1): A; = Ao; at level a; using
the critical region K(ai) = Cy E Y : F, == F ( a ; ; q;, N; - Ic;)}, where
(Ai - Xoi)’(seR;(x~!xi)-’R:)-’(Ai
- AOi)
F,=
4;
with Ai = Rib;,
s’ = II(I,vi -X;(XiX;)-’Xi)~;11~/(N;-Ic;),and F ( a ; ; q;, N i - k ; ) is
the 1 - a; percentile of the Fisher distribution with (q;, N; - k ; ) degrees of freedom.
The ai’s can be chosen so that a; = a.Then the level a critical region for an
induced test of H i ’ ) is U
,: K (a;).
If we wish to test H i 2 ) at level a, we simply have to build rn confidence regions
at level 1 - a; for A; which are defined by
C;(Y;, a ; )= {x E [Wq : (A; - x)’(s’R;(xlxi)-’R:)-l(A;
- x)
I qiF(ai; qi, Ni - k ; ) }
in the A; space, and reject whenever n,:
C i b i , ai) = 0, with ai = a.
Note that, under assumption ( A l ) , the induced procedure for a test of Hh’) can
be improved by taking into account the independence of the regressions. In Section
11, we showed that the rejection region associated with an induced test of HA’) is
Ukl R(a;),where K ( a; )is the critical region for a test of /I; = /I0 at level ai.
Under ( A l ) , we have
Pe
L1
U KW
I = 1 - pe ny\K(ai)
I =1- n
m
i=l
Under Hh’) we have Pe[J’\K((a;)] = 1 - a;, Thus by choosing the a;’s so that
P~[Y\R(~;>I
nk1(1- a ; ) = a, we get a test of HA’) which has level a. If a; = ao, i =

1 , 2 , .. . rn, we must have a0 = 1 - ( 1 - ( a ) ’ ’ m .
480 DUFOUR
AND TORR~S
Unfortunately, the independence assumption (Al) is not helpful when we turn

to the test of Hf). But the procedure of Section 1I.B remains valid. To see why it
is difficult to exploit the independence of the regressions, consider the case where
rn = 2 and k = 1. We can build two confidence intervals Ci(y;,ao) = [Bi -
Bi +
ci ( j i , ao), ci ( j i , ao)], with a0 = a/2. According to our rejection criterion (see
+
Section II.B.l), we reject Hi2) at level a when 181 - 8 2 1 > cl 011, ao) c z ( y 2 , ao).
It is quite difficult to find the size of this critical region.
Consider now model (5)where assumption ( A l ) is not imposed. This model has
+
m(m 1)/2 + ELl E,:
ki parameters and Ni observations. In this case, no usual
finite-sample or asymptotic test procedure appears to be available for comparing the
coefficients of the different regressions. But the induced test method allows one to
test HA’) and H f ) relatively easily.
B. Some Examples
We now present some illustrations of the procedure described in the previous sec-
tions.
I, Testing Restrictions in a System of Demands for Inputs

The first example we consider is taken from Berndt (1991,pp. 460462). We consider
the problem of testing restrictions on the parameters of a generalized Leontieff cost
function. We assume that the production technology has constant returns to scale
and incorporates only two inputs, capital ( K ) and labor (L),whose prices are PK and
PL respectively. If we denote the output by Y and the total cost by C, the generalized
Leontieff cost function is
If the producer has a cost-minimizing strategy, it can be shown that the demands for
factors K and L are given by
A stochastic version of this model would consist in the two-equation SURE model
kt = ak + bkPtk + k
Ut 1, = al + hip, -+-
1 1
Ut
where uk and U’ are two Gaussian random vectors with zero mean and covariance
matrices O ~ Z N and CJ:ZN, respectively, and N = 25 is the sample size for each vari-
able of the model. A restriction imposed by the theory is b k = bl, which will be
our null hypothesis. To test Ho, the procedure described in Section II.B.l is partic-
ularly well suited since we have no a priori information on the relation between the
0.3992
0.1659
Figure I 97.5%confidence ellipsoids and intervals in the Berndt example. Confidence el-
lipsoid for ( a k , bk)’: -; confidence ellipsoid for (all bl)’: ---;
confidence intervals for b k and
ul appear on the left and right vertical axes, respectively.
random variables U: and U:. Using the data provided in Berndt (1991), which are
described in Berndt and Wood (1975),we performed separate tests of the following
null hypotheses:
The results of the estimation are
k, = 0.0490 + 0.00342 pf + iif

( .000125) ( .000084)
I , = -0.OMM + 0.28295 pi + Cf
(.OO1621) (.OO2350)
where the standard errors are given in parentheses. In Figure 1, we show the two
97.5%level confidence ellipsoids required for testing H l . It is straightforward to
see that we can reject both null hypotheses at level 5%because none of the regions
intersect. Similarly, the 97.5%level confidence intervals for b k and bl are respec-
tively (-0.01869,0.02539) and (0.1659,0.3992), and so do not intersect.
Since no information on the joint distribution of U: and U: is available, usual
GLS procedures cannot be applied in this context. However, suppose that we assume
482 DUFOUR
AND TORR~S
that (U!, U;, . . . , uss, U:, U:, . . . , ui5)’ is a Gaussian random vector with variance
matrix
as is usually done in SURE models. Using standard GLS techniques, the estimate of
( a k , b k , a l , bi)’ is (0.05100,0.00235,-0.04886,0.28804)’
and the F-statistics for
testing Ho and H: are 27.61 and 938.37, respectively. Since the corresponding 5%
asymptotic critical values are 4.05175 and 3.19958, the null hypotheses are both
rejected. However, one may prefer the empty intersection test procedure, because
it makes a weaker assumption on the error distribution. Moreover, GLS-based tests
only have an asymptotic justification.
2. Testing Restrictions on Returns t o Schooling

This example is taken from Ashenfelter and Zimmerman (1993). The study considers
the following SURE model:
where K; and Xi; represent the log wage and the schooling of the ith brother in the jth
family. These equations are the reduced form of a structural model which expresses
the relationship between the wage and years of schooling:
+
where F is a family specific component. We must have 0i = /?; A;,i = 1 , 2 .
The structural model has been estimated over a sample of 143 pairs of brothers.
The estimates reported by Ashenfelter and Zimmerman (1993, Table 3) are given
below, with standard errors in parentheses:
41 = 0.052, Xi = 0.018, 82 = 0.068, X2 = 0.006

(0.015) (0.020) (0.019) (0.015)
A natural hypothesis to test here is Ho : (B1, h.1)’ = (82,h.2)’. This can eas-
ily be tested from the estimated structural model, since Ho is equivalent to H; :
(01, AI)’ = ( 0 2 , A2)’. Here, we will use the hyperrectangle technique, because
Ashenfelter and Zimmerman (1993) do not provide the full estimated covariance
matrix for each regression. We first find a confidence interval with level 1 - a / 4 for
each one of the mean parameters in the structural model, and check whether the
two rectangles so obtained overlap, in which case we accept the null hypothesis.
This is done for a = 5%. Each event [0.0140,0.0900] 3 81, [-0.0326,0.0686] 3
A I ,[0.0199,0.1161] 3 Q 2 , [ -0.0320,0.0440] 3 h.2 occurs with probability 0.9875.
We accept the null hypothesis at level 5%, since the two boxes [0.0140,0.0900] x
[ -0.0326,0.0686] and [0.0199,0.1161] x [ -0.0320,0.0440] have a nonempty in-
tersection, which is [0.0199,0.0900] x [ -0.0320,0.0440].
V. E X A C T INFERENCE IN LINEAR REGRESSION

MODELS WITH MA(q) ERRORS
In this section, we show that the procedures developed in Section I1 can be useful
for inference in some dynamic models.
A. A Test on the Mean of a General MA(q) Model

In this section, we consider models of the form
yl = ml + ut, ut = Q ( B ) E ~ ,t E T = { 1 , 2 , .. . , 7’)
E EZ7,. . . , E o , E l , . . . , ET)’ - N ( 0 , a2ZT+(J
(6)
where Q ( z ) = $0 + +
$12 $2z2 + +.* $ f q Z q , $0
*
K
1, mt = x k = l x l k b k = xib,
b = ( b l , b 2 , . . . , b K ) ’ is a vector of unknown coefficients, and x l = ( x l l , x12, . . . ,
xl~)’,t = 1 , 2 , . . . , T , are vectors of fixed (or stricly exogenous) variables. In model
with
-
(6),y N ( m , a), where m = ( E m i , Em2, . . . , EmT)’ and = ( W ~ , . J ~ ,,..., ~ =T ~
, ,~
(7) shows the key feature of model (6): observations distant by more than q periods
from each other are mutually independent. Then, we are naturally led to consider
model (6)for subsamples obtained as follows. Define subsets of T, J; = {i, i ( q + +
+ + + +
l ) ,i 2(q l ) , . . . , i ni(q l)},where ni G I[(T - i ) / ( q l ) ] (I[.] denotes +
+
the integer part of x), i = 1 , 2 , . . . , q 1, and consider the q 1 equations +
Equation (8) belongs to the class of model (1). In each equation, the error term sat-
isfies the assumptions of the linear regression model, so that it is possible to apply
usual inference procedures to test restriction on b , Ho : b E Cp. This null hypothesis
+
can be seen as the intersection of q 1 hypotheses H o , ~each, of which restricts the
+
mean of the ith subsample to be in @, i = 1 , 2 , . . . , q 1. The methods presented
in Sections I1 and I11 are perfectly suited to such situations. We build q 1 critical +
+
regions with level a/(q 1) to test each one of the hypotheses Ho,i, and reject the
484 DUFOUR
AND TORR~S
null hypothesis at level a! if the vector of observations belongs to the union of these
regions. Note we did not make any assumption on the roots of W ( t ) . In particular,
we did not restrict the MA process { * ( B ) E ~: t E T} to be invertible.
In the next subsection we apply the procedure to a MA(1) process with a con-
stant and provide comparisons with some alternative procedures such as asymptotic
tests and bounds tests.
B. Exact Inference in the Context of a MA( I) Process
I. An Induced Test on the Mean

Consider the model described by (6), with q = 1, K = 1 and x t = 1, V t E T:
yt = B + + Et @&t-l, ~t -
ind
N ( 0 , a2), t E T (9)
The vector of parameters is 8 = (B, @, a2)'.The null hypothesis we consider is Ho :
8 E 0 0 , O o = (0 E 0 : B = 0). According to our procedure, assuming T is even,
we form two subsamples of size T/2, ( y t ,t E J i ) , where J 1 = { 1,3,5, . . . , T - 1)
and J2 = { 2 , 4 , 6 , . . . , T ) . For each subsample, we make inference on B from the
regression equation
yt =B + Ut, t E J;, U ; = ( ~ : tt E 1;)' , 1,2

N ( O , a t l ~ p i) =
(10)
A natural critical region with level a!/Z for testing B = 0 is then given by
where bi is the OLS estimator of B and P@;) the usual unbiased estimator of the
variance of /!Ii from regression (10) using sample (yt : t E J i ) ; t ( T / 2 - 1; 4 4 ) is the
upper 1 - a/4 percentile of Student's t distribution with T/2 - 1 degrees of freedom.
We reject Ho : B = 0 at level a! ify E Wl(a/2)U W2(a!/2).
2. Alternative Procedures
We compared this procedure with two alternatives. The first one consists in testing Ho
using bounds proposed by Hillier and King (1987), Zinde-Walsh and Ullah (1987),
Vinod (1976), and Kiviet (1980); see also Vinod and Ullah (1981, Chap. 4). The
latter are based on standard least-squares-based tests statistics for testing /? = 0
obtained from the complete sample, such as the t-statistic or its absolute value. Since
the distributions of the latter depend on the unknown value of the moving average
parameter @, one finds instead bounds &(a) and t " ( a ) which do not depend on the
parameter vector 8 and such that
for all 0 E 0 0 , a E (0, 1). Then the decision rule that consists in rejecting Ho
when TCy) > t"(a)and accepting Ho when T O ) < &(a)has level a. An incon-
venient feature of such procedures is that they may be unconclusive (when T O ) E
[t'(a),t"(a)]). Obviously, to avoid losses of power, the bounds should be as tight as
possible.
In all the above references on bounds tests, the bounds are derived assuming
that the MA parameter is known, so that they depend on it, even under the null
hypothesis. Therefore we will denote by t i ( a )and t i ( a )the lower and upper bounds
on te(a).But as @ is unknown, we have to find the supremum, t"(a),over the set
{ t i ( a ): ,$(a)L t o ( @ ) , VO E Oo}, to make sure that the test based on the rejection
region
V ( a )= cy E y : TCy) > t " ( a ) }

satisfies the level constraint
Since the moving-average parameter is not restricted by Ho, the set of admissible
values for @ is R. The upper bound is then likely to be quite large.
In the context of model (9), T O ) is typically the usual t-statistic, its square
or its absolute value. Since under Ho, its distribution only depends on @ (and the
sample size), we write te, t ; , and tfL instead of t g , t i , and t i , respectively.
Here, we only use the bounds of Zinde-Walsh and Ullah (1987) and Kiviet
(1980),denoted by t;,$(a) and t i , $ ( a ) because
, they are respectively tighter than
those of Hillier and King (1987)and Vinod (1976).The supremum t i ( a )of ti,+(a)
for @ E R is difficult to establish, but Kiviet (1980,Table 6, p. 357), gives the values
of the bounds for @ E {.Z, .3, .5, .9},and it can be seen that t i , , 9 ( a 2 ) t i , + ( a )for
,
@ E { 2 , .3, .5, .9}.We note that these bounds increase with @, and we suspect that
the supremum is arbitrarily large, possibly infinite when @ # 1. Nevertheless, we
will use t i 9(a)as the relevant upper bound in our simulations. Zinde-Walsh and
Ullah (1987)derived bounds on the Fisher statistic (or on the square of the t-statistic
in our case). t ; , + ( ( ~ is
) proportional to the ratio )Lmax(@)/hmin(@) of the highest and
lowers eigenvalues of the covariance matrix ofy:
We need to make here a remark about the accuracy of Zinde-Walsh and Ullah's
bound. Their test rejects Ho at level a when [TCy)I2 > s ~ p ~ ~ ~ t $ , t;(a). ~ ( a )
The critical value t i ( a )is not easy to determine analytically, so instead of finding
the maximum of t i , $ ( a )on R, we reckoned ti,$(0.05)for some values of @ in the
interval [-1, 21. We found a maximum at @ = 1, and a minimum at @ = -1, for
every sample size we considered. Although ti,,(0.05) 5 t;(0.05), we used this
486 DUFOUR
AND TORRCS
Table I Zinde-Walsh and Ullah‘s Bounds

Sample Size T 25 50 75 100
t& (0.05) 1 164.1972 4 254.3396 9 291.4222 16 274.6855
value as the upper bound. Doing so gives more power to the Zinde-Walsh-Ullah test
than it really has, because it may reject Ho more often than it would do if we used
t;(0.05). Despite this fact, t;, (0.05) is so large (see Table 1)that the power of the
test is zero everywhere on the set of alternatives we considered, for any sample size
and for any @ (see Section V.B.3).
The second alternative consists of using asymptotic tests. In this category, we
considered three commonly used test. The first category includes tests based on a
GLS estimation of (9). In the first step, one finds a consistent estimator h of G!and
P such that P’P = h-’ . In the second step, we multiply both sides of (9) by P and
apply ordinary least squares (OLS) to that transformed model. In the last step, we
test Ho using the standard F-test. We examine two estimation procedures that lead
to a consistent estimator of /?,resulting in two test statistics. The first one is detailed
in Fomby, Hill, and Johnson (1984, pp. 220-221). We denote it by GLS-MM because
in the first step of GLS, we estimate the MA parameter @ by the method of moments.
@ is estimated by minimizing the distance (in the sense of the Euclidean norm on
R) between the sample and true first-order autocorrelations. The second estimation
procedure uses exact maximum likelihood in the first step of GLS and will be denoted
by GLS-ML.*
The third test we consider is motivated by a central limit theorem (Brockwell
and Davis 1991, p. 219) which establishes the following property: if a process, with
mean B, has an infinite-order MA representation with IID error terms and MA coef-
ficients @i, i = . . . , -2, - 1 , O , 1,2, . . . , satisfying the conditions
00 00
then the sample mean of the process is asymptotically normally distributed, with
mean /3 and variance T - ’ y ( k ) , where y ( k ) is the autocovariance at lag k .
Note that the last condition on the @i’s is not satisfied for the MA(1) process (9) with
@ = -1, but as @ is unknown, we might not be aware of this fact or ignore it. Then a
*For further discussion of ML estimation in this context, see Tunnicliffe Wilson (1989)and Laskar and
King (1995).
Table 2 Size and Critical Values of 5% Level Asymptotic Tests
T = 25 T = 50
Sample size $* Size (%) ACV CCV $* Size (%) ACV CCV
GLS-MM -0.5 19.22 4.25968 30.664 -0.5 18.59 4.0384 59.555

GLS-ML -0.5 27.87 4.25968 37.979 -0.5 15.06 4.0384 14.615
NW 1 15.03 3.840 8.459 1 10.25 3.840 5.789
T = 75 T = 100
Sample size +* Size (%) ACV CCV $* Size (%) ACV CCV
GLS-MM -0.5 16.98 3.97024 64.502 -0.5 14.98 3.9371 38.789

GLS-ML -0.5 10.13 3.97024 6.396 -0.5 7.84 3.9371 4.983
NW 1 8.82 3.840 5.243 1 8.08 3.840 4.907
natural way of testing Ho is to estimate B by the sample mean LTand the asymptotic
variance by the consistent estimator proposed in Newey and West (1987):
where r T ( k ) is the sample autocovariance at lag k . Then, if Ho is true, the statistic
has an asymptotic x'

distribution with 1 degree of freedom. We will denote this
procedure by NW.*
Before presenting the results of our simulations, we wish to insist on a very
important condition one has to impose when comparing the relative performance
of two tests. In the Neyman-Person approach to the problem of testing a null hy-
pothesis Ho against an alternative H I , it is meaningless to say that a test A has a
higher power than a test B , if the two tests do not have the same level. A test of
Ho : 0 E 0 with critical region W has level a if supoEooPo(W) 5 a, and it has size
*Of course, the list of the methods considered in the present simulation is not exhaustive. For example,
possible variants of the NW method include the covariance matrix estimators proposed by Wooldridge
(1989).Bayesian methods (Kennedy and Simons 1991) and marginal likelihood methods (King 19%)
could also be used in this context. But space and time limitations have precluded us from including all
proposed methods in our simulations.
488 DUFOUR
AND TORRB
(a) T=25
80
beta beta
(c) T=75 (d) T=lOO
beta
Figure 2 Rejection frequencies of Ho : = 0 in model (4) with II/ = 1, T =

25,50, 75, 100; induced test (-), tests based on subsamples (yc : t E Ji) and (yc : t E J2)
( - .. and - . -).
a if supo,=-ooPo(W) = a;see Lehmann (1986,Section 3.1). When the probability of

rejection Po(W) under the null hypothesis (6 E 0 0 ) is not constant, controlling the
level of a test requires one to ensure that Pe(W) 5 a for all 6 E 0 0 , and controlling
its size involves ensuring that the maximum (or supremum) of the rejection probabil-
ities over 8 E 00is equal to a.Of course, this may lead to a difficult search over
the parameter space. When the distribution of a test statistic depends on a nuisance
parameter (the unknown value of $r, in the present case), correcting the size of a test
requires one to find a critical value such that the maximum probability of rejection
under the null hypothesis (irrespective of $r) is equal to a.A way of doing this is to
detect the value of $r, $* say, for which the discrepancy between the level and the
size is maximum. For that value, we simulate S times the test statistic, Z'$* (y) .* We
+
then take the observation of rank (95 x S)/100 1 of the statistic as our corrected
5% level critical value: we reject Ho at level 5% when T+*(y) is larger than or equal
to that value. Table 2 reports $*, the size (in %), the 5% asymptotic critical value
*For all asymptotic test procedures, we set S = 10,000.

(a) T=25
100
1
0 1
beta beta
(c) T=75 (d) T=lOO
beta beta
Figure 3 Rejection frequencies of Ho : = 0 in model (4) with II. = 0.5, T =

25,50, 75, 100; induced test (-), tests based on subsamples (yt : t E J l ) and (yL : t E 52)
(. . . and - . -).
(ACV), and the 5% corrected critical value (CCV), for each sample size T , and each
of the three asymptotic procedures.
3. Simulations
In our simulations, we proceeded as follows. For 1c/ E {-1, -.5,0, .5. 1) and T E
{25,50,75, loo}, we considered a grid of /? values around = 0. In each case,
1000 independent samples (yt, t = 1 , 2 , . . . , 7‘) were generated and the follow-
ing statistics were computed: (1)the t-statistic based on the whole sample; (2) the
t-statistic based on the two subsamples (yt : t E J,) and (yL: t E J 2 ) containing
the odd and even observations respectively; (3) the GLS-MM and GLS-ML based
F-statistics; (4) the c,””
-statistic. Using these statistics, the following tests were
implemented at level 5% and the corresponding rejection frequencies were com-
puted: (a) Zinde-Walsh and Ullah’s bounds test; (b) Kiviet’s bounds test;* (c) GLS-
*Because Kiviet (1980)does not provide the upper bound for T = 75 and T = 100, we did not investigate
the behaviour of Kiviet’s test for these values of T.
490 DUFOUR
AND TORR~S
(a) T=25 (b) T 4 0
0 1 0 1
beta beta
. . T=75
(c) (d) T=lOO
0 1
beta beta
Figure 4 Rejection frequencies of Ho : B = 0 in model (4)with @ = 0, T =

25,50,75, 100; induced test (-), tests based on subsamples (yl : t E A ) and (yl : t E J 2 )
(. ’ . and - . -).
MM asymptotic test (corrected and uncorrected for size); (d) GLS-ML asymptotic test
(corrected and uncorrected for size); (e) NW asymptotic test (corrected and uncor-
rected for size); (f) the induced test which combines the standard t-tests based on
the two subsamples (yt : t E ,II)and (yt : t E 1 2 ) ; (g) the separate tests based on the
subsamples (yt : t E J 1 ) and (yt : t E 1 2 ) . The results are presented in Figures 2 to
6 and Tables 3 to 7 .
As it became clear in the description of the induced test, when applying such
a procedure to model (9),one is led to split the sample in two, and make two tests
at level a/2. At first sight, the procedure displays features that may seem quite
unattractive. First, it splits the available sample in two, and second it combines two
tests whose levels are only a/2 (instead of a).From these two remarks, one may ex-
pect the procedure to lack power. But we should keep in mind that, since the two
“subtests” have level a/2, the resulting induced test has level certainly greater than
a/2 (although not greater than a).Furthermore, this test actually uses the informa-
tion contained in the whole sample. Then it becomes less clear whether the induced
test procedure automatically leads to a loss of power relatively to other alternative
(a) T=25 (b) T 5 0
beta beta
. . T=75
(c) (d) T=lOO
0
beta beta
Figure 5 Rejection frequencies of Ho : B = 0 in model (4)with I) = - . 5 , T =

25, 50, 7 5 , 100; induced test (-), tests based on subsamples (yL : t E J I )and (yL : t E J.r)
(. . . and - . -).
tests. Two questions arises from these remarks: (1) Is combining preferable to not
combining? i.e., should our decision at level a rely on an induced test procedure or
on a test based on one of the subsample only? (2) How does the induced test compare
with the procedures mentioned in Section V.B.2?
Figures 2 to 6 answer the first question. They show that the power of the in-
duced test (solid line) is generally higher than that of an a-level test based on one
of the two subsamples (dashed lines). In other words, combining is preferable to not
combining. When it is not the case (when the true value of the MA parameter is
unity, @ = 1, see Figures 2a to Zd), the power loss from using the induced test is
very small, so that one would usually prefer the sample-split procedure that uses all
the observations.
Tables 3 to 7 report the estimated probability of a rejection of Ho : B =
0 for different sample sizes (7' E (25, SO, 75, 100)) and true values for p ( p E
{ - 1, -.8, - .5, -.2,0, 2 , .5, .8, 1 }), for each one of the test procedures of Section
V.B.2. If we first consider bounds tests, we note that the Kiviet test is dominated by
the induced test, except for $ = .5 and $ = 1. We already mentioned in Section
492 DUFOUR
AND TORR~S
(a) T=25 (b) T=50
‘51 0 1
beta beta
(c) T=75 (d) T=100
0 1
beta beta
Figure 6 Rejection frequencies of Ho : = 0 in model (4)with $ = -1, T =

25, 50, 7 5 , 100; induced test tests based on subsamples (y, : t E 11)and (yt : t E 1 2 )
(-)?
(. . . and - . -).
V.B.2 that the bound which has been used here, namely ti,,(0.05), is not appropri-
ate because we do not know whether this value satisfies the level constraint:
In other words, a critical region based on Kiviet’s bounds has an unknown level.
Moreover, what makes the induced test more attractive relatively to Kiviet’s test is
that it avoids the calculation of a bound that changes with the sample size. Finally,
because Zinde-Walsh and Ullah’s upper bounds are so large (see Table l),the power
of their test is zero for all @. These are not reported in Tables 3-7.
The most surprising result which emerges from our Monte Carlo study can
be seen in Tables 3’4, and 5. Once the asymptotic critical values used for the GLS-
MM and GLS-ML tests have been corrected so that the corresponding critical regions
have the desired level, our procedure becomes more powerful than these alternatives
for many plausible values of @. The difference between estimated power functions
grows as @ increases, but diminishes when the sample size T gets larger. The GLS-
MM method seems to be the worst of all the asymptotic procedures studied here,
Table 3 Rejection Frequencies of Ho : fi = 0 in Model (9)with $r =1
Sample size T = 25 Sample size T = 75
/? GLS-ML GLS-MM NW Kiv. Ind. test GLS-ML GLS-MM NW Ind. test
-1.0 0.80 3.40 56.00 66.70 59.80 97.70 0.70 99.10 98.80
-0.8 0.20 1.10 37.80 46.60 40.00 84.90 0.10 91.50 89.70
-0.5 0.00 0.20 17.00 20.10 16.60 40.00 0.00 52.00 48.00
-0.2 0.00 0.00 6.10 6.20 5.20 6.60 0.00 14.30 9.90
-0.1 0.00 0.00 4.50 5.00 3.50 2.20 0.00 7.00 4.40
0.0 0.00 0.00 4.30 4.30 3.50 1.10 0.00 4.30 2.70
0.1 0.00 0.00 5.20 5.10 4.30 2.10 0.00 6.50 3.80
0.2 0.00 0.00 6.50 7.60 5.30 5.80 0.00 11.80 9.80
0.5 0.10 0.60 18.10 21.20 18.10 38.60 0.00 55.20 48.90
0.8 0.10 1.80 38.80 46.40 40.50 83.60 0.20 91.10 89.40
1.o 0.30 3.90 55.00 64.90 59.60 %.70 0.80 98.90 98.80
- 1.0 46.00 0.20 92.40 93.30 90.70 99.50 25.60 99.80 99.70
-0.8 21.50 0.00 77.30 78.30 72.30 97.00 6.70 97.70 %.90
-0.5 3.50 0.00 40.10 40.20 32.90 61.30 0.20 69.10 63.20
-0.2 0.20 0.00 9.60 10.00 6.80 11.90 0.00 17.40 12.20
-0.1 0.00 0.00 4.90 4.50 3.50 4.60 0.00 7.90 5.00
0.0 0.10 0.00 3.90 3.90 2.40 3.40 0.00 4.40 2.80
0.1 0.10 0.00 4.30 4.70 3.50 5.60 0.00 7.40 3.90
0.2 0.20 0.00 8.20 9.20 6.60 11.60 0.00 15.10 10.30
0.5 3.70 0.00 34.90 40.60 33.70 62.80 0.30 67.00 60.40
0.8 21.00 0.10 74.00 78.60 72.70 96.50 4.60 97.30 96.60
1.o 45.10 0.40 89.70 95.00 90.30 99.90 22.90 99.90 99.80
494 AND TORR~S
DUFOUR
Table 4 Rejection Frequencies of Ho : = 0 in Model (9) with I) = .5
GLS-ML GLS-MM NW Kiv. Ind. test GLS-ML GLS-MM NW Ind. test

- 1.0 4.80 10.70 79.40 85.30 83.90 100.00 8.50 100.00 100.00
-0.8 2.50 4.70 61.20 67.30 66.30 98.80 1.oo 99.50 99.40
-0.5 0.60 1.oo 27.20 30.30 29.00 65.50 0.10 77.50 76.70
-0.2 0.00 0.00 6.70 6.80 7.00 11.30 0.00 19.40 18.10
-0.1 0.00 0.00 4.60 3.90 4.60 2.80 0.00 8.40 7.00
0.0 0.00 0.00 4.30 3.50 4.20 0.80 0.00 4.30 2.90
0.1 0.00 0.00 5.60 4.70 4.70 3.10 0.00 7.80 6.40
0.2 0.10 0.10 7.80 8.10 8.30 10.40 0.00 19.30 17.10
0.5 0.50 0.60 27.80 30.90 30.20 64.70 0.10 77.70 75.60
0.8 1.60 4.60 60.50 66.80 64.70 97.80 1.10 99.50 99.40
1.o 5.00 12.10 78.30 84.80 82.80 100.00 7.20 100.00 100.00
-1.0 83.00 2.40 99.30 99.80 99.40 100.00 65.10 100.00 100.00
-0.8 50.80 0.20 93.80 94.30 94.00 99.80 26.80 99.90 99.90
-0.5 10.80 0.00 58.30 59.20 55.80 86.80 1.10 90.90 89.00
-0.2 0.40 0.00 12.70 11.50 11.90 20.50 0.00 24.70 22.50
-0.1 0.10 0.00 6.50 4.10 5.60 6.00 0.00 10.90 7.90
0.0 0.10 0.00 3.30 2.90 3.60 3.40 0.00 4.30 3.50
0.1 0.20 0.00 5.70 4.30 5.20 7.40 0.00 8.70 6.70
0.2 0.40 0.00 13.00 12.20 11.90 19.10 0.00 22.30 20.70
0.5 10.40 0.00 59.70 59.30 57.40 88.10 1.oo 89.70 88.10
0.8 50.60 0.50 94.80 96.00 94.70 100.00 25.20 100.00 99.90
1.0 82.10 3.50 99.40 99.70 99.50 100.00 63.30 100.00 100.00
Table 5 Rejection Frequencies of Ho : ,8 = 0 in Model (9)with I)= 0
/9 GLS-ML GLS-MM NW Kiv. Ind. test GLS-ML GLS-MM NW Ind. test
-1.0 35.30 45.80 97.80 96.20 97.00 100.00 70.30 100.00 100.00
-0.8 19.30 23.10 89.30 81.70 85.20 100.00 28.20 100.00 100.00
-0.5 6.20 5.10 48.40 31.70 45.30 96.90 0.80 98.50 94.30
-0.2 2.50 0.90 8.90 2.80 9.90 25.80 0.00 33.70 28.20
-0.1 1.20 0.30 3.30 0.90 6.60 7.20 0.00 11.80 10.40
0.0 1.oo 0.20 3.10 0.40 4.70 1.40 0.00 2.70 4.50
0.1 1.40 0.60 5.10 0.80 7.00 6.90 0.00 10.60 9.50
0.2 1.90 0.70 9.40 3.OO 11.oo 23.40 0.00 32.10 27.10
0.5 7.00 5.40 48.30 31.80 45.70 96.10 0.80 98.00 94.30
0.8 20.80 23.80 87.50 80.50 85.90 100.00 26.70 100.00 100.00
1.o 37.70 45.30 98.00 97.20 97.00 100.00 70.70 100.00 100.00
-1.0 99.70 40.80 100.00 100.00 100.00 100.00 99.80 100.00 100.00
-0.8 95.00 13.20 100.00 99.90 99.60 100.00 93.20 100.00 100.00
-0.5 45.10 0.50 88.60 73.30 80.20 99.80 20.80 99.60 98.30
-0.2 2.70 0.00 21.40 6.90 19.40 42.50 0.00 47.50 36.30
-0.1 0.50 0.00 7.50 1.40 6.50 12.30 0.00 14.60 12.50
0.0 0.30 0.00 2.40 0.10 4.70 3.90 0.00 3.30 5.20
0.1 0.90 0.00 6.90 1.60 8.00 12.60 0.00 12.70 11.50
0.2 2 -60 0.00 20.80 7.10 18.30 43.50 0.20 43.60 35.40
0.5 43.80 0.60 89.00 72.90 80.20 100.00 17.40 99.60 98.60
0.8 95.10 13.30 99.90 99.70 99.60 100.00 94.00 100.00 100.00
1.0 99.60 41.30 100.00 100.00 100.00 100.00 99.90 100.00 100.00
496 DUFOUR
AND TORR~S
Table 6 Rejection Frequencies of Ho : p = 0 in Model (9) with @ = -0.5
?!/ GLS-ML GLS-MM NW Kiv. Ind. test GLS-ML GLS-MM NW Ind. test
-1.0 93.50 91.70 100.00 97.50 97.20 100.00 100.00 100.00 100.00
-0.8 82.80 76.30 99.80 77.30 84.90 100.00 98.50 100.00 100.00
-0.5 49.80 41.70 82.70 10.10 40.30 100.00 56.70 100.00 96.90
-0.2 18.00 18.20 6.30 0.00 10.80 82.90 14.70 65.10 24.60
-0.1 8.00 8.60 1.40 0.00 6.40 29.60 9.30 10.80 9.90
0.0 4.50 4.30 0.10 0.00 4.40 3.60 5.30 0.40 4.00
0.1 10.60 9.70 1.90 0.00 6.20 29.20 9.60 12.20 9.40
0.2 18.90 19.50 8.70 0.00 11.10 81.50 1.5.30 67.10 23.80
0.5 50.00 41.80 81.40 10.50 42.80 100.00 58.00 100.00 97.60
0.8 82.10 76.00 99.80 76.60 84.40 100.00 98.10 100.00 100.00
1.o 93.10 92.40 100.00 97.90 97.90 100.00 100.00 100.00 100.00
~~ ~~
- 1.0 100.00 97.90 100.00 100.00 100.00 100.00 100.00 100.00 100.00
-0.8 100.00 86.90 100.00 100.00 100.00 100.00 100.00 100.00 100.00
-0.5 98.60 38.90 100.00 66.10 81.80 100.00 97.00 100.00 99.90
-0.2 33.50 16.80 39.70 0.10 18.30 94.90 20.60 84.60 32.30
-0.1 10.30 10.30 4.60 0.00 7.00 45.50 9.40 18.60 11.00
0.0 2.90 5.00 0.10 0.00 4.40 5.00 4.40 0.40 4.30
0.1 9.50 9.10 5.60 0.oo 7.10 44.40 9.30 16.80 10.10
0.2 33.90 15.80 37.20 0.00 17.30 95.30 19.70 82.70 32.00
0.5 99.00 41.30 99.90 67.00 80.20 100.00 97.40 100.00 99.80
0.8 100.00 87.70 100.00 99.90 99.90 100.00 100.00 100.00 100.00
1.0 100.00 98.10 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Table 7 Rejection Frequencies of Ho : B = 0 in Model (9)with $r = -1

-
j3 GLS-ML GLS-MM NW Kiv. Ind. test GLS-ML GLS-MM NW Ind. test
-1.0 99.50 93.10 100.00 79.90 85.20 100.00 100.00 100.00 100.00
-0.8 98.20 80.10 100.00 34.80 64.00 100.00 99.20 100.00 100.00
-0.5 89.80 58.30 89.10 0.00 28.70 100.00 82.30 100.00 83.60
-0.2 49.90 40.80 0.10 0.00 9.00 100.00 51.60 84.60 17.10
-0.1 2.20 4.30 0.00 0.00 5.90 98.80 47.00 0.10 8.10
0.0 0.00 0.00 0.00 0.00 4.80 0.70 0.00 0.00 3.80
0.1 2.30 2.80 0.00 0.00 6.10 99.10 47.00 0.10 8.10
0.2 52.40 41.40 0.40 0.00 9.30 100.00 51.40 87.90 17.90
0.5 90.70 59.40 88.10 0.40 28.50 100.00 82.00 100.00 85.50
0.8 98.10 8 1.30 100.00 34.30 64.70 100.00 99.30 100.00 100.00
1.o 99.60 93.10 100.00 79.40 87.20 100.00 100.00 100.00 100.00
-1.0 100.00 98.30 100.00 100.00 100.00 100.00 100.00 100.00 100.00
-0.8 100.00 92.60 100.00 99.20 99.00 100.00 100.00 100.00 100.00
-0.5 100.00 69.30 100.00 13.10 58.20 100.00 99.20 100.00 98.30
-0.2 96-60 46.00 29.10 0.00 12.10 100.00 6 1.20 99.70 21.70
-0.1 75.00 42.10 0.00 0.00 6.20 100.00 48.60 0.50 8.50
0.0 0.00 0.00 0.00 0.00 3.SO 1.10 0.00 0.00 3.80
0.1 76.00 4 1.90 0.00 0.00 6.50 100.00 48.90 0.40 8.70
0.2 %.40 45.80 28.20 0.00 10.80 100.00 61.90 99.50 20.70
0.5 100.00 69.30 100.00 12.10 59.20 100.00 99.00 100.00 98.50
0.8 100.00 92.30 100.00 99.50 98.70 100.00 100.00 100.00 100.00
1.o 100.00 98.30 100.00 100.00 100.00 100.00 100.00 100.00 100.00
498 DUFOUR
AND TORR~S
whereas GLS-ML appears to benefit from the asymptotic efficiency property of max-
imum likelihood estimators. But for nonnegative values of @, the sample size has to
be T = 100 for the GLS-ML test to have a probability of correctly rejecting the null
as high as the induced test. The GLS-MM test is still dominated for some negative
values of +(@ = -.5), irrespective of the sample size. Only when @ is close to -1
does this procedure become admissible.
While the two comm‘only used asymptotic inference procedures, GLS-MM and
GLS-ML, cannot be recommended on the ground of our Monte Carlo study, the con-
clusion is less negative for the NW method. Except for small sample sizes ( T = 25)
and large values of the MA parameter (@ = 1, .5), it does better than the induced
test procedure. This result is somewhat unexpected because the Newey-West estima-
tor of V ( ~ Tdoes
) not take into account the autocovariance structure of the process.
However, although the induced test is conservative, it is more powerful than NW test
for alternatives close to the null hypothesis when @ is negative. Furthermore, it is
important to remember that the NW test suffers from level distortions (see Table 2)
that are not easy to correct in practice.
4. An Example: An Induced Test on the Mean of the Canadian

Per Capita GDP Series
We now apply our procedure to test the nullity of the mean of a process that has a
MA(1) representation. Our series is the first difference of the Canadian per capita
GDP, denominated in real 1980 Purchasing Power Parity-adjusted US dollars, ob-
served yearly from 1901 to 1987. It is taken from Bernard and Durlauf (1995). Fig-
ure 7 plots the series. Using standard Box-Jenkins procedure (autocorrelation and
partial autoconelation functions), we identified a MA( 1) process for the series (see
Table 8).
B
We then consider a model like (9). ML estimation of (9) gives = 136.1810
and IJ = 0.4211 with estimated variances 945.1919 and 0.0095, respectively.
6)
The estimated COV(B, is 0.0834 and the sample variance of the residuals is
401 17.5725.
To implement an induced test for the nullity of the mean parameters, /3, at level
5%, we split the sample in two parts, {yL : t E J i } , i = 1 , 2 , which contain respec-
tively the odd and the even observations, and make two 2.5% tests of /3 = 0, using the
statistics ti = &‘lY;(/si, where = ri x / n i , s? = (EjEJ,(% - r i ) 2 ) / ( n i- l ) ,
and ni is the size of subsample i, i = 1, 2. We reject the null hypothesis when
t l > t(a/4,U I ) or t 2 > t(a/4,UZ), where t(a/4, U) is the 1 - ( a / 4 )percentile of
Student’s t distribution with U degrees of freedom. We also perform both GLS-MM
and GLS-ML asymptotic tests. Our results are reported in Table 9. is the two step
estimator of B, $ the estimator of @ that has been obtained in the first step to esti-
mate the error covariance matrix, and t the test statistic, whose distribution will be
approximated by a student’s t-distribution with 86 degrees of freedom. Both subtests
Year
Figure 7 First differences of the Canadian per capita GDP. (From Bernard and Durlauf
1994.)
reject the null hypothesis at level 2.5%. Hence the induced test rejects the nullity
of the mean at level 5%. The two asymptotic tests also reject the null hypothesis, if
we admit that the asymptotic critical value is a good approximation when the sample
size is 87. Our findings are consistent with the results of the Monte Carlo study of
Section V.B.3. For similar sample sizes ( T = 75 or T = 100) we found that the
GLS-MM test produces larger values of the test statistic than the GLS-ML test does.
This is what we have here with T = 87.
If we decide to include a linear trend in the mean of the MA(1) process, our
induced test procedure still applies. The per capita GDP series now admits the rep-
resentation
able 8 Sample Autocorrelations of the Canadian Per Capita GDP Series
aa 1 2 3 4 5 6 7 8 9 10 11 12
utocorrelation .41 .19 .10 -.04 .OS .07 .12 .04 -.04 .09 .08 .20
tandard error .ll .12 .13 .13 .13 .13 .13 .13 .13 .13 .13 .13
jung-box 15.4 18.8 19.8 19.9 20.2 20.6 22.1 22.3 22.5 23.3 24.0 28.3
@statistic
500 DUFOUR
AND TORR~S
Table 9 Induced and Asymptotic Tests. Model yt = + et + $ret-,

i=l i=2 GLS-MM GLS-ML
s(v)
t
127.6836
4.1892 (43)
125.4406
3.5076 (42)
122.2522
7.5112 (86)
123.6574
6.9345 (86)
p-value 0.00014 0.00109 0.00000 0.00000
$ - - 05298 0422 1
for t = 1 , 2 , . . . , T . We consider the three hypotheses
For each hypothesis, we perform the induced test as well as the asymptotic test.
Results appear in Table 10. We note that only one of the subtests rejects the pres-
ence of a linear trend. However, according to our decision rule, this is enough to
reject HA’). Both GLS-MM and GLS-ML unambiguously reject this hypothesis. But
we know from our simulations that the asymptotic tests tend to reject the null too
often when it is true. For the parameters p;, j = 1,2, we also report two confidence
intervals and I;, each with level 97.5%, based on the two subsamples (yt : t E J l )
and (yt : t E J 2 ) . The intersection Z{ n Zi gives the set of values y E R such that the
hypothesis H ; ( y ) : 3/; = y is not rejected at level 5% by the induced test. These
intervals are
Table I0 Induced and Asymptotic Tests. Model yt = PO + B1 t + + $rct-,

E~
i=l i=2 GLS-MM GLS-ML

~~~~-~
~~~-~~~~
-37.0695 -6.298 1 -22.5258 -22.5578

3.7444 2.9941 3.3507 3.3554
-0.6832 (42) -0.0903 (41) -0.6726 (85) -0.6577 (85)
0.49823 0.92849 0.50303 0.51251
3.5058 (42) 2.1674 (41) 5.0392 (85) 4.9319 (85)
0.00110 0.03606 0.00000 0.00000
17.2244 (2,42) 9.0421 (2, 41) 37.9250 (2, 85) 39.3875 (2,85)
0.00000 0.00056 0.00000 0.00000
0.3536 0.3253
tj, j = 1 , 2 , and F denote Student’s t - and Fisher’s F-statistics used for testing H$),j = 1 , 2 , and Ho,
respectively.
Table I I Induced and Asymptotic Tests. Model yl = fi + E~ + +$2~t-p

i=l i=2 i=3 GLS-MM GLS-ML
B 122.4644 130.7687 126.4918 128.9405 129.7032

t(V) 2.6573 (28) 4.1328 (28) 2.9150 (28) 3.6828 (86) 3.2812 (86)
p-value 0.01286 0.00029 0.00692 0.00040 0.00149
$1 - - - 0.4096 0.3931
$2 - - - 0.2354 0.1037
t denotes the statistic used for testing Ho : /3 = 0.
Zp = [-163.2002,89.0612] Zi = [-168.5787, 155.98251

I: = [1.2616,6.22721, Zi = [-0.2201, 6.20831
yielding the following 95%confidence intervals for PO and pi:
/I0 E [-163.2002,89.0612], /I1 E [1.2616,6.2083]
These entail that is not significantly different from 0, while B1 is.
When we apply the induced test procedure, we implicitly assume that we have
correctly identified a MA(1) process. An interesting issue is to look at what we get
if, instead of the true MA(1) representation, we use a MA(2) model to build our test
statistics. In this case, we split the sample in three parts, make three tests at level
5/3%,and reject the null Ho : /I = 0 when our sample falls in one of the three critical
regions. The results are in Table 11. We first note that as two of the subsample based
tests reject HO at level 5/3%,we reject the null at level 5%.We also note that both
asymptotic tests reject the null hypothesis. But we know that using 5% critical values
obtained from asymptotic distributions leads to a probability of making a type I error
larger than 5%. Therefore, although asymptotic tests and tests based on the sample
split yield the same decision of rejecting Ho, we put more confidence in the sample
split procedure.
VI. CONCLUDING REMARKS
In this chapter we proposed a set of inference methods for comparing and pooling
information obtained from different data sets, which simply use separate tests (or
confidence sets) based on the different data sets. The methods described are based
on a systematic exploitation of Boole-Bonferroni inequalities and can yield exact
tests and confidence sets without the need to specify at all the relationship between
the data sets, even with small sample sizes. As a result, they are quite versatile and
usually easy to implement. The general problems studied include (1) combining sep-
arate tests based on different data sets for an hypothesis of interest (more precisely,
502 DUFOUR
AND TORRh
for the intersection of similar hypotheses), to obtain more powerful tests; (2) compar-
ing parameters estimated from the different data sets (e.g., to test their equality); (3)
combining confidence regions based on different samples to obtain a more accurate
confidence set. For problem 1,we were led to consider Bonferroni-type induced tests;
for problem 2, we proposed empty intersection tests; and for problem 3, we suggested
taking the intersection of separate confidence sets with appropriate levels.
We also showed that the methods proposed can be quite useful in various mod-
els where usual inference procedures based on a complete sample involve difficult
distributional problems (e.g., because of nuisance parameters), but for which distri-
butional properties of test statistics computed on appropriately selected subsamples
are simpler. This leads to an interesting form of sample-split (SS) method. One first
splits the sample into several subsets of observations from which separate inferences
(tests of confidence sets) are obtained. Then these results are recombined, using the
general union-intersection (UI) methods already described, to obtain a single infer-
ence which uses the full sample. The way the data is split depends on the model
considered. In some situations the structure naturally suggests the division. This
is for example true when the model contains several equations. In other cases, the
division is based on more elaborate arguments, as in moving average models.
The UI/SS methods proposed can be applied to a wide spectrum of econometric
situations and models. We discussed and illustrated their applications in two cases
where only asymptotic methods are typically available, namely inference in SURE
models and linear regressions with MA errors. In the latter case, we also presented
an extensive Monte Carlo study comparing the UI-SS method for testing an hypoth-
esis about a mean with other available approaches. Two main conclusions emerged
from these results: first, they provided further evidence on the size distortions asso-
ciated with usual asymptotic procedures; second, they showed that UI-SS tests not
only have the predicted levels, but enjoy good power properties. In view of the fact
that these methods involve splitting the sample and lead to conservative procedures,
hence leading one to expect a power loss, this is indeed quite remarkable. Our re-
sults show that the Bonfen-oni-based recombination of the evidence obtained from
the different subsamples apparently makes up for the loss. For another application
of UI-SS approach to autoregressive and other types of dynamic models, the reader
may consult Dufour and Torrks (1995).
Before closing, it is worthwhile noting a few other points. First, the U1 (or UI-
SS) procedures are often simpler to implement than usual asymptotic procedures.
For SURE models and linear regressions with MA errors, they only require critical
values from standard distributions. For MA models, they avoid the task of estimating
MA coefficients. Second, they offer some extra robustness to model specification, as
illustrated by SURE where no assumption on the relationship between the different
equations is needed. Third, although we stressed here the derivation of finite sample
methods, there is nothing that forbids the application of U1 (or UI-SS) methods to
situations where only asymptotically justified tests or confidence sets are available
UNION-INTERSECTIONAND SAMPLE-SPLIT METHODS 503
from the separate data sets. In such cases, the methods are applied in exactly the
same way. This feature may be especially attractive for gaining robustness to model
specification. We think all these properties make this UI-SS approach an attractive
and potentially quite useful addition to the methods available to applied econome-
tricians.
ACKNOWLEDGMENTS
We thank Eugene Savin, the editor, David Giles, and an anonymous referee for sev-
eral useful comments. This work was supported by grants from the Social Sciences
Research Council of Canada, the Natural Sciences and Engineering Council of Canada,
and the Government of Qukbec (Fonds FCAR).
REFERENCES
Alt, F. B. (1982),Bonferroni Inequalities and Intervals, in Encyclopedia ofStatistical Sciences,

Vol. 1 (N. L. Johnson and S. Kotz, eds.), Wiley, New York, 294-301.
Angrist, J. D. and A. B. Krueger (1994),Split Sample Instrumental Variables, Technical Work-
ing Paper 150, NBER, Cambridge, MA.
Arnold, S. F. (1981),The Theory of Linear Models and Multivariate Analysis, Wiley, New York.
Ashenfelter, 0. and A. Krueger (1992), Estimates of the Economic Return to Schooling from
a New Sample of Twins, Working Paper 304, Industrial Relations Sections, Princeton
University.
Ashenfelter, 0. and D. J. Zimmerman (1993), Estimates of the Returns to Schooling from
Sibling Data: Fathers, Sons, and Brothers, Working Paper 318, Industrial Relations
Section, Princeton University.
Bernard, A. B. and S. N. Durlauf (1995), Convergence in International Output, Journal of
Applied Econometrics, 10,97-108.
Berndt, E. R. (1991), The Practice of Econornetrics: Classic and Contemporary, Addison-
Wesley, Reading (MA).
Berndt, E. R. and D. 0. Wood (1975),Technology, Prices and the Derived Demand for Energy,
Review of Economics and Statistics, 53,259-268.
Birnbaum, A. (19S4),Combining Independent Tests of Significance, Journal ofthe American
Statistical Association, 49, 559-575.
Brockwell, I! J. and R. A. Davis (1991), Time Series: Theory and Methods, Springer-Verlag,
New York.
Burnside, C. and M. Eichenbaum (1994), Small Sample Properties of Generalized Method of
Moments Based Wald Tests, NBER Technical Working Paper, No 155.
DeJong, D. N., J. C. Nankervis, N. E. Savin, and C. H. Whiteman (1992), The Power of Unit
Root Tests in Time Series with Autoregressive Errors, Journal of Econometrics, 53,
323-343.
504 AND TORR~S
DUFOUR
Dufour, J.-M. (1990), Exact Tests and Confidence Sets in Linear Regressions with Autocorre-
lated Errors, Econometrica, 58,479-494.
Dufour, J.-M. (1997), Some Impossibility Theorems in Econometrics, with Applications to
Structural and Dynamic Models, Econometrica, forthcoming.
Dufour, J.-M. and J. Jasiak (1993, Finite Sample Inference Methods for Simultaneous Equa-
tions and Models with Unobserved and Generated Regressors, Discussion Paper, CRDE,
Universitede Montreal.
Dufour, J.-M. and 0. Torrhs (1995), Two-sided Autoregressions and Exact Inference for Sta-
tionary and Nonstationary Autoregressive Processes, Discussion Paper, CRDE, Univer-
sitkde Montrkal.
Folks, J. L. (1984), Combination of Independent Tests, in Handbook of Statistics, Volume 4,
Nonparametric Methods (P. R. Krishnaiah and P. K. Sen, eds.), North Holland, Amster-
dam, pp. 113-121.
Fomby, T. B., R. C. Hill and S. R. Johnson (1984),Advanced Econometric Methods, Springer-
Verlag, New York.
Gourikroux, C. and A. Monfort (1989),Stntistique et Modbles kconom&triques,vol. 2, Econom-
ica, Paris.
Hedges, L. V. and I. Olkin (1985),Statistical Methodsfir Meta-Analysis, Academic Press, San
Diego.
Hillier, G. H. and M. L. King (1987), Linear Regression with Correlated Errors: Bounds on
Coefficient Estimates and t-values, in Specijcation Analysis in the Linear Model (M. L.
King and D. E. A. Giles, ed.), Routledge & Kegan Paul, London, Ch. 4.
Kennedy, P. and D. Simons (1991), Fighting the Teflon Factor: Comparing Classical and Bayesian
Estimators for Autocorrelated Errors, Journal of Econometrics, 48, 15-27.
King, M. L. (1996), Hypothesis Testing in the Presence of Nuisance Parameters, Journal of
Statistical Planning and Inference, 50, 103-120.
Kiviet, J. F. (1980), Effects of ARMA Errors on the Significance Tests for Regression Coeffi-
cients: Comments on Vinod’s Article; Improved and Additional Results, Journal ofthe
American Statistical Association, 75, 353-358.
Laskar, M. R. and M. L. King (1995), Parameter Orthogonality and Likelihood Functions, in
Proceedings of the 1995 Econometrics Conference at Monash, Melbourne (Australia),
(C. S. Forbes, P. Kofman, and T. R. L. Fry, eds.), Monash University, 253-289.
Lehmann, E. L. (1986), Testing Statistical Hypotheses, Wiley, New York.
McCabe, B. P. M. (1988), A Multiple Decision Theory Analysis of Structural Change in Re-
gression, Econometric Theory, 4, 499-508.
Miller, R. G. J r (1981), simultaneous Statistical Inference, Springer-Verlag, New York.
Miyazaki, S. and W. E. Griffiths (1984),The Properties of Some Covariance Matrix Estimators
in Linear Models with AR(1) Errors, Economics Letters, 14,351-356.
Nankervis, J. C. and N. E. Savin (1987), Finite Sample Distributions o f t and F Statistics in
a n AR(1) Model with a n Exogenous Variable, Econometric Theory, 3,387-408.
Nelson, C. R., R. Startz, and E. Zivot (1996), Valid Confidence Intervals and Inference in the
Presence of Weak Instruments, Technical Report, Department of Economics, University
of Washington.
Newey, W. K. and K. D. West (1987), A Simple, Positive Semi-Definite, Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix, Econometrica, 55, 703-708.
Park, R. E. and B. M. Mitchell(1980), Estimating the Autocorrelated Error Model with Trended
Data, Journal of Econometrics, 13, 185-201.
Phillips, G. D. A. and B. P. M. McCabe (1988), Some Applications of Basu's Independence
Theorems in Econometrics, Statistica Neerlandica, 42, 37-46
Phillips, G. D. A. and B. P. M. McCabe (1989), A Sequential Approach to Testing Econometric
Models, Empirical Economics, 14, 151-165.
Roy, S.N. (1953), On a Heuristic Method of Test Construction and Its Use in Multivariate
Analysis, Annals of Mathematical Statistics, 24,220-238.
Savin, N. E. (1984), Multiple Hypothesis Testing, in Handbook of Econometrics (Z. Griliches
and M. D. Intriligator, eds.), North-Holland, Amsterdam, Chap. 14.
Savin, N. E. and A. Wurtz (1996),The Effect of Nuisance Parameters on the Power of LM Tests
in Logit and Probit Models, Technical Report, Department of Economics, University of
Iowa, Iowa City, IA.
Srivastava, V. K. and D. E. Giles (1987), Seemingly Unrelated Regression Models, Marcel
Dekker, New York.
Staiger, D. and J. H. Stock (1997), Instrumental Variable Regression with Weak Instruments,
Tippett, L. H. C. (1931), The Methods ofStatistics, Williams & Norgate, London.
Tunnicliffe Wilson, G. (1989), On the Use of Marginal Likelihood in Time Series Model Esti-
mation, Journal of the Royal Statistical Society, Series B , 51, 15-27.
Vinod, H. D. (1976), Effects of ARMA Errors on the Significance Tests for Regression Coef-
ficients, Journal of the American Statistical Association, 71,929-933.
Vinod, H. D. and A. Ullah (1981), Recent Advances in Regression Methods, Marcel Dekker,
New York.
Wang, J. and E. Zivot (1996), Inference on a Structural Parameter in Instrumental Variables
Ikgression with Correlated Instruments,' Technical Report, Department of Economics,
University of Washington.
Wooldridge, J. M. (1989), A Computationally Simple Heteroskedasticity and Serial Correla-
tion Robust Standard Error for the Linear Regression Model, Economics Letters, 31,
239-243.
Zinde-Walsh, V. and A. Ullah (1987), On the Robustness of Tests of Linear Restrictions in
Regression Models with Elliptical Error Distributions, in Time Series and Econometric
Modelling (I. B. MacNeill and G. J. Umphrey, eds.), Reidel, Dordrecht.
Modeling Economic Relationships
with Smooth Transition Regressions
Timo Terasvirta
Stockholm School of Economics, Stockholm, Sweden
1. INTRODUCTION
The assumption of linearity has long dominated macroeconometric model building.

If linear approximations to economic relationships had not been successful in empir-
ical work they would no doubt have been abandoned long ago, but this has not been
the case. On the contrary, very powerful concepts such as that of cointegrated vari-
ables (Granger 1981) have been built upon the idea of linear relationships between
variables. The statistical theory of linear models with cointegrated variables is well
established and enables the analysis of linear systems of cointegrated variables; for
recent overviews see Banerjee et al. (1993),Hatanaka (1996),and Johansen (1995).
A coherent modeling strategy built upon the multivariate normal distribution, its
conditional means, and the idea of moving from general to more specific models has
been proposed and successfully put into practice in a number of applications. The
idea has been to merge economic theory and statistical considerations in search of a
congruent model. Hendry (1995)contains a detailed account of these principles.
Nevertheless, there exist situations in which the underlying economic theory
is strongly nonlinear, and the econometric methods have reflected that fact. Mod-
eling economies with rationing constitutes an example. Maddala (1986) surveyed
the resulting switching regression models that were disequilibrium models with the
“minimum condition.” The class of switching regression models also includes mod-
els that assume a finite number of linear regimes but without the minimum condition.
Which regime generates the observations at any given point of time in these models
depends instead on a so-called switching variable that can either be observable or
unobservable. In the former case, the switchpoint itself is often assumed unknown;
507
508 TE~SVIRTA
see Goldfeld and Quandt (1972, Section 9.2). Piecewise regression models, see for
example Ertel and Fowlkes (1976) and the references therein, belong to this cate-
gory. The general idea forms a basis for testing parameter constancy against a struc-
tural break at an unknown point as in Quandt (1960)and Andrews and Ploberger
(1994). Goldfeld and Quandt (1973) first discussed the case in which the sequence
of unobserved switching variables is an irreducible first-order Markov chain, and
Lindgren (1978) derived the maximum likelihood estimators of parameters of such
hidden Markov models.
Switching regression models may be generalized in such a way that the tran-
sition from one extreme regime to the other is not discrete but smooth. Such gen-
eralizations are the topic of this chapter. Instead of a usually (small) finite number
of regimes there exists in these generalizations a continuum of them. Bacon and
Watts (1971) first suggested such a model and coined the term “smooth transition”
to illustrate how a locally linear equation changes from the one extreme linear pa-
rameterization to the other as a function of the continuous transition variable. In the
econometrics literature, Goldfeld and Quandt (1972, pp. 263-264) presented a sim-
ilar idea. Suppose one has the following switching regression model of two variables
yt and xt:
yt = a111 - D ( z t ) ) + a 2 W t ) + 1B1 (1 - D(zt>)+ 82D(Zt>)Xt

(1)
+ 11 - D(ZJU1, + D(Ztb2t
where uit - nid(0, a:), i = 1 , 2 , are the error terms of the two regimes, zt is the
transition variable, and D ( z t ) is the Heaviside function:
Goldfeld and Quandt pointed out that the estimation of the parameters of (1)includ-
ing c in (2) is complicated and that the problem can be simplified by using a feasible
approximation of (2). Their suggestion was to define
i.e., to assume that the transition function is the cumulative distribution function
of the normal ( c , a2)variable and that ay = a; in (1). Replacing (2) by (3) in the
switching regression (1) defines a smooth transition regression. In the time series
literature, Chan and Tong (1986)made a similar suggestion in order to generalize a
univariate switching autoregressive or “threshold autoregressive” model, albeit not
for computational reasons. Tong (1990) contained a detailed account of threshold
autoregressive models. Maddala (1977, p. 396) proposed that (3) be replaced by the
logistic function
~ ( z ,=
) (1 + exp(S1 + ~22,})-’ (4)
MODELING
ECONOMIC RELATIONSHIPS 509
which is one of the alternatives to be considered in this chapter; see also Granger
and Terasvirta (1993) and Terasvirta (1994).
A smooth transition between two extreme regimes may be an attractive param-
eterization because the resulting smooth transition regression (STR) model is locally
linear and thus often allows easy interpretation. Also from the point of view of the-
ory, the assumption of a small number (usually two) of regimes may sometimes be
too restrictive compared to the STR alternative. For instance, instead of assuming
that an economy just has two discrete states, expansion and contraction, say, it may
be more convenient and realistic to assume a continuum of states between the two
extremes. Another argument is that economic agents may not all act promptly and
uniformly at the same moment; their response to news requiring action may con-
tain delays. Nevertheless, these two viewpoints are not competitors. As is already
obvious from Goldfeld and Quandt (1972, pp. 263-264), the two-regime switching
regression model is a special case of an STR model and can therefore be treated in
that framework.
Furthermore, an STR model may be used in the same way as the switching
regression to serve as an alternative against which to test parameter constancy in
a linear model. The alternative to parameter constancy in this framework is a con-
tinuous change in parameters, which often is statistically a more convenient case to
handle than just a single structural break. Lin and Terasvirta (1994)discussed this
possibility which requires defining zt = t in (4) while the STR model has the form
where xt is a vector of explanatory variables and and 0 are parameter vectors. If

the null of parameter constancy, D ( t ) = 0 in (5), is rejected one can estimate the
alternative including the parameters in D ( t ) and find out how the parameter vec-
+
tor p 0 D ( t ) is changing over time. For multivariate applications, see for exam-
ple Albaek and Hansen (1995), Coutts, Mills and Roberts (1995), Heinesen (1996),
Jansen and Terasvirta (1996),Liitkepohl, Terasvirta, and Wolters (1995)and Wolters,
Terasvirta, and Liitkepohl(l997).
Generally, the transition variable of an STR model is an observable economic
variable in the tradition of Fair and Jaffee (1972) who applied a switching regression
model to modeling the demand for and supply of housing. Models with an unob-
served transition variable will not be discussed any further here. The application
in Section V is also related to the housing market. Jansen and Terasvirta (1996)
applied the general STR model in various ways as a vehicle for testing super exo-
geneity; see, for example, Hendry (1995, Chap. 5) or Ericsson (1992) for definitions
of this and other types of exogeneity. Other applications of the STR model to-date
include Granger, Terasvirta, and Anderson (1993), Semmler and KoCkesen (199S),
and Chiarella, Semmler, and KoCkesen (1996).
The approach in this chapter is classical. However, Bacon and Watts (1971)
used the Bayesian approach for estimating the parameters of their smooth transition
5 I0 TERASVIRTA
model. They first obtained the joint marginal posterior distribution for the parame-
ters of the transition function; for example in (3) those are c and 1/02.Then they
estimated the remaining parameters conditionally on the mean of this joint posterior
distribution. Later, Tsurumi (1982) applied the same idea to his “gradual switching”
simultaneous-equation model. More recent Bayesian treatments have been mainly
concerned with switching regression or threshold autoregressive models. Pkguin-
Feissolle (1994) who discussed the smooth transition autoregressive model consti-
tutes an exception. Published work on threshold autoregressive models includes Pole
and Smith (1985),Geweke and Terui (1993),Pfann, Schotman, and Tschernig (1996),
and Cook and Broemeling (1996).
This chapter is organized as follows. The STR model is defined and its prop-
erties and potential in applications are discussed in Section 11. Section I11 considers
statistical inference in STR models. This includes testing linearity which should
precede any nonlinear modeling. The modeling cycle consisting of specification, es-
timation, and evaluation of STR models is described in Section IV. Section V con-
tains an application to a U.K. house price equation considered in Hendry (1984),
and Section VI concludes.
II. S M O O T H T R A N S I T I O N REGRESSION M O D E L
Consider the nonlinear regression model
wherex, = (1, xll, . . . , xpt)‘ = (1, y,-1, . . . , y,-k; z l t , . . . , z,,)’ withp = k+m is
the vector of explanatory variables, cp = (cpo, cpl , . . . , cpp)’ and 8 = (80,81,. . . ,8,)’
are parameter vectors, and {U,} is a sequence of independent, identically distributed
errors. Some of the parameters qi and 8, may be zero a priori or the restriction cpi =
-8; may hold for some i. In (6), G is a bounded continuous transition function; it is
customary to bound G between zero and unity, and S, is the transition variable. It may
be a single stochastic variable, for example, an element of xt,a linear combination
of stochastic variables or a deterministic variable such as a linear time trend. By
writing (6)as
it is seen that the model is locally linear in x, and that the combined parameter vector
+
q 8 G is a function of the transition variable sI. If G is bounded between 0 and 1,
+
the combined parameters fluctuate between cp and cp 8. In dynamic modeling, this
property makes it possible, for example, to characterize an economy with dynamic
properties in expansion being different from those in contraction. Terasvirta and
ECONOMIC RELATIONSHIPS 5 I I
MODELING
Anderson (1992) contained univariate examples of this kind. Model (6) is an example
of the smooth transition regression (STR) model discussed in the Introduction.
The practical applicability of (6)depends on how G is defined. A few defini-
tions have been suggested in the literature; see, for example, Granger and Terasvirta
(1993, Chap. 7). If G has the form
G d y ; c, S t ) = (1 + exp{-y(st - c)H-', y >0 (8)

then the STR model (6) is called the logistic STR of LSTRl model. The transition
function (8)is a monotonically increasing function of st . The restriction y > 0 is an
identifying restriction. The slope parameter y indicates how rapid the transition from
zero to unity is as a function of st and the location parameter c determines where the
transition occurs. If y + 00 in (8),(6) becomes a two-regime switching regression
model with the switching variable s t . In this special case, st = c is the switchpoint
+ + +
between the regimes y, = xi(o U , and yt = x ; ( q 19) u t .
Monotonic transition may not always be a satisfactory alternative in applica-
tions. A simple nonmonotonic alternative is
where the restrictions on y , c1, and c2 are identifying restrictions. This transition
function is symmetric about (cl +
c2)/2, and G(y, c; s t ) + 1 for st + f w . The
minimum value of G remains between 0 and 1/2, the upper limit holds for cl = c2.
On the other hand, when y + 00, G ( y , c; s,) + 0 for cl 5 st 5 cg; for other values
G ( y , c; s,) + 1. This is a special case of a three-regime switching regressions model
in which the two outer regimes are equal. The STR model (6)with transition function
(9) is called the LSTR2 model.
Jansen and Terasvirta (1996) suggested (9) as a generalization of the expo-
nential STR (ESTR) model discussed in the literature; see Granger and Terasvirta
(1993, Chap. 7). The transition function of an ESTR model is defined as
G(y, c; s t ) = 1 - exp{-y(s, - c ) ~ } , y >0 (10)

which is closely related to the case c1 = c2 in (9). The transition function is sym-
metric about c and G(y, c; s,) + 1 for s, + f00.However, when y + 00, the
transitiOn function G ( y , c; s t ) + 1 except that for S, = c the limit is 0. Thus for
large values of y it is difficult in practice to distinguish an ESTR model from a lin-
ear model. By introducing another parameter as in (9) one obtains an STR model
with more useful limiting properties as y + 00. An LSTR2 model with three ex-
treme regimes and fairly sharp transitions between them achieved by (9) may in fact
be quite an interesting alternative in practice; see for example the case discussed in
Anderson (1997). In all theoretical derivations of the chapter, however, it is assumed
that y is finite.
5 I2 TERASVIRTA
Defining st = t yields an important special case of the STR model. Then (7)
becomes
Model (11)can be interpreted as a linear model whose parameters change over time.
It contains as a special case the presence of a single structural break which has
been the most popular alternative to parameter constancy in econometric work. This
special case is obtained by completing (6) by (8)with st = t and letting y -+ 00 in
(8).Lin and Terasvirta (1994)defined another nonmonotonic transition function (see
also Jansen and Terasvirta 1996)
where y > 0, cl 5 c2 5 c3. In fact, Lin and Terasvirta (1994) defined the exponent
of (12) directly as a third-order polynomial without requiring the roots to be real. As
we shall see, this does not make any difference as far as testing parameter constancy
is concerned. On the other hand, if an STR model with (12) is to be estimated, re-
stricting the roots to be real alleviates the potential problem of very high correlation
between the estimator of 0 on the one hand and that of y and possibly c l , c2, and c3
on the other. At the same time one does not give up too much generality in the sense
that (12) still allows quite a lot of flexibility in the transition function.
Many parameter constancy tests explicitly or implicitly assume the alternative
to parameter constancy to be a single structural break. If this null hypothesis is re-
jected it is often not obvious what to do next. More information can be obtained by
using recursive tests: see, for example, Hendry (1995, Chap. 16). An advantage of
testing parameter constancy in the STR framework is that any rejection of the null
hypothesis is a rejection against a parametric alternative. In case of a rejection the
parameters of the alternative can be estimated, which helps obtain information about
where and how parameter constancy breaks down if it does. This information in turn
is helpful in deciding how the specification of the model should be improved in order
to obtain a model with constant parameters.
It is of course possible to define (12) also for other transition variables than
st = t. However, macroeconomic time series are usually not overly long. When
modeling with such series it may be advisable to restrict the order of the exponent
of the transition function to two to avoid excessive difficulties in parameter estima-
tion unless there is economic theory suggesting a higher order. This is done here
excepting the case st = t. In that case the experience has shown that there are less
problems: a heuristic explanation to that is that for st = t, superconsistency makes
the estimation of the parameters in the exponent easier in small samples than if s t is
stationary.
MODELINGECONOMIC RELATIONSHIPS5 I3
111. INFERENCE IN SMOOTH T R A N S I T I O N

REGRESSION MODELS
A. Testing LinearityAgainst STR
The first question an applied econometrician is facing in considering a nonlinear
model is: is it worthwhile? If the economic relationship to be modeled can be ad-
equately characterized by a linear model, then working with a nonlinear model is
waste of time. Besides, as will be seen, fitting certain types of nonlinear models such
as STR models to data is not a statistically feasible undertaking if the data-generating
process is linear. Econometric modeling with STR models thus has to begin with test-
ing linearity against STR. In order to discuss testing statistical hypotheses within
the STR model, additional assumptions about (6)are necessary. The stochastic vari-
ables among z l t , . . . , Z k t are assumed stationary whereas the nonstochastic ones are
dummy variables. Furthermore, st is assumed to be a stationary variable (if it is not
a time trend), and all cross-moments Ezitzjt, Eziis!, Eyt-is:, and EyL-lZjt, k 5 3,
are assumed to exist. Note, however, that some of the variables zjt may be stationary
linear combinations of I(1) variables. The parameters of such combinations may be
treated as known because their least squares estimates are super consistent, and the
inference to be discussed is still valid. Finally, the errors are assumed uncorrelated
with xt and s t .
A minor redefinition of the transition functions will allow convenient notation
in this section. Let G* = G - 1/2 where G is any of the transition functions defined
in the previous section. Rewrite (6) as
-
(although q~ and 8 are changed the previous notation is retained). In order to de-
rive the test statistic, assume that uc nid(0, a'). The conditional log-likelihood
function of the model is
T T
T
- logo'
2
--
1
20'
ZU; (14)
1=1 i=1
The null hypothesis of linearity in (13)is Ho: y = 0 against H I : y > 0. It is assumed

that the roots of the lag polynomial 1- q,Lj are outside the unit circle when the
null hypothesis holds. It is seen from (13) that the linearity hypothesis can equally
well be expressed as Hoe: 8 = 0. This is an indication of an identification problem in
(13):the model is identified under the alternative but not under the null hypothesis.
This poses the following statistical problem. Suppose that the investigator wants to
apply a likelihood ratio test to testing Ho. In order to do that (13) has to be estimated
both under Ho and H I . However, the parameters of (13) cannot be consistently es-
5 I4 TERASVIRTA
timated under Ho because 0 and c are then nuisance parameters whose values do
not affect the value of the likelihood. For this reason, the likelihood ratio statistic
does not have its standard asymptotic x’
distribution under the null hypothesis. The
two other classical tests, the Lagrange multiplier and the Wald test share the same
property.
Davies (1977, 1987) first discussed solutions to this problem; for later con-
tributions in the econometrics literature, see for example Shively (1988),King and
Shively (1993), Lee, White, and Granger (1993), Andrews and Ploberger (1994),
and Hansen (1996).It occurs in connection with many nonlinear models which nest
a linear model such as the STR model, the switching regression model with an un-
known switchpoint or the Hidden Markov model of Goldfeld and Quandt (1973)and
Lindgren (1978).A common feature in much of the literature is that under the null
hypothesis there is a single nuisance parameter in the model. When the STR model
is concerned there are at least two such parameters, whichever way one formulates
the null hypothesis. A way of solving the identification problem by circumventing
it is discussed for example in Granger and Terasvirta (1993, Chap. 6) and is also
considered here. It is based on the work of Saikkonen and Luukkonen (1988) and
Luukkonen, Saikkonen, and Terasvirta (1988a).
To discuss this idea, take the logistic transition function GT = G I - 1/2 and
its Taylor series approximation with the null hypothesis y = 0 as the expansion
point. The latter can be written as
Tl = 60 + 61% + M y , c ; 4 (1 5)
where R I is the remainder and 60 and 61 are constants. Substituting Tl for Gy in (13)
yields
yt = x:Bo + (xtd’B1 + U,* (16)

where U: = u t + (xi0)Rl ( y , c; s t ) and
is a ( p + 1) x 1 parameter vector. The zero-
and first-order terms emerging due to the substitution merge with the corresponding
terms in (13) leading to (16).This approximation may be viewed either as an approx-
imation to the STR model (the conditional mean of yc) or to the log-likelihood (14).
Use of this Taylor expansion in (13) or (14)amounts to giving up information about
the structure of the alternative in order to circumvent the identification problem and
obtain a simple test of the null hypothesis. Parameter vector B1 has the property
/31 = yB1; for details see Luukkonen, Saikkonen, and Terasvirta (1988a).Thus the
null hypothesis Ho: y = 0 in (13) implies Hh: = 0 and H i : # 0 within (16).
Since (16)is linear in parameters and U: = ut when HL holds, one can test this null
hypothesis by a straightforward LM-type test. Under Ho,
ECONOMIC RELATIONSHIPS 5 i 5
MODELING
where h ? ~= Et=, Et=,

T z c z i , fio1 = fiio = T ztwi,iiI1l= Et=,
T W C W ~ , = ~ ~
(l/T) cf, fi:, tit is the residual estimated under the null hypothesis, zt = xt and
wt = x t s t , has an asymptotic x2 distribution with p + 1 degrees of freedom when
the moments and cross-moments implied by (17) exist. For detailed derivations, see,
for example Luukkonen, Saikkonen, and Terasvirta (1988a, 1988b) or Granger and
Terasvirta (1993, Chap. 6).
The above notation requires that st is not an element of xt. If it is then the
auxiliary regression becomes
where Zt = ( x l t , . . . , xpt)’ and is a p x 1 vector. Furthermore, wt = Ztst in

(17) and the asymptotic null distribution of the test statistic thus has p degrees of
freedom. This is probably the most common case in practice except when st = t ,
and in the following the notation will conform to it,
The LM-type statistic (17) seems to have good power already in small sam-
ples (see for example Luukkonen, Saikkonen, and Terasvirta 1988a and Petruccelli
1990), but in a special occasion it only has trivial power against H i . This is the case
when st is an element of xt and 8 = (8o,O, . . . , 0)’, 80 # 0. In other words, the
only nonlinear element in (13)is the intercept. Then = 0 in (17) even under H;’
and the test thus has no power. To remedy this situation, Luukkonen, Saikkonen, and
Terasvirta (1988a) suggested a third-order Taylor approximation to GT. This can be
written as
where RS is a remainder and a,, j = 0, 1 , 2 , 3 , are constants. When (19) is substi-

tuted for GT in (13)one obtains
+
yt = x : s o + (ats;)’Bz + ( P t s Q ) ’ B 3 + U:
(atdB1 (20)
where U: = uc + (xiB)Rs(y,c ; s t ) and B j = y b j , j = 1 , 2 , 3 . The LM-type test
of HA: pj = 0, j = 1, 2, 3, against H i : “at least one Bj # 0” can be constructed
as before. The test statistic is (17) with wt = (Pis,,ZLs?, Pis:)’ and the number of
degrees of freedom in the asymptotic x2 distribution under HO is 3 p . This result
requires the existence of all the moments implied by w tand (17).
When xt has a large number of elements, the auxiliary null hypothesis will
sometimes be large compared to the sample size. In that case the asymptotic x2
distribution is likely to be a poor approximation to the actual small sample dis-
tribution. It has been found out (see Granger and Terasvirta 1993, Chap. 7) that
an F-approximation to (17) works much better (the empirical size of the test re-
mains close to the nominal size while power is good). The test can be carried out
in stages:
5 I6 TERASVIRTA
1. Re ress yt on xt and compute the residual sum of squares SSRo = (1/T)

F
Ct=16;.
2. Regress fit (or y t ) on x,,ztst, f , s : , and z t s ; and compute the residual sum
of squares S S R ~= ( 1 / ~ ) 6;.
3. Compute
Under Hb: 81 = 82 = 83 = 0, F has approximately an F-distribution

with 3 p and T - 4p - 1 degrees of freedom.
The above theory works when { s t } is stationary. It continues to work when st = t . In

that case, t is not an element of xc.The auxiliary regression when testing parameter
constancy of a linear model against STR with transition function (12) (first-order
Taylor approximation), is
(Lin and Terasvirta 1994).The F statistic corresponding to (21) thus has 3 ( p 1) +

and T - 4p - 4 degrees of freedom.
The parameter constancy test can also be carried out for any subset of param-
eters. In many modeling situations considering various subsets is advisable when
one wants to obtain a clear idea of which parameters may actually be nonconstant if
the model appears to have unstable parameters. This is done by assuming that the
appropriate elements 8; = 0 in (13) in which case the corresponding elements of
p;,j = 1 , 2 , 3 , also equal zero a priori and are not included in the null hypothe-
sis. The empirical example in Section V will illustrate the possibilities of this ap-
proach. The same also applies to testing linearity against STR. In some occasions,
economic theory behind the model or statistical considerations may suggest such a
priori parameter restrictions. A dummy variable with all values equal to zero except
one constitutes an example of a variable whose coefficient must always be assumed
constant.
The above theory covers the LSTRl model, whereas testing linearity against
the LSTR2 model has not been discussed separately. The first-order Taylor approx-
imation of GZ at the expansion point y = 0 eventually leads to the auxiliary re-
gression (20) with 8 3 = 0. Thus the null hypothesis y = 0 is transformed into Hh:
= 82 = 0. This indicates that the LM-type test of Hh: 81 = 82 = 8 3 = 0 within
(20)also has power against the LSTR2 model. If the alternative to the linear model is
an ESTR model defined by (10)the auxiliary regression based on the first-order Tay-
lor approximation to the transition function is identical to that of the LSTRZ model.
A similar remark holds for (22). Note in particular, that (22) can also be regarded as
a first-order Taylor approximation to (13)with (12).
Sometimes it may not be clear from economic theory which variable should be
taken to be the transition variable under the alternative. Suppose, however, that the
ECONOMIC RELATIONSHIPS 5 I 7
MODELING
choice is between the elements of Z l . Then one can define the linear combination
a‘Zl where a = (0, . . . , 1 , 0 , 0 , . . . , 0)’ is a p x 1 vector with the only unit element
corresponding to the true but unknown transition variable and substitute it for st in
(8). Proceeding as in Luukkonen, Saikkonen, and Terasvirta (1988a) leads to the
auxiliary regression
P P P P
P P
with the linearity hypothesis HA: Bli; = 0, i = 1 , 2 , . . . , p ; j = i, i + 1, , . . ,p ;

B 2 i j = p 3 ~ ;= 0, i, j = 1, 2, . . . , p . The LM-type test statistic is formed accordingly
and the number of degrees of freedom in the asymptotic x2 distribution under H(,
is p ( p+ +
1)/2 2 p 2 . This quickly grows large with increasing p and the authors
therefore suggested an “economy version” based on
P P P
The number of degrees of freedom in the test based on (24) is p ( p + 1)/2 +p,
the null hypothesis of linearity being HA: B l i j = 0, i = 1 , 2 , . . . , j ; j = i, i +
1 , .. . , p ; /33j = 0, j = 1 , 2 , . . . , p . Of course, the choice of potential transition
variables may be restricted to only a subset of variables in z l . In that case, the null
hypothesis must be modified accordingly and the relevant coefficients in (23)or (24)
set equal to zero a priori.
The above statistical theory has the advantage that the asymptotic null dis-
tributions are standard and the tests can be carried out just by using ordinary least
squares. Although they are designed against STR they are also sensitive to other
types of nonlinearity. After rejecting linearity it may therefore not be clear what to
do next. However, one may have decided to consider STR models if linearity is re-
jected. In that case, tests based on the above auxiliary regression may be used only
for testing linearity but also, in case of a rejection, for the specification of an STR
model. This argument is elaborated in Section 1V.B.
The LM-type test statistics continue to have reasonable power in small sam-
ples when y + 00, at least when the alternative is an LSTRl model; see Luukkonen,
Saikkonen, and Terasvirta (1988a) and Hansen (1996). But then, if the alternative is
a switching regression model (it is assumed a priori that y is infinite) the above theory
does not work. Hansen (1996) recently considered a general framework for hypoth-
esis testing when the model is only identified under the alternative. His results have
bearing for switching regression models as well. Let v E N be the vector of nuisance
5 I8 TE~SVIRTA
parameters and S T ( U )a test statistic which is a function of the nuisance parame-

ters. Hansen showed how to obtain critical values for statistics such as supuENS T ( V )
(Davies 1977, 1987) or aveuENST(u) = S T ( U )dW(u) (Andrews and Ploberger
1994) by simulation. Hansen’s technique is not restricted to the case where there
only exists a single nuisance parameter, but the amount of computations required to
perform any of the possible tests increases with the number of nuisance parameters.
He applied the test to testing linearity against a two-regime threshold autoregres-
sive model with an unknown transition variable (delay). In general, the supremum
and average tests should be preferred to the LM-type tests discussed above if one
is able to exclude smooth transition as an alternative to linearity in favor of a dis-
crete switch. This is the case, for example, if one is investigating structural change
and knows beforehand that the only alternative to constant parameters during the
observation period must be a single structural break.
B. Misspecification Testing in STR Models

In the previous section the focus was on testing linearity against STR. This section
deals with the situation in which the parameters of the STR model have been es-
timated and the validity of the assumptions of the model is checked in the light of
the results. The misspecification tests to be discussed here are those Eitrheim and
Terasvirta (1996) recently derived in a univariate setting. As the authors remarked,
generalizing them to STR models is straightforward. An assumption one has to make
is that the parameters of the STR model have been estimated consistently and that
the estimates are asymptotically normal; see Wooldridge (1994) and Escribano and
Mira (1995)for discussions of conditions for this.
Estimation of an STR model is carried out, among other things, under the as-
sumption of no error autocorrelation and that of parameter constancy. These two as-
sumptions thus have to be tested, and procedures to that effect are discussed here.
Furthermore, it is of interest to try and find out whether or not the estimated STR
model captures all nonlinear features present in the data. Statistical inference for
treating this problem is another topic of this section. To retain the same order of
presentation as in Section 1II.A it is considered before tests of parameter stability.
I . Test of No Error Autocorrelation

Consider first the test of no error autocorrelation in the STR model (assume for sim-
plicity that the transition variable s1 is an element of Z t ) . More generally,
yt = W x 1 ; @) + ut
u1 = a’vt + Et ( t = 1, . . . , T ) (25)
where M is at least twice continuously differentiable with respect to the parame-
-
ters, a = ( a l , . . . , aq)’ is a parameter vector, vt = ( u l - l , , . . , ut-q)’, and cl
IN(0,d).
ECONOMIC RELATIONSHIPS 5 I9
MODELING
The null hypothesis is Ho: a = 0, and H I : a # 0. The conditional log-

likelihood, given the fixed starting values yo, y-1, . . . , ~ - ~ + and
l xg, x-1, . . . , x P q + l ,
has the form
L=c--ha
2 202
t=l
The information matrix related to (26) is block diagonal such that the element cor-
responding to the second derivative of (26) with respect to o2forms its own block.
The variance a2 can thus be treated as a fixed constant in (26) when deriving the
test statistic. The first partial derivatives of the log-likelihood with respect to a and
II/ are
Furthermore, in (28)
and
The exponential terms in (29) and (30) are bounded for y < 00. It follows that the
existence of the necessary fourth moments is required for consistent estimation of
parameters.
520 TERASVIRTA
The test statistic is (17) with wt = at = ( a t - l , . . . , fit-y)’

and tt = tt =
8 M ( x t ; $)/a+, and its asymptotic null distribution under the null hypothesis is x2
with q degrees of freedom. The indicates consistent estimates under No. The
approximate F-version of the test may be computed in three stages (Eitrheim and
Terasvirta 1996):
1. Estimate the LSTRl model by NLS under the assumption of uncorrelated
errors and compute the residual sum of squares SSRo = CtE1
T
fi:.
2. Regress f i t on and tt and compute the residual sum of squares, SSRl .
3. Compute the test statistic FI,M= {(SSRo - SSR1)/q}/{SSRI/(T - n -
q ) } ,where n is the dimension of the gradient vector Z t .
In small samples, the F-statistic with q and T -n-q degrees of freedom is preferable
to the asymptotic x2 variant because its superior size properties. To carry out the
corresponding test when G = G2 (LSTR2 model) requires an obvious modification
of zt and the assumption of the existence of the appropriate sixth moments in (17).
A computational detail is worth mentioning here. The auxiliary regression in
Stage 2 contains q lags of f i t . A standard way of proceeding in such a situation is to
trim all the time series by omitting the first q observations in them. However, if this
is done, EL,,, f i t & # 0 so that the trimming affects the empirical size of the test.
In small samples the effect may not be negligible. A better solution is not to shorten
the series but rather replace the missing observations in the beginning of the series
of lagged residuals by zeros. Also, if the STR model in question happens to be a very
difficult one to estimate it may be that the NLS estimation algorithm is not able to do
a perfect job; that is,E;‘=, fit.& # 0. To remedy that, step 1 may be replaced by
1’. Estimate the LSTRl model by NLS under the assumption of uncorrelated
errors. Regress the residuals fit on 2, and compute the residual sum of
squares SSRo = T U,cl=, A*2
.
Step 1guarantees CiEl
T
ci,*tt= 0 and prevents size distortion of the test. This
extension of step 1can also be recommended for the other two tests discussed in this
section.
2. Test of N o Remaining Nonlinearity
Since a purpose of the STR model is to give an adequate characterization to the

nonlinear features in the data it is of particular interest to find out how successful the
estimated model is in this respect. In order to test the adequacy of the STR model,
one may extend the model in some suitable way. Eitrheim and Terasvirta (1996)
suggested such an extension for STAR models. Following their approach, define the
additive STR model as
MODELINGECONOMICRELATIONSHIPS 52 I
This STR model has two additive nonlinear components, and the transition func-
tion H where rt is assumed an element of x L may be defined analogously to (8)or
(9). Since testing the presence of the additional component is discussed, assume
H ( 0 , c z , r t ) = 0 for notational simplicity. When adequacy of the standard STR
model is the issue it can be investigated by testing Ho: y2 = 0 in (31). Because
of the parameterization, the additive STR model (31) is not identified under this null
hypothesis. This problem may be solved (Eitrheim and Terasvirta 1996) in the same
way as in Section 1II.A. This means that the transition function H is replaced by its
third-order Taylor approximation
T3(y2, c2; rt) = 60 + 61rt + 6 2 r ~+~63r: + R 3 ( Y 2 , c2; rt)
in (31). Doing that and rearranging terms yields
+
where /3j = ypp;, j = 1 , 2 , 3 , and U ; = ul (x:$)Rs(y2, cp; r t ) .The null hypoth-
esis is H;): 81 = 8 2 = 8 3 = 0 and when it holds, U : = u l . Deriving the appropriate
test statistic with the asymptotic x2 null distribution is straightforward. In practice,
the test can be carried out in the three stages described above. In the present case,
8, = (%irt,Zir:, 2;r-f)’ in stage 2, whereas & is the same as before. The degrees of
freedom in the F-statistic are 3p and T - 4p - 1, respectively. If G zs 0 in (32), the
test collapses into the linearity test discussed in Section II1.A.
In the above it is assumed that all elements of $ in (30) are nonzero. This is
not necessary, and the elements of pL included in the second nonlinear component
+
may be selected freely. On the other hand, write z i p = x’llp~ xk1p2 in (13) and
consider the case in which the parameters have been estimated under the restriction
p 2 = 0. Then (32) can be written as
and the null hypothesis of no remaining nonlinearity HA: 802 = 0, 81 = B p =

83 = 0. One can also separately test II((: 802 = O in (33) assuming = 82 =
83 = 0. This is simply an LM test of validity of the restrictions 9 2 = 0 in (13).
Furthermore, one may not want to be specific about the transition variable in the
second nonlinear component. In that case one can generalize the relevant linearity
tests to this situation as discussed in the previous section. Such a generalization
based on the Taylor expansion of the appropriately parameterized transition function
is straightforward.
Sometimes a model builder may not want to be specific about the parametric
form of the remaining nonlinearity except that it is of additive type. In that case the
maintained model can be written as
yl = pfxl + w’xL)w,
c; + ax,)
+ sl) Ut (34)
522 TERXSVIRTA
If the STR does not adequately characterize the nonlinearity in the data then K ( x , )
is a nonlinear function. To investigate that possibility assume that K ( x , ) is at least
three times continuously differentiable with respect to x, and expand K ( x , ) into a
Taylor series about the expansion point xt = xe. The third-order expansion has the
form
where R K ( ~ ,is) the remainder term. Approximating K ( x , ) in (34) by (35)yields

P P
+
where U; = ut R K ( ~ ,The
.
) . null hypothesis of no remaining nonlinearity is HA:
.
Kij = 0 , i = 1 , . . . , p ; J = 1 , . . . , P ; K i ; e = 0 , i = 1 , . . . , p ; J = i, . . . , P ; l =
j , . . . , p . The test can be carried out as before as an F-test if all the necessary mo-
ments (sixth) for x, exist. Because the maintained model is very general, the null
hypothesis is large. As a result, the test is likely not to have very good power in small
samples if p is not small. Note that if p = 1, the test is equivalent to the corre-
sponding test against STR based on the auxiliary regression (33) with r, = x l t and
8 3 = 0.
3. Parameter Constancy
Parameter constancy is one of the key assumptions of an STR model. Testing it is
therefore as important as it is in linear models. In this chapter the alternative to
parameter constancy is a set of smoothly changing parameters. Following Eitrheim
and Terasvirta (1996), the definition of parameter change is based on the idea of
smooth transition, and the developments in this section are just a generalization of
results in Lin and Terasvirta (1994).Consider the STR model
where zf contains those po 5 p +

1 elements of xt whose coefficients are not as-
sumed zero a priori, and the p1 x 1 vector x:, p1 5 p +
1, is defined in the same
way for the nonlinear part of the model. G is defined by (8)or (9).Let qo(t)= qo +
h l H ( t ; y1, c1) where qoand hl are po x 1 vectors, andOO(t) = 0° + h2H (y l, c l ; s,)
where Oo and A2 both are p l x 1. As in Section III.A, let H be either
MODELING
Hz(y1, c l ; S A = (1 + exp{-yl(t - Cll)(t - c12>1)-’ - i1 (38)

1
H3(y1, c l ; S t ) = (1 + exp{-yl ( t - c l l ) ( t - c12)(t - c13)])-’ - i (39)
where y1 > 0 and c l l 5 c12 5 c13. These definitions accord with those in Jansen
and Terasvirta (1996). Using them, the null hypothesis of parameter constancy in
(36) is Ho: y1 = 0. Note that the parameters in the exponent of G in (36) are as-
sumed constant a priori. The three transition functions H,, j = 1 , 2 , 3 ,allow a lot of
flexibility in characterizing parameter instability. For instance, choosing H I and al-
lowing y1 + 00 yields a single structural break. Doing the same in H2 if c11 < c12
yields two structural breaks: the second one restores the original parametrization.
Transition function H3 is the most general one and allows nonsymmetric and non-
monotonic parameter nonconstancy.
To discuss testing parameter constancy against qo(t)and @ ( t ) in (36),let the
transition function be H 3 . The other two functions, H I and H 2 , are special cases of H 3
and need not be discussed separately. The null hypothesis of parameter constancy is
Ho: y1 = 0 which is tested against H I : yl > 0. The parameterizations of q O ( t )and
O o ( t ) have the property that (39) is unidentified when y1 = 0. This by now familiar
complication which precludes the use of standard tests is dealt with in the same way
as before. In order to construct a viable test, approximate H3 by its first-order Taylor
expansion about y1 = 0. This yields
H3(yI,c1; t ) = 6 0 + 61(t - c l l ) ( t - c12)(t - c13) + R3()/1,CI;t )
= 6;; + s;t + 6;t’ + 6 ; s + R3(y,, c1; t )
= M y 1 7 c1; t ) + M y l , c1; t ) (40)

Substituting 2‘3(y1,cl; t ) for (39) in (36) and rearranging terms leads to the
approximation
y, = ( X y p o + (&’81 + ( x l ) t 2 ) / B 2+ (xl)t3)‘B3
+ {(&’BLt + (&’Bs + ( x f t 2 > ’ B 6 + (x,‘t3>’B7}G(v,c; + st> U: (41)
wherep, = y l p j , j = 1 , 2 , 3 ,5 , 6 , 7,andu; = u l + { ( x ~ ) ’ q O + ( x f ) ’ O o G ( y l ,c ; s t ) }
R ~ ( Y I C tI );. The null hypothesis of parameter constancy thus becomes Hb: =
82 = 8 3 = 0, ps = 86 = 8 7 = 0. Since aG/ay = g,(t) and aG/ac = g,(t) are
bounded everywhere, the results in Lin and Terasvirta (1994) generalize to this sit-
uation (Eitrheim and Terasvirta 1996).From this fact it follows that the test statistic
(17) has an asymptotic x2 distribution under HA; in this case
2, = r c x y , (x$(P, 2; S,))’? g y w , gm)’
524 TERASVIRTA
and
Again, an F-version of the test is preferred to the asymptotic theory. The degrees of
+ +
freedom of the F-statistic are 3(po p l ) and T - 4 ( p o p l ) , respectively.
If the alternative to parameter constancy is characterized by H2 then 8 3 = 0
and 87 = 0 in (41), and the null hypothesis is reduced accordingly. If only monotonic
change is considered to be the alternative to stable parameters, then the transition
function is the logistic function (37)and one has 8 2 = 8 3 = 0 and 86 = 8 7 = 0 in
(41).Even this test may be restricted to cover only certain parameters of the model
while the remaining ones are constant. This is done by rewriting (36)as
in obvious notation. The maintained nonconstancy is restricted to the coefficients of

xp2 and x:2. Carrying out tests corresponding to appropriate block divisions of xp
and x: one can gain information about which parameters in (36)may be nonconstant
if not all of them are. The empirical example in Section V illustrates the usefulness
of testing the stability of subsets of parameters in STR models; see also, for example,
Lutkepohl, Terasvirta, and Wolters (1995)and Wolters, Terasvirta, and Lutkepohl
(1997).
Finally, the LM test of no autoregressive conditional heteroskedasticity (Engle
1982,McLeod and Li 1983)in the error process of an STR model can be carried out
in the standard fashion exactly as in linear models. The same is true for the Lomnicki-
Jarque-Bera test (Lomnicki 1961,Jarque and Bera 1980)of normality of the errors.
4. Additional Remarks
The tests discussed in this and the preceding subsection as well as the linearity tests
of Section 1II.A have been designed against parametric alternatives. There exists a
wide selection of other tests that are used in connection with nonlinear modeling and
testing for structural change; see for instance Granger and Terasvirta (1993,Chap.
6)for an account. Many of those tests do not have a specific alternative hypothesis
although some of them, such as RESET (Ramsey 1969)may be interpreted as LM
tests of linearity against a parametric nonlinear alternative. They have rather been
intended as general tests of either linearity or parameter constancy in linear models.
Nonparametric tests surveyed in Tjflstheim (1994)form a large subset of this class
of tests. Because the focus in this chapter is on econometric modeling with STR
models these other tests have not received the attention that they otherwise would
deserve.
MODELINGECONOMIC RELATIONSHIPS525
C. Testing the Granger Noncausality Hypothesis

Testing the null hypothesis of Granger noncausality between economic variables is
normally carried out in the linear framework; for the original causality definition,
see Granger (1969), and for a survey of the literature see Geweke (1984). The STR
model offers a possibility to do that in a parametric nonlinear framework. For a non-
parametric test, see, for example, Bell, Kay, and Malley (1996). This is a potentially
useful extension, because the functional form assumed for the test may affect the
outcome as Hendry (1995, p. 176) argued. It is also a straightforward one because
the bivariate test is just another variant of the test of no additive nonlinearity dis-
cussed in Section III.B.2. Following Skalin and Terasvirta (1996), who considered
this extension, modify (31) slightly such that
where w t = (1, ~ ~ - 1. ., . , yt-,,)’ and u1 = ( ~ ~ - 1 .,. . , x ~ - ~ )and

’ 6 and 11/. being two
q x 1 parameter vectors, q 2 1. Testing the hypothesis that the stationary variable
x t does not cause the other stationary variable yl is equivalent to testing Ho: 6 = 0
and y2 = 0 in (43). Note that here it cannot be assumed that the delay e in H is
known. Luukkonen, Saikkonen, and Terasvirta (1988a) discussed this situation, and
an LM-type test of the present null hypothesis may be obtained following their ideas;
see Skalin and Terasvirta (1996)for details. The restriction that the contribution of
lagged x to the predictability of y be of STR-type may seem a strong one for some. It
is possible, however, to adopt a general functional form for x as in (34) and proceed
from there. A drawback of that approach is that the dimension of the null hypoth-
esis increases quickly with q . If it is assumed a priori that y1 = y2 = 0 the test
collapses into the customary linear bivariate single-equation test. The application in
Skalin and Terasvirta (1996)shows that the results from the STR-based test and the
bivariate linear test applied to the same data set may indeed be very different.
IV. T H E MODELING CYCLE

A. Introduction
When STR models form an alternative to linear models in econometric modeling
there are several practical questions to be answered before one can fit any STR model
to data and more to come after the estimation. First, as mentioned, one has to find
out whether or not a linear model provides an adequate description of data. If it does,
nonlinear models are not needed. Second, if nonlinearity seems to be present in the
data, economic theory may not be explicit about the parametric form of the STR
model, and the dynamic structure of the model may not be completely specified a
priori either. For instance, it may not be obvious which of the independent variables
526 TERASVIRTA
should be the transition variable. Third, after estimating an STR model, the ade-
quacy of the model has to be checked. In order to deal with these issues, Terasvirta
(1994) proposed a modeling cycle for univariate STAR models consisting of specifi-
cation, estimation, and evaluation stages. This was an application of ideas in Box and
Jenkins (1970), who developed such an approach to constructing ARIMA models.
Granger and Terasvirta (1993, Chap. 7) extended the STAR modeling cycle to STR
models. This cycle is the topic of the present section. The misspecification tests of
estimated STR models introduced in Eitrheim and Terasvirta (1996)and discussed
in the previous section have not become available until recently and are thus new
compared to previous presentations of the STAR or STR modeling cycle. The three
main stages of the cycle will be considered separately. The use of the encompassing
principle to compare an STR model with its rivals explaining the same phenomenon
is another addition not discussed before.
B. Specification of STR Models

The specification of STR models consists of finding answers to the following ques-
tion: (i) Is a linear model adequate? (ii) If it is not and STR models are considered as
an alternative, which variable should be selected to be the transition variable? (iii)
Should one choose an LSTR1, LSTRZ or possibly an ESTR model? To answer the
first question, linearity is tested against STR assuming each of the potential transi-
tion variables in turn to be the transition variable in the test. If linearity is rejected
against an STR with the transition variable xjt, this alternative is tentatively ac-
cepted. Otherwise linearity is accepted in which case no further nonlinear modeling
is necessary. If linearity is rejected against STR for more than one transition vari-
able, one selects the variable giving the strongest rejection (lowest p-value) to be
the transition variable. Terasvirta (1 994) provided a heuristic justification for this
decision rule. In STR modeling, the linearity testing thus has a dual purpose in the
sense that if linearity is rejected the test results are also used for finding the right
transition variable for the STR model. A false rejection of linearity is likely to be
discovered at some later stage of the modeling cycle anyway. It may be pointed out
that by performing several individual tests one is not in control of the overall sig-
nificance level of the linearity test. This is not crucial if the main purpose of testing
is to help model building. On the other hand, the control of the overall significance
level is important if the main purpose is to test an economic theory. In that case one
should use such a variant of the test that does not assume that the transition variable
is known a priori. These have been discussed in Section 1II.A.
After making a decision about the transition variable the next step is to choose
the type of the model. The decision rule can be based on the auxiliary regression (20)
with the appropriate transition variable as follows. Define the following sequence of
null hypotheses within (20):
MODELING
carry out the tests and apply the following decision rule. If the reject’on of Ho3 S
the strongest one, choose an LSTR2 model (or an ESTR) model, otherwise select an
LSTRl model. The coefficient vectors p j , j = 1, 2 , 3 , are functions of the parameters
of the original STR model and they depend on the type of the model. The selection
rule is based on this fact; for details see Granger and Terasvirta (1993, Chap. 7) or
Terasvirta (1994).
There exists another, computationally slightly heavier but still very practica-
ble strategy. Consider the original STR model (6).Giving fixed values to the parame-
ters in the transition function makes (6)linear in parameters. These parameters can
be estimated by OLS. Construct a two-dimensional (LSTRl) and three-dimensional
(LSTRZ) grid of y and c and estimate the other parameters for these combinations
of y and c . In order to be able to choose a meaningful set of values of y , the expo-
nent of the transition function should be standardized; see the discussion in the next
subsection. A reasonable set of values of c may be selected between the observed
minimum and maximum values of the transition variable. Estimate the models for
both LSTRl and LSTR2 and select between these alternatives after comparing the
fit of the best-fitting LSTRl and LSTRZ models. This procedure can also be used to
obtain initial estimates for the NLS estimation and to reduce the size of the model
by imposing exclusion restrictions. An illustration can be found in the empirical ex-
ample of Section V. Finally, the choice between LSTR2 and ESTR can be made after
estimating an LSTR2 model by testing c1 = c2 within that model.
C. Estimation of STR Models

Estimation of the STR model (6) with (7), (8), or (9) is carried out by NLS which
is equivalent to the maximum likelihood estimation in the case of normal errors.
Wooldridge (1994)and Escribano and Mira (1995)recently discussed conditions for
obtaining consistent and asymptotically normal estimates. The grid estimation men-
tioned above can be used to obtain sensible initial values. Hendry (1995, Appendix
5.5) contains a useful updated overview of numerical optimization techniques for
maximizing the likelihood; for other accounts see, for instance, Judge et al. (1985,
Appendix B) and Quandt (1984). One should mention that there are sometimes nu-
merical problems in the estimation of LSTRl models and that they are related to
estimating the slope parameter y of the transition function. First, y is not a scale-
free parameter as its value depends on the magnitude of the values of the transition
variable s l . In order to reduce this dependence it is advisable to standardize the
exponent of the transition function by dividing it by the sample standard deviation
528 TERXSVIRTA
(LSTR1 model) or the sample variance (LSTR2 model) of s t . This standardization

may bring y close to the other parameters of the model in magnitude unless y is very
large, which is numerically an advantage in the estimation. It also makes it easier to
find sensible initial estimates for this parameter.
Another problem is that if y is large the STR model is very close to a switching
regression model. This makes the estimation of y difficult in small samples because
accurate estimation requires a sufficient number of observations of the transition
variable st in a small neighborhood of c (LSTRl model) or c1 and c2 (LSTR2 model).
Because it is unlikely in small samples that there exists a large cluster of obser-
vations of st sufficiently close to these parameters, estimating y with satisfactory
precision is a problem; see, for example, Bates and Watts (1988, p. 87), Seber and
Wild (1989, pp. 480-481) and Terasvirta (1994).This in turn may cause problems in
carrying out the tests discussed in Section I1I.B. The partial derivatives (29) and(30)
reflect this difficulty. If y and then most likely also the NLS estimate p are large
and there exist no observations of st that are quite close but not extremely close to
c, (29) will be practically zero for all t . Thus the moment matrix of any of the test
statistics in Section 1II.B will be near-singular. In (30), observations very close to c
(LSTRl) will cause a blip in the series whose observations otherwise are practically
zero. If none of the observations is sufficiently close to zero then the whole vector will
be practically zero and the moment matrix of the test statistic again near-singular.
Rescaling (29) and (30)does not help. A feasible solution is to omit these variables
from 2t when carrying out the test so that the results are numerically stable. This
hardly affects the results one would obtain if one had enough precision to compute
them. When the omission is accompanied by stage 1’ in the testing sequence any
potential size distortion (usually negligible) due to the omission is eliminated.
The estimation of the STR models in the example of Section V is carried out
by applying a variant with numerical derivatives of the Broyden-Fletcher-Goldfarb-
Shanno (BFGS) algorithm as implemented in the OPTMUM routine of GAUSS 3.1.
While the exponent of the transition function is standardized as discussed above, it
will be seen that the size of y does not cause any difficulties in that example.
D. Evaluation of STR Models

After estimating the parameters of an STR model it is necessary to test the basic as-
sumptions underlying the estimation. Tests of no error autocorrelation, no remaining
nonlinearity and parameter constancy have been considered in Section 1II.B. There
are also other, informal ways of checking the adequacy of the model. For instance,
inflated standard errors for the parameter estimates of cp; and 8; at certain lags are
usually an indication of overspecification. As a remedy, one should try imposing one
of the restrictions cp; = 0,8; = 0 and cpi = -8i. Which one to choose is most
often an empirical matter and is best settled by imposing each restriction in turn,
reestimating the model and comparing the results. An estimated E (LSTRl), or 21 or
MODELING
22 (LSTR2), far outside the observed range of the transition variable is often a sign
of convergence to an infeasible local minimum and thus an indication of an inade-
quate model. In that case, however, the model usually also fails at least some of the
misspecification tests discussed above. However, in the case of an LSTR2 model, if
either i.1 or 22 lies far outside the observed range of the transition variable while p
is not small, it may also indicate that an LSTRl model is a better choice than an
LSTR2 one.
If the model fails the test of no error autocorrelation, respecification seems the
only feasible solution. This is probably the most common route to follow also when
the model badly fails the tests of no additional nonlinearity. When the STR model
does not have constant parameters, respecification is an obvious solution as well.
But there exists at least one special case in which one might actually want to have
another STR component to accommodate and parameterize such nonconstancy. This
is when the model contains seasonal dummies and their coefficients seem to change
over time. There may not be economic reasons for such a change but seasonality
may change slowly anyway because of changing institutions. This is not an uncom-
mon situation in macroeconometric models. It has not been frequently accounted for
in practice, perhaps because the econometricians have generally seen a single struc-
tural break as the most interesting alternative to parameter constancy. The paper by
Farley, Hinich, and McGuire (1975) was an early exception to this rule. The STR
framework offers a possibility of parameterizing such a continuous change; see the
example of the next section. For other macroeconometric applications of this idea,
see for example Jansen and Terasvirta (1996)and Lutkepohl, Terasvirta, and Wolters
(1995). Gradual changes in institutions affecting other things than seasonality may
also be modelled using the STR approach as in Heinesen (1996).
E. Encompassing Other Models

STR models may be constructed in areas where there already exist rival models esti-
mated for the same time period and explaining the same economic phenomenon. In
such a case, even when the estimated STR model passes the above misspecification
tests it is useful to find out whether the model is an improvement over the previous
quantitative explanations or not. This is the case if the STR model explains the re-
sults obtained by the rival models while the converse is not true. The STR model is
then said to encompass its competitors; see, for example, Hendry (1995, Chap. 14)
for a formal definition and a thorough discussion of this concept.
The encompassing property may be investigated by statistical tests. Assume
for simplicity that there exists a single rival model to the STR one and that it is a
linear single-equation model. Then it may be possible to construct a minimal nest-
ing model (MNM) nesting the two competitors. This is done by extending the STR
model by an additive linear component that contains those variables in the rival
model that do not linearly enter the STR model. The MNM is thus an STR model. If
530 TE~SVIRTA
the rival model is also an STR model then the MNM may be an additive STR model.
The M N M trivially encompasses the two original models because they are nested
in it; see Hendry (1995, p. 511). If an MNM can be constructed, one can apply a
simplification encompassing test (Mizon and Richard 1986)to see if the STR model
parsimoniously encompasses the MNM. If the rival model is linear then the test con-
sists of testing the null hypothesis that the corresponding linear component does not
additively enter the nonlinear MNM model. Such a test was discussed in Section
III.B.2 when the idea was to test exclusion restrictions on coefficients of the linear
component of an STR model. Accepting this hypothesis is equivalent to accepting
that the STR model encompasses the MNM. Finding out if the linear model parsimo-
niously encompasses the MNM is tantamount to testing linearity within an M N M of
STR type. This test has to be carried out using the techniques discussed in Section
1II.A because an STR type MNM is not identified under the null hypothesis. Suppose
that this is done and the null hypothesis rejected. Then the conclusion is that the lin-
ear model does not encompass the MNM. If the STR model does, then the transitivity
property of encompassing implies that the STR model encompasses its linear rival.
The use of this testing procedure requires that both the STR equation and its single-
equation rival are valid models in that they can be analyzed without knowledge of the
rest of the system. An example of simplification encompassing tests can be found in
Section V.
V. APPLICATION
A. Background
This section contains an example of the modeling cycle consisting of specification,
estimation and evaluation of STR models. It is based on the data set and results in
Hendry (1984), who modeled house prices in the United Kingdom in 1960-1981
using error-correction models. For another econometric analysis of this data set, see
Richard and Zhang (1996).The purpose here is not to present a new model for UK
house prices: in order to do that the first thing would be to extend the time series
as close to the present time as possible. The main objective is instead to use the
equation for house price expectations Hendry (1984) specified and estimated as a
benchmark and see if the STR approach based on that equation and the same ob-
servation period leads to any new insight or yields an improved specification. This
provides an opportunity to show how the: STR modeling strategy works in practice.
The period Hendry (1984)considered was eventful. The nominal house prices
increased 12-fold and the real prices by over 50%. The nominal prices were also
clearly more volatile than the ordinary retail price index. These and other features of
the observed time series are discussed in Hendry (1984). The paper also contains a
description of the way the housing market functioned in the United Kingdom during
MODELINGECONOMIC RELATIONSHIPS 53 I
the observation period. The theoretical model in the paper is based on the assumption
that the housing stock H , evolves intertemporally according to
where 6, is the depreciation rate and C, denotes net additions. At any time, C, is
very small relative to H,-l so that H,-1 is taken to be the fixed supply of housing in
the short run. As a result, the fluctuations in demand translate into fluctuations in
the price of housing, Ph,. Thus (see Hendry 1984, pp. 224-225 for a more complete
argument), the demand equation is the one determining the price of housing. Its
general form is postulated as
H D = f(Ph/P, Y , p , R ,R m , M , T , N , F )
- + - - - + Y + ?
where P is the general price level, Y the real income, p the real rental rate, R the
market interest rate, and Rrn is the mortgage rate of interest, which for institutional
reasons may differ substantially from the market interest rate. Furthermore, M is the
stock of mortgages, T the tax rate, N the size of the population, and F the average
family size. Changes in the real price of housing, the real rental rate and the interest
rates have a negative effect on demand. Changes in the real income, the stock of
mortgage, and the size of the population have the opposite effect. The tax rate and the
average family size also affect the demand for housing, but the sign is indeterminate.
In the following the focus will be solely on modeling price expectations which
Hendry (1984)discussed in detail. The expectations are assumed unbiased:
where A is the difference operator, ph, the logarithmic nominal price of housing (low-
ercase letters denote logarithms), ph; the corresponding expectation, and uLan inno-
vation with respect to the information used in predicting Aph,: E u , = 0, var(u,) =
o:,cov(u,, U,) = 0, .r # t . Equation (11) in Hendry (1984, p. 228) gives the expec-
tations the following parametric form:
2
Aphr = CjAphL-1 - q ( R " - A4p)L-I + cq.Am,-~
i= 1
+
- ~ 5 ( p h h - p - y - C,L),--I + c6(m - ph - h - cb)i-l
where RP is the after-tax interest rate, c,, and C b are constants, and c , , i = 1 , . . . , 6 ,
are unknown parameters. The expectation equation contains two error correction
terms. The first one requires the nominal value of housing to stand in constant pro-
portion to nominal income in the long run. The second error correction term implies
+
a constant long-run ratio of mortgage ( m )to own equity (ph h); H is the housing
stock.
532 TERASVIRTA
The estimated counterpart of (45) considered here is somewhat different from

(45) itself. It is Eq. (18) in Hendry (1984, p. 237) and contains more short-run dy-
namics than (45). The equation has been estimated for the period 1959(1)-1982(2)
and has the following estimated form*
Apht = 0.23 Apht-2 + 13.1( A p h t - , ) 3+ 0.51 A2(Ayt)
(0.085) (4.3) (0.13)
+ 0.17 ( m - p h - h)t-l + 0.43 (y -h)t-1
(0.028) (0.072)
4- 0.82 F13(p) + 0.53 F 1 3 ( n - p ) - 0.21 Re-3
(0.14) (0.13) (0.10)
- 0.51 ARY-l - 0.53 - 0.0006 Qlt + 0.025 Q2t

(0.20) , (0.12) (0.0057) (0.0049)
+ 0.019 Q 3 t - 3.7 D:L - 2.2 D;t +at (46)

(0.0047) (1.5) (0.63)
T = 94, AZC = -3.80, R2 = 0 . 7 8 , 6 = 0.0143, b2/&;= 0.94, sk = 1.3, ek =
4.8, LJB = 117(4 x 10-26)
where the figures in parentheses below the coefficient estimates are estimated stan-
dard deviations, 6’ = (2‘ - k ) - * fif is the residual variance (k is the number
of estimated regression coefficients), a;, is the residual variance of the correspond-
ing linear model (Apht-l not cubed), sk is skewness and ek excess kurtosis of the
residuals, T is the number of observations, R2 is the coefficient of determination,
LJB is the Lomnicki-Jarque-Bera normality test and the value in parentheses is the
p-value of the test statistic. The small p-value is due to a few large residuals. In
(4% A d 4 = (1/3)(3% +2+1 +x,-zL Fl3(4 = A(%-1 + x t - d , h ( m - p ) =
A{(m - p)t-r +( m - p ) t - 3 ) , x : - 3 = (1/2)(R:-;3 +RI)-4), Q i t , i = 1 , 2 , 3 , are
the three seasonal dummy variables, and DyI and D:t are two dummy variables dis-
cussed in Hendry (1984, pp. 241-242). The errors of (46) are not autocorrelated,
see Table 1. Tests of no autoregressive conditional heteroskedasticity (Engle 1982)
in Table 1 do not indicate any problems either.
Equation (46) is a result of a specification search. Ratio “real value of housing
to income” in (45) has become “real income to stock of housing.” The lag structure
has been simplified by omitting lags and imposing coefficient restrictions defined by
A 2 , F13, and E. An interesting feature is the cubic price difference ( A p h t - 1 ) 3whose
*The estimated equation is not exactly the same as Eq. (18)in Hendry (1984).The differences are due to
the fact that the data set Hendry (19%)used was no longer available in its original form. The data set
used in this chapter 15 as close to the original data a\ possible.
MODELING
Table I p-Values of the LM Test of No Error Autocorrelation against an AR(q)

and MA(q) Error Process and the LM Test of No Autoregressive Conditional
Heteroskedasticity against ARCH(q) in model (46)
Maximum lag q
Test 1 2 3 4 6
~~~~~~
No error autocorrelation 0.55 0.83 0.75 0.88 0.38

No ARCH 0.99 0.99 1 .oo 1.00 0.99
coefficient estimate is significant and which Hendry (1984, p. 228) describes as a

local approximation to a more complicated lag structure. When it is introduced, a
linear first-order lag Aph,-l is no longer needed in the equation.
B. Specification and Estimation of the STR Model

The presence of the cubic price lag in (46) may be interpreted as an indication of
possible nonlinearity. It can be viewed as a first-order Taylor approximation to a
certain ESTR equation with Aphl-l as a transition variable. Instead of directly es-
timating such an equation a more general approach is adopted. First, one has to find
out whether the expectations model is nonlinear, the alternative to linearity being the
STR. If the null hypothesis of linearity is rejected, the next step is to build an ade-
quate STR model to characterize the nonlinearity. When this is attempted, the linear
parameter restrictions imposed in (46)are retained because the focus is on potential
nonlinearity of the price expectations. Omitting the restrictions and starting with a
richer parameterization than (46)has not been considered.
To test linearity against STR, the cubic lag ( A p h , _ ~in) ~
(46) is replaced by
Apht-l so that the basic model is linear not only in parameters but also in variables.
This model forms the null hypothesis in testing linearity as discussed in Section
1II.A. It is assumed that the dummy variables only enter the model linearly and do
not thus under any circumstances appear in the nonlinear part of the model. For
Dyl and D:t with very few nonzero values this is the only feasible assumption, but
it also covers the seasonal dummies. All the nondeterministic variables are a pri-
ori regarded as potential transition variables. Therefore, a separate test sequence is
carried out, assuming each of them in turn being the transition variable under the
alternative hypothesis. The results appear in Table 2. The null of linearity is strongly
rejected when Aph,-l is the transition variable. This accords with the specification
of (46),although the test sequence (Ho4 and H02 rejected more strongly than H03)
suggests an LSTRl rather than an LSTR2 or an ESTR model. The other strong re-
jection occurs when F13(rn - p ) involving changes in the real value of mortgage is
assumed to be the transition variable. Since the first rejection is the strongest one
534 TE&VIRTA
Table 2 p-Values of Tests of Linearity of the Linear (Aph,-l not cubed) U.K. Housing Price Equation against STR, Transition Variable
Assumed Known
F 0.00014 0.057 0.22 0.47 0.41 0.82 0.00037 0.76 0.80

F4 0.017 0.34 0.0036
F3 0.066 0.74 0.069
F2 0.0013 0.0025 0.037
“The p-values for the whole sequence of tests are given only if the p-value of the general test ( F ) lies below 0.1.
MODELING
Table 3 p-Values of Parameter Constancy Tests of the Linear

(Apht-l not cubed) U.K. House Price Equation against
STR-Type Nonconstancy
Null hypothesis
Parameter
constancy test (1) (2) (3) (4)
0.090 0.026 0.85 0.42
0.19 0.056 0.54 0.50
0.37 0.24 0.78 0.53
(1): Ho: “All parameters except the coefficients of and 11: are constant.”
(2): Ho: “Intercept and the coefficients of seasonal dummy variables are con-
stant.”
(3): Ho: “Coefficients of Aph,-l and Aph,-p are ronstant.”
(4):Ho: “Coefficients of ‘exogenous’ variables are constant.”
Notes: (1)The parameters not under test are assumed constant also under the
alternative. (2) Test 4 is a test against an STR model with transition
function H,, j = 1, 2 , 3 ; see Section III.B.3 and definitions (23)-(25).
and as (46) may be interpreted as an approximation to an STR model with AphL-l

as the transition variable it is tentatively assumed that the data have been generated
by such an LSTRl model. Furthermore, the constancy of parameters in the linear
equation was tested applying the tests discussed in Section 1II.A. The results can
be found in Table 3. The overall test vaguely indicates that the equation may not
have constant parameters. The main reason for this seems to be that seasonality has
been changing over time. Nevertheless, the evidence against parameter constancy is
clearly weaker than that against linearity, and the nonlinearity is therefore dealt with
first. It can be mentioned that if the parameter constancy of (46) is investigated by
the same tests (results not reported here), the rejection is stronger than in the case
of the linear equation.
After choosing an LSTRl model for Aph, one has to specify the parame-
ter structure of the model. In order to do that one first assumes a fully parameter-
ized model and estimates the parameters using ordinary least squares and the two-
dimensional grid for y and c as proposed in Section 1V.B. The results appear in
Table 4. As discussed, they can be used to specify the parameter structure of the
STR model. Noting that the t-values are conditional on p and E , it is seen that only
few coefficients in the nonlinear part of the equation may be nonzero. Somewhat ar-
bitrarily first retaining those with It1 2 1.6estimating the corresponding STR model
and removing the redundant variables leads to the result that only the intercept and
the error-correcting variable (rn - ph - have significant nonlinear coefficients
(they had the highest t-ratios already in Table 4). One thus obtains the following STR
model:
536 TE~~SVIRTA
Table 4 Grid Estimation of the Fully Parameterized LSTRl

Model for the U.K. Housing Prices, 1959(1)-1981(2)
Linear Standard
coefficient of Estimate deviation t-Value
-0.065 0.092 -0.71

0.2 1 0.073 2.8
0.53 0.10 5.1
0.13 0.023 5.5
0.37 0.061 6.1
0.82 0.13 6.5
0.52 0.11 4.9
-0.26 0.082 -3.2
-0.54 0.21 -2.6
Intercept 0.38 0.092 4.1
911 -0.,00098 0.0045 -0.22
Q2t 0.022 0.0037 5.8
Q3t 0.014 0.0036 3.9
-2.16 0.47 -4.6
-3.78 1.08 -3.5
Nonlinear
coefficient of
-0.17 0.43 -0.41

1.39 0.83 1.7
-1.01 2.18 -0.46
0.37 0.12 3.1
-2.18 1.42 -1.5
-2.95 4.26 -0.69
-0.87 1.81 -0.48
4.83 3.19 1.5
-7.95 3.87 -2.1
3.09 1.23 2.5
P 9.5
E 0.07
0.87
0.104
MODELING
Table 5 p-Values of the LM Test of No Error Autocorrelation against an AR(y)

and MA(q) Error Process and the LM Test of No Autoregressive Conditional
Heteroskedasticity against ARCH(q) in (47)
Maximum lag y
Test 1 2 3 4 6
No error autocorrelation 0.34 0.30 0.38 0.51 0.48

No ARCH 0.65 0.85 0.53 0.69 0.80
APhc = 0.22 Apht-2 + 0.48 A2(Ayt) + 0.13 ( m - ph - h)c-l

(0.074) (0.11) (0.027)
-I- 0.34 (y - h)t--l f 0.70 F13(p) f 0.43 F13(m - p )
(0.060) (0.12) (0.10)
- 0.21 Ry-3 - 0.51 ARY-, - 0.38 - 0.0012 Qit
(0.084) (0.21) (0.11) (0.0047)
+ 0.024 Q2t + 0.0174 Q3t - 3.6 Dyt - 2.2 D:t
(0.0040) (0.0040) (0.24) (0.53)
+{ 3.0 + 0.60 (n- ph - h)t-l}
(1.2) (0.25)
x [l + exp{ -2.5 (Aph,-l - 0.088) / & ( A p h t - l ) } ] - ’
(0.79) (0.0085)
+ rzt (47)
T = 94, AZC = -8.63, R2 = 0.85,& = 0.0123, a’/&: = 0.69, sk = 0.93, ek =
4.0, LJB = 76(4 x l O - I 7 )
where &(Aph,-l) is the sample standard deviation of Aph,-l and 8; the residual
variance of the corresponding linear model. The residual variance of (47) is only
about 70% of that of the corresponding linear model. Results of the LM test of no error
autocorrelation in Table 5 do not indicate autocorrelation, nor is there any evidence
of ARCH. Large skewness and excess kurtosis estimates are mainly due to a large
positive residual in 1964(1);see Figure 1.
C. Interpretation and Evaluation of the STR Model

The residuals of (47)are graphed in Figure 1 together with the residuals of (46).The
main difference in the fits of the two models is due to different characterizations of the
538 TERASVIRTA
-0.04 4 *.
* ..
1959(1)
. ' ' .
1963( I ) 1967( 1) 1971(1)
. . . . . . . ' . . . . . . . .. . . .
1975(1)
- -
1979( I )
.. ... ..
Quarter
Figure I Residuals from the linear model with a cubic lag (46)and from the STR model (47)
for the first differences of the logarithmic UK housing price index, 1959(1)-1982(2).
price turbulence in 1973.The linear equation strengthened by the cubic lag does not
explain the features of the housing price boom and its immediate aftermath as well as
(47).Apart from that period both models have almost identical fits. It seems obvious
that the nonlinear specification is mainly required to model the exceptional increase
in house prices in 1973. This is also seen from Figure 2. It shows that the transition
function obtains values close to zero most of the time. A comparison between (46)
and the linear part of (47)indicates that they are quite similar. For those periods (46)
and (47)thus may be expected to have rather similar residuals.
1 .o
0.8
0.6
0.4
0.i
Figure 2 Values of the transition function of the STR model (47), 1959(1)-1982(2).
MODELING
Figure 3 Values of the error-correction term (48) in equation (47), 1959(1)-1982(2).
It is instructive to find out how the nonlinear error correction (nec,) works in
1973. From (47),
nec, = 0.38 + 3.OGl(P,2; Aphl-l)
+ (0.13 + O.6OG1(P, 2; Aphl-l)}(rn - ph -
where
The graph of (48) over the observation period appears in Figure 3. When a large
price shock arrives (Aph, obtains a large value) it causes a sharp increase in (48)
+
one period later through the combined intercept 0.38 3.OG1. The error-correcting
combination (rn - ph - h ) has negative values throughout so that an increase in
the value of the transition function initially weakens the error correction. This initial
effect is soon offset by a large change in (rn - p h --h),-1 as the value of the mortgage
stock does not follow the rapid increase in the value of the housing stock. Because
of the large positive nonlinear coefficient (0.60)of (rn - ph - h)l-l in (48)the pull
toward the equilibrium increases dramatically and eventually suppresses the price
boom.
The STR model (47) seems to explain the dynamics of the unusually large
increase in housing prices but does it explain all nonlinearity in the data? Table 6
contains the results of the tests of no additive nonlinearity against an additive STR
model considered in Section III.B.2. Note that Aph,-l is included in the second
nonlinear component although it only appears in the transition function of (47).When
the test is carried out with Aph,-l as the transition variable, the p-value of the test
equals 0.11. This indicates that the nonlinearity causing a very low p-value for the
540 TERASVIRTA
Table 5 p-Values of Tests of No Additive N ~ n l i n e ~ rin ~ LSTRl Model (47) for a Set of Transition Variables
i t the
F 0.11 0.90 0.011 **a **a 0.65 0.17 0.72 0.89

F4 0.011
F3 0.056
F2 0.57 0.43 0.79
“Test not compu~eddue to near-singu~arityof the moment matrix. In that case, results of Fz (test against additive nonlinearity of LSTRl type) are shown.
ECONOMIC RELATIONSHIPS 54 I
MODELING
Table 7 p-Values of Parameter Constancy Tests of the LSTRl Model (47)

against STR-Type Constancy
Null hypothesis
Parameter
constancy test (1) (2) (3) (4) (5)
Fl 0.022 0.017 0.61 0.0078 0.71
F2 0.19 0.15 0.84 0.038 0.87
F3 0.30 0.19 0.96 0.12 0.71
~ ~~~~~~
(1): Ho: “All parameters except the coefficients of DY and ,!I

:are constant.”
(2): Ho: “All parameters in the linear part of the model except the coefficients of DY
and D: are constant.”
(3): Ho: “All parameters in the nonlinear part of the model are constant.”
(4):Ho: “Intercepts and coefficients of the seasonal dummy variables are constant.”
(5): Ho: “All parameters in the linear part of the model except the dummy variables
are constant.”
Notes: (1) The parameters not under test are assumed constant also under the al-
ternative. (2) Test 4 is a test against an STR model with transition function
H,, j = 1,2,3; see Section III.B.3 and definitions (23)-(25).
corresponding linearity test (Table 3) has been dealt with in a satisfactory manner.
Another result pointing at the same direction is that the test with F13(rn - p ) as the
transition variable has p-value 0.17, whereas the corresponding linearity test had a
low value. On the other hand, the test with A2(Ayt) as the transition variable now
has a p-value close to 0.01, but this order of magnitude is considerably higher than
that of the lowest p-values in the linearity tests. As at the same time all the other
tests have p-values exceeding 0.1, this result does not cause too much concern. As a
whole, it can be concluded that the STR model (47) explains most of the nonlinearity
present in the data.
Parameter constancy tests of the linear model indicated some nonconstancy
although the result could also have been interpreted as an effect of neglected non-
linearity on these tests. Table 7 contains results of the parameter constancy tests de-
scribed in Section 1II.B. The results suggest that despite careful parametrization of
nonlinearity the parameter nonconstancy is still a problem. It seems obvious that sea-
sonality in U.K. house prices has been changing over time. Furthermore, the change
seems to have been monotonic during the observation period because Fl is the test
with the strongest rejection of the null hypothesis just as it was in the linear model.
D. Respecification and Reestimation of the Model

To capture this parameter change one can estimate the parametric alternative to pa-
rameter constancy, which in this case means estimating an additive STR model. In
542 TERASVIRTA
this model the second transition function has time as the transition variable. A spec-
ification search indicated that the second and the third quarters have to be included
in the additional nonlinear component of the model. The estimated equation has the
form (note that the second transition variable t / T is standardized between 0 and 1)
Apht = 0.23 Apht-2 + 0.58 AZ(Ayt) + 0.12 ( m - ph - h)t-l

(0.070) (0.11) (0.025)
+ 0.33 (y - h)t-l + 0.55 F13(p) + 0.31 F i 3 ( m - p )
(0.061) (0.12) (0.10)
- 0.32 ZQ-, - 0.48 AR:-, + 0.38 - 0.0035 Q i t
(0.082) (0.19) (0.098) (0.0045)
- 0.24 Q2t + 0.23 Qst - 3.9 DYt - 1.6 Dit
(0.57) (0.54) (1.1) (0.48)
+{ 2.9 + 0.59 ( m - ph - h)t-i}
(1.2) (0.24)
x [l + exp{- 2.8 (AphL-l - 0.088 )/i3(Aphl-1)}]-'
(0.94) (0.0075)
+ { 0.53 Q z ~+ 0.50 Qstj
(1.1) (1.1)
x [l + exp{- 0.13 ( t / T - 0.49 )/&(t/T)}]-'
(0.28) (0.16)
+ at (49)
T = 94, AZC = -8.75, R2 = 0.88,6 = 0.0114,i32/&~

= 0.59, sk = 0.86, ek =
3.4, LJB = 56 (7 x 10-13)
where 3 ( t / T ) = {(l/T) C y = l ( t / T - 1/2)'}'/2. The residual variance of (49) is

about 60% of that of the corresponding linear model. Testing the residuals against
error autocorrelation and ARCH does not indicate any model misspecification (the
test results are not shown). The fit of (49) is very similar to that of (47). AZC has
decreased compared to (47).
Figures 4 and 5 illustrate the effect of the second transition function. Figure 4
shows that during the observation period, the estimated transition function is prac-
tically a straight line. This explains the large standard deviations of the estimates
of the coefficients of both the linear and the nonlinear seasonal dummy variables
and the slope parameter of the transition function y. Information about where the
logistic transition function is bending is necessary for accurate estimation of the pa-
MODELINGECONOMIC RELATIONSHIPS 543
I .o
0.8
0.8
0.4
0.1
0.a
19!59(1) I963(1)
LYC'.....' ..''...'.......'..,..,.........
1967(1) 1971(1) 1975( 1) 1979(1)
Ouorter
Figure 4 Values of the transition function of seasonal dummy variables in the STR model
(49), 1959(1)-1982(2).
rameters involved in the characterization of seasonal effects. That this informa-

tion is scarce in the sample is seen from Figure 4. A similar situation is described
in Jansen and Terasvirta (1996). Figure 5 shows the values of the time-varying
seasonals (-0.24 +
0.53H(P, E ; L / T ) } Qand +
~ ~ (-0.23 O.SOS(P, E ; t / T ) } & .
The figure shows that seasonality in the house prices increases from 1959(1) to
1982(2)and that the effect can be ascribed to the second and the third quarters of
the year.
0'015
0.040 I
0 035
0 030
0 025
0 020
0015
0010
0 005
0 000 Y
Ouorter
Figure 5 Values of the second and third quarter time-varying seasonal effects (second quar-
ter = solid line, third quarter = dashed line) according to Eq. (49), 1959(1)-1982(2).
544 TERASVIRTA
In order to find out whether or not the large uncertainty in the estimates of sea-
sonal parameters is mainly due to overparameterization the second logistic transition
function in (49) is replaced by the linear approximation ( t / T ) and the parameters
reestimated. This yields
+ 0.33 (y - h)t-l + 0.55 F13(p) + 0.31 Fi3(m - p)

(0.060) (0.12) (0.10)
- 0.32 Re-3 - 0.48 ARY-, + 0.38 - 0.0035 Q l t
(0.080) (0.20) . (0.096) (0.0045)
+ 0.0067 Q2t + 0.0016 Q3L - 3.9 Dyt - 1.6 D:t

(0.0063) (0.0060) (1.1) (0.51)
+{ 3.0 + 0.59 ( m - ph - h)L-l)
(1.2) (0.24)
x [l + exp{- 2.8 (Aph,_l - 0.088 )}]-I
(0.93) (0.0077)
+ 0.035 (t/T )Q 2t + 0.033 ( t / T ) Q 3 t + fit (50)

(0.010) (0.0094)
T = 94, AZC = -8.79, R2 = 0.88,6 = 0.0113, 6’/&;, = 0.58, sk = 0.86,ek =

3.4, IJB = 56(7 x lO-I3).
The estimates of other parameters than the seasonal dummies remain practically
unchanged. The seasonality seems indeed to be changing over time. The amount of
uncertainty in the parameter estimates is considerably less than in (49) and AZC
smaller. The seasonals in the linear part of (50) could even be removed altogether;
that is, in the beginning of the period there has been little or no seasonality in house
prices. For illustration, however, those variables have been retained in the model.
The test results for models (49) and (50) are very close to each other. Those
for (SO) are reported. Results of testing the hypothesis of no remaining nonlinear-
ity against the same alternative as previously can be found in Table 8. They are
rather similar to those for model (47). The only conspicuous difference is that the
test against LSTR with (rn - ph - h)c-l as the transition variable now has a rel-
atively small p-value. Because most of the other tests have clearly higher p-values
(even the one with A2(yL) as the transition variable) the current specification is ten-
tatively accepted.
Results of the parameter constancy tests can be found in Table 9. The null
of constancy is not rejected in any of the tests. The parameterization of changing
MODELING
Table 8 p-Values of Tests of No Additive Nonlinearity in the LSTRl Model (SO)for a Set of Transition Variables
F 0.27 0.91 0.031 **a **a 0.13 0.16 0.25 0.44

F4 0.015
F3 0.37
F2 0.24 0.016 0.65
‘Test not computed due to near-singularity of the moment matrix. In that case, result of F2 (tests against additive nonlinearity of LSTRl type) are shown.
546 TERASVIRTA
Table 9 p-Values of Parameter Constancy Tests of the LSTRl Model (50)

against STR-Type Nonconstancy
Null hypothesis
Parameter
constancy test (1) (2) (3) (4) (44 (5)
Fl 0.70 0.61 0.58 **a 0.97 0.50

**a 0.81 0.60 **a 0.85 0.73
F2
F3 **a 0.71 0.81 **a 0.96 0.32
(1): Ho: “All parameters except the coefficients of Dy and D; are constant.”
(2): Ho: “All parameters in the linear part of the model except the coefficients of DY and
Di are constant.”
(3): Ho: “All parameters in the nonlinear part of model are constant.”
(4): Ho: “Intercepts and coefficients of the seasonal dummy variables are constant.”
(4a): Ho: “Linear intercept and coefficients of the seasonal dummy variables are constant.”
(5): Ho: “All parameters in the linear part of the model except the coefficients of the dummy
variables are constant .”
“Test not computed due to near-singularity of the moment matrix.
Notes: (1) The parameters not under test are assumed constant also under the alternative.
(2) Test I$ is a test against an STR model with transition function H,, j = 1, 2 , 3 ;see
Section III.B.3 and definitions (23)-(25).
seasonality has removed the variation in the coefficients of seasonal dummy vari-
ables. According to (50),U.K. housing price changes now have a trending seasonal
component. However, it is not very realistic to extend the conclusion far outside the
sampling period. In fact, model (49) should be preferred to (50) as far as the in-
terpretation of the “trend” is concerned. The change one observed just happens to
lie within an interval for.which the logistic transition function is almost linear. Of
course, even the interpretation that Eq. (49) offers may turn out to be incorrect in
the light of any new data, but at any rate, the assumption of a long-run linear trend
in seasonality is hardly a plausible one.
E. Encompassing Tests
An objective of this application is to find out if the specification of the price ex-
pectations equation (46) can be improved by applying STR models. Equation (50)
is an STR model, and the question is whether or not it can be viewed as an im-
provement over (46).This can be investigated by encompassing tests as discussed
in Section 1V.E. The task is to investigate whether the STR model (50) encompasses
(46) or not and vice versa. To find out if (50) encompasses (46) the first step is to
construct an MNM. This is done by augmenting (50) linearly by the cubic lag of
the price change ( A & - I ) ~ . The augmented model trivially encompasses (50) be-
MODELING
cause the latter model is nested in it. In order to see whether (50) parsimoniously
encompasses the M N M one estimates the parameters of the M N M and computes the
likelihood ratio statistic
where Lo is the likelihood of (50)and L1 that of the M N M . Under Ho: “Coefficient of

( A ~ h , - l equals
)~ zero in the MNM,” (51) has an asymptotic x2(1)distribution. In
this case one obtains LR = 0.0299 with p-value 0.86. The conclusion is that (50)
parsimoniously encompasses the M N M .
The following computational remark is in order. Equation (46) and the STR
model (50)are competitors, of which the latter is nonlinear even in parameters. Non-
linear least-squares estimation of the M N M required to carry out the test of parsimo-
nious encompassing may sometimes be difficult because the parameterization of the
M N M may turn out to be very rich. In the present case the convergence was achieved
in the estimation when the initial estimates of the parameters of the M N M (save the
~ ) the same as the final estimates in (50). If the initial
coefficient of ( A p h , - ~ )were
estimates were chosen further away from those values, the N L S algorithm (BFGS)
either converged to a local minimum or did not converge at all.
Testing the null hypothesis that (46) parsimoniously encompasses the M N M
is not equally straightforward because the latter equation is nonlinear and is not
identified if linearity in parameters (the null hypothesis) holds. This situation was
discussed in Section 1V.E and it was pointed out the hypothesis can be tested by
applying linearity tests described in Section 1II.A. The transition variable of the al-
ternative needed in the test is Aph,-l. The linearity test statistic is the usual one
and the variables in the auxiliary regression based on the M N M include the ordi-
nary ones in (50) but also ( A p h , - ~ ) ‘ Note
? . that in accordance with (46), Apht-l
only appears as the transition variable and is not present in the null model. The
F-approximation to the likelihood ratio test results in F(27, 52) = 2.42(0.0031),
whereas Fz(9, 70) = 2.51(0.0051) (FL is a test against LSTR1). This implies a
rejection of the null hypothesis, and the conclusion is that Eq. (46) does not par-
simoniously encompass the MNM. As (50) does, one has the result that (50) also
encompasses (46) and thus contributes to the general understanding of the mecha-
nism generating U.K. house price expectations during 1960-1981.
VI. FINAL REMARKS
In this chapter the emphasis is on showing how STR models can be applied to model-
ing problems in time series econometrics. It is demonstrated how the actual modeling
is carried out in a systematic fashion through a modeling cycle. This cycle can be
repeated until an adequate model passing the available diagnostics is found. Alter-
natively, the cycle may be terminated by concluding that the family of STR models is
548 TERASVIRTA
not an appropriate one for empirical modeling of the economic relationship in ques-
tion. The central role of hypothesis testing in STR modeling becomes clear from the
text. First, it is important to single out the linear cases from nonlinear ones. But test-
ing is also an essential part of model specification and evaluation. The validity of
assumptions underlying the STR model are investigated by tests after estimating the
parameters of the model as well as the question whether or not an estimated STR
model is an improvement over previous models in the literature.
The STR can be used, as in the example of Section V, for modeling economic
relationships between variables. Another role of the STR model is that it constitutes
a feasible alternative to important null hypotheses concerning linear models. The
null hypothesis of parameter constancy is one: in that case the alternative to con-
stant parameters are continuously changing parameters. As Section V shows, this
is a useful alternative, for example, in testing the constancy of the pattern of sea-
sonal fluctuations. Furthermore, although this possibility has not been discussed in
this chapter in any detail, the STR model provides a convenient framework for joint
testing of weak exogeneity and a restricted form of invariance, that is, for testing su-
perexogeneity. Finally, the additive STR model may be used for testing for Granger
causality in the presence of STAR-type nonlinearity.
The STR model considered in this chapter is a single-equation model. In the-
ory, the idea of smooth transition can be extended to systems of equations. This can
be done in many ways of which the ones with practical significance still need to be
sorted out. Anderson and Vahid (1995) recently searched for common nonlineari-
ties between variables, using a vector STAR model. In general, however, there is as
yet little empirical experience available of such systems and more is needed. In the
meantime, additional applications of the single-equation STR model are also neces-
sary to learn more about how the proposed modeling strategy works in practice and
to find out ways of improving it and developing it further.
ACKNOWLEDGMENTS
Research supported in part by the Swedish Council for Research in Humanities

and Social Sciences and the Copenhagen Business School. Material from the chap-
ter, including the empirical example, has been discussed at courses at the Copen-
hagen Business School, Universidad Carlos I11 de Madrid, University of Helsinki,
and GREQAM-Universite d’ Aix-Marseille I1 & 111. Comments from participants as
well as those from Rolf Tschernig are gratefully appreciated. I am particularly in-
debted to David Hendry for making his U.K. housing market data set available to
me. The responsibility for any errors and shortcomings remains mine.
MODELING
REFERENCES
AlbEk, K. and H. Hansen (1995), Estimating Aggregate Labour Market Relations, mimeo.,
Department of Economics, University of Copenhagen.
Anderson, H. M. (1997),Transaction costs and Nonlinear Adjustment towards Equilibrium in
the US Treasury Bill Market, Oxford Bulletin of Economics and Statistics, forthcoming.
Anderson, H. M. and F. Vahid (1995),Testing Multiple Equation Systems for Common Non-
linear Components, mimeo., Texas A & M University.
Andrews, D. W. K. and W. Ploberger (1994), Optimal Tests When a Nuisance Parameter Is
Present Only under the Alternative, Econometrica, 62, 1383-1414.
Bacon, D. W. and D. G. Watts (1971), Estimating the Transition Between Two Intersecting
Straight Lines, Biometrika, 58, 525-534.
Banerjee, A., J. Dolado, J. W. Galbraith, and D. F. Hendry (1993), Co-integration, Error-
Correction, and the Econometric Analysis of Non-stationary Data, Oxford University
Press, Oxford.
Bates, D. M. and D. G. Watts (1988).Nonlinear Regression Analysis a n d h Applications, Wiley,
New York.
Bell, D., J. Kay, and J. Malley (1996), A Nonparametric Approach to Nonlinear Causality
Testing, Economics Letters, 51, 7-18.
Box, G. E. P. and G. M. Jenkins (1970),Time Series Analysis, Forecasting and Control, Holden-
Day, San Francisco.
Chan, K. S. and H. Tong (1986),On Estimating Thresholds in Autoregressive Models, Journal
of Time Series Analysis, 7, 178-190.
Chiarella, C., W. Semmler, and L. Koqkesen (19%), The Specification and Estimation of a
Nonlinear Model of Real and Stock Market Interaction, mimeo., Department of Eco-
nomics, New School for Social Research, New York.
Cook, P. and L. D. Broemeling (1996),Analyzing Threshold Autoregressions with a Bayesian
Approach, in T. Fomby (ed.), Advances in Econometrics, Vol. 11, Part B: Bayesian Meth-
ods Applied to Time Series Data, JAI Press, Greenwich CT, 89-107.
Coutts, J. A., T. C. Mills, and J. Roberts (1995),Parameter Stability in the Market Model: Tests
and Time Varying Parameter Estimation with U.K. Data, mimeo., Sheffield University
Management School.
Davies, R. B. (1977),Hypothesis Testing When a Nuisance Parameter Is Present Only under
the Alternative, Biometrika, 64,247-254.
Davies, R. B. (1987),Hypothesis Testing When a Nuisance Parameter Is Present Only under
the Alternative, Biometrika, 74, 33-44.
Eitrheim, 0.and T. Terasvirta (1996),Testing the Adequacy of Smooth Transition Autoregres-
sive Models, Journal of Econometrics, 74, 59-75.
Engle, R. F. (1982),Autoregressive Conditional Heteroskedasticity with Estimates of the Vari-
ance of the U.K. Inflation, Econometrica, SO, 987-1008.
Ericsson, N. R. (1992),Cointegration, Exogenctity and Policy Analysis: An Overview, JournaZ
of Policy Modeling, 14,251-280.
Ertel, J. E. and E. B. Fowlkes (1976), Some Algorithms for Linear Spline and Piecewise
Multiple Linear Regression, Journal of the American Statisticul Association, 71,
640648.
550 TERASVIRTA
Escribano, A. and S. Mira (1995), Nonlinear Time Series Models: Consistency and Asymptotic
Normality of NLS under New Conditions, Universidad Carlos I11 d e Madrid, Statistics
and Econometrics Series 14, Working Paper 95-42.
Fair, R. C. and D. M. Jaffee (1972), Methods of Estimation for Markets in Disequilibrium,
Econometrica, 40,497-514.
Farley, J. U., M. Hinich, and T. W. McGuire (1975), Some Comparisons of Tests for a Shift
in the Slopes of Multivariate Linear Time Series Model, Journal of Econometrics, 3,
297-3 18.
Geweke, J. (1984), Inference and Causality in Economic Time Series Models, in Z. Griliches
and M. D. Intriligator (eds.), Handhook of Econometrics, Vol. 2, North-Holland, Ams-
terdam, 1101-1 144.
Geweke, J. and N. Terui (1993),Bayesian Threshold Autoregressive Models of Nonlinear Time
Series, Journal of Time Series Analysis, 14, 441-454.
Goldfeld, S. M. and R. E. Quandt (1972), Nonlinear Methods in Econometrics, North-Holland,
Amsterdam.
Goldfeld, S. M. and R. E. Quandt (1973), A Markov Model for Switching Regression, Journal
of Econometrics, 1, 3-16.
Granger, C. W. J. (1969), Investigating Causal Relations by Econometric Models and Cross-
Spectral Methods, Econometrica, 37, 424-438.
Granger, C. W. J. (1981), Some Properties of Time Series Data and Their Use in Econometric
Model Specification, Journal ofEconometrics, 16, 121-130.
Granger, C. W. J. and T. Terasvirta (1993, Modelling Nonlinear Economic Relationships, Ox-
ford University Press, Oxford.
Granger, C. W. J., T. Terasvirta, and H. Anderson (1993), Modeling Nonlinearity over the
Business Cycle, in J. H. Stock and M. W. Watson (eds.), Business Cycles, Indicators,
and Forecasting, University of Chicago Press, Chicago, 31 1-325.
Hansen, B. E. (1996), Inference When a Nuisance Parameter Is Not Identified under the Null
Hypothesis, Econometrica, 64,413-430.
Hatanaka, M. (1996), Time-Series-Based Econometrics. Unit Roots and CO-integration, Oxford
Heinesen, E. (1996), The Tax Wedge and the Household Demand for Services, mimeo., Insti-
tute of Local Government Studies, Copenhagen.
Hendry, D. F. (1984), Econometric Modelling of House Prices in the United Kingdom, in D. F.
Hendry and K. F. Wallis (eds.), Econometrics and Quantitative Economics, Blackwell,
Oxford, 211-252.
Hendry, D. F. (199S),Dynamic Econometrics, Oxford University Press, Oxford.
Jansen, E. S. and T. Terasvirta (1996), Testing Parameter Constancy and Super Exogeneity in
Econometric Equations, Oxjord Bulletin ($Economics and Statistics, 58, 735-763.
Jarque, C. M. and A. K. Bera (1980), Efficient Tests for Normality, Homoscedasticity, and
Serial Independence of Regression Residuals, Economics Letters, 6,255-259.
Johansen, S. (1995),Likelihood-Based Inference in Cointegrated Vector Autoregressive Models,
Oxford University Press, Oxford.
Judge, G. G., W. E. Griffiths, R. C. Hill, H. Liitkepohl, and T.-C. Lee (1985), The Theory and
King, M. L. and T. S. Shively (1993), Locally Optimal Testing When a Nuisance Parameter Is
Present Only under the Alternative, Review of Economics and Statistics, 75, 1-7.
ECONOMIC RELATIONSHIPS 55 I
MODELING
Lee, T.-H., H. White and C. W. J. Granger (1993), Testing for Neglected Nonlinearity in Time
Series Models: A Comparison of Neural Network Methods and Alternative Tests, Jour-
nu1 of Econometrics, 56,269-290.
Lin, C.-F. and T. Terasvirta (1994), Testing the Constancy of Regression Parameters Against
Continuous Structural Change, Journal ojEconometrics, 62,211-228.
Lindgren, G. (1978), Markov Regime Models for Mixed Distributions and Switching Regres-
sions, Scandinavian Journal of Statistics, 5,81-91.
Lomnicki, Z. A. (1961), Test for Departure from Normality in the Case of Linear Stochastic
Processes, Metrika, 4 , 3 7 4 2 .
Lutkepohl, H., T. Terasvirta, and J. Wolters (199S), Investigating Stability and Linearity of a
German M 1 Money Demand Function, Stockholm School of Economics, Working Paper
Series in Economics and Finance No. 64.
Luukkonen, R., P. Saikkonen, and T.Terasvirta (1988a), Testing Linearity against Smooth
Transition Autoregression, Biometrika, 75, 491-499.
Luukkonen, R., P. Saikkonen, and T. Terasvirta (1988b), Testing Linearity in Univariate Time
Series, Scandinavian Journal ofstatistics, 15, 161-17s.
Maddala, D. S. (1977), Econometrics, McGraw-Hill, New York.
Maddala, D. S.(1986), Disequilibrium, Self-Selection, and Switching Models, in Z. Griliches
and M. D. Intriligator (eds.), Handbook ofEconometrics, Vol. 3, North-Holland, Ams-
terdam, 1633-1688.
McLeod, A. I. and W. K. Li (1983), Diagnostic Checking ARMA Time Series Models Using
Squared Residuals, Journal oj’Time Series Analysis, 4,269-273.
Mizon, G. E. and J.-F. Richard (1986). The Encompassing Principle and Its Application to
Non-nested Hypothesis Tests, Econometrica, S4,657-678.
Peguin-Feissolle, A. (1994), Bayesian Estimation and Forecasting in Nonlinear Models. Ap-
plication to an LSTAR model. Economics Letters, 46, 187-194.
Petruccelli, J. D. (1990), A Comparison of Tests for SETAR-Type Non-linearity in Time Series,
Journal of Forecasting, 9,25-36.
Pfann, G. A., P. C. Schotman, and R. Tschernig (1996), Nonlinear Interest Rate Dynamics and
Implications for the Term Structure, Journd of Econometrics, 74, 149-176.
Pole, A. M. and A. F. M. Smith (1985), A Bayesian Analysis of Some Threshold Switching
Models, Journal ofEconometrics, 29,97-119.
Quandt, R. E. (1960), Tests of the Hypothesis That a Linear Regression System Obeys
Two Separate Regimes, Joiirnd oj’ the American Statistical AssocicLtion, 55, 324-
330.
Quandt, R. E. (1984), Computational Problems and Methods, in Z. Griliches and M. D.
Intriligator (eds.), Handbook of Econometrics, Vol. 1, North-Holland, Amsterdam,
699-764.
Ramsey, J. B (1969), Test for Specification Errors in Classical Linear Least-Squares Regres-
sion Analysis, Joiwnd ojthe Royal Statistical Society B, 31, 350-371.
Richard, J.-F. and W. Zhang (1996), Econometric Modeling of UK House Prices Using Ac-
celerated Importance Sampling, (Ixford Bulletin ofEcon.omics and Statistics, 58,601-
613.
Saikkonen, P. and R. Luukkonen (1988), Lagrange Multiplier Tests for Testing Non-linearities
in Time Series Models, Scandinclvim Joiirnal of Statistics, 15, 55-68.
Seber, G. A. F. and C. J. Wild (1989), Nonlinear Regression, Wiley, New York.
552 TERASVIRTA
Semmler, W. and L. Koqkesen (199S),Testing for Nonlinear Financial-Real Interaction Using

a Smooth Transition Regression Model, mimeo., Department of Economics, New York
School for Social Research, New York.
Shively, T. S. (1988), An Analysis of Tests for Regression Coefficient Stability, Journal of
Skalin, J. and T. Terasvirta (1996), Another Look at Swedish Business Cycles, 1861-1988,
Stockholm School of Economics, Working Paper Series in Economics and Finance,
No. 130.
Terasvirta, T. (1994),Specification, Estimation, and Evaluation of Smooth Transition Autore-
gressive Models, Journal of the American Statistical Association, 89,208-218.
Terasvirta, T. and H. M. Anderson (1992), Characterizing Nonlinearities in Business Cycles
Using Smooth Transition Autoregressive Models, Journal of Applied Econometrics, 7,
S119-S136.
Tjgstheim, D. (1994), Non-linear Time Series: A Selective Review, Scandinavian Journal of
Tong, H. (1990), Non-linear Time Series. A Dynamical System Approach, Oxford University
Press, Oxford.
Tsurumi, H. (1982), A Bayesian and Maximum Likelihood Analysis of a Gradual Switching
Regression in a Simultaneous-equation Framework, Journal of Econometrics, 19,165-
182.
Wolters, J., T. Terasvirta, and H. Liitkepohl(1997), Modeling the Demand for M 3 in the Unified
Germany, Review of Economics and Statistics. forthcoming.
Wooldridge, J. M. (1994), Estimation and Inference for Dependent Processes, in R. F. Engle
and D. L. McFadden (eds.), Handbook of Econometrics, Vol. 4, North-Holland, Ams-
terdam, 2649-2738.
16
Modeling Seasonality in Economic
Time Series
Philip Hans Franses

Erasmus University Rotterdam, Rotterdam, The Netherlands
1. INTRODUCTION
This chapter surveys issues concerning seasonality in economic time series. An elab-
orate discussion on a formal definition of seasonality is given in, e.g., Hylleberg
(1986, 1992).Here I loosely refer to seasonality as the variation in time-series data
that displays a certain regularity corresponding with the measurement interval. For
example, for quarterly data one may consider the annually recurring positive or neg-
ative peaks in certain quarters as seasonal fluctuations. Furthermore, the observa-
tion that stock returns on Mondays seem more volatile than those on other weekdays
concerns seasonality too, that is, seasonality in variance.
In many cases, seasonality in economic time series is due to weather or insti-
tutional factors. An example of the latter is that school holidays are fixed by local
governments, and hence one may expect tourism spending to be high in the cor-
responding season. Another example is that the deadline for companies to publish
their annual reports can be dictated by law. Right after that deadline, one may expect
more volatility in stock markets in case the news differs from the expectations, and
one may also observe changes in key macroeconomic figures such as consumer confi-
dence indicators. Hence, part of the seasonal variation may be roughly constant over
time, since for example it is unlikely that Christmas will move to other months, but
another part of seasonality may change because of changes in institutional factors.
Finally, seasonal patterns can also change because economic agents start to behave
in a different way. In fact, if Mondays would always display high volatility, one would
be able to make money through derivatives. Hence, because of the so-called week-
end effect, one may expect a high volatility on Mondays, but this feature is unlikely
553
554 FRANSES
to be constant over time, see, e.g., Franses and Paap (1995). Another example is that
seasonal labor supply may make the unemployment rate to display more seasonality
in the expansion stage.
In this chapter I will focus on statistical models that can describe and forecast
economic time series with seasonal variation which changes over time. The running
examples in this chapter to be used for illustration are taken from macroeconomics,
tourism, marketing, and finance. A dominant approach in macroeconomics is to sea-
sonally adjust the data prior to analysis; that is, one assumes that seasonality is not an
interesting data feature and should be removed. In most fields of economics, however,
seasonality is considered important since it can convey information on, for example,
the behavior of economic agents (Ghysels 1994a). In this chapter I concur with this
view and I will therefore not consider seasonal adjustment, and confine myself to
models that explicitly incorporate descriptions of seasonality. I refer the reader in-
terested in seasonal adjustment to the surveys in Hylleberg (1992), Bell and Hillmer
(1984),and Maravall (1995)inter alia.
The outline of this chapter is as follows. Section I1 gives some summary statis-
tics for four sample series. These statistics mainly show that seasonality does not
appear constant over time, and that any changes do not occur quickly. Hence, sea-
sonality seems to change slowly over time. Section I11 reviews the two approaches
which are nowadays commonly used in many applications, i.e., univariate and multi-
variate models that incorporate seasonal unit roots and seasonal parameter variation.
Due to space limitations, I only highlight some of the key features of these models
and refer the interested reader to the surveys in Hylleberg (1994), Franses (1996a,
b), and to the specific studies mentioned here. In Section IV, I discuss further topics
of research. Section V concludes with some remarks.
II. SOME EXAMPLE SERIES
In this section I discuss some time series features of four sample series. The first time
series is real consumption nondurables in the United Kingdom, which is observed
quarterly for 1955.1-1988.4. The source of these data is described in Osborn (1990).
The graph of this series is displayed in Figure 1. It is clear that the data display an
upward-moving trend, which seems sometimes hampered by shocks, especially those
around 1974 and 1979. Furthermore, seasonal variation seems a dominant source of
variation. As is usual for economic time series, these data are transformed by taking
natural logarithms.
The second quarterly time series is depicted in Figure 2 and it concerns the
unemployment rate in Federal Germany for 1962.1 to 1991.4. The source of these
data is the OECD Main Economic Indicators. The graph in Figure 2 shows that un-
employment seems to increase rapidly around the recession periods 1967,1974, and
1980-1982. The decline in unemployment occurs more slowly, and hence this series
SEASONALITY IN TIMESERIES
MODELING 555
, , , , l , , . , i , , , ~ i ~ , , ,, , , , , , , , , , , ,
1955 1960 1965 1970 1975 1980 1985
Figure I Nondurable consumption in the United Kingdom.
12.5
10.o
7.5
5 .O
2.5
0.o l ~ l ' i ' i I ' ' l ~ i ~ l 1 ' 1 l' 1 0 1 ' 1 ' 1 '
2 64 66 68 70 72 7 4 76 78 80 82 84 86 88 90
Figure 2 Unemployment rate in Germany.

556 FRANSES
displays nonlinear features. Again, seasonal fluctuations seem important sources of

variation. Since this unemployment series is a percentage, the log transformation is
not applied.
A sample series for the tourism sector is taken from Gonzalez and Moral (1995).
It concerns the monthly real tourism and travel receipts for Spain (in billions pe-
setas), for the sample 1979.01-1993.12. From Figure 3 one can observe that (as
expected) seasonal variation is quite pronounced, where the peaks correspond with
August and the dips with December, approximately. There is an upward trend until
1989 and a slightly different trend after 1990.
Finally, the fourth sample series concerns marketing data; i.e., these are ex-
penditures on radio advertising in The Netherlands, measured four-weekly from
1984.01 to 1994.13.The data source is the Bureau for Budget Control, Amsterdam.
Its graph in Figure 4 shows that the second half of the sample displays strmg sea-
sonal variation.
For convenience, I denote all four time series as yL. The index t runs from 1
to n. The four series are observed S times per year, where S equals 4 for consump-
tion nondurables and unemployment, 12 for tourism in Spain, and 13 for the radio
advertising data. The variables concern N years of data, where the index T is used
to indicate years (with T = 1 , 2 , . . . , N ) . Hence, n equals S N . The seasonal index
is denoted by s, with s = 1 , 2 , . . . , S. Finally, I denote the conventional seasonal
dummy variables by Ds,L, +
which take a value 1 in season s for which t = S ( T - 1) s
for T = 1,2, . . . , N , and zero otherwise.
A useful first step in analyzing the properties of seasonal time series is to
calculate the autocorrelation function (ACF) of (transformations of) the data. These
transformations concern the application of certain differencing filters to remove any
nonstationary features. For example, when a time series has a stochastic trend, its
ACF when estimated along standard lines cannot be interpreted since the variance
of such a process depends on time. A stochastic trend can be removed by taking
first differences A I , where A1 is defined by Akyl = (1 - Bk)yt = yL - yl-k for
k = 1 , 2 , . . .. An analogous version of A1 for seasonal time series is A s = 1 - B S .
In that case one says that yt has seasonal stochastic trends. When seasonality is
approximately deterministic, one may consider the residuals from the regression of
A,yl on S seasonal dummies, since in that regression one removes the seasonal con-
stants from A 1yL.Finally, it may be that seasonal and nonseasonal stochastic trends
somehow interact, and hence that one considers the A1 As filter, as is advocated by
Box and Jenkins (1970).Table 1 reports on the ACFs for the various transformed
+
time series. To save space, only the autocorrelations at lags, 1, 2, S - 1 , S, S 1,
and 2s are given.
The results in Table 1 can be said to be typical for economic time series with
seasonality. For all four sample series, the ACF of yL dies out only very slowly. The
ACFs of the A1 transformed data all show peaks at lags S and 2s. When the A ,
MODELINCSEASONALITY IN TIME
SERIES 557
17.o
16.5
16.O
15.5
15.04 , , , , , , , , , ,
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93
Figure 3 Tourism and travel receipts in Spain.
Figure 4 Advertising expenditures on the radio in the Netherlands.

558 FRANSES
Table I Autocorrelations of Sample Seriesa
Variable Lag 1 2 s-1 S s+1 25

Consumption nondurables (n = 136, S = 4)
0.928* 0.900* 0.876* 0.891* 0.823* 0.788*
Yt
-0.463* -0.014 -0.481*. 0.947* -0.438* 0.910*
A1 yt
s p,D,,t
*
A I R - Cs=I -0.074 -0.359* -0.034 0.554* 0.023 0.491*
0.779* 0.625* 0.449* 0.248* 0.238* -0.014
ASYt
-0.164 0.050 0.048 -0.444* 0.236* 0.023
AI ASYt
Unemployment rate (12 = 120, S = 4)
0.969* 0.946* 0.934* 0.934* 0.891* 0.831*
Yt
-0.195* -0.234* -0.224* 0.874* -0.254* 0.798*
A1 yt
yt - c$,fl,Ds,t 0.220* 0.290* 0.078 0.457* -0.077 0.188*
0.919* 0.772* 0.593* 0.401* 0.250* -0.002
ASYt
0.422* 0.189* 0.072 -0.239* -0.239* -0.089
A 1 ASYt
Tourism and travel receipts (n = 180, S = 12)
0.833* 0.590* 0.756* 0.860* 0.728* 0.721*
Yt
0.219* 0.091 0.330* 0.733* 0.276* 0.658*
AlYL
Alyt - E,=]
s ,iLsDs,, -0.433* -0.164* 0.169* 0.056 -0.134* -0.042*
0.304* 0.278* 0.234* -0.106 0.141 0.067
Asyt
-0.481* -0.051 0.307* -0.428* 0.135* 0.002
A1 Asyt
Advertising expenditures, radio (n = 143, S = 13)
0.876* 0.719* 0.662* 0.745* 0.649* 0.483*
Yt
0.131 -0.332* 0.155 0.753* 0.187* 0.588*
A1 yt
Alyt - cL, 0.081 -0.216* 0.121 0.506* 0.226* 0.298*
0.631* 0.465* 0.093 -0.037 0.059 -0.162
ASYt
-0.277* 0.026 0.180* -0.311* 0.262* 0.004
A1 ASYt
*Significant at the 5% level.
"All variables are in logs, except the unemployment rate.
S
data are regressed on seasonal dummies (A1yl - psDs,l),some of the large
values at S and 2s reduce to small and insignificant values. This shows that one
should always account for seasonal constants, since neglecting these may lead to
spuriously large ACF values at seasonal lags. The ACFs of the AsyLseries seem to
die out reasonably quickly, and one may want to estimate low-order autoregressive
(AR) models for these transformed series. In the last row of each panel, one can
observe (and this is common to many economic series) that the ACF of A1 AsyLhas
significant values at lags 1, S - 1, S, and S 1. +
MODELING
SEASONALITY I N TIMESERIES 559
The latter feature for the A1 A s transformed series is also recognized in Box
and Jenkins (1970), who propose to analyze the so-called airline model
AiAsyt = (1 - 81B)(1 - 8sBs)&l (1)
where E t is a standard white noise series. Clearly, the theoretical ACF of (1) corre-
+
sponds with nonzero values for the ACF at lags 1, S - 1, S, and S 1. In the case
where 81 = 0s = 1, which amounts to overdifferencing since then (1 - B)(1 - B s )
cancels from both sides of (l),the corresponding ACF values are -0.5,0.25, -0.5,
and 0.25, respectively. Most textbooks in time series advocate the use of (1). Fur-
thermore, for many different example series model (1)fits well. Its key advantage is
that it only contains two moving-average (MA) parameters to be estimated.
As an example for the consumption nondurables data it appears that the follow-
ing model passes the usual LM-type diagnostic checks for residual autocorrelation
at lag 1 and at lags 1-to-S:
AiA4yl = 0.0002 + E, - O.634Et-4
(0.0010) (0.069)
where the parameters are estimated using Micro TSP routines, and where standard
errors are reported in parentheses. Notice that (2) effectively contains only one pa-
rameter, which (together with the double-differencing filter) appears sufficient to re-
move the autocorrelation in yl.
As another example, for the four-weekly advertising data, the following model
is found to pass the diagnostic checks for residual autocorrelation:
AlA13yt = 0.0012 + Et - 0.3942,-1 - O.388it-2
(0.027 7) (0.077) (0.072)
(3)
- O.4962,-13 - O.304Et-14
(0.069) (0.076)
Due to the disaggregation level of these data, one should expect the need for several
parameters to whiten the errors. For the MA part of the model, the solutions to its
characteristic polynomial are six pairs of complex roots with absolute values 0.981,
0.963,0.943,0.930, 0.928, and 0.921, and two real roots 0.993 and 0.612. Hence,
given that 13 of the 14 solutions are close to the unit circle, it may be that the A13
filter on the left-hand side of (3) is redundant. Given the fact that MA parameters
are typically estimated away from the unity boundary, one may now be tempted to
conclude that the A I filter for advertising should be sufficient. In fact, if so, with
seasonality being approximately constant, Bell (1987) shows that the solutions to
the MA polynomial for (1) approach the unit circle.
It is worthwhile to note that the solutions to 1 - z s = 0, or equivalently
to exp(Si4) = 1 (where i2 = -1) are { 1,cos(27tkl.S) + i sin(27tklS)) for k =
560 FRANSES
1 , 2 , . . . , yielding S different solutions which all lie on the unit circle. In other words,
+ +
the A1 As filter in (1)to (3) assumes S 1 unit roots, i.e., S 1independent sources
of stochastic and nonstationary variation. Even though the MA component in models
as (1) seems to “repair” in some sense the possibly overestimated number of unit
roots, one may question the notion that economic time series are governed by such
a large number of trends.
Consider for example the so-called seasonal random walk process Asyl = E~
and also consider the S time series Y s , ~which , are the annually observed data on
yt in season s = 1 , 2 , . . . , S . Given the seasonal random walk process, it is clear
that the individual Ys,r series are annually observed random walks which are in-
dependent. The observations in the various seasons are not tied together somehow,
and they wander through time without any restriction. Strictly speaking, this means
that “summer” can become “winter.” Of course, the MA component in (1) will pre-
vent such changes to occur rapidly, but in principle it is possible. Given the graphs
in Figures 1 through 4, it seems that for economic data the seasons are somehow
Table 2 Changing Seasonality
Sample Lags Min 8, Max 8, SD
Consumption nondurables (S = 4)
1957.3-1 973-4 2,4,5,8,9 -0.054 (81) 0.041 (82) 0.042
1974.1-1988.4 2,4,5,8,9 -0.078 (81) 0.039 (84) 0.055
Unemployment rate (S = 4)
1963.3-1976.4 1,475 -0.679 (82) 0.590 (81) 0.548
1977.1-1991.4 1,495 -0.451 (82) 0.288 (81) 0.322
Tourism and travel receipts (S = 12)
1979.07-1986.12 1,2,3,4,5 -0.459 (812) 0.569 ( 8 8 ) 0.375
1987.01-1993.12 1,2,3,4,5 -0.438 (812) 0.434 (8,) 0.299
Advertising expenditures, radio (S = 13)
1980.02-1988.13 1,2, 3, 4, 5, 13 -0.114 (81) 0.086 ( 8 6 ) 0.059
1989.01-1994.13 1,2, 3, 4, 5, 13 -0.355 (8,) 0.297 (82) 0.179
~ ~~~~
The auxiliary regression is
where the number of lags is based on LM tests for residual autocorrelation at lags 1 and 1-to-S.
SD is the standard deviation of the estimated 8,. Note that a formal test for the equality of the
8, across the two subsamples is only valid when this regression model is valid. For example,
unemployment may not be linear, and perhaps filters as As are needed instead of A I , see
Sections I11 and IV.
MODELINGSEASONALITY IN TIME
SERIES 56 I
tied together. In other words, although the seasonal fluctuations seem to change over
time, they do so not that quickly.
To obtain a tentative impression of how seasonality can change over time, I
consider the estimates of 6,, s = 1 , 2 , . . . , S, from the auxiliary regression
where the values of i are set such that there is no autocorrelation in E t . In Table 2,
I report minimum and maximum values of 8, and the standard deviation of the S 8,
parameters for two subsamples which are roughly similarly large. For consumption
nondurables one can observe that the maximum value for 8, is obtained for the sec-
ond quarter in the first sample, while it corresponds with the fourth quarter in the
second sample. For unemployment, these extreme 8, values correspond to the same
seasons, although the standard deviation decreases with about 60%.For the Span-
ish tourism data, August loses its importance to July toward the end of the sample.
For illustrative purposes, I depict the estimates for the twelve 6, parameters in the
two subsamples in Figure 5. One can observe that tourism in Spain seems to shift
slightly from the summer months more toward the winter months January to March,
suggesting structural shifts in the behavior of tourists.
Figure 5 Estimates of monthly dummy parameters for 1979.01-1986.12 (Part 1) and for
1987.01-1993.12 (Part 2).
562 FRANSES
Finally, for radio advertising expenditures in Table 2 it seems that indeed

“winter” becomes “summer”; i.e., the dip in January moves to about July/August,
while the peak changes from June/July to about February. Furthermore, as could
be expected from Figure 4, the variance of seasonality increases from 0.06 to
0.18, i.e., with 300%.
Notwithstanding the results for the advertising series, it seems that the sea-
sonal fluctuations in economic time series change over time only at a slow pace.
This implies that the airline model in (1) may assume too many changes in season-
ality since it assumes a large number of unit roots. In the next section I will discuss
two important approaches that may overcome the conceptual problems with (l),and
which can yield descriptions of economic data that allow for slowly changing sea-
sonality.
111. M O D E L I N G STRATEGIES
Broadly speaking, there are two general strategies to model seasonal economic time
series which display slowly changing seasonality. The first is to impose (depending
on formal test results) one or a few seasonal unit roots on the AR part of a model, and
the second is to allow the AR parameters to vary across the seasons. For brevity, I call
the first approach “seasonality in dynamics,” and discuss this in Section 1II.A and
call the second approach “seasonality in parameters” (Section 1II.B). Recent studies
in Boswijk, Franses, and Haldrup (1997) and Ghysels, Hall, and Lee (1996) seem
to suggest that combinations of the two approaches do not obtain much empirical
support. This section only highlights the main issues, and for further more detailed
overviews the reader is referred to, for example, Franses (1996a, b).
A. Seasonality in Dynamics
Consider the AR(p) model for a time series with seasonal frequency S:
yt = 41yt-1 + 42Yt-2 + +
* * . 4pYt-p + Et (5)
where typically p exceeds S. First, I focus on univariate yt processes, while the sec-
ond part of this section allows yL to be an rn x 1 vector process.
I. Univariate Time Series

When the AR polynomial in (5) denoted as @,(B) can be decomposed as
as in (l),the yt process is said to be both integrated (because of 1 - B ) and sea-

sonally integrated (because of l - Bs). In this case, yt contains S + l unit roots.
MODELING
SEAsONALlTY IN TIME
SERIES 563
The polynomial 1 - BS can be decomposed as (1 - B)(1 B + + +

BS-'), and
yt is an 1(2) process since its AR model contains the component ( 1 - B)2 and it
has S - 1 so-called seasonal unit roots. The latter unit roots correspond with the
stochastic trends at the seasonal frequency, and hence can be relevant in case the
seasonal pattern in yt changes over time. Notice that these changes are permanent.
For example, consider the simple process
yt = -yt-1 +EL t 7)
which in case S = 4 is called a process with a seasonal unit root at the biannual
frequency, see Hylleberg et al. (1990). Substituting lagged yt in (7) results in
yt = E t - Et-1 + Et-2 - Et-3 + * * (8)
and hence the variance of yt at t equals to2,which is the same as that for a standard
random walk. Shocks E~ in (7) cause permanent changes to the pattern of yt because
their effect does not die out as seen from (8), and because of the seasonal unit root,
the seasonal pattern in yt changes permanently.
Given the conceptual impact of the assumption of seasonal unit roots, and
given that unit roots lead to wider forecasting intervals and to more involved model-
ing strategies for multivariate time series in the next step, it is important to test for
the number of seasonal and nonseasonal unit roots in a univariate time series. The
two commonly applied test procedures are proposed in Osborn et al. (1988)and in
Hylleberg et al. (1990)[HEGY]. The OCSB method investigates whether the #,(I?)
polynomial contains the components 1 - B , 1 - B", both, or none of these. The test
regression is
q(mA1 Asyt = Cct + Y1 Asyt-1 + Y2AlYt-s + Et (9)

where q ( B ) is an AR polynomial, and
s-1 s- 1
Cct = a0 + asDs,t + Pot + PsDs,t t
s= 1 s= I
where t reflects a deterministic trend. OCSB consider the case with S = 4 and where
POto Ps-1 are set equal to zero. Critical values for other cases are given in Franses
and Hobijn (1997). When y1 = y2 = 0, the A1 A s filter is appropriate, when y1 = 0
and y2 is not, the A1 filter is needed, when y2 = 0 and y1 is not the A s filter, and
no filter is required when both y1 and y2 are unequal to zero. When the B's in (10)
are not set equal to zero, and when these also appear useful in a model for A1 or
A s filtered yt, this yt series in fact shows increasing seasonal variation. Such models
may be useful when one does not want to use the logarithmic data transformation
(e.g., Bowerman, Koehler, and Pack 1990).
As an example, an application of this OCSB method to the unemployment data,
where S = 4, q ( B ) contains one lag, and all P's in pt are set to zero, yields the t -
values t ( P 1 ) = 0.006 [-2.111 and t ( P 2 ) = -4.384 [-3.751, where 5% critical
564 FRANSES
values are given in brackets. Other choices for pcdo not change the conclusion that
yt appears to contain the nonseasonal unit root 1 since P1 is insignificant and that
unemployment does not need the double-differencing filter.
An application to radio advertising expenditures, where now a trend is in-
cluded in (9) and q ( B ) = 1, results in t ( P 1 ) = -4.913 [-2.781 and t ( P2 ) =
-5.506 [-5.811. Hence, for this four-weekly series it is found that the As filter may
be needed since t ( P 2 ) does not appear significant. In sum, for both series the A l a s
filter assumes to many unit roots. It should be mentioned that the empirical work in
Osborn (1990)convincingly shows that there are only very few quarterly time series
for which the A1 A 4 filter is required to obtain stationarity.
Since the As filter assumes S - 1 seasonal unit roots, it is now relevant to test
how many of these roots are present for a certain time series yL.A commonly applied
method for this purpose is (a variant of) the HEGY approach. For S = 4 this method
concerns the auxiliary regression
where M ( B ) = 1 B + + + +
B2 B3 and A(B) = -(1 - B)(1 B 2 ) . HEGY give
asymptotics and critical values for S = 4 and pcwith Bs = O(s = 1 , . . . , S - l),
Smith and Taylor (1995) allow for seasonally varying trends in case S = 4, and
Franses and Hobijn (1997) consider additional cases. The key focus is on the n l to
n 4 parameters in (11) since nl = 0 implies 1 - B, n 2 = 0 implies 1 + B, and
+
n 3 = n 4 = 0 implies 1 B2, where the latter two seasonal unit roots f zcorrespond
with the annual frequency. Typically, one considers t-tests for n1 and n 2 and a joint
F-test for n3 and n 4 . One may also use the joint F-tests for nl to n 4 or n 2 to n 4 ,
where the latter concerns all seasonal unit roots (Ghysels, Lee, and Noh 1994).
An application of the HEGY method to the German unemployment data (where
q ( B ) in (11) includes two lags, and pcdoes not include any trends) results in t(fi.1) =
1.292 [-2.831, t ( f i 2) = -1.914 [-2.831, and F ( f i . 3 , 3 4 ) = 8.380 [6.62], where
again 5%critical values are given in brackets. Hence, to remove the nonstationar-
+
ities in this series, one needs the (1 - B)( 1 B ) = 1 - B2 filter. In other words,
one may describe the changing seasonal variation in unemployment using a seasonal
unit root at the biannual frequency.
For the Spanish tourism data, the application of an extension of the HEGY
regression (ll),with Bo in (10) unequal to zero and q ( B ) = 1, results in t ( f i 1 ) =
-1.161 [-3.291, t(fi.2) = -4.799 [-2.761, and F ( f i 1 , ., iz l2 ) = 18.835 [4.46].
Hence, for this monthly series only the A , filter is required. This implies that the re-
sults in Table 1 and Figure 5 can be interpreted safely since the auxiliary regression
model cannot be rejected by the data.
An application of the HEGY method to many quarterly U.K. time series (also
including consumption nondurables) in Osborn (1990)yields that often only one or
two seasonal unit roots are present. This finding, which appears typical across em-
MODELING
SEASONALllY IN TIME
SERIES 565
pirical applications of the HEGY method (see, for example, Hylleberg, Jgrgensen,
and Sgrensen 1993, Ghysels, Lee, and Siklos 1993, and Lee and Siklos 1993), im-
plies that the double-differencing filter A ] As as assumed in (1) imposes too many
unit roots on the AR polynomial in (5). However, as shown in Ghysels, Lee, and
Noh (1994),in case (1) is the true data-generating process in simulation exercises,
the empirical size of the HEGY test largely exceeds the nominal size, implying that
one is inclined to find not enough unit roots. Furthermore, it can be shown that in-
appropriate lag augmentation in regressions as (1 1) can have a large impact on the
empirical outcomes, Finally, if p in (5) is smaller than S, the HEGY and OCSB ap-
proaches may be difficult to apply. Additional research is needed to fully under-
stand the theoretical and empirical properties of both test methods. For example,
Ghysels, Lee, and Siklos (1993) examine the effect of lag selection of HEGY test
results.
2. Multivariate Time Series

In case the yl process in (5) is an rn x 1 vector time series, and hence when the
41 to q5p are M x m. matrices, and the components, say, y ~to, ym,t ~ contain one or
more seasonal or nonseasonal unit roots, one may wish to decrease the number of
parameters in a vector AR (VAR) by investigating common properties across time
series. For example, if two quarterly series y1 , 1 and ~ 2 have , ~a common seasonal unit
+
root -1, i.e. when they both have a 1 B component in their univariate AR mod-
els, these two series are seasonally cointegrated at the biannual frequency if a linear
combination Gl(B)yl,, - c ~ G z ( B ) y z(where
,~ +
the G i ( B ) do not contain 1 B) does
not have this seasonal unit root (Engle et al. 1993). When both variables require
+
the A4 filter, the polynomials G1 ( B ) and G z ( B ) equal 1 - B B2 - B 3 . When for
example ~ 1 only, ~ requires the 1 - B2 filter, G1 ( B ) equals 1 - B. Hence, the trans-
formations G ; ( B ) (i = 1, 2) can differ across time series, and they usually depend
on the HEGY test results for univariate series.
In principle, the test method for seasonal cointegration in EGHL concerns
regressions of G1 (B)yl,, on p l and G ~ ( B ) ~and L , ~an, analysis of the seasonal and
nonseasonal unit root properties of the estimated residuals. Lee (1992) extends the
Johansen method to test for cointegration to the seasonal case (S = 4) by investi-
gating the ranks of the rn x rn matrices 7rl to n4 in a VAR model similar to (11).
An application to several data sets for various countries is given in Kunst (1993).
In Franses and Kunst (1995) the impact of seasonal dummies on the properties of
seasonally cointegrated variables is discussed. It can be shown that such seasonal
dummies can imply the data to have seasonally varying trends. Since this feature
may not be present in many log-transformed time series, Franses and Kunst (1995)
propose to check the cointegration properties at each of the frequencies separately.
Finally, a Granger representation theorem for seasonal cointegration is given in Jo-
hanseri and Schaumburg (1996).
566 FRANSES
B. Seasonality in Parameters
The second currently dominant approach in modeling and forecasting economic time
series with seasonality concerns so-called periodic models. Since the studies in Os-
born (1988) and Osborn and Smith (1989), these models have become increasingly
popular in economics. This section surveys some key aspects of periodic models for
univariate and multivariate time series. A full account of the literature and of recent
developments in periodic models for nonstationary seasonal time series is given in
Franses (1996b). For ease of notation, I confine most of the discussion to quarterly
time series (S = 4).
I, Univariate Time Series
Consider the simple first-order AR model with seasonal varying parameters

4 4
s= 1 s= 1
where ps are seasonal intercepts and the 4~7 denote AR parameters that are allowed
to vary with the season. Model (12) represents a periodic AR process (PAR) of order
1. This PAR(1) is the simplest case of the more general PAR(p) class of models, but
for the present discussion expression (12) suffices to highlight some of the properties
of periodic AR models. Franses (1996b) focuses at great length on more elaborate
models.
For the periodic process in (12) with 4.s# 4 for all s = 1, 2, 3 , 4 , it is clear
that the observations in each of the four seasons are described by a different model.
Denoting Ys,r as the observation on yc in quarters in year T , (12) implies that Y ~ , J= -
p4 -I- 4 4 Y 3 , T +
E ~ , Tand for example Y I , T = pl + +
@ ~ Y ~ , J - - I E I , T . These expres-
sions show that the models for the annual time series Y s , have ~ constant parameters.
In fact, model (12) can be written as
1 0 0 -
[
-42 1 0
0 -43 1
0 0 -44
where B Y ~ , =J Y ~ , J - - IThe
. model for the 4 x 1 vector process YJ- containing Y ~ , to
T
Y ~ , Tis convenient to analyze the unit root properties of yl since (13) represents the
same time-series data as (12) does. The characteristic equation of (13)is
implying that yl can contain a single unit root when 6 1 4 2 4 3 4 4 = 1. Obviously,

when all (6, = 1, which is the standard random walk case, yLhas the unit root 1, and
also the YT process has one unit root. When all @,7 = -1, yl has the seasonal unit
MODELING
SEASONALln I N TIMESERIES 567
root -1, and the YT process again has one unit root. Hence, seasonal unit roots in
yl correspond with regular unit roots in the YT process. Given this correspondence,
it seems natural for a PAR process first to test whether @1@2@344 = 1 and next
to test restrictions on the 4sparameters. Boswijk and Franses (1996) show that the
first test follows a Dickey-Fuller distribution, and that the second step involves just
x2 asymptotics. Franses and Paap (1996a) show through simulation and through
forecasting empirical series that this two-step method yields useful results. Boswijk,
Franses, and Haldrup (1997) extend this method to allow for more general structures
involving 1(2)processes and more seasonal unit roots. Ghysels, Hall, and Lee (1996)
focus on testing for restrictions as 4s = - 1 in (12 ) in one step.
In case $1$2@344 = 1in (12) and $.v # 1or -1 for all s = 1,2,3,4, the yl pro-
cess is said to be periodically integrated (Osborn 1988, Franses 199613). Otherwise
formulated, yl requires a periodic differencing filter 1 - q5sB with $1&?@3$4 = 1
to remove the stochastic trend. Notice that these q5s parameters have to be estimated
from the data, and that some 4svalues will exceed 1. Typical values for $s for quar-
terly data are within the range of 0.8 and 1.2. For example, the estimation results
for a PAR(1) with the restriction 41@2$344 = 1 for the U.K. consumption durables
sample series are $1 = 1.003(0.008),$2 = 0.932 (0.007),$3 = 1.030(0.008), and
$4 = 1.039 (0.008),with estimated standard errors in parentheses. Note that these
standard errors underestimate the true standard errors since the 4sparameters are
estimated superconsistently, see Boswijk and Franses (1996).An F-test for the re-
striction 4s= 1 obtains the value of 31.61 7, and this is clearly significant at the 5%
level. Hence, U.K. consumption nondurables appears to be a periodically integrated
process; see also Osborn (1988) where it is shown that this finding is consistent with
a modified economic theory and Franses (1996b) for some results on forecasting such
senes.
For many sample series considered in Franses (1996b) it is found that the pe-
riodic differencing filter 1 - 4, B is appropriate to remove the unit root. This finding
appears robust to data transformations and structural breaks. One of the main impli-
cations of periodic integration is that seasonality changes, which can be illustrated
by writing (12) without the p, as
4 4
s= 1 s= 1
4
+ 4s4s-14s-2Ds,lEl-3
s= 1
which after taking first differences and with @~T values close to unity can be approx-
imated by
4
s= 1
568 FRANSES
When model (15) is estimated under the assumption that qs = q for all s, it is very
similar to the airline model in (1). Hence, periodically integrated time series may
seem adequately described by the airline model. Note that this does not hold the
other way around. In fact, for the consumption series, such a model is estimated in
(2). In general, it can be shown that neglecting periodic parameter variation increases
the lag order (Osborn 1991, Tiao and Grupe 1980);i.e., there appears to be a trade-
off between lags in nonperiodic models and the number of intrayear parameters in
periodic time-series models. Furthermore, neglecting periodicity reduces the power
of nonseasonal unit root tests (Franses 1996b), and it may lead to the finding of
spurious seasonal unit roots (Boswijk and Franses 1996).
The application of periodic integration is not necessarily restricted to periodic
models as (12). In fact, Bollerslev and Ghysels (1996)introduce the stationary peri-
odic GARCH model to describe seasonally observed financial time series. Franses
and Paap (1995) extend this model to allow for persistence of volatility shocks in
order to describe the stylized fact that for many daily financial time series the vari-
ance appears to change slowly over time. Franses and Paap (1995) fit the following
PAR(p)-PIGARCH(1, 1) model to daily returns on the Dow-Jones index (for about
4000 daily observations):
P
i= 1
Et cv “0, of)
of = 0, + CKSEf-, + pot-,
2
(s = 1,2,3,4,5)
under the restriction that
5
n(a, + j3) = 1 with C I , ~# a (s = 1 , 2 , 3 , 4 , 5 ) (19)
s= 1
It is found that for example B1 +6

= 1.008 and &2 +6
= 0.956, reflecting the
well-documented importance of Monday stock returns over Tuesday returns.
A second important aspect of periodic integration is that the trend/cycle com-
ponent and the seasonal component in an economic time series interact. This is ob-
vious from the fact that the parameters in the differencing filter 1 - 4sBdiffer
@~$
across the seasons; see Franses (1996~) for more details. Hence, seasonal adjust-
ment methods as Census X-11, which treat all seasons in an equal fashion, do not
remove the intrinsic periodicity in a time series. As shown in Franses (1996b) it
is possible to fit periodic models to seasonally adjusted data if the underlying time
series shows dynamic periodicity. Strictly speaking, it does not make sense to sea-
sonally adjust periodic time series since the key assumption for seasonal correction
is that one can isolate the seasonal from the nonseasonal component. As a possi-
ble consequence of this conceptual problem, there is evidence that the NBER peaks
MODELING
SEASONALITY I N TIMESERIES 569
and troughs display seasonality, even though these dates are set using seasonally ad-
justed data (Ghysels 1994b). In fact, Franses (1996b) shows that seasonally adjusted
periodically integrated time series can generate such features. Additional evidence
for the apparent link between seasonal fluctuations and the trend/cycle is presented
in Barsky and Miron (1989),Beaulieu, MacKie-Mason, and Miron (1992), Canova
and Ghysels (1994),Franses (1995),and Miron (1996).
Finally, a feature of periodic integration in connection with the intercept pa-
rameters p, in (12) is that this allows a description of variables with increasing sea-
sonal variation without taking logs. This may be useful for such data as the trade
balance, and other economic data that may take negative values. A key drawback
of periodic AR models, however, is that the number of parameters increases quite
rapidly if S and p increase. For example, a PAR(2) model as (12) for monthly data
involves the estimation of 24 parameters. Hence it may be useful to impose certain
parameter restrictions (Anderson and Vecchia 1993).
2. Multivariate Time Series

Obviously, the number of parameters in multivariate periodic models increases quite
rapidly, see Lutkepohl(l991) and Ula (1993)for periodic VARs for time series with-
out unit roots. For example, when yt in (12) is rn x 1, and all parameters are allowed to
vary with the season, this periodic VAR(1) contains 4rn2 parameters. An additional
drawback of this representation is that it is quite complicated to investigate the unit
root properties of the vector yL process since the expressions for the characteristic
equations similar to (14) are not as simple.
One possibility is to resort to an alternative representation of periodic VARs,
as is done in Kleibergen and Franses (1995). This representation assumes that long-
run cointegration relations across variables obey certain restrictions. If one wants
to allow for more flexibility, one may consider the second possibility, which is to
impose more structure on the dynamics and also to focus on only a single equation
(Birchenhall et al. 1989, Boswijk and Franses 1995).A simple example of a resultant
model for a bivariate time series is
where the equilibrium and adjustment parameters are allowed to vary with the sea-
son s = l , 2, 3 , 4. Boswijk and Franses (1995) propose an empirical strategy for
this so-called periodic cointegration model. Franses (1996b) illustrates that models
such as (20) can also generate time series for which the trend/cycle and the seasonal
fluctuations are dependent.
At present, the literature on multivariate periodic models is not extensive, and
much more research is needed into the properties of such models and into the design
of useful empirical methods for more general cases than (20).
570 FRANSES
IV. FURTHER RESEARCH TOPICS
The issue of investigating seasonal variation in economic data has gained much in-
terest in the last few years, and this has resulted in the two approaches discussed
in the previous section. There are also several studies in which seasonality is in-
corporated explicitly into economic theory, e.g., Osborn (1988),Hansen and Sargent
(1993),Todd (1990), and Miron and Zeldes (1988).There are, however, many issues
for current and future research, and in this section I will highlight only a few of these.
A. Structural Breaks
The time-series models in the previous section assume that seasonality changes be-
cause of shocks; i.e., the changes are stochastic. It may, however, be that the changes
are deterministic. Institutional changes may make economic agents to start behaving
differently in certain seasons. For example, allowing country regions to fix their own
school holiday periods may reduce variation in tourism spending. Another example
is the introduction of a new TV and radio broadcasting channel that can change the
structure of the market for advertising expenditures. Such a new radio channel was
introduced in The Netherlands in 1989, around observation 154 in Figure 4. It is
clear from this graph that seasonality in advertising expenditures changes dramati-
cally. The source of this change appears to be a new pricing policy, which in turn is
due to changing pricing policies for TV. This change is seasonality can be said to be
deterministic.
Following the arguments in Perron and Vogelsang (1992), it can be expected
that neglecting changing parameters in deterministic seasonal dummies biases sea-
sonal unit root tests toward nonrejection; see Franses and Vogelsang (1997) for formal
results. Furthermore, such changes bias tests for periodicity toward the alternative;
i.e., too much periodicity is found, as shown in Franses (1996b).As an example, for
the advertising series, when the OCSB regression in (9)is enlarged with 13 seasonal
dummies for the period from observation 154 onward, the t-tests obtain the values
t ( P 1 ) = -2.497 [-2.861 and t ( P 2 ) = -1 1.358 [-7.501, where the 5% critical val-
ues are from Franses and Hobijn (1997). Hence, the previous conclusion that radio
advertising needs a A13 filter changes to the necessity of only the A1 filter. All 1 2
seasonal unit roots disappear when one allows for deterministic shifts.
For many economic time series the location of a break that affects tests for
seasonal unit roots is unknown. If one suspects such breaks, one may then use the
extreme values of the various t - and F-tests to search for a break data. Asymptotic
theory for this approach is presented in Franses and Vogelsang (1997). If one does
not like the idea of searching for possibly inconveniently located breaks, one may
wish to use this method in order to investigate the robustness of the outcomes of
seasonal unit root tests. In fact, Paap, Franses, and Hoek (1997) show through simu-
lation experiments that making mistakes in either direction yields highly imprecise
forecasts.
MODELING
SEASONALITY I N TIME
SERIES 57 I
Further research is needed to investigate the robustness of empirical findings

on periodicity and seasonal unit roots to structural shifts in seasonal means. Since
Ghysels and Perron (1996) document that seasonal adjustment tends to obscure
structural breaks, it may be that the approach in Franses and Vogelsang (1997) can
also be usefully applied for unadjusted variables to test for nonseasonal unit roots
in the presence of seasonal unit roots and mean shifts. Finally, in case determinis-
tic seasonal mean shifts occur quite frequently, it seems relevant to investigate the
robustness of the empirical regularities summarized in Miron (1996).
B. Time-Varying Parameters
Allowing for structural mean shifts can be helpful to detect exactly when seasonality
changes, if it does. Seasonal unit root and periodic integration models may not be
very helpful to decide in which part of the sample the seasonal variation changes.
Hence. in order to be able to understand more clearly what economic behavior causes
such changes, one may consider models that are in between models with seasonal
unit roots and deterministic shifts. For example, it may be useful to consider
4
s= 1
where uc is some ARMA process and 6,s,Lare time-varying parameters for the sea-
sonal dummies. The 6.s,lcan be made functions of time, of lagged 6S,l or of economic
variables. Consequently, one can extend the HEGY approach to allow for more flex-
ible structures in the pulterm in (10).
Recent examples of flexible structures for seasonal variation are given in An-
dersen and Bollerslev (1994), Harvey and Scott (1994), and Canova and Hansen
(1995). Hylleberg and Pagan (1997) put forward the so-called evolving seasonals
model that amounts to an intermediate case between the models in (4) and the sea-
sonal unit root models. This model appears useful to explain the simulation results
in Hylleberg (1995),where the HEGY test appears better in some cases and the
Canova-Hansen test in others. The two null models in these tests are both spe-
cial cases of the evolving seasonals model. Further research is needed to inves-
tigate the practical usefulness of more flexible structures with respect to those in
Section 111.
C. Nonlinear Modeling
It may be possible and important to introduce even more flexible structures by allow-
ing yt to be described by nonlinear time-series models, while taking care of seasonal-
ity. For example, Ghysels (199413) finds that regime shifts in the business cycle tend
to occur more frequently in some seasons than in others. Additionally, Canova and
Ghysels (1994) and Franses (199s)find that some macroeconomic variables show
572 FRANSES
0 2 Q3
Figure 6 Quarterly output of hidden layer units.
different seasonality in expansions than in recessions, which in turn may depend on

when the regime switches occur. In these studies the business cycle is defined by
the NBER indicators, but one may want to estimate these dates where seasonality
changes using the same data. Given the periodic structure of the business cycle, it
is proposed in Ghysels (1993)to extend the Hamilton-type Markov switching model
to allow for seasonality in the transition probabilities.
Such nonlinear models may also be helpful to decide when seasonality is
changing. For example, Franses and Draisma (1997) propagate the linear neural
network model with q hidden layers as
Yt = Pt + 41yt-1 + + 4pyt-p
* * *
j= 1
where G is the logistic activation function; see Kuan and White (1994) for an overview
of neural network models. Although the parameters in the nonlinear component can-
not be interpreted, Franses and Draisma (1997) use the B;GG(.) components (for each
of the seasons) to investigate the contribution of the hidden layer components. For
example, consider again the unemployment data for Germany. When pt in (22) is
as in (10) with 8 2 to 84 set to zero and p in (22) equals 4, the Schwarz criteria for
q = 0, 1 , 2 , 3 , are -756.3, -790.8, -809.5, and -789.0, respectively. Hence the
neural network model with q = 2 hidden layer units is selected, and German unem-
ployment shows nonlinear features.
In Figure 6, I depict the impact of the two hidden layer units (HI and H2) in
each of the four quarters. The dotted line is the time series in each of the seasons
( Y s , ~and
) , the straight line is the contribution of the relevant hidden layer output to
the final output yt. It is clear that H I is active only in quarters 2 and 4 (Q2 and Q4),
MODELING
SEASONALITY IN TIMESERIES 573
where for 42 this activity ends around 1975 and for Q 4 it ends around 1985. Hidden
layer 2 appears active only in quarters 1 and 4. These results show that the time
series in quarter 3 seems linear and that changes in certain seasons occur around
1975 and 1985.
It is now interesting to see whether the nonlinear structures in these German
unemployment data are robust to seasonal adjustment. For this purpose, I estimate
the neural network model in (22) for the official adjusted data. For q = 0, 1,2,3, I ob-
tain Schwarz criteria values of -918.7, -892.1, -867.7, and -855.9, respectively.
Hence, seasonal adjustment appears to affect the apparent nonlinear features of this
time series, at least when represented by the highly parameterized neural network
model in (22). See Granger and Terasvirta (1993) for more parsimonious nonlinear
models. Ghysels, Granger, and Siklos (1996) document that Census X-1 1 seasonal
adjustment appears to introduce nonlinear features in otherwise linear data. Finally,
Franses and Paap (1996b) show that seasonal adjustment may leave nonlinearity in-
tact but that it changes some key parameters. Obviously, it seems useful to study the
properties of seasonal adjustment with respect to nonlinearity. Also, it seems of great
importance to consider models that explicitly describe nonlinearity and seasonality
at the same time, possibly along the lines of Lewis and Ray (1996).
D. Economics
A fourth important topic for further research considers how one can design economic
models that can describe the behavior of economic agents which cause macroeco-
nomic aggregates to show slowly changing seasonality over time. As mentioned, sev-
eral studies incorporate seasonality somehow, but to my knowledge there are no stud-
ies that deal with the question of why seasonality changes; i.e., why do economic
agents endogenously change their seasonal behavior? Furthermore, can (seasonal or
periodic) unit root models generate time series that really mimic economic behavior?
One possible route to follow may be to focus on consumer expectations. An em-
pirical analysis of confidence indicators in Ghysels and Nerlove (1988)and Franses
(1996b) shows that, even when agents are asked to remove seasonality by focusing on
annual trends, the indicators display marked seasonality. In fact, for most countries
one can nowadays obtain only seasonally adjusted consumer confidence indicators,
which somehow may seem counterintuitive.
Another research strategy, which may involve game-theoretic aspects, is to
investigate if seasonality changes because economic agents mistake a set of large
shocks as a precursor for “new” seasonality, and hence start to behave like that.
Many important changes for, e.g., Germany, occurred in the fourth quarter (1966.4:
economic crisis, 1973.4 and 1979.4 dramatic increase in oil prices and 1989.4:uni-
fied Germany). Hence, theoretical models where the arrival rate of shocks can have
an impact on economic behavior may be useful.
574 FRANSES
V. CONCLUDING REMARKS
This chapter surveys recent and possible future research issues in modeling eco-
nomic time series with seasonality. There is a growing interest in modeling season-
ality and not in removing it through some seasonal adjustment methods. In areas as
marketing and tourism, seasonality itself is the focus of investigation. In macroeco-
nomics there is a tendency to apply Census X-11 type methods to remove seasonal-
ity. However, recent empirical studies have shown the large number of drawbacks of
seasonally adjusted data. Furthermore, other studies have shown that it is not that
difficult to model seasonality explicitly.
From a statistical point of view, there appears a consensus that not too many
stochastic trends can be found in economic data. In other words, although seasonal
variation changes over time, it changes quite slowly. The tools for analyzing seasonal
time series can be refined in the direction of nonlinear models or alternative flexible
structures. From an economic point of view, and especially in case of macroeco-
nomics, there is a need to understand why such seasonal patterns change over time.
ACKNOWLEDGMENTS
I thank the Royal Netherlands Academy of Arts and Sciences for its financial sup-
port. Dick van Dijk, Richard Paap, and an anonymous referee provided helpful com-
ments.
REFERENCES
Andersen, T. G. and T. Bollerslev (1994), Intraday Seasonality and Volatility Persistence in

Foreign Exchange and Equity Markets, Working Paper 186, Northwestern University,
Evanston, IL.
Anderson, P. L. and A. V. Vecchia (1993), Asymptotic Results for Periodic Autoregressive
Moving-Average Models, Journal of Time Series Analysis, 14, 1-18.
Barsky, R. B. and J. A. Miron (1989), The Seasonal Cycle and the Business Cycle, Journal of
Political Economy, 97, 503-535.
Beaulieu, J. J., J. K. MacKie-Mason, and J. A. Miron (1992), Why Do Countries and Industries
with Large Seasonal Cycles Also Have Large Business Cycles?, Quarterly Journal of
Economics, 107,621-656.
Bell, W. R. (1987), A Note on Overdifferencing and the Equivalence of Seasonal Time Series
Models with Monthly Means and Models with (0, 1, 1)12 Seasonal Parts When 0 = 1,
Journal of Business and Economic Statistics, 5, 383-387.
Bell, W. R. and S. C. Hillmer (1994), Issues Involved with the Seasonal Adjustment of Eco-
nomic Time Series (with discussion), Journal of Business and Economic Statistics, 2,
291-320.1984.
MODELING
SEASONALITY IN TIMESERIES 575
Birchenhall, C. R., R. C. Bladen-Hovell, A. P. L. Chui, D. R. Osborn, and J. P. Smith (1989),

A Seasonal Model of Consumption, Economic Journal, 99, 837-843.
Bollerslev, T and E. Ghysels (19%), Periodic Autoregressive Conditional Heteroskedasticity,
Journal of Business and Economic Statistics, 14, 139-151.
Boswijk, H. P. and P. H. Franses (1995), Periodic Cointegration: Representation and Infer-
ence, Review of Economics and Statistics, 77,436-454.
Boswijk, H. l? and P. H. Franses (1996), Unit Roots in Periodic Autoregressions, Journal of
Time Series Analysis, 17,221-245.
Boswijk, H. P., P. H. Franses, and N. Haldrup (1997), Multiple Unit Roots in Periodic Autore-
gression, Journal of Econometrics, 80, 167-193.
Bowerman, B. L., A. B. Koehler, and D. J. Pack (1990), Forecasting Time Series with Increas-
ing Seasonal Variation, Journal of Forecasting, 9,419-436.
Box, G. E. P. and G . M. Jenkins (1970), Time Series Analysis; Forecasting and Control, Holden-
Day, San Francisco.
Canova, F. and E. Ghysels (1994),Changes in Seasonal Patterns: Are They Cyclical?, Journal
of Economic Dynamics and Control, 18, 1143-1171.
Canova, F. and B. E. Hansen (1995), Are Seasonal Patterns Constant over Time?, A Test for
Season Stability, Journal of Business and Economic Statistics, 13,237-252.
Engle, R. F., C. W. J. Granger, S. Hylleberg, and H. S. Lee (1993), Seasonal Cointegration:
The Japanese Consumption Function, Journal of Econometrics, 55,275-298.
Franses, P. H. (1995), Quarterly USA Unemployment: Cycles, Seasons, and Asymmetries,
Empirical Economics, 20, 717-725
Franses, P. H. (1996a), Recent Advances in Modelling Seasonality, Journal of Economic Sur-
veys, 10,299-345.
Franses, P. H. (19%b), Periodicity and Stochastic Trends in Economic Time Series, Oxford
Franses, P. H. (1%6c), Multi-step Forecast Error Variances for Periodically Integrated Time
Series, Journal of Forecasting, 15,83-95.
Franses, P. H. and G. Draisma (1997), Recognizing Changing Seasonal Patterns Using Artifi-
cial Neural Networks, Journal of Econometrics (to appear).
Franses, P. H. and B. Hobijn (1997), Critical Values for Unit Root Tests in Seasonal Time
Series, Journal of Applied Statistics, 2 4 , 2 5 4 7 .
Franses, P. H. and R. Kunst (1995), On the Role of Seasonal Intercepts in Seasonal Cointe-
gration, Institute for Advanced Studies, Vienna.
Franses, P. H. and R. Paap (1995).Modeling Changing Day-of-the-Week Seasonality in Stock
Returns and Volatility, Econometric Institute Report 9556, Erasmus University Rot-
terdam.
Franses, P. H. and R. Paap (1996a), Periodic Integration: Further Results on Model Selection
and Forecasting, Statistical Papers, 37,33-52.
Franses, P. E. and R. Paap (1W6b), Does Seasonal Adjustment Change Inference from Markov
Switching Models?, Econometric Institute Report 9615, Erasmus University Rotter-
dam.
Franses, P. H. and T. J. Vogelsang (1997), On Seasonal Cycles, Unit Roots and Mean Shifts,
Review of Economics and Statistics (to appear).
Ghysels, E. (1993), A Time Series Model with Periodic Stochastic Regime Switching, Unpub-
lished manuscript, University of Montreal.
576 FRANSES
Ghysels, E. (1994a),On the Economics and Econometrics of Seasonality, in C. A. Sims (ed.),

Advances in Econometrics, Sixth World Congress of the Econometric Society, Cambridge
University Press, Cambridge, UK.
Ghysels, E. (1994b),On the Periodic Structure of the Business Cycle, Journal ofBusiness and
Economic Statistics, 12,289-298.
Ghysels, E., C. W. J. Granger, and P. L. Siklos (1996), Is Seasonal Adjustment a Linear or
Nonlinear Data-Filtering Process?, Journal of Business and Economic Statistics, 14,
374-386.
Ghysels, E., A. Hall, and H. S.Lee (1996), On Periodic Structures and Testing for Seasonal
Unit Roots, Journal of The American Statistical Association, 91, 1551-1559.
Ghysels, E., H. S.Lee, and J. Noh (1994), Testing for Unit Roots in Seasonal Time Series,
Journal of Econometrics, 62,415-442.
Ghysels, E., H. S.Lee, and P. L. Siklos (1993), On the (Mis)Specification of Seasonality and
Its Consequences: An Empirical Investigation with US Data, Empirical Economics, 18,
747-760.
Ghysels, E. and M. Nerlove (1988), Seasonality in Surveys; A Comparison of Belgian, French
and German Business Tests, European Economic Review, 32,81-99.
Ghysels, E. and P. Perron (1996), The Effects of Linear Filters on Dynamic Time Series with
Structural Change, Journal of Econometrics, 70,69-98.
Gonzdez, P. and P. Moral (1995), An Analysis of the International Tourism Demand in Spain,
International Journal of Forecasting, 11,233-251.
Granger, C. W. J. and T. Terasvirta (1993),Modelling Nonlinear Economic Relationships, Ox-
ford University Press, Oxford.
Hansen, L. P. and T. J. Sargent (1993), Seasonality and Approximation Errors in Rational
Expectations Models, Journal of Econometrics, 55,21-56.
Harvey, A. and A. Scott (1994), Seasonality in Dynamic Regression Models, Economic Jour-
naZ, 104,1324-1345.
Hylleberg, S.(1986), Seasonality in Regression, Academic Press, Orlando.
Hylleberg, S.(1992), Modelling Seasonality, Oxford University Press, Oxford.
Hylleberg, S.(1994), Modelling Seasonal Variation, in Hargreaves, C. P. (ed.), Nonstationary
Time Series Analysis and Cointegration, Oxford: Oxford University Press.
Hylleberg, S.(1995),Tests for Seasonal Unit Roots: General to Specific or Specific to General?,
Hylleberg, S., R. F. Engle, C. W. J. Granger, and B. S.Yoo (1990), Seasonal Integration and
Cointegration, Journal of Econometrics, 44,215-238.
Hylleberg, S., C. Jdrgensen, and N. K. Sdrensen (1993), Seasonality in Macroeconomic Time
Series, Empirical Economics, 18,321-335.
Hylleberg, S. and A. R. Pagan (1%7),Seasonal Integration and the Evolving Seasonals Model,
International Journal of Forecasting (to appear).
Johansen, S.and E. Schaumburg (1996),Inference for Seasonally Cointegrated Autoregressive
Processes, Mimeo, University of Copenhagen.
Kleibergen, F. and P. H. Franses (1995), Direct Cointegration Testing in Periodic Vector Au-
toregression, Econometric Institute Report 9524, Erasmus University Rotterdam.
Kuan, C.-M. and H. White (1994), Artificial Neural Networks: An Econometric Perspective,
Econometric Reviews, 13, 1-91.
MODELINGSEASONALITY IN TIME
SERIES 577
Kunst, R. M. (1993), Seasonal Cointegration in Macroeconomic Systems: Case Studies for

Small and Large European Countries, Review of Economics and Statistics, 75, 325-
330.
Lee, H. S. (1992), Maximum Likelihood Inference on Cointegration and Seasonal Cointegra-
tion, Journal of Econometrics, 54, 351-365.
Lee, H. S. and P. L Siklos (1993), The Influence of Seasonal Adjustment on the Canadian
Consumption Function, 1947-1991, Canadian Journal of Economics, 26,575-589.
Lewis, I? A. and B. K. Ray (1996), Modeling Periodic Threshold Autoregressions Using
TSMARS, Unpublished manuscript, Department of Mathematics, New Jersey Institute
of Technology.
Lutkepohl, H. (1991),Introduction to MulTiple Time Series Analysis, Springer-Verlag, Berlin.
Maravall, A. (1995), Unobserved Components in Economic Time Series, in H. Pesaran, P.
Schmidt, and M. Wickens (eds.), Handhook of Applied Econometrics, Basil Blackwell,
Oxford.
Miron, J. A. (1996), The Ecorsomics oJ’Seasona1 Cycles, MIT Press, Cambridge, MA.
Miron, J. A. and S.P. Zeldes (1988), Seasonality, Cost Shocks, and the Production Smoothing
Model of Inventories, Econometrica, 56,877-908.
Osborn. D. R. (1988), Seasonality and Habit Persistence in a Life-Cycle Model of Consump-
tion, Journal ojAppZied Econometrics, 3, 255-266.
Osborn. D. R. (1990), A Survey of Seasonality in 1JK Macroeconomic Variables, Znternational
Journal of Forecasting, 6,327-336.
Osborn, D. R. (1991),The Implications of Periodically Varying Coefficients for Seasonal Time-
Series Processes, Journal of’Econometrics, 48, 373-384.
Osborn, D. R., A. P. L. Chui, J. P. Smith, and C. R. Birchenhall (1988), Seasonality and the
Order of Integration of Consumption, Oxjhrd Bulletin of Economics and Statistics, 50,
36 1-3 77.
Osborn, D. R. and J. P. Smith (1989), The Performance of Periodic Autoregressive Models in
Forecasting Seasonal UK Consumption, Journal of Business and Economic Statistics,
7,117-127.
Paap, K., P. H. Franses, and H. Hoek (1997), Mean Shifts, Unit Roots and Forecasting Sea-
sonal Time Series, International Journal ojForecasting (to appear).
Perron, P. and T. J. Vogelsang (1992), Nonstationarity and Level Shifts with a n Application to
Purchasing Power Parity, JournaZ oJ’Business and Economic Statistics, 10, 301-320.
Smith, R. J. and A. M. R. Taylor (1995), Additional Critical Values and Asymptotic Repre-
sentations for Seasonal Unit Root Tests, Discussion Paper 9S/43, Department of Eco-
nomics, University of York, 1995.
Tiao, G. C. and M. R. Grupe (1980), Hidden Periodic Autoregressive-Moving Average Models
in Time Series Data, Biometrika, 67, 365-373.
Todd, R. (1990),Periodic Linear-Quadratic Methods for Modeling Seasonality, Journal of Eco-
nomic Dynamics and Control, 14, 763-795.
Ula, T A. (1993), Forecasting of Multivariate Periodic Autoregressive Moving-Average Pro-
cesses, Journal of Time Series Analysis, 14,645-657.
17
Nonparametric and Semiparametric
Econometrics of Panel Data
Aman Ullah and Nilanjana Roy

1. INTRODUCTION
Longitudinal or panel data refers to data where we have observations on the same
cross section of individuals, households, industries, etc., over multiple periods of
time. Often the panel data is short in the sense that the cross-sectional units are
available for a period of 2 to 10 years. Some panel data sets are rotating panels where
a proportion of cross-sectional units is kept for revisits, while the remaining is re-
placed by new cross-sectional units. Among the most analyzed panel data are the
Michigan Panel Study of Income Dynamics (PSID) and the National Longitudinal
Surveys of Labor Market Experience (NLS) in the United States, the International
Crops Research Institute for the Semi-Arid Tropics Village Level Studies (ICRISAT
VLS) in Hyderabad, India, and the Living Standards Surveys (LSS) in C6te d’Ivoire.
Recent years have witnessed a significant growth in the availability of the panel
data (Borus 1982,Ashenfelter and Solon 1982, Deaton 1994). The drive behind this
growth stems from the fact that the panel data helps to study dynamics of the indi-
vidual cross-sectional units. It is useful for studying intertemporal and intergenera-
tional behavior of the cross-sectional units. From the inference point of view panel
data leads to efficiency gains in the econometric estimators. For details on advan-
tages and problems with the panel data, see Hsiao (1985), Klevmarken (1989), and
Deaton (1994).
An important difference between the panel data econometric models with ei-
ther cross-sectional or time-series models is that it allows for the cross-sectional
and/or time heterogeneity. Within this framework two types of model are mostly esti-
mated, one is the fixed-effects model where one makes inferences conditional on the
579
580 ULIAH AND ROY
cross-sectional units in the sample while the other is the random-effects model which
is used if we want to make inferences about the population generating the cross-
sectional units. Both the models control for heterogeneity. There is no agreement
in the literature as to which one should be used in the applied work; see Maddala
(1987) for a good set of arguments on the fixed- versus random-effects models. The
econometrics of the fixed-effects model was initiated in the works of Mundlak (1961)
and Hoch (1962), while the work on the random effects model was introduced by
Balestra and Nerlove (1966) and developed further by Wallace and Hussain (1969),
Maddala (1971), Nerlove (1971), and Fuller and Battese (1973). Since then a volumi-
nous econometric literature has developed; for details see the excellent monographs
by Heckman and Singer (1985),Hsiao (1986),Dielman (1989), and Baltagi (1995);
surveys by Chamberlain (1984), Maddala (1987), and Baltagi (1996); a journal vol-
ume by Raj and Baltagi (1992); a recent handbook of approximately 900 pages by
MAtyAs and Sevestre (1996); and the recent work on dynamic panel data models by
Pesaran and Smith (1995) and Harris and MBtyAs (1996).
The motivation for this chapter is based on a simple observation that this volu-
minous literature has been largely confined to the linear parametric models; although
see MQtyiis and Sevestre (1996)and Baltagi (1996) for the references on some of the
recent works on the nonlinear parametric regression and the latent variable models.
It is, however, well known that the misspecified linear or nonlinear parametric mod-
els may lead to inconsistent and inefficient estimates and suboptimal test statistics.
With this in view, the modest aim of this chapter is to systematically develop the
nonparametric estimation of both the fixed- and random-effects panel models which
are robust to the misspecification in the functional forms. Some new estimators are
proposed. Further, the estimation of semiparametric models are also considered. For
the nonparametric regression analysis based on either cross-sectional or time-series
data and the usefulness of their application, see Hardle (1990) and Pagan and U1-
lah (1996). Muller (1988) has considered the nonparametric longitudinal models
with fixed-design regressors and has discussed its applications to growth and other
'
biomedical models. Generally the econometric regressors do not have fixed-design
structure, and his work does not consider the random- and fixed-effects models. Only
the static models are considered here, and it is hoped that in the future the results
could be extended to various other nonparametric econometric models such as the
dynamic models, limited dependent variable models, and duration models.
Another objective of this chapter is to explore the application of the nonpara-
metric panel models to study the calorie-income relationship based on the ICRISAT
VLS panel data. The usefulness of this application stems from the issue of possi-
ble nonlinearity in the calorie-income relation raised in Ravallion (1990). There has
been an ongoing debate in the calorie-income literature in the context of develop-
ing countries on the magnitude of the income elasticity of calorie intake; see Strauss
and Thomas (1990), Bhargava (1991),Bouis and Haddad (1992), and Subramanian
and Deaton (1996), among others. All these authors have considered the per capita
ECONOMETRICSOF PANEL DATA 581
household level data on calorie intake. In an important contribution Behrman and

Deolalikar (1990)considered the individual level panel data on calorie intake and
estimated the income elasticity by the parametric fixed-effects method to control for
the heterogeneity. We use the same individual level data as used by Behrman and De-
olalikar (1990)and compare our nonparametric and parametric elasticity estimates
with their parametric estimate. For this data set it has been found that the parametric
and nonparametric elasticities are quite similar except that the nonparametric one
gives us the added information at the various income levels. We note that Subrama-
nian and Deaton also calculated the nonparametric estimate, but their analysis is
based on a single cross section of households, and hence they could not control for
the heterogeneity.
The plan of this chapter is as follows. In Section I1 we first review the estima-
tion of pooled nonparametric panel data model and then present our proposed non-
parametric and semiparametric estimators for the fixed-effects and random-effects
models. The proposed nonparametric estimators are then used to analyze the calorie-
income relationship in Section 111.
II. NONPARAMETRIC PANEL DATA MODELS

A. Pooled Nadaraya-Watson Kernel Estimation
Let u s consider the nonparametric regression model as
where xic is a vector of q regressors, rn(xiC)= E(yic I xic),E(u;, I x i t ) = 0, V(uic 1

xic) = o 2 ( x i C )and
, (yic,x i c )are assumed to be i.i.d. We consider the usual panel data
situation where n is large and T is small. The nonparametric estimate of rn(x), the
conditional mean at a point x, is the smoothed average of y values which correspond
to the xic values in a small interval of x such that xic -x = O(h),where h is the width
of the interval known as window width and it tends to zero as n + 00. In a least-
squares sense, this amounts to fitting a constant to the data around x. More explicitly,
+
we use Taylor expansion of m(xiC)around x as m ( x i t ) = m ( x ) O(x;, - x), where
O(xiC- x) represents the remainder term of order at most x;,- x or O(h).The term
O(h)can be added with u ; so ~ that for large n it is still the case that the expectation
of the combined error is zero. With this in mind, but not explicitly writing it, one can
write (1)as
or, more compactly, as

582 ULLAH AND ROY
and minimize CnC T ( y i t- m ( ~ ) ) ~ K ( (-x ); /, h ) = u'K(x)u with respect to m ( x ) ;

l n is
~an nT x 1 vector of unit elements and K ( x ) is the diagonal matrix with the
diagonal elements K ( ( x i , - x ) / h ) = K i t , which is the kernel or weight function
taking low values for x;,far away from x but high values for xit close to x. The least-
squares (LS) solution of m ( x ) is
fitx) = (r~TK(x>rnT)-lr~TK(x)y (3)
which is the Nadaraya (1964)and Watson (1964)kernel regression estimator. This
estimator provides local averaging or smoothing, and it is essentially a local constant
least-squares estimator.
The pointwise estimator of the partial derivative of m ( x ) ,p ( x ) = am(x)/ax =
+
[m(x h/2) - m(x - h/2)]/h,is then
fi(x + h/2) - &(x - h/2) a
B(.> =
h
= -fi(x)
ax (4)
as given in Rilstone and Ullah (1989).Expression (4)is a numerical derivative.
Vinod and Ullah (1988)have considered the analytical derivative of &(x) in (3),
which is approximately identical to (4).
The estimation of the average derivative, p = E p ( x ) , can be done by using
Rilstone (1991)and Hardle and Stoker (1989).Following Rilstone (1991)a direct
estimator for p is
n T
BR=C):---
nT
B(xit)
(5)
i=l t=l
where B(xil) is obtained from (4). For the Hardle and Stoker estimator we assume the
density f ( x ) vanishes at the boundary of its support, and using integration by parts
we write B = E B ( x ) = -E[y( f (')(x)/'(x))], where f(')(x) is the first derivative
of f ( x ) . Then the estimator B is
where
is a kernel density estimator and f ( ' )is its first derivative.

The details on the asymptotic properties of the above estimators and the choice
of kernel and h can be found in Hardle (1990)and Pagan and Ullah (1996). Specifi-
cally we note that, under the smoothness conditions on the kernel K , Eluit12+S < 00
for some S > 0, and assuming h + 0, nhq+2s+ 00 and nhq+' + 0 as n + 00:
ECONOMETRICS OF PANEL DATA 583
where s represents the sth derivative and o:(x) = V ( u 1 x); for s = 0, rit(')(x) =
rit(x), and for s = 1, rit(')(x) = B ( x ) . In practice o:!,) can be replaced by its
consistent estimator a :
(
.
) ( x ) Z ~E , is the vector of
= ( L ~ ~ K ( x ) ~ , T ) - ' L / , ~ Kwhere
the squared nonparametric residuals iit = (yit - rit(~;,))~. We can then use (7)
to calculate the confidence intervals. The conditions for the asymptotic normality
of B R and ,!IHs are very similar to the conditions of the asymptotic normality of the
pointwise B ( x ) . It follows from Rilstone (1991)and Hardle and Stoker (1989)that,
as n + 00,
In practice, for the confidence intervals and the hypothesis testing, an estimator of
C can be obtained by replacing 0% (x),f ( ' ) ( x ) ,and B ( x ) by their kernel estimators
a;(%),f")(x), and B ( x ) and then taking the sample averages. Then for q = 1,
The second component of C is essentially the variance of the limiting distribution of

the sample average E"E' B ( x i , > / n T ,and it will be zero if the true m ( x ) is linear
in x. We also observe that the asymptotic variance does not depend on the window
width h, kernel K , and number of regressors q . This is quite different compared to
the pointwise asymptotic variance result in (7).
Now we turn to the small sample behavior of the bias and mean-square error
(MSE) of the above estimators. First, considering &(x) and taking its expectation
conditional on xit we get
E ( f i ( x ) I X l l ) = (c'K(x)L)-IL'K(x)m* (10)
where 1 = I,,T and m* = [ m ( x l l ) ,. . . , n ( x , ~ ) ] Further
.
V ( f i ( x ) I x ; t ) = (1'K(x)1)-lc'~(x)1(c'~(x)1)-I (1 1)
where !G? (x) = K ( x ) C I K ( x ) and C1 is a diagonal matrix with the diagonal elements
a;(xi,) = E ( U ; ~I xit). If 21 = a;Z, a(%) = o ? K 2 ( x ) . In practice a:(x) can be
consistently estimated by $E(.) If we expand m* by the Taylor series and consider
n to be large then it can be shown that (Ruppert and Wand 1994, Pagan and Ullah
1996),up to O(h2),
584 ULLAH AND ROY
and, up to O( llnhq),
S
where p2 = q 2 K ( Q ) dQ < 00. We note that the expressions for the condi-
tional bias and the conditional variance do not depend on xit. Thus the unconditional
bias, E ( f i ( x ) ) - m ( x ) ,up to O(h2)and the unconditional variance, V(h(x)), up to
O(l/nhq) also are the same as given in (12) and (13), respectively. Similarly, we can
show that (Pagan and Ullah 1996)
and
Essentially, (14) is the derivative of (12) with respect to x.

It follows from (12) and (13) that the optimal h values which minimize the
integrated MSE of f i ( x ) and B ( x ) , respectively, are
ho o< n-'/('7+4) and h l o< n-l/(q+')
(16)
For q = 1, hl a n-'/7. This optimal h is much bigger than the optimal h o< n-2/7
for the average derivative estimator BHS,
see Hardle and Stoker (1989), where they
give approximate MSE of BHS.
B. Pooled Local Linear Kernel Estimation
An alternative to (2) is to fit a line locally-that is, to write (1)as a linear approxi-
mation
or
where Z ( x ) is an n T x ( q +
1) matrix with itth element [ l xit - x] and G(x) =
+
[m(x)B'(x)]' is a ( q 1) x 1 parameter vector. Again minimizing u'K(x)u we get
8 (x) = ( Z ' ( x )K ( x ) Z(x)) - ' Z ' ( x ) K (x)y (19)

which amounts to doing LS of &yi, on & and K ( x i , - x) where K;, =

K ( ( x ; , - x)/h). The local least-squares estimators of m ( x ) and B ( x ) are then
f i ( x ) = (1 0)6“(x), $(x) = (0 L q ) 6 ” ( X ) (20)
Note that Z ( x ) = L,T for the Nadaraya-Watson estimator, which amounts to doing
LS of &yi, on &.
Estimator f i ( x ) was introduced by Stone (1977) in the time-series context; also
see Cleveland (1979). Stone (1980,1982) generalizes f i ( x ) and its derivatives for the
higher-order polynomials and provides their optimal rates of convergence. Cleveland
and Devlin (1988) discuss practical implementation of f i ( x ) . Muller (1987) develops
asymptotic properties of f i ( x ) when y = 1 and xit are nonstochastic and follow a
“regular” grid design. Muller shows that at interior points &(x) is asymptotically
equivalent to h ( x ) . Fan (1992, 1993) studied the asymptotic bias and variance of
iii(x) when q = 1 but xit are stochastic and demonstrated the superior behavior
of f i ( x ) compared to the Nadaraya-Watson estimator r i t ( x ) . In particular, he shows
that f i ( x ) has an important minimax property. Further, unlike the estimator rit(xj,
the bias and variance of f i ( x ) near the boundary of the support of f ( x ) are of the
same order of magnitude as in the interior; also see Fan and Gijbels (1992). This
is an important property since many economic data may be thin in the tails. In an
important work Ruppert and Wand (1994)extended Fan’s results for the q-regressors
and local quadratic fits. For y = 1, they also analysed the bias and variance of B ( x ) ,
and investigated general polynomial fits. We note here that $(x) may not be identical
to
This proposed estimator is analogous to Rilstone and Ullah B ( x ) in (4). An alternative

is to calculate analytical derivative of f i ( x ) in (20). The calculation of B ( x ) may be
simpler as it turns out directly from the local least-squares regression (19).
The asymptotic normality of 8 ( x ) = [ f i ( x ) ,&(*)I’ has been studied by Kneis-
ner and Li (1996). Li, Lu, and Ullah (1996) have considered the estimator of average
of a(%), 6 = E 6 ( x ) , as
s=-
nT
C8(XiL)
l n T
i = l ,=1
and established its asymptotic normality. This result also provides the asymptotic
normality of the average derivative
where B;(xit) = (0 i q ) 8 ( x i t )
586 ULLAH AND ROY
To see the behavior of the above estimators more explicitly, we consider first
the asymptotic results. From Kneisner and Li (1996), as n + 00,
If xit contains p lag dependent variables then T is replaced by T - p in (23) and

(24). The above results hold under some smoothness conditions on the kernel K ,
i.i.d. assumption of {yit,% i t } , existence of fourth moments of xi, and U i t , and that
h + 0, nhq+2 -+ 00 and nhqf4 + 0 as n + 00. Further, following Rilstone and
Ullah (1989),
In practice one can replace a',(x) and f (x) by their consistent estimators &:(x) and
m. The above results indicate that while the asymptotic variances of f i ( x ) and
PI(%)are the same as those of the Nadaraya-Watson estimators &(x) and the Ril-
stone-Ullah estimator B ( x ) , respectively, the asymptotic variance of estimator is B(x)
different from (x) and B ( x ) . For the standard normal kernel, however, the asymp-
totic variance of P ( x ) is the same as that of B1 (x) and B ( x ) . This is because py2
q 2 K 2 ( Q ) dQ = $(K"'(Q))2 dQ = 1/4fi.
Regarding the average derivative in (22) we note from Li, Lu, and Ullah
(1996)that
where X is as in (9).Thus, the local linear average derivative estimator has the same
asymptotic variance as BR and BHS
(8).The Monte Carlo analysis in Li, Lu, and
Ullah (1996),however, indicates that in small samples fi performs better than both
BR and BHS B
in terms of the MSE. Further the MSE of is minimum when h O( n-2/7.
We now turn to the small-sample behavior of &(x) and B ( x ) compared to A(x)
and f i ( x ) . Conditional on %it,
E ( ~ ( xI)X i t ) = (z'(x)K ( x ) z ( x ) > z ' ( ~ ) K ( x ) ~ * (27)

v(J(x) 1 xi, = ( z ' ( ~ ) K( x > z(x)) ~ ' ( x~2)( X ) Z ( ~ ' ( xK)( X > Z ( ~ ) >-
(28)
where m* and Q ( x ) = K ( x ) X K ( x ) are as given in (10) and (11). In practice, C
can be consistently estimated by using ir;(x) for a',(,), where ir;(x) can either be
obtained by the Nadaraya-Watson kernel estimator described above or by the local

linear estimator (1 0) Z’(x)K(x)Z(x))-lZ’(x)K(x)G; is the vector of local linear
squared residuals.
If we consider the Taylor expansion of rn* and take n to be large, then it can
be shown that (Ruppert and Wand 1994),up to O ( h 2 ) ,
h2
E ( f i ( x ) I xJ - m ( x ) = - / 4 r n ( 2 ) ( x ) (29)
2
and
where p 4 is the fourth moment of the kernel around zero. Using (29) it can also be
verified that
Again, these conditional bias results are the same as the unconditional results. Fur-
ther, the variance of f i ( x ) , up to O(l/nhq),and variances of B ( x ) and Bl(x), up to
O(l/nhq”), are the same as their asymptotic variances given above. It follows from
these results that the optimal h’s which minimize the integrated MSE of %(x), B ( x ) ,
and B1 (x) are the same as in (16) except for the proportionality constants.
Comparing the bias of Nadaraya-Watson estimator h ( x ) with that of the local
linear estimator f i ( x ) we note that while the bias of %(x) depends on the curvature
behavior m ( 2 ) ,the bias of h ( x ) depends on rn(2) as well as m ( ’ ) f ( ’ ) / fdue to the
local constant fit. When (m(’)I is large or when f ( ’ ) / f is large, especially in highly
clustered data, the bias of h ( x ) is large. Even when the true regression is linear h
is biased but f i is unbiased. The asymptotic variances of h and f i are the same.
Thus, one might expect f i to perform better than R. Fan (1992,1993) reports good
finite-sample MSE performance of % compared to h. He shows that f i is the best
among all linear smoothers, including orthogonal series and splines. Further f i has
100%efficiency among all linear smoothers in a minimax sense and a high minimax
efficiency among all possible estimators. Fan and Gijbels (1992)have reported better
performance of f i ( x ) near the boundary of the support o f f , although see a word of
caution by Ruppert and Wand (1994).
The comparison of the bias of b(x)
with the bias of B ( x ) is, however, not so
clear. Both are seriously affected by m(’), rn(3),f(’), and f . In contrast, the bias of
our proposed estimator Bl(x) is much simpler and smaller compared to B ( x ) and
B ( x ) . In fact, when the true rn is linear both b ( x ) and B ( x ) are biased but f i l ( x ) is
unbiased. Though not studies here, it is conjectured that the MSE performance of 81
will be much superior to both and B.
588 ULLAH AND ROY
The estimator s"(x) in (19) can also be obtained by the local nonparametric
estimation of the linear parametric model
yit =a + XiLp + UiL (32)

in the sense of minimizing E"E'
U;~K((- ~ ;x,) / h ) with respect to a and /3. This
gives the estimator of p, which will be the same as B ( x ) , and &(x) will be &(x) +
x B ( x ) . The advantage of viewing the problem in this way is that one can extend it
to the local estimation of the parameters of any nonlinear parametric model g(xi, p).
This idea stems from the recent literature on the local maximum likelihood estima-
tion of the parameters of a parametric density (Hjort and Glad 1995). The same idea
is also used in Robinson (1989b), who considered the problem of estimating /?over
+
time in the time series regression? = xc/3 u t . Robinson obtained the estima-
tor B ( t * ) = (ET
KL(t*)x;)-' K t ( t * ) x t y tby minimizing ET u;K, (t*), where
K , ( t * ) = K((Tt* - t ) / T h ) for t* in [0, 11. An interpretation of (32) can be given
by using the small-a expansion. For this consider the small o expansion of rn(xit)
+
in (1)around p, instead of small h around x, as m(x;,)= n ( p ) (xit - p ) p ( p ) +
+
O(xit - p)' where O(xic- p)' = O(o2) by writing xit = p oeit. Thus, assuming
o -+ 0, we get
Yit =m(p) + b i t - P ) B ( P ) + uic (33)
which is (32);a = m ( p ) - p p ( p ) .Thus while the local linear model (17) is based
on the small h approximation of rn, the model (32) may be interpreted as the small
cr amroximation of rn.
I I
We point out that when h = CO the local minimization of E"ET

U ; ~ K ( (-~ ; ~
x)/h) = K ( 0 ) E"E'
ufL becomes the global minimization of E"ET
uft. This
gives the usual parametric pooled LS estimators of a and p.
The idea of local linear estimation of m ( x ) ,the conditional mean, can be eas-
ily extended to the local estimation of the higher central moments, p r = E[(y;, -
rn(xit))' I xit] = E ( U ; ~1 xit). This will be the coefficient of a
in the regression
of &ii:L on & and &(xit - x), where ii;, is obtained from (17) or (32). It
will be interesting to compare the properties of pr so obtained with those obtained
by the Nadaraya-Watson-type local constant estimator f i r ,which is obtained by re-
gressing &&It
on &, where &;, is the residual from (2). The comparison will
especially be useful for r = 2, the conditional heteroskedasticity model often used
in macroeconomics and finance.
C. Nonparametric Fixed-Effects (FE) Model

A nonparametric FE model can be written as
where ai is the individual specific fixed parameters, and uit is i.i.d. with mean zero
and constant variance at.A more general specification of (34)is discussed in Section
1I.D. When m(x;,) = xit/?, model (34)is the well-known linear parametric FE model
studied in the literature very extensively. An important reason for the popularity of
the linear parametric model is that there exists a class of transformations of (34)
which eliminates ai so that /? can be estimated by either a simple LS or a generalized
least-squares (GLS) estimator. It is not straightforward to get transformations which
will remove ai from (34)when m(xit)is of unknown form. However, if we reformulate
the problem of estimating rn and its derivative in terms of local linear estimation of
(17), then it is possible to implement some of the existing transformations. These are
proposed below. First, following (17),
which gives
The local FE estimator of B ( x ) can then be obtained by minimizing
This gives, for q = 1, our proposed estimator as
and for q 2 1,
where X is an nT x q matrix, D = I n @ L T is an nT x n matrix, and Mu = I - DIY/T.

) essentially the LS of 6 ( y i t - yi,)on &(xit
The estimator B ~ , y ( x is - Xi,).If
h = 00, then K ; , = K ( 0 ) and ,81..s(x) becomes the well-known parametric FE
estimator FE. The asymptotic properties of this can be worked out in the same way
as that for B ( x ) in (20). However, conditional on xit,
E ( B F E (x)) = (X’MDK (x)h ! D x ) - ’ X’MDK (x)Mum* (38)

V(BF&)) = 02,(X’M&x)MuX)-’ (X’M,K’(x)MnX)
x (X’MDK(x)h!DX)-’
provided ~ : ~ ( =
x )of.For a feasible version of the variance of BFR(X) one needs
to replace O: by sz = E”C c iil*,2/nT,where is the residual from the regression
590 ULLAH AND ROY
in (36). If of(x) # a: we need to modify (38) as in (11).Also, as in Sections 1I.A

and II.B, expanding m” by Taylor series and taking n to be large we can obtain the
asymptotic bias and variance of PFE(X). This, along with the asymptotic normality,
will be the subject of a future study.
The FE estimator (37) can also be interpreted as the local estimation of the
+ +
parametric FE model y;, = a; xi,#? ui, after it has been transformed into yit -
-
+
yi. = (xi,- Xi.)#? uit - iii.. This will be analogous to the estimation procedure
in Section 1I.B. One can also obtain a FE estimator of #? by using first-difference
transformation yi, - y;,-1 = (xi, - x;,-l)#? + uit - U;,-] and then minimizing
E”ET(u;, - ~ i , - l ) ~ K ( ( x-i ,x ) / h ) . This gives
with
for given xi,; A is the matrix which transforms y into the vector
- of yi, - y;,-l. It is
not clear, unlike in the linear parametric regression, whether FE is the same as BFE
for T = 2 or 3. This seems true even if we had estimated #? by the weighted GLS
procedure, which takes into account the moving-average nature of the error term.
An alternative way to estimate #?(x) is to write
and do the LS of &Ay;, on &, a ( x ; , - x), and &(x;,-l - x - l ) , where

K ; , is now K ( -( x ; , - x ) / h , (xi,-] - ~ - 1 ) / h )The
. relationship of the estimator so
obtained with BFE BFE
and remains unknown. Li and Stengos (1995)propose writing
+
(40) as Ayit = m(x;,,x i t - ] ) Au;, and then estimating rfz by the nonparametric
kernel method. Roy (1996) has considered the estimation of (34) by the Newey-type
series estimator.
Following the results in Section 1I.B the average derivatives can be estimated
after getting the pointwise derivatives from above.
D. Nonparametric Random-Effects (RE) Model

A general nonparametric RE model can be considered as
ECONOMETRICS OF PANEL DATA 59 I
which can be written, after expanding rn around the mean values of xit = p and
ai = 0, as
where rn(p)- p@(p)= a,vi = a;p1( p ) ,and @ and @1 are derivatives with respect
to xit and ai respectively. The model (41) is the RE version of the local linear model
in (33)based on small-a expansion. The RE version of the local linear model in (17)
can also be written as (42) with p replaced by x:
We consider uit to be i.i.d as in Section II.C, and assume vi also to be i.i.d with mean
zero and variance of.
The local nonparametric RE estimator of M and /3 in (43) can be obtained by
minimizing
where y* = Q-'/'y, Z * ( x ) = Q - ' / ' Z ( x ) , and a-'/' = Z,,T - (1 - h'/')DD'/T; D

is as in (37).Z ( x ) and S ( x ) are as defined in (18), y: = yit - (1 - hl/*)yi,, =
zit - (1 - h1/')2i.,and h = oE/(a', +TO:). This amounts to doing the LS regression
of a y ; on &z: = &(x1", - x*)] which gives our proposed
estimator as
&E -'
(x) = ( Z * ' ( x ) K ( x ) Z *(x)) Z * ' ( x ) K (x)y* (44)
When h = 00, K ( x ) = K ( 0 ) and we get the well-known parametric RE estimator
given in econometric texts (Baltagi 1995, Hsiao 1986).
A feasible estimator of 8 ~ ~ ( xis )obtained by replacing h with its estimator
X +
= 3:/(3: Tb:). 3; is obtained as
i=l t=1
where p ~ ~ (isx the) nonparametric FE estimator given in (37).The estimator 3: is

defined as 35 = 3: - 8 : / T , where 3: = E(?;, - fi(Xi.))'/n and &((xi.) = %(x)
at x = is obtained by performing local least-squares estimation on the model
+ + +
pi. = m ( x ) ((xi. - x ) p ( x ) Vi El..
592 ULLAH AND ROY
As described in Section II.B, the estimators of a and /3 in (42) are also given
by the estimator in (44). For given xi,,
W R E (x)>
z
= 0;( Z *’(x)K (x) * (x)) - z*’(x)K 2(x)z*(x)( Z *’(x)K (x)2* (x))-
(46)
The detailed study of the asymptotic and finite-sample properties of b ( x ) and
its comparison with the FE estimators will be the subject of future research.
E. Semiparametric FE and RE Models

Here we first consider the model of the following type:
y;, = X i t p + rn(z;,) + +
zli U;, (47)
where t i t is a vector of p regressors. For zli = 0, the model is considered by Li and
Stengos (199S), who propose the &z consistent estimation of p by Robinson’s (1988)
procedure. This is given by transforming (47) with vi = 0 as
R r = R::@ + (48)
where RT,’ = yi, - E(yi, I zi,) and R;: = nil - E(xi, I z i t ) and then applying the LS
method. This gives
l n T \-l n T
/ i=l ,=I
where S P represents semiparametric pooled estimator. For a feasible version of this

estimator, one needs to replace E(yi, I z;~) and E(%;,I z i t ) by their respective non-
parametric estimators. Li and Stengos propose the Nadaraya-Watson kernel estima-
tor for this, but an alternative estimator can be obtained by using the local linear
estimation in Section 1I.B. The estimators of m ( z ) and its derivative can then be ob-
tained by performing local linear regression of yi, -?it on r n ( z i l ) , where = xitBsp.
Kneisner and Li (1996) look into the estimation of @ (with U; = 0) when zit contains
the lagged variables of yi,. They also propose the estimation of the rn(zi,) and its
derivative by doing the local least squares procedure for yi, - ?it = rn(z;,) ui,. +
In the case when U;is nonzero but fixed, the Li and Stengos (1995) estimation of
B is not affected. Once p is estimated the estimation of the derivative of m ( z ) ,y(z),
+ +
can be carried out by analyzing y: = y;, - Tit = rn(z;,) a; UiL with the meth-
ods described in Section I1.C. From (37), our proposed semiparametric FE (SFE)
estimator of the derivative of rn(zi,) for p = 1 can be written as
and for p >_ 1,
PSFE(2) =(Z'MDK(z)M~Z)-'Z'M~K(z)~~y*
It is important to be able to estimate the derivative since that is the parameter of
interest in most economic applications.
When vi is random, the model becomes the semiparametric RE (SRE) model
considered in Li and Ullah (1996), who propose a &-consistent GLS estimator of
BY
where R;i"* = RTlz - (1 - h112)R:z, R:;* = R:: - (1 - h ' 1 2 ) ~ : zand

, h is as in
Section I1.D. Using their estimator, one can develop the estimation of y: = m(t;,) +
+
v; uit by procedures given in Section 1I.D; y: = yit - x i t f i ~ = ~ ~yiL; - ?it. This
gives the estimators of m ( z ) and its derivative y ( z ) from
as h ( z ) = (1 O ) ~ S R E ( Zand z ) , S^SR~;(Z) is a vector

) ~ . w ; ( z )= (0 ~ ) ~ s R E ( where
of ~ ~ R E ( andz ) P S R E ( Z ) ,rt7* = (1 z;*1* - t * * )
t:*, = zit - ( 1 - h'/'))zi,,z** =
h'I2z, y*: = yz - (1 - h 1 / 2 ) y iand . , Ki, = K ( ( z i , - z ) / h ) .
To get the feasible estimators of f i . 5 ~and~ $ . ~ R E ( Z )one needs to replace h with
iand X respectively where i= 6:/(6: T6:) and X =+ iiE/(iiE + Tii?).Then
where fi,~~
is the FE estimator from the model RTtZ= R::P
2
U; + +
uit. The estimator
6:is given by 6; - 6 : / T , where 6; = C" -xz
f i ~
(R!,' - #?eRi.) / n and is the between
A
estimator obtained by performing LS estimation on the model R:z = Ri. B v;+U;,.

-x z
+
We note that our proposed feasible estimator , 8 L y based~ ~ on (52) is different from Li
and Ullah (1996),where 6; is obtained by using the LS estimator of /?instead of FE
estimator of /?.
Similarly, X is obtained with 3: = E"
CTjy; -7; - ~ L ~ ~ ~ ( z -) ~( zi .i )L) ~ / n T ,
~ ; 5; = E(?; - G ~ ( z ) ) ~ / n ,
) as in (50) with y: = yit - . x i t / 3 s ~and
where ~ S F E ( Z is
+
where & ( z ) is obtained by doing local LS on the model yF = m ( z ) ( ~ i . - z ) y ( z ) +
Ui + iii;
In a special case of Li and Ullah, when t i t = T;.,
the model becomes a semi-
parametric extension of the Mundlak (1978) model where m(xi.)is linear in xi,.An-
other semiparametric model one can consider is where m(z;,) is zero but V(ui, \
594 ULIAHAND ROY
xit) = Li and Stengos (1994) suggest a two-step GLS estimator of p for such
a model. One can also consider V(ai I xi.) = a2(x;,) and develop a two-step GLS
estimator. The situation where both a2(xi.)and a 2 ( x i t )are present has not been con-
sidered in the literature. In a recent paper Horowitz and Markatou (1996) consider
the nonparametric kernel density estimation of at and a: and then propose the max-
imum likelihood estimation of 6 in (47) with no m ( t ) . The issues of unit roots and
serial correlation in the errors remain the subjects of future research.
One disadvantage of the semiparametric model in (47) with an unknown func-
tion of regressors, or the purely nonparametric model in Section II.A, is the “curse
of dimensionality.” This refers to the fact that the rate of convergence of the non-
parametric estimator of m ( z ) decreases drastically with the increase in the number
of regressors. One solution explored in the literature is to use the generalized ad-
ditive models of Hastie and Tibshirani (1990) and Berhame and Tibshirani (1993),
which estimate the p-dimensional m ( z ) at the convergence rate of one dimensional
nonparametric estimator. Essentially the model (47) is written as (assuming 0 and
U ; = 0) yit = E;==, +
m ; ( ~ ; ; ~ )U i t , and m;(z;it) is estimated by a nonparametric
method; see Linton and Neilson (1996)for the kernel method of estimation.
There is an extensive semiparametric literature on the estimation of limited
dependent variables models such as single-index models, censored models, and se-
lection models (Melenberg and Soest 1993,Pagan and Ullah 1996). Essentially these
semiparametric models can be considered as special cases of (47) with 2ii = 0 where,
+
for example, in the single-index case x i t p m(zit)is a function of single index, say,
t i t s , that is, m ( z i t S ) . In the censored case m(zil)in (47) is m(zit6),which becomes
the inverse Mill’s ratio under the assumption of the normality of u i t . When u i t is
nonnormal, vi # 0 but is fixed, and m(z;,) = 0, the estimation of (47) by the least
absolute deviation (LAD) method has been discussed in Honor6 (1992)and Keane
(1993),among others.
F. Specification Testing
There is an extensive literature on various specification testing in the parametric FE
and RE models (Hsiao 1986, Baltagi 1995, 1996, Greene 1993).Here we mainly
look into the recent work in the context of nonparametric panel models.
Considering the nonparametric pooled model in Section II.A, we note that the
pointwise hypothesis testing for the linear restrictions on the derivatives can be done
by using the asymptotic normality results for the local linear or Nadaraya-Watson
given there (Ullah 1988, Robinson 1989a, Lewbell995, and Pagan and Ullah 1996
give more details and references). In fact, the pointwise asymptotic test for various
+
misspecifications in the local linear model yit = m ( x ) (xit - x ) p ( x ) U ; , may +
follow from the corresponding tests in the linear parametric models. The global tests
based on comparing the restricted residual sum of squares (RRSS) with the unre-
stricted RRSS or based on the conditional moments are developed, among others, in
Fan and Li (1996), Li and Wang (1994), and Bierens (1990). Their results cover the
tests for linearity, exclusion of regressors and semiparametric specification, and can
+
be extended for yit = r n ( x i L ) uit, where xit and uit are i.i.d. across i and t .
Frees (1995) considered the problem of testing cross-sectional correlation in
the parametric panel model and noted that the Breusch and Pagan (1980) measure
does not possess desirable asymptotic properties for the practical situation where n
is large but T is small. In fact, he showed that the asymptotic distribution depends
on the parent population even under the hypothesis of no cross-sectional correlation.
In view of this, he introduced a distribution-free statistic which does not have this
problem. An extension of this to nonparametric and semiparametric models will be
useful. Li and Hsiao (1996) have considered the semiparametric model (47) and
developed a LM-type test for the null hypothesis that uit is white noise against the
alternative that uit has the RE specification.
An important question in the nonparametric panel data analysis is whether to
pool the data. A conditional moment test for this problem, Ho : r n l ( x i 1 ) = mz(x;z)
against H1 : m l ( x i l ) # mn(x;z)assuming T = 2 here for simplicity, is proposed
in Baltagi et al. (1995).If the Ho is accepted then one can pool the data and use the
results of Section 1I.A. If Ho is rejected then the estimates for two different periods
can be pooled to obtain a more precise estimate (Pinske and Robinson 1996).
111. A N APPLICATION
Here we present an empirical example based on the methodology discussed and de-
veloped in the previous sections. For a long time now there has been a debate in
the nutrition-income literature in developing countries on the response of nutrition,
more specifically calorie intake, resulting from a rise in income. Some of the recent
articles that have engaged in this debate are by Behrman and Deolalikar (1990),
Strauss and Thomas (1990), Bhargava (1991), Bouis and Haddad (1992), Subrama-
nian and Deaton (1996), and Grimard (1995). For them estimating the income elas-
ticity is important because it has serious policy implications on how best to reduce
malnutrition. If the elasticity turns out to be close to zero, the implication is that
improvement in the income of the poor will have little impact on the extent of mal-
nutrition. Then the development policies aimed at improving nutrition will have to
use policy instruments which attack malnutrition directly rather than relying solely
on rising income.
Behrman and Deolalikar (1990) used individual level ICRISAT VLS panel
data for two years from three villages in south central India and estimated a linear
parametric FE model. In this section we consider both the standard parametric panel
models and the nonparametric panel models discussed in Section I1 to study the
calorie-income relationship based on the data set used by Behrman and Deolalikar
(1990). For details on ICRISAT VLS data, see Binswanger and Jodha (1978), Ryan
596 ULlAH AND ROY
et al. (1984),and Walker and Ryan (1989).Our contribution to the existing literature
on calorie-income relationship is that we are able to take into account both the func-
tional form and the heterogeneity while modeling the calorie-income relationship,
and this we are able to do using the results in Section 11. Note that previous works
have considered one or the other (i.e., the heterogeneity or the functional form) but
never both together. While we recognize the fact that there are other variables which
influence individual calorie intake, we choose to use income as our only regressor
since it is undoubtedly the most influential factor in individuals’ consumption de-
cisions and some other authors in this literature have done the same. For example,
Subramanian and Deaton (1996) studied the regression of calorie intake on expen-
diture. A nonparametric regression analysis of calorie intake with other variables
besides income included as regressors will be the subject of a future study. We think
that in the multivariate case, the semiparametric method described in Section II.E,
rather than a pure nonparametric analysis, may be a better way to study the calorie-
income relationship.
We consider three types of nonparametric models: constant intercept, fixed-
effects, and random-effects models. These correspond to model (34)with a;equal to
a constant, ai as an individual fixed effect, and a;as a random effect respectively.
Similarly we consider the same three types of models with linear parametric speci-
fication r n ( x i t ) = x ; , B . The dependent variable, yit, in all the models represents the
logarithm of individual calorie intake for the ith individual in the tth time period, the
explanatory variable, xit, represents the logarithm of per capita real income, and ai
represents the combined effects of unobserved individual characteristics, household
characteristics, etc., which can be considered to be fix<,.-! or random, as may be the
case.
The results are all based on a total number of observations of 730, that is, 365
individuals each observed over two years. For the nonparametric regression analysis
the kernel used is the normal kernel given as
and h, the window width, is taken as ~ s n - ’ ’ ~where

, c is a constant, s is the stan-
dard deviation of the variable x, and n is the number of observations. For details on
choosing window width and kernel, see Marron (1988),Hardle (1990), and Pagan
and Ullah (1996).
Our parametric estimates of the income elasticity of calorie intake from differ-
ent versions of model (34) with rn(xit) = xilB are presented in Table 1. They are all
positive and significant. Our result is in contrast to Behrman and Deolalikar’s result
that the elasticity estimates from the parametric FE model are zero and hence not sig-
nificant. Perhaps one reason is that they had other regressors besides income in their
model. Also, the sign of the current income elasticity of calorie intake was negative
Table I Parametric Income Elasticity of Calorie Intake
Model Beta coefficient t-Ratio
Constant intercept 0.138 7.104

Fixed effect 0.108 4.261
Random effect 0.126 6.673
Table 2 Nonparametric Income Elasticity of Calorie Intake
Model Beta at the mean t-Ratio
Constant intercept 0.136 4.789

Fixed effect 0.115 4.931
Random effect 0.117 5.906
Table 3 Nonparametric Income Elasticity of Calorie Intake
Model Mean beta Minimum Maximum
Constant intercept 0.132 0.008 0.187

Fixed effect 0.121 0.062 0.134
Random effect 0.107 0.018 0.164
in their constant-intercept model, which may be due to the presence of collinearity

since they had both the current and permanent income as regressors. We also calcu-
lated the t-ratios by using the cluster-corrected standard errors (Deaton 1994),but
still the estimates remained significant.
Note that the magnitude of the estimated current income elasticity from the
constant-intercept model is higher than that from the FE or the RE model. The re-
sult from the F-test rejects homogeneity of the intercept, and hence the coefficient
estimate of the constant intercept model which fails to take into account the het-
erogeneity is biased. The magnitudes of elasticities from the FE and the RE model
are quite similar. In fact, the Hausman specification test failed to reject the null hy-
pothesis of no systematic difference in the two coefficients. Hence, it is up to the
researcher to decide whether to use a FE model or a RE model, and this decision
will depend on whether he or she wants to make inferences based on the sample
or on the population (Hsiao 1986). However, in our case, the conclusion about the
elasticity will not change much depending on whether one uses a FE or a RE model.
From this parametric regression analysis, the overall picture is that rising income
598 ULLAH AND ROY
will affect calorie intake but only very slightly. What the parametric analysis does
not tell us is whether the elasticity is significant at all levels of income, and if so what
is its magnitude. Of course, one can do parametric regression analysis by percentile
groups (for example, the elasticity for the bottom and the top deciles, say), but still
one cannot get the elasticity at each income level. This question can be answered
from the nonparametric regression analysis.
Table 2 gives the nonparametric elasticity estimates at the mean value of the
regressor to make it somewhat comparable with the parametric elasticity estimates
from Table 1. The results from the nonparametric models are similar to our paramet-
ric model results as can be seen from Table 2.
Given that the nonparametric specification gives us elasticity estimates at dif-
ferent income levels, we also report in Table 3 the mean, the minimum, and the max-
imum values of the elasticity for the different models. The mean elasticities were
calculated by using h = ~ s n - ~as/ ~ indicated
, in Section 1I.A. Note, however, that
h a n-2/7 is known to be optimal for the constant-intercept model only (Li and U1-
lah 1996),but the optimal h values for the FE and the RE models are not yet known.
It can be seen from the table that the elasticity can be quite different for different
income levels, and looking at just the mean elasticity estimate can be misleading.
We find for all three models (constant intercept, FE, and RE) and, on average,
the elasticity is higher for poorer households compared to richer households (Fig-
ures 1, 2, and 3). For our constant-intercept nonparametric model the elasticity is
--..
t
0.3
0.2 4
---.
-4.-
... --..
\
\
4.1 --
\
\
\
\'
'.
Figure I Elasticity of calorie intake with respect to per capita real income from pooled non-
parametric model using local linear estimation method.
leoofprruo1tlM-
Figure 2 Elasticity of calorie intake with respect to per capita real income from a nonpara-
metric FE model.
T
0.2 '.
0.1s 9 '
-..-._.
). 0.1 -.
F]
......ubond
I 4.5 5 5.5 6 6.S 7 7.5 '.,8 8:5
-. .
leoofprrrqlcdheonr
Figure 3 Elasticity of calorie intake with respect to per capita real income from a nonpara-
metric RE model.
600 ULLAH AND ROY
significant everywhere except at the tails, but the tail behavior of nonparametric
estimators is generally not very good. For FE and the RE models, the elasticities are
significant everywhere except at the upper tail.
Thus our results suggest that the income elasticity of calorie intake is signif-
icant but is rather low. The small magnitude of elasticity, however, does not neces-
sarily mean that income is an ineffective policy instrument to reduce undernutrition,
as has been effectively demonstrated in Ravallion (1990).Both parametric and non-
parametric regression analyses give us a similar result, except that the nonparamet-
ric one gives us the added information that the elasticity gradually declines as one
moves up the per capita income distribution.
ACKNOWLEDGMENTS
We are grateful to B. Baltagi, Q. Li, L. MAtyAs, B. Raj, and M. Ravallion for their
useful comments and suggestions. We are also thankful to the seminar participants
at McGill University and UCR for their comments. We are especially grateful to Anil
Deolalikar for the discussions on the subject matter of this chapter and for providing
the ICRISAT VLS data. Any remaining errors are ours. The first author gratefully
acknowledges research support from the Academic Senate, UCR.
REFERENCES
Ashenfelter, 0. and G. Solon (1982), Longitudinal Labor Market Data-Sources, Uses and Lim-
itations, in What’s Happening to Americun Labor Force and Productivity Measurements?
Proceedings of a June 17,1982, conference sponsored by the National Council on Em-
ployment Policy, W. E. Upjohn Institute for Employment Research.
Balestra, P. and M. Nerlove (1966), Pooling Cross-Section and Time-Series Data in the Estima-
tion of a Dynamic Model: The Demand for National Gas, Econornetrica, 34, 585-612.
Baltagi, B. H. (1995), Econometric Analysis ($’Panel Data, Wiley, New York.
Baltagi, B. H. (1997), Panel Data Methods, in Handbook ojApplied Economics Statistics (A.
Ullah and D. E. A. Giles, eds.), Marcel Dekker, New York.
Baltagi, B. H., J. Hidalgo, and Q. Li (1996), A Nonparametric Test for Poolability Using Panel
Data, Journal of Econometrics, 75(2), 345-367.
Behrman, J. R. and A. B Deolalikar (1990). The Intrahousehold Demand for Nutrients in Rural
South India: Individual Estimates, Fixed Effects and Permanent Income, Journal of
Human Resources, 25(4), 665-696.
Berhame, K. and R. J. Tibshirani (1993), Generalized Additive Models for Longitudinal Data,
Manuscript, [Jniversity of Toronto.
Bhargava, A. (1991), Estimating Short and Long Run Income Elasticities of Foods and Nutri-
ents for Rural South India, Journal ojthe Royal Statistical Society, 154, 157-174.
Bierens, J. H. A. (1990), Consistent Conditional Moment Tests of Functional Form, Econo-
metricu, 58, 1443-1458.
ECONOMETRICS OF PANEL DATA 60I
Binswanger, H. P. and N. S. Jodha (1978), Manuals of Instructionsfor Economic Investigators

in ICRISAT's Village Level Studies, Vol. 2, Village Level Studies Series, Economics
Program, ICRISAT, Patancheru.
Borus, H. E. (1982), An Inventory of Longitudinal Data Sets of Interest to Economists, Review
of Public Data Use, 10, 113-126.
Bouis, H. E. and L. J. Haddad (1992), Are Estimates of Calorie-Income Elasticities too High?
A Recalibration of the Plausible Range, Journal of Development Economics, 39(2),
333-364.
Breusch, T.S. and A. R. Pagan (1980), The Lagrange Multiplier Test and Its Applications to
Model Specification in Econometrics, Review of Economic Studies, 47,239-253.
Chamberlain, G. (1984), Panel Data, in Handbook of Econometrics (Z. Griliches and M. In-
triligator, eds.). North-Holland, Amsterdam, 1247-1318.
Cleveland, W. S. (1979), Robust Locally Weighted Regression and Smoothing Scatter Plots,
Journal of the American Statistical Association, 78(368), 829-836.
Cleveland, W. S.. and S. Devlin (1988), Locally Weighted Regression: An Approach to Re-
gression Analysis by Local Fitting, Journal of the American Statistical Association, 83,
596-616.
Deaton, A. (1994), The Analysis of Household Surveys: Microeconometric Analysis for Develop-
ment Policy, manuscript, Princeton University.
Dielman, T. E. (1989), Pooled Cross-Sectional and Time Series Data Analysis, Marcel Dekker,
New York.
Fan, J. (1992), Design Adaptive Nonparametric Regression, Journal ofthe Americurz Statisti-
cal Associution, 87,998-1024.
Fan, J. (1993), Local Linear Regression Smoothers and Their Minimax Efficiencies, Annuls
of Statistics, 21, 196-216.
Fan, J. and 1. Gijbels (1992),Variable Bandwidth and Local Linear Regression Smoothers,
Annuls of Statistics, 20,2008-2036.
Fan, Y. and Q. Li (1996), Consistent Model Specification Tests: Omitted Variables and Semi-
parametric Functional Forms, Econometrica, 64, 865-890.
Frees, E. W. (1995), Assessing Cross-Sectional Correlation in Panel Data, Journal of Econo-
metrics, 69, 393-414.
Fuller, W. A. and G. E. Battese (1973), Transformations for Estimation of Linear Models
with Nested Error Structure, Journal of the American Statistical Association, 68,626-
632.
Greene, W. H. (1993), Econometric Analysis, Prentice Hall, Englewood Cliffs, New Jersey.
Grimard, F. (1995), Does the Poor's Consumption of Calories Respond to Changes in Income?
Evidence from Pakistan, manuscript, McGill University, Canada.
Hardle, W. (1990), Applied Nonparametric Regression, Cambridge University Press, Cam-
bridge.
Hardle W. and T. M. Stoker (1989), Investigating Smooth Multiple Regression by the Method
of Average Derivatives, Journal ofthe American Statistical Association, 84,986-995.
Harris, M. R. and L. MBty6s (1996), A Comparative Analysis of Different Estimators for Dy-
namic Panel Data Models, Department of Economics Monash University, Australia,
Working Paper.
Hastie, T. J. and R. J. Tibshirani (1990), Generalized Additive Models, Chapman and Hall,
London.
602 ULLAH AND ROY
Heckman, J. J. and B. Singer (1985),Longitudinal Analysis of Labor Market Data, Cambridge

Hjort, N. L. and I. K. Glad (1995), Nonparametric Density Estimation with a Parametric Start,
Annals of Statistics, 23,882-904.
Hoch, I. (1962), Estimation of Production Function Parameters Combining Time-Series and
Cross-Section Data, Econometrica, 30,34-53.
Honorb, B. E. (1992),Trimmed LAD and Least Squares Estimation of Truncated and Censored
Regression Models with Fixed Effects, Econometrica, 60, 533-565.
Horowitz, J. and M. Markatou (19%), Semiparametric Estimation of Regression Models for
Panel Data, Review of Economic Studies, 63, 145-168.
Hsiao, C. (1985), Benefits and Limitations of Panel Data, Econometric Reviews, 4, 121-174.
Hsiao, C. (1986),Analysis of Panel Data, Cambridge University Press, Cambridge.
Keane, M. P. (1993), Simulation Estimation of Panel Data Models with Limited Dependent
Variables, in Handbook of Statistics (G. S.Maddala, C. R. Rao, and M. D. Vinod, eds.),
North-Holland, Amsterdam, 545-57 1.
Klevmarken, N. A. (1989), Panel Studies: What Can We Learn from Them? Introduction,
European Economic Reviews, 33,525-529.
Kneisner, T. and Q. Li (19%), Semiparametric Panel Data Model with Dynamic Adjustment:
Theoretical Considerations and an Application to Labor Supply, manuscript, University
of Guelph.
Lewbel, A. (1995),Consistent Nonparametric Hypothesis Tests with an Application to Slutsky
Symmetry, Journal of Econometrics, 67,379-340.
Li, Q. and C. Hsiao (1996),Testing Serial Correlation in Semiparametric Panel Data Models,
manuscript, University of Guelph.
Li, Q., X. Lu, and A. Ullah (1996), Estimating Average Derivatives by Local Least Squares
Method, manuscript, University of Guelph.
Li, Q. and T. Stengos (1994), Adaptive Estimation in the Panel Data Error Component Model
with Heteroskedasticity of Unknown Form, International Economic Review, 35, 981-
1000.
Li, Q. and T. Stengos (19%), Semiparametric Estimation of Partially Linear Panel Data Mod-
els, Journal of Econometrics, 71(1-2), 389-397.
Li, Q. and A. Ullah (1996), Estimating Partially Linear Panel Data Models with One-way
Error Components Errors, manuscript, University of Guelph.
Li, Q. and S.Wang (1994), A Simple Consistent Bootstrap Test for a Parametric Regression
Function, manuscript, University of Guelph.
Linton, 0. and J. P. Neilson (1996), A Kernel Method of Estimating Structured Non-parametric
Regression Based on Marginal Integration, Biometrike, 82,93-100.
Maddala, G. S.(1971),The Use of Variance Components Models in Pooling Cross-Section and
Time-Series Data, Econometrica, 39,341-358.
Maddala, G. S. (1987), Recent Developments in the Econometrics of Panel Data Analysis,
Transportation Research, 21,303-326.
Marron, J. S. (1988), Automatic Smoothing Parameter Selection, Empirical Economics, 13,
66-86.
MAty6s L. and P. Sevestre (ed.) (1996), The Econometrics of Panel Data: A Handbook of the
Theory with Applications, Kluwer, Dordrecht.
Melenberg, B. and A. V. Soest (1!993), Semiparametric Estimation of the Sample Selection

Model, manuscript, Tilburg University.
Miiller, H. G. (1987), Weighted Local Regression and Kernel Methods for Nonparametric
Curve Fitting, Journal of the American Statistical Association, 14,457-471.
Miiller, H. G. (1988), Nonparametric Regression Analysis of Longitudinal Data, Springer-
Verlag, New York and Berlin.
Mundlak, Y. (1961), Empirical Production Function Free of Management Bias, Journal of
Farm Economics, 4 3 , 4 6 5 6 .
Mundlak, Y. (1978),On the Pooling of Time Series and Cross-Section Data, Econometrica, 46,
69-85.
Nadaraya, E. A. (1%4), On Estimating Regression, Theory of Probability and Its Applications,
10, 186-190.
Nerlove, M. (1971), A Note on Error Components Models, Econometrica, 39,359-382.
Pagan, A. and A. Ullah (1996), Non-parametric Econometrics, manuscript, Australian Na-
tional University, Australia.
Pesaran, M. H. and R. Smith (1995), Estimating Long-Run Relationships from Dynamic Het-
erogenous Panels, Journal ($Econometrics, 68, 79-1 13.
Pinske, C. A. P. and P. M. Robinson (1W6), Pooling Nonparametric Estimates of Regression
Functions with a Similar Shape, in The Econometrics of Panel Data: A Handbook of the
Theory with Applications (L. MAtyiis and P. Sevestre, eds.), Kluwer, Dordrecht, 172.
Raj, B. and B. Baltagi (1992), Panel Data Analysis, Physics-Verlag, Heidelberg.
Ravallion, M. (1990), Income Effects on Undernutrition, Economic Development and Cultural
Change, 38,487-515.
Rilstone, P. (1991), Efficient Instrumental Variable Estimation of Nonlinear Dependent Pro-
cess, manuscript, University of Laval.
Rilstone P. and A. Ullah (1989), Nonparametric Estimation of Response Coefficients, Com-
munications in Statistics, 18, 2615-2627.
Robinson, P. M. (1988), Root-N Consistent Semiparametric Regression, Econometrica, 56,
931-954.
Robinson, P. M. (1989a), Hypothesis Testing in Semiparametric and Nonparametric Models
for Economic Time Series, Review ojEconomic Studies, 56, 51 1-534.
Robinson, P. M. (1989b), Nonparametric Estimation of Time-Varying Parameters, in Statisti-
cal Analysis and Forecasting ofEconomic Structural Change (P. Hackl, ed.), Springer-
Verlag, New York and Berlin.
Roy, N. (1996),Series Estimation of Income Elasticity of Calorie Intake Across Income Classes,
Journal of Quantitative Economics, 12(2), 77-84.
Ruppert, D. and M. P. Wand (1994), Multivariate Locally Weighted Least Squares Regression,
Annals of Stat istics, 22, 1346- 1370.
Ryan, J. G., P. D. Bidinger, P. Pushpamma, and N. Prahlad Rao (1984), The Determinants
ojlndividual Diets and Nutritional Status in Six Villages of Southern India, Research
Bulletin No. 7, ICRISAT, Patancheru.
Stone, C. J. (1977), Consistent Nonparametric Regression, Annals of Statistics, 5, 595-644.
(1977).
Stone, C. J. (1980). Optimal Rates of Convergences for Nonparametric Estimation, Annals of
604 ULlAH AND ROY
Stone, C. J. (1982), Optimal Global Rates of Convergence for Nonparametric Regression, An-
nals ofStatistics, 10, 1040-1053.
Strauss, S. and D. Thomas (1990), The Shape of the Calorie Expenditure Curve, Economic
Growth Centre Discussion Paper no. 595, Yale University.
Subramanian, S. and A. Deaton (1996),The Demand for Food and Calories: Further Evidence
from India, Journcd of Policital Economy, 104(l), 133-162.
Ullah, A. (1988), Semiparametric and Nonparametric Econometrics, Physics-Verlag, Heidel-
berg.
Vinod, H. D. and A. Ullah (1988), Flexible Production Estimation by Nonparametric Kernel
Estimators, in Advances in Econometrics: Robust and Nonparametric Statistical Infer-
ence (T. B. Fomby and C. F. Rhodes, eds.), JAI Press, Greenwich.
Walker, T. S. and J. G. Ryan (1989), Village and Household Economies in India’s Semiarid
Tropics, The Johns Hopkins University Press, Baltimore.
Wallace, T. D. and A. Hussain (1969), The IJse of Error Components Model in Combining
Cross-Section with Time-Series Data, Econometrica, 37, 55-72.
Watson, G. S. (1964), Smooth Regression Analysis, Sankhya, Series A , 26,359-372.
18
On Calibration
Adrian Rodney Pagan

The Aus tralian National University, Canberra, Australia
1. INTRODUCTION
Econometric methodologies generally develop spasmodically, frequently spurred on

by advances in computing power, and it is generally the case that they are codified
long after they have been in use for some time. For the past two decades macro-
econometricians have tended to classify econometric methodologies as those associ-
ated with three schools of thought: the “LSE approach,” VAR analysis, and Bayesian
methods. Microeconometricians have tended to choose from a smaller set, and much
of this research remains of the type that features quite complex statistical analysis
laid over a thin veneer of economic theory. Perhaps the best examples of the latter
are in a field such as labor econometrics, where one needs to be knowledgable about
a myriad of estimators designed to handle the statistical issues that arise with such
data. Regardless of the field, however, the passage of these decades has witnessed
the growth of a different philosophy concerning the nature of modeling and how to
go about it, an approach that its practitioners have increasingly given the rubric of
“calibration.” This literature is now very diverse. To support that contention consider
the following list of studies, each of which claims to be dealing with models that are
“calibrated.”
1. Macroeconomic Policy Models

Bank of Canada (1994) model (Black et al. 1994)
New Zealand Treasury (1994)model (Murphy 1995)
MSG model (McKibbin and Sachs 1991)
G3 model (McKibbin and Wilcoxen 1992)
2. Macroeconomic Analysis Models
King, Plosser, and Rebelo (1988)(real business cycles models)
605
606 PAGAN
3. Term Structure Models and Asset Pricing Models

Canova and Marrinan (1993)
Boudoukh (1993)
Fisher (1994)
Rudebusch (1995)
4. Macroeconomic Forecasting Models
Ingram and Whiteman (1994)(VAR model with RBC priors)
5. Computable General Equilibrium Models
6. Monetary Models
Nason and Cogley (1994)
Gilles et al. (1993)
7. International Monetary Models
Gruen and Gizycki (1993)
Bansal et al. (1995)
8. Industry Models
Rosen, Murphy, and Scheinkman (1994)
Jovanovic and MacDonald (1994)
So it is clear that the practice of “calibration” is a very popular activity. It is

therefore important that some investigation be made into the question of what it is
and what difficulties one would face when attempting to use it when persuing an an-
swer to a question that relies on quantitative measurement. In the following section
we provide a definition and discuss what this definition means. This is followed by
a more extensive treatment of some of the issues. Section IV notes that those who
regard themselves as calibrators are fierce proponents of it, and this ferocity can
sometimes border upon intolerance for other approaches. yet there are many unre-
solved questions about whether calibration really is a satisfactory answer to the quest
for quantitative information. We point out some of these and ask whether they have
yet been satisfactorily answered within that literature. In some instances, where the
answer is in the negative, the technology exists that should enable one to be able to
go at least part of the way toward a resolution of these questions.
II. A DEFINITION OF CALIBRATION
One issue that rises is how all the authors listed above can regard themselves as in-
volved in calibration-do they mean the same thing when describing what they do
in this way? To get to the heart of this we really need some definition of the term.
That is not easy to come by. Definitions, such as “. . . calibration . . . is not estima-
tion” (Kydland and Prescott 1996, p. 74) are negative rather than positive, while
alternative explanations, such as Kydland and Prescott (1991), tend to be rather too
diffuse. Nevertheless, I would agree with the negative statement above, in the sense
ON CALIBRATION 607
that the primary focus of calibrators is not really estimation, even though they may
perform estimation as part of their task. My definition is
Calibration is the process whereby data is employed in order to measure speci-

fied characteristics of a system.
There are three key words or phrases in the definition-“process,” “data em-
ployed,” and “measure specified characteristics,” and we need to dwell a little on
the last two elements, concentrating upon how one would carry out the process dis-
tinguished in the definition. The employment of data distinguished calibration from
exercises with models in which unknown parameters are just replaced with some
values in order that they might be simulated-what King (1996) refers to as quanti-
tative theory. The characteristics that investigators wish to measure are multifold-
examples drawn from the literature include
0 The moments of variables

0 The density of a variable
0 Derivatives of one variable with respect to another
In order to make these measurements it is necessary to describe the system

which is being used. In turn this requires
0 A description of relations between variables

0 An assignment of values to parameters in relations
0 A point at which the measurement is made
How to describe relations between variables invokes the question of the role
of theory in econometric analysis. For this reason we defer discussion of it until later.
This leaves the issues of how parameter values in a relation are to be assigned and
what methods can be used for evaluating any model derivatives, etc., that pertain to a
specific point. The last issue arises in many computable general equilibrium (CGE)
models. Because the aim of experiments in such models is to perform the equivalent
of comparative static analysis, it was important to have a data set that represented
an equilibrium position from which the derivative could be computed. Construction
of such a data set was termed “benchmarking,” and the process is generally part of
the “calibration” strategy of CGE researchers. Again, this points to the elusiveness
of the concept. In few of the other areas in which “calibration” is practiced would
we find benchmarking to be an important part of the analysis.
111. ASSIGNMENT OF PARAMETERS
This leaves us with the topic of parameter assignment. There is no one way of doing
estimation that is common to all who describe themselves as calibrators. Instead,
608 PAGAN
the whole gamut of estimation procedures is represented, ranging from the somewhat
vague (and possibly inconsistent) prescriptions of Kydland and Prescott:
Thus data are used to calibrate the model economy so that it mimics the world as
closely as possible along a limited, but clearly specified, number of dimensions.
(Kydland and Prescott 1996, p. 74)
It is important to emphasize that the parameter values selected are not the ones
that provide the best fit in some statistical sense. (Kydland and Prescott 1996,
P- 74)
to maximum likelihood (e.g., McGrattan 1994), GMM (e.g., Christian0 and Eichen-
baum 1992, Fkve and Langot 1994), and indirect estimation (e.g., Smith 1993, Bansal
et al. 1995).Indirect estimation is an interesting approach in that it brings together
those who are primarily interested in fitting statistical models to data with those con-
cerned with having a theoretical model as the way of organizing the facts. In indirect
estimation, as set out in Gourieroux et al. (1993)and Gallant and Tauchen (1996),
the parameters of the theoretical model are derived from the estimated parameters
of the statistical model. The method works from the observation that, if the theoret-
ical model is correct, then, from the principles of encompassing, one can predict
what the parameters of the statistical model should be. Hence, if we reverse the nor-
mal encompassing methodology, we can recover estimates of the parameters of the
theoretical model from those of the statistical model.
It is worth asking why one does this rather than fit the theoretical model di-
rectly to the data, as would be the practice with those doing MLE, i.e., “direct” esti-
mation. It turns out that there may be some gains to doing indirect rather than direct
estimation. To see this we look at a small stochastic equilibrium model set out in In-
gram (1995). In this model the system consists of equations describing the evolution
of the log of the capital stock, k,, productivity, a,, and the real interest rate r,, of the
form (Ingram 1995, p. 20)
where A1 is a function of the discount factor /3 and the cost of adjustment coefficient
6, and the equation describing the evolution of the capital stock comes from solving
the Euler equations.
Now let us play some games in which we describe the results from being a cali-
brator, who is either performing direct estimation or employing an indirect estimator
as a way of measuring any unknown parameters. We will assume that the “theorist,”
ON CALIBRATION 609
who provides the model to be estimated, makes some errors, and we ask how robust
the estimation methods are to these mistakes. The parameters to be estimated will be
p 1 , 6 , /3, and p r . One can always get consistent estimators of pa and pr from the sec-
ond and third equations of the system, allowing u s to concentrate on the estimation
of the first equation as that relevant to producing estimates of 6 and B. The statistical
model we choose for indirect estimation is
Case I : Calibrationist Invalidly Assumes p I = pr

Direct estimator ofp,, 6 , /3
The direct estimator estimates k , = const + +

(AI pa)k,-l - pahlkt-2 error. +
There is clearly a specification error since r,- I has been invalidly excluded. Assum-
ing that least squares is used on each of the first two equations pn will be consistently
estimated, but the estimator of h l , and hence 6 and /3, will be inconsistent.*
Indirect estimator
The idea behind indirect estimation is to find what values of 6, /3 are implied by
b l , b2, and b3. In the statistical model bl , b z , and b3 are all consistently estimated;
i.e., hl + pa and - p n h l , are consistently estimated. Hence, the estimation of B
and 6 will be consistent. Thus the use of the general statistical model as the way
of inferring estimates of the parameters of the theoretical model has protected u s
against a misspecification that comes from making an incorrect assumption about
the parameters of the latter.
Case 2: Modeler Assumes po = 0

Direct estimators
The direct estimators of 6, ,t? are inconsistent as the equation estimated is
and this still involves a specification error in that there are incorrect restrictions
imposed between the coefficients of r,-l, k , - l , and k t - q .
*Of course we cannot recover both B and 6 from a single parameter h i , hut the variance of the error term
also contains B and 6 and it can be consistently estimated if A1 can be.
610 PAGAN
Indirect estimators
The indirect estimator will also be inconsistent for the same reason as the direct es-
timator, i.e., even though b l , bz, and b3 are consistent, the theoretical model imposes
an incorrect relation between them.
Even though indirect estimation might be more robust than direct estimation, I
do not feel that one should oversell this idea, and my presumption would be that there
is likely to be little gain from doing indirect estimation, at least in regards to avoiding
the consequences of specification error. It does seem though that, in many instances,
indirect estimation may be an easier way to do estimation, in that information on
good statistical models of data is plentiful, and it is frequently easy to simulate from
theoretical models, which is the modus operandi of indirect estimation. An example
would be models of exchange rates. These can become very complex when allowance
is made for intervention points, etc., and direct estimation may be very difficult.
There is an extensive literature on the type of GARCH models that fit such data, so
it makes sense to use these models to estimate the parameters of some underlying
theoretical model of exchange rates. One might even argue that it is a philosophy
that is ideally suited to calibration endeavors in that it provides the rationale for
a division of labor between those designing good statistical models to fit the data
and those interested in generating economic models. It is likely to be rare that any
individual has skills in both of these areas and the indirect estimation principle
therefore provides a way to reap the benefits of specialization when estimating the
parameters of economic models.
IV. W H A T IS T H E DEBATE ABOUT?
With so much agreement one might wonder what the argument is about? I think that
there are three major areas in which calibrators have a distinctive stance and these
revolve around
0 Weak versus strong theory

0 Weak versus strong data consistency
0 Role of statistics
A. Preeminence of Theory
The following quotations provide what might be regarded as the polar cases in atti-
tudes toward theory. At one level is the “LSE approach” to econometrics. It is not the
case that such a methodology eschews theory, but it sees theory as just one element
in modelling, as witnessed by the following statements.
ON CALIBRATION 61 I
. . . there is nothing that endows an economic theory with veracity a priori, and so
coherence of an econometric model with an economic theory is neither necessary
not sufficient for it to be a good model. (Mizon 1995, pp. 115-116)
I think it is important to emphasize that the issue is not theory versus “no
theory,” even though many representatives of the calibrationist approach do seem
to make such a stark contrast. Empirical work in most methodological traditions to-
day is sensitive to the need for theory. Indeed, even in the “systems of equations”
approach most despised by Kydland and Prescott (1991) there are few models nowa-
days that do not have a strong theoretical core; see Hall (1995, pp. 980-983) for
a brief review of this fact and Murphy (1988)and Powell and Murphy (1995)for a
working model. The issue is more “how much theory” or “what type of theory” rather
than “no theory.”
In contrast to the position just advanced is a statement by Kydland and Prescott
about the role of theory that seems to be shared, to different degrees, by most of those
who see themselves as calibrators. It is my belief that it is this belief in thepreemi-
nence of theory that distinguishes a calibrator from a noncalibrator.
The degree of confidence in the answer depends on the confidence that is placed
in the economic theory being used. (Kydland and Prescott 1991, p. 171)
A belief in the preeminence of theory carries with it the stance that consistency
with data is of secondary importance; i.e., “strong” data consistency is not necessary
when working with models. One might ask if such a stance is reasonable. I think it
is if all we are doing with the model is demonstrating thefeasibility of a particular
outcome, i.e., we are doing quantitative theory. A good example of this would be
the debate over the validity of uncovered interest parity as an essential element of
many macroeconometric models. Defining the log of the spot exchange rate as S,, its
expected rate one period in the future as S;+], and the forward rate as F,, covered
interest parity yields
Invoking rational expectations so that SL+.l = SE+, + et+l, uncovered interest parity
eventuates as
A&+l = B ( K - S,) + e,+1

where B = 1. Now regressions of AS,+] on the forward discount F, - S , frequently
B
produce values of < 0, seemingly repudiating uncovered interest parity. Hence, it
is interesting to see if we can construct models that would produce 2 0. Gruen and
Gizycki (1993)produce such a calibrated model. In this instance it does not seem
relevant to ask whether the model produces other known characteristics of AS,, such
as GARCH and leptokurtosis; the aim of the model is just to show that it is possible
to replicate a striking feature of the data with a plausible set of assumptions in a
theoretical model.
612 PAGAN
B. Consistency with Data
Of course, this raises the issue of the credibility of a model when used for a range
of issues rather than just the replication of a single fact. It is rare to see a calibrated
model whose originator is not aiming to say something about the real world, from
statements that business cycles are largely due to supply side shocks to the idea that
cycles reduce welfare by very small amounts. I really find it impossible to believe
that anyone can take such conclusions or prescriptions seriously if they derive from
models whose credentials have not been established by measuring them against the
data. It is therefore fascinating, and troubling, to look at the first of the two polar
attitudes that I discern in the literature and which are reproduced below.
If the theory is strong and the measurements good, we have confidence that the
answer for the model economy will be essentially the same as for the actual
economy. (Kydland and Prescott 1996, p. 83)
The second position is certainly the opposite:

If an econometric model is to be taken and used seriously then its credentials
must be established. Two important ways to do this are to demonstrate that the
model is coherent with the available relevant information, and that it is at least
as good as alternative models of the same phenomenon. (Mizon 1995, p. 115)
The statement by Kydland and Prescott comes very close to blaming the data if the
calibrator’s model fails to fit. It is breathtaking because of our lack of strong theory.
We have theory, but to think it is this “strong” is truly amazing. The idea that a model
should be used just because the “theory is strong,” without a demonstration that it
provides a fit to an actual economy, is mind-boggling.
One might argue that few calibration exercises fail to include some evidence on
their fit to data. There are however two defects currently in such presentations which
hamper my acceptance of the proposition that the credibility of the maintained mod-
els has been established. One of these is the selective nature of the facts upon which
fit is to be judged. In some exercises this seems to come down to a single index, e.g.,
the correlation between hours and productivity. In others (e.g., Burnside et al. 1993),
the attempts at model evaluation are far more respectable, in that quite a number of
features are examined for their correspondence with the data. Nevertheless, it is the
case that few of these attempts are holistic. What is to be regarded as holistic de-
pends upon the nature of the problem, but when the models involve restrictions upon
a VAR, as is typical of most RBC and monetary models, it seems appropriate to test
aZZ the restrictions, and not just a subset of them.* King and Watson (1995) make
*Anderson (1991)makes the same point in commenting on Kydland and Prescott’s (1991)paper.
O N CALlBRATiON 6 I3
this same point, although they prefer the comparison of impulse responses. As ex-
plained in Pagan (1995)I do not think impulse response comparisons are as good
a method of assessing fit as a VAR, albeit it may be that any discrepancy between
model and data might be usefully expressed in terms of a discrepancy between the
model and data based impulse responses. In Canova et al. (1994) this philosophy
was put into action with respect to the model of Burnside et al. (1993). That model
looks good when judged by a limited number of features, but very poor when it is
forced to address all aspects of the VAR.
Even if one abstracts from the proper way to evaluate a model, one is still
left with the question of how we are to assess the magnitude of any inconsistency
between data and model. Early on in calibration studies, “eyeball” tests seemed to
predominate as measures of the size of the deviation between the model and real-
ity. Consequently, the chosen metric was very fuzzy, leaving one to despair at any
agreement being reached over whether a model is satisfactory. Just as beauty is in
the eye of the beholder, some of the judgments rendered concerning fit seemed quite
remarkable; e.g., a glance at the woeful (to me) match between model predictions
and data in either Figure 5 of Hansen and Prescott (1993)or Figure 4 in Jovanovic
and MacDonald (1994) leaves one with a sense of wonderment when reading that
the authors described the graphical evidence as supportive of the model. Given the
tendency for these same authors to pull out the measuring stick of predictive perfor-
mance when judging other methodologies, e.g., “one reason for its demise was the
spectacular predictive failure of the approach” (Kydland and Prescott (1991, p 166),
a legitimate question would seem to be why such a benchmark should not be uni-
versal rather than particular. Fortunately, some discipline has begun to emerge in
this literature, mostly through variants of statistical hypothesis testing (e.g., Burn-
side et al. 1993),and it is therefore time to turn to the question of the role of statistics
within the calibration agenda.
C. Role of Statistics
In many papers written by calibrationists there is a clear hostility to the use of purely
statistical models of data. This is most apparent in Kydland and Prescott’s (1991)
paper where the statistical models are identified with the “systems of equation ap-
proach.” In that paper, Frisch’s name is invoked as someone who heartily disap-
proved of this type of work. Actually, a closer reading of the paper they cite does not
support that interpretation. Frisch was certainly worried about how much informa-
tion there was in time-series data, and he was very much in favor of investigators
“going down coal mines” in order to collect and understand the workings of the in-
stitutions they were studying, but the whole paper that is quoted so approvingly by
Kydland and Prescott, is directed against the use of theoretical models that are not
closely connected with modeling features seen in the real world-what he terms
614 PAGAN
“playometrics.”* His stance on this vividly reminds one of some parts of the calibra-
tionist school of modelling as to be worth recording:
In too many cases the procedure followed resembles too much the escapist pro-
cedure of the man who was facing the problem of multiplying 13 by 27. He was
not very good at multiplication but very proficient in the art of adding figures,
so he thought he would try to add these figures. He did and got the answer 40,
which mathematically speaking was the absolutely correct answer to the prob-
lem as he had formulated it. But how well does the figure 40 tell us about the
size of the figure 351?
Nevertheless, despite a suspicion about statistical analysis, calibrators do increas-

ingly use statistics, and, when they do, I do believe there are some attendant diffi-
culties. These stem from the observations that
All model economies are abstractions and are by definition false. (Kydland and
Prescott 1991, p. 170)
It is pointless to test all the strong restrictions implied by this simple model: it
is known to be wrong in its details, and formal statistical rejections of the null
would tell us no more than we already know. The more interesting question is,
How wrong is it? (Rosen et al. 1994, p. 482)
If one takes this proposition seriously then it calls into question our ability to easily
decide whether a model fits the data based on some specified metric. In particular
the type of analysis described below becomes problematic.
. . . first a set of statistics that summarizes relevant aspects of the behaviour of
the actual economy is selected. Then the computational experiment is used to
generate many independent realizations of the equilibrium process for the model
economy. In this way, the sampling distribution of this set of statistics can be
determined to any degree of accuracy for the model economy and compared
with values of the set of statistics for the actual economy. (Kydland and Prescott
1996, p. 75)
To see what the problem is assume that the model described by the calibrator
has the form
where z: are the “latent” (unobserved) variables of the theoretical model, et are
shocks that drive the theoretical model, and 0 are the parameters of the model. The
random variable whose realizations are the data will be z t , and it will be definitional
*It is therefore rather ironic to read Kydland and Prescott’s (19%) comment that “here by theory we do
not mean a set of assertions about the actual economy’’ (p. 72).
ON CALIBRATION 615
that z, = 2: + U,, where the properties of the “observation errors” U, are unknown.
This is simply a formal description of a misspecified model. To complete the analysis
we suppose that the calibrationist computes some quantity, g ( z , , 8); e.g., this could
be the sample variance of (say) output. Kydland and Prescott’s proposal is to treat
g(zt, 6 ) as fixed and to study the distribution of g ( z : , 6 ) , locating where g ( z , , 6)
lies in this distribution. In order to find the distribution of g(zr, 6) one only needs to
be able to simulate from the theoretical economy. Despite its seductive appeal, the
procedure is an invalid one, unless it is assumed that ZI, = -zr, as only then is it
true that 2, would remain constant as different realizations of z: are made. Otherwise,
we need to make some assumptions about how U, varies with 2.: By far the simplest
solution, used by most investigators, is to make zt = 2.: Then the quantity of interest
will be taken to be a function of 8 alone and g ( z t ,6) can be compared to the value
predicted by the model. However, this presumes that the model is correctly specified
and contradicts Kydland and Prescott’s fundamental premise about such models.*
What does one do in the face of this problem? My answer, described in more
detail in Pagan (1994,pp. 8-9; 1995, p. 51), would be to take the theory of misspec-
ification of econometric models seriously when judging the significance of a value
of g ( z , , 6). The theory for completing this task is now very well developed (e.g., see
White 1994),as are the computational methods. Perhaps the major obstacle to im-
plementing the theory is to arrive at a description of how the data is generated that is
independent of the theoretical model. In some cases, such as the analysis of macro-
economic data, a VAR would seem to be appropriate, but in other instance one may
need to fit quite complex models, e.g., as in Bansal et al. (1995).This is how I see
both statistical and theoretical models being used in a way that benefits from spe-
cialization. Unlike indirect estimation, which assumes the validity of the theoretical
model and then estimates its parameters from a statistical model after generating re-
alizations from the theoretical one, the scheme above reverses the steps, simulating
data from the statistical model to be used for studying estimators and statistics that
are associated with the theoretical model. Diebold et al. (1995)apply such a scheme
when evaluating the quality of the “cattle cycle” model in Rosen et al. (1994).
ACKNOWLEDGMENTS
This chapter was the basis of my comments made in the Calibration Symposium at
the 7th World Congress of the Econometric Society in Tokyo, August 1995. Some of
the ideas are drawn from Canova et al. (1994),Kim and Pagan (1995),Pagan (1994),
and Pagan (1995).
*Thispoint is also relevant to those proposals for a Bayesian rather than classical assessment of the quality
of the model e.g., De Jong et al. (1996).
616 PAGAN
REFERENCES
Anderson, T. M. (1991), Comment of F. E. Kydland and E. C. Prescott, The Econometrics of

the General Equilibrium Approach to Business Cycles, in S. Hylleberg and M. Paldam
(eds.), New Approaches to Empirical Macroeconomics, Blackwell, Oxford, 51-56.
Bansal, R. A., R. Gallant, R. Hussey, and G. Tauchen (1995), Nonparametric Estimation of
Structural Models for High-Frequency Currency Market Data, Journal of Econometrics,
66,251-287.
Black, R., D. Laxton, D. Rose, and R. Tetlow (1994), The Bank of Canada’s New Quarterly
Projection Models. Part 1.The Steady State Model: SSQRM, Technical Report No. 72,
Bank of Canada.
Boudoukh, J. (1993), An Equilibrium Model of Nominal Bond Prices with Inflation-Output
Correlation and Stochastic Volatility, Journal of Money, Credit and Banking, 25,636-
665.
Burnside, C., M. Eichenbaum, and S. Rebelo (1993, Labor Hoarding and the Business Cycle,
Journal of Political Economy, 101,245-273.
Canova, F., M. Finn, and A. R. Pagan (1994), Evaluating a Real Business Cycle Model, in
C. P. Hargreaves (ed.), Nonstationary Time Series Analysis and Cointegration, Oxford
Canova, F. and J. Marrinan (1996), Reconciling the Term Structure of Interest Rates with
the Consumption Based ICAP Model, Journal of Economic Dynamics and Control, 20,
709-750.
Christiano, L. and M. Eichenbaum (1992), Current Real Business Cycles Theories and Ag-
gregate Labor Market Fluctuations, American Economic Review, 82,430-450.
DeJong, D. N., B. F. Ingram, and C. H. Whiteman (1996), Beyond Calibration, Journal of
Economic Dynamics and Control, 14, 1-9.
Diebold, F. X., L. Ohanian, and J. Berkowitz (1995), Dynamic Equilibrium Economies: A
Framework for Comparing Models and Data, NBER Technical Paper No. 174.
Fitve, P. and F. Langot (1994), The RBC Model Through Statistical Inference: An Application
with French Data, Journal of Applied Econometrics, 9 , Sll-S35.
Fisher, S. J. (1994), Asset Trading, Transaction Costs and the Equity Premium, Journal of
Applied Econometrics, 9, S71-S94.
Frisch, R. (1970), Econometrics in the World Today, in W. A. Eltis, M. F. G. Scott, and J. N.
Wolfe (eds.), Induction, Growth and Trade: Essays in Honour of Sir Roy Harrod, Claren-
don Press, Oxford, 152-166.
Gallant, A. R. and G. Tauchen (19%), Which Moments to Match?, Econometric Theory, 12,
657-68 1.
Gilles, C., J. Coleman and P. Labadie (1993), Identifying Monetary Policy with a Model of the
Federal Reserve, Finance and Economics Discussion Paper Series, Board of Governors
of the Federal Reserve System 93-24.
Gourieroux, C.,A. Monfort and E. Renault (1993), Indirect Inference, Journal of Applied
Econometrics, 8, S854118.
Gruen, D. W. R. and M. C. Gizycki (1993), Explaining Forward Discounting Bias: Is It An-
choring?, Reserve Bank of Australia Research Discussion Paper, No. 9307.
Hall, S. (1995), Macroeconomics and a Bit More Reality, Economic Journal, 105,974-988.
ON CALIBRATION6 I 7
Hansen, G. D. and E. C. Prescott (1993), Did Technology Shocks Cause the 1990-1991 Re-
cession, American Economic Review Papers and Proceedings, 83,280-286.
Ingram, B. F. (1995),“Recent Advances in Solving and Estimating Dynamic Macroeconomic
Models,” in K. D. Hoover (ed.). Macroeconometrics: Developments, Tensions and Pros-
pects (Kluwer Academic Publishers. Boston) 1546.
Ingram, B. F. and C. H. Whiteman (1994),Towards a New Minnesota Prior: Forecasting Macro-
economic Series Using Real Business Cycle Model Priors, Journal of Monetary Eco-
nomics, 47,497-510.
Javanovic, B. and G. MacDonald (1994),The Life Cycle of a Competitive Industry, Journal of
Political Economy, 102,322-347.
Kim, K. and A. R. Pagan (1995),The Econometric Analysis of Calibrated Macroeconomic
Models, in M. H. Pesaran and M. R. Wickens (eds.) Handbook ofApplied Econometrics,
Blackwell, Oxford, 356-390.
King, R. G. (1995),Quantitiative Theory and Econometrics, Federal Reserve Bunk of Rich-
mond Economic Quarterly, 81/3,53-105.
King, R. G., C. I. Plosser, and S. T. Rebelo (1988),Production, Growth and Business Cycles.
I: The Basic Neoclassical Growth Model, Journal of Monetary Economics, 21, 195-
232.
King, R. G. and M. W. Watson (1995),On the Econometrics of Comparative Dynamics, mimeo,
University of Virginia.
Kydland, F. E. and E. C. Prescott (1991),The Econometrics of the General Equilibrium Ap-
proach to Business Cycles, Scnndinavian Journal of Economics, 93, 161-178.
Kydland, F. E. and E. C. Prescott (1996),The Computational Experiment: An Econometric
Tool, Journal of Economic Perspectives, 10,6945.
McCallum, B. T. (1994),A Reconsideration of the Uncovered Interest Parity Relationship,
Journal of Monetary Economics, 33, 105-132.
McGrattan, E. R. (1994),The Macroeconomic Effects of Distortionary Taxation, Journal of
Monetary Economics, 33, 573-601.
McKibbin, W. J. and J. D. Sachs (1991),Global Linkages, Brookings Institution, Washington,
DC.
McKibbin, W. J. and P. J. Wilcoxen (1993),G-Cubed: A Dynamic Multi-sector General Equi-
librium Growth Model of the Global Economy, Brookings Discussion Puper in Znterna-
tional Economics No. 98, Brookings Institution, Washington, DC.
Mizon, G. M. (1995),Progressive Modelling of Macroeconomic Time Series: The LSE Method-
ology, in K. D. Hoover (ed.), Macroeconometrics: Developments, Tensions and Prospects,
Kluwer, Boston, 107-170.
Murphy, C. W. (1988), An Overview of the Murphy Model, in M. Burns and C. W. Murphy
(eds.), Macroeconomic Modelling in Australia (supplementary conference issue of AUS-
tralian Economic Papers), 61-68.
Murphy, C. W. (1995),A Model ofthe New Zealand Economy, New Zealand Treasury, Welling-
ton.
Nason, J. M. and T. Cogley (1994),Testing the Implications of Long-Run Neutrality for Mon-
etary Business Cycle Models, Journal of Applied Econometrics, S37-S70.
Pagan, A. R. (1994),Calibration and Economic Research: An Overview, Journal of Applied
Econometrics, 9, S1-Sl0.
618 PAGAN
Pagan, A. R. (1995), Some Observations on the Solution, Estimation and Use of Modern
Macroeconometric Models, in K. D. Hoover (ed.), Macroeconometrics: Developments,
Tensions and Prospects, Kluwer, Boston, 47-55.
Powell, A. A. and C. W. Murphy (1995), Inside a Modern Macroeconometric Model: A Guide
to the Murphy Model, Springer-Verlag, Berlin and New York.
Rosen, S., K. M. Murphy, and J. A. Scheinkman (1994), Cattle Cycles, Journal of Political
Economy, 102,468-492.
Rudebusch, G. D. (1995), Federal Reserve Interest Rate Targeting, Rational Expectations,
and the Term Structure, Journal of Monetary Economics, 35,245-274.
Smith, A. A. (1993), Estimating Non-linear Time-Series Models Using Simulated Vector Au-
toregression, Journal of Applied Econometrics, 8,563-584.
White, H. (1994), Estimation, Inference and Spec$cation Analysis, Cambridge University
Press, Cambridge.
Aggregation Autoregressive (AR) process, 299,
consistency, 125 386387,566
exact, 180,199-208 periodic, 566
condition, 201 spatial moving average (SARMA), 251
Gorman, 183-191
additive aggregation, 184-185
Bayesian analysis, 368,379,427,605
affine homotheticity, 197
Best linear unbiased predictor (BLUP),
polar form (GPF), 179, 187
294
quasi-homotheticity, 187
BFGS algorithm, 528
Stone-Geary structure, 187-188
Binary response model, 75
joint, 219
Boole-Bonferroni inequalities (see Union
for commodities and agents, 227
intersection)
Klein, 220-222
Bootstrap, 145, 164, 419, 436
Klein-Nataf structure, 182,220-222
for confidence intervals, 424,425
multiple fixed inputs, 225-226
percentile-t, 436
of fixed and variable inputs,
refinements, 428,429-432
219-220,226-227
simple, 421
vintage models, 225-226
size correction, 433
Muellbauer, 179, 191-196, 199
Bounds test
Almost ideal demand system (AIDS), 179,
of Kiviet, 489-491
196
of Zinde-Walsh and Ullah, 486,489,
quadratic, 179
492
Analysis of variance (ANOVA), 293,298,
303,305
Approximate slope analysis, 394 Calibration, 605
Atkinson-Bourguigon condition, 131 definition, 606
Atkinson’s index, 6,8,9,47,48, 140 direct estimator, 609
Autocorrelation (see also Serial indirect estimator, 609
correlation) Calorie-income relationship, 595
function (ACF), 556 Calorie intake, 595,596,598
of errors, 394,413,434, 528-529,559, Cambridge controversy, 181
561 Census X-11, 568,573,574
spatial (see Spatial autocorrelation) Central limit theorem, 23,253
619
620 INDEX
Cluster fixed effects (see Panel data, fixed Error component model, 293,299-302,
effects) 305
Cluster sampling (see Sampling, cluster) Error correction model, 5 3 0 4 3 1 , 5 3 9
Coefficient of variation, 7,9, 11, 124, 333, Estimator, direct, 609
335
Cointegration, 507 Fixed difference two stage least squares
seasonal, 565 (FD-2SLS),303,304
Collinearity, 374, 378,597 Fixed effects model (see also Panel data),
Conditional heteroskedasticity (see 293-295,298,299,301,303,305,
Autoregressive conditional 306,308-310
heteroskedasticity) Fourier series, 398400
Confidence set estimation (see Union Full information maximum likelihood
intersection) (FIML) estimator (see Simul-
Cook’s distance, 458 taneous equations models)
Covariates, 29-32
Cram&-Rao lower bound, 294
Galtonian model, 139
Critical point, 474
Gastwirth bounds, 53, 55
Cross-entropy principle, 368
Gauss-Newton regression, 384
Generalized additive models, 594
Debreu-Sonnenschein-Mantel (DSM),
Generalized entropy indices, 6 , 9 , 10, 16,
124,125,127
179,313
Diagnostic testing, 239,383-414 Generalized least squares (GLS), 259,
for GARCH effect, 390,610 293-295,302,313,331,337,341,
347,350,486,492,499
Hausman test, 384
higher moments, 389 Generalized method of moments (GMM),
information matrix test, 384-385 82,258-260,295,303,307,310,
311,374,467
J-test, 387,433
Geographic effects, 74, 82
multiple, 390
Geographic model, 67
of conditional mean, 386
variable addition, 384 Gibbs phenomenon, 399
Gibrat’s law, 139
Dimensionality, curse of, 594
Disequilibrium model, 507 Giffen paradox, 181
Gini coefficient, 8, 10, 11, 17,45,46,49,
Durbin’s h test, 277
53-56
Durbin-Watson test, 265,267,387
Hart’s mobility index, 139

Elasticity, 112-1 13, 301, 595-598, Hausman specification test, 296,345,
600 384,597
Empty intersection test (see also Union Hausman’s test, 295,308
intersection hypothesis testing), Head count ratio, 56
502 Herfiridahl coefficient, 49, 124
Encompassing Hessian matrix, 371
principle, 608 Heterogeneity, 312, 579, 596, 597
test, 54.6-547 and market demand, 213
Equilibrium model, 608 income, 214-219
INDEX 621
[Heterogeneity] Least absolute deviation method, 594

preference, 214-2 19 Leverage values, 457
Heteroskedasticity, 82,260,295, Likelihood ratio (LR) test, 264,271,273,
299-300,314,413,423 276,278,280,514
autoregressive conditional (ARCH), LIMDEP package, 280
390,524, 532,588 Limited dependent variables, 282,
generalized (GARCH), 390,610,611 307-308,314
consistent covariance matrix, 414 Limited information maximum likelihood
consistent standard errors, 82,424 (LIML) estimator (see
High-leverage points, 456-457 Simultaneous equations models)
detection of, 456 Linear expenditure system (LES), 177,
Homogeneity, 124,597 207
Locally mean most powerful test (LMMP),
Income gap ratio, 56 298
Inequality, 1 Locational simultaneity, 241
between groups, 3, 14, SO Logistic transition, 514, 544
income, 14,40,333,335 Logit model, 308
relative, 3,5-7 multinomial (see also Treatment choice
within group, 3, 14, SO models), 76,97, 104
Influential observations, 458-459 Longitudinal data, 579
detection of, 458 Lorenz curve, 2, 10, 12, 15, 18,26,27,29,
Information theory, 124 40-44,55,57,120, 122,132,136
Instrumental variables, 82,258-264, dominance, 16,26,27,43,121
301-303,387 generalized, 11, 12, 17, 18, 123, 141,
Integration 143
by parts, 582 dominance testing, 26
numerical, 308,314 equivalence testing, 26,27
periodic, 5 6 8 4 6 9 , 5 7 1 nonparametric estimators, 25
Inverse Mill's ratio, 314,594 parametric method, 21,22,24,29
semiparametric methods, 30
Jackknife, 430,431 nonparametric estimator of, 2.5,28,29
order, 41,42,44
Jarque-Bera test, 532
parametric method 21,22,24,29
semiparametric methods, 30
Kelejian-Robinson test, 268,280
Kernel
density estimator, 337, 582, 594 Malnutrition, 595
normal, 586, 5% Markov transition process, 133-134
Khintchine's theorem, 393 Maximum entropy
Kolmogorov-Smirnov tests, 141-142 aggregators, 122
Kurtosis, 537,611 estimation, 368-374
Maximum equalization percentage, 46,47
Lagrange multiplier (LM) test, 270, Maximum likelihood estimation (MLE),
297-298,314,514-515,518,524, 20,21,25,32,255-258,260-261,
533 276,294,305-306,308,311-313,
Latent variable model, 580 345,492,486,498-500
622 INDEX
Mean squared error, 296,333-335,376 Noncausality (Granger), 525

Measurement of inequality axiomatic Nonlinear parametric model, 588
approach, 48 Non-nested models, 387
M-estimation, 371 Non-normality, 332,335
Meta-analysis, 468 Nonparametric
Method of simulated moments, 311 estimation, 25-26,28-30,346
Milliken-Graybill theorem, 388 fixed effects model, 588-590,596
Minimal nesting, 529 average derivative, 590
Minimum norm quadratic unbiased generalized least-squares estimator,
estimation (MINQUE), 294,305 589,590
Minimum sufficient statistic, 308 local estimator, 589,590,591
Minmax property, 585 optimal h, S98
Missing data, 313 random effects model, 590-592, 596
Mobility, 67, 73, 83, 130 feasible estimator, 591
and gender, 146 local estimator, 591
and income level, 150 optimal h, 598
by education, age and race, 155
indices, 120, 121, 138
Omitted variables, 400-401
matrices, 134
Ordinary least squares (OLS), 81,256,
welfare ranking of, 133
275,278,294,302,313,345,387,
Model, augmented, 390
393,422,434,486
Monotonic transition, 511
Outlier detection, 441-461
Monte Carlo experiment, 271,278,294,
LMS (least median of squares), 454
296,309-311,374-378,
Mahalanobis distance, 447-450
400412,427,492
Studentized residual, 453
Moran’s I test, 265,266,270,280
tests, 455
Moving average (MA) process, 299,478,
502,560
of order 1,466,469 Panel data, 291,579
of order q, 469,483 dynamic model, 84,298,302-304
MSZ indices, 121,122, 128-131,139, fixed effects model (see also Fixed
140 effects model), 80,81,291,
300-301,579,581,595
cluster, 98, 101, 104, 106
Nadaraya-Watson kernel estimator, incomplete, 304-307
581-584 nonparametric, 588-590
asymptotic normality, 585 semiparametric, 592-594
average derivative, 585, 586 random effects model (see also Random
bias, 583 effects model), 293,301,580-581
mean square error, 583 nonparametric, 590-592
optimal h, 584 semiparametric, 592-594
pointwise estimator, 582 rotating panels, 579
squared residuals, 583, 588 unbalanced, 298
variance, 584 VAR model, 299
Newey-West covariance estimator, 414, Pareto improvement, 4
498 Pietra index, 47,49
INDEX 623
Poolability restriction, 296 [Rank of demand system]

Pooled nonparametric panel models, weak axiom of revealed preference
581-588 (WARP), 181
asymptotic normality, 585 RATS package, 280
average derivative, 585,586 Reduced form model, 370
bias, 587 Representative agent
local linear kernal estimator, 338, Afriat’s theorem, 198
584-588 consumer, 178-179,189,194,196-199
mean square error, 587 GARP test, 198
optimal h, 587 Monte Carlo method, 197-198
squared residuals, 587 producer, 178, 189,219
Poor areas, 63-65 revealed preference method, 196-197
Posterior probability, 371 Slutsky matrix, 196-197
Poverty, 55,56,71,74,84 test, 180, 196-199
Price-independent generalized linearity RESET test, 384,395-398,401-414,524
(PIGL), 179, 196,200,201,208, Fourier (FRESET), 384,398
213 Fourier-linear (FRESETL), 400-414
Price-independent generalized Fourier-sinusoidal (FRESETS),
logarithmic (PIGLOG). 179, 180, 400-4 14
199-201,208,213 Residual resampling, 423
Principle of transfers 15, 125 Returns to schooling, 482
Pigon-Ralton, 15,41, 136-139, 143
Robin Hood, 41,46,48,50
Probit model, 309, 313 Sample split, 502
Proxy variables, 119,385,393 Sampling
Pseudopanel, 306 cluster, 325,347,356,357
random, 325,327-339
simple random, 328,339,341
Quadratic expenditure system (QES), 207, stratified, 325,339,346,353,357
208 systematic, 325426,339,351,353,
Quadratic loss, 374 357
Scale measures, 5
and inequality indices, 43,44,120, 132
Random effects model (see also Panel quantile based, 6
data), 293-295,301,302,309, standard deviation of income, 6
313 variance of income, 6
Random sampling (see Sampling, random) Score test (see Lagrange multiplier test)
Rank of demand system, 181,208-213 Seasonality, 553
price-independent proportional and dummy variables, 542,544,558,
budgeting (PIPB), 212-213 565
proportional budgeting (PB), 212-2 13 and structural breaks, 570
rank 2,194,199,208-213,217 HEGY test, 562-565
rank 3,196,208-209 in dynamics, 562
rank 4,209 in house prices, 544
test, 208-210 in parameter, 566
uniform curvature (UC), 21 1 OCSB, test 562-565
624 INDEX
Seemingly unrelated regression equations [Smooth transition regressions (STR)]

(SURE) model, 300,433,467, specification, 526
478483, 502 Social welfare, 3
Semiparametric panel models, 592-594 convex, 44
between estimator, 593 economic measures of, 3-5
feasible estimators, 593 function, 14, 16,34, 119, 120, 124
fixed effects estimator, 592 Schur concave, 123, 137
generalized least squares estimator, 593 Sociometrics, 244
pooled estimator, 592 Spatial, 66
Serial correlation (see also autocorrelation, 237,239-252
Autocorrelation), 240, 300, 314 dependence, 237-281
Series estimator, 590 econometncs, 237-28 1
SHAZAM package, 280,398,405,424 effects, 68,69
Shrinkage estimator, 427 error dependence, 237-281
Simple random sampling (see Sampling, lag operator, 245-246
simple random) test, 239,264
Simultaneous equations models, 365 variation, 66,96
DLS, 367 weights, 243-245
FIML, 367 Specification testing, 594-595
GCE (generalized cross entropy), for linearity, 595
370-3 79 global, 594
GME solutions, 370-374,376 search, 426-428
k-class, 367 Standard errors
LIML, 367,373-374 cluster-correlated, 597
maximum-entropy estimator, 368 heteroskedasticity-consistent,82,424
sampling design, 375 Stochastic dominance, 1, 12,13, 15-18,
traditional estimators, 366 121,135,141,142,333
2SLS, 260,367,373 Stratified sampling (see Sampling,
3SLS, 367,375,379 stratified)
Simulation, 489 Structural break, 523, 570-571
Barnard method, 432-433 Super exogeneity, 509
Skewness, 537 SURE model (see Seemingly unrelated
Small h approximation, 588 regression equations model)
Small-o expansion, 588 Switching regression model, 508-509, 517
Smooth transition regressions (STR), 507, Markov, 572
511 Systematic sampling (see Sampling,
additive (STAR), 520,526,548 systematic)
application, 530 System of demands, 480
estimation, 527
evaluation, 528
linearity test, 513 Taylor-series expansion, 398,413,
LM-type test statistic, 515, 517 515-516,533,581,583,587
minimum nesting models (MNM), Theil index, 49,51, 124, 126, 140
529-530 3SLS (see Simultaneous equations models)
misspecification testing, 518, 520, 522, Time-varying parameters model, 81,
525 57 1-573
INDEX 625
Tobit model, 314 Union intersection, 465

Transfer sensitivity 18, 50 Boole-Bonferron inequalities, 467,473,
Transition matrix, 120, 132, 135-137, 50 1
140,145 confidence set estimation, 4 7 7 4 7 8
Translog, 204,396 hypothesis testing (see also Empty
function, 296-398 intersection test), 1 4 1 , 4 6 9 4 7 6
system, 204-205 variance stabilization, 432
Treatment choice models, 100 Unit root, 139,299, 303
multinominal logit, 96,97, 101, 104 seasonal, 563-564,570,571
2SLS (see Simultaneous equations models)
Wald test, 264,271,273,276,278,434,
Unconditional likelihood function, 308 467,514
Uniformly most powerful (UMP) test, 297, Weekend effect, 553-554
388,390,392 Window width, 581,596

Hand Book of Applied Economic Statistics

Uploaded by

Copyright:

Available Formats

Hand Book of Applied Economic Statistics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hand Book of Applied Economic Statistics

Uploaded by

Copyright:

Available Formats

HANDBOOK OF APPLIED

Handbook of applied economic statistics/edited by Aman Ullah, David E. A.

This book is printed on acid-free paper.

Eastern Hemisphere Distribution

World Wide Web

Copyright 0 1998 by MARCEL DEKKER, INC. All Rights Reserved.

Current printing (last digit):

PRINTED IN THE UNITED STATES OF AMERICA

The most important objective of this volume is to cover the developments in

with outliers, which is an alternative to computer-intensive quantile estimation tech-

Part I Applied Economic Statistics

I. Economic Statistics and Social Welfare Comparisons: A Review I

Part 2 Econometric Methodology and Data Issues

6. Aggregation and Econometric Analysis of Demand and Supply I77

9. Econometric Analysis in Complex Surveys 325

Part 3 Model Specification and Simulation

I5. Modeling Economic Relationships with Smooth Transition

Badi H. Baltagi, Ph.D. Professor, Department of Economics, Texas A&M University,

Anil K. Bera, Ph.D. Professor, Department of Economics, University of Illinois,

Robert V. Breunig Graduate Student, Department of Economics, University of Cal-

Chia-Hui Chiu Graduate Student, Department of Economics, University of Cali-

James B. Davies, Ph.D. Professor and Chair, Department of Economics, University

Anil B. Deolalikar, Ph.D. Professor, Department of Economics, University of Wash-

Jean-MarieDufour, Ph.D. Professor, C.R.D.E. and Department of Economic Sci-

Philip Hans Franses, Ph.D. Associate Professor, Department of Econometrics, Eras-

David E. A. Giles, Ph.D. Professor, Department of Economics, University of Victo-

Amos Golan, Ph.D. Visiting Associate Professor, Department of Agricultural and

David A. Green, Ph.D. Department of Economics, University of British Columbia,

Ali S. Hadi, Ph.D. Professor, Department of Statistics, Cornell University, Ithaca,

George Judge, Ph.D. Professor, University of California at Berkeley, Berkeley, Cal-

Walter Kramer, Dr. rer. pol. Professor, Department of Statistics, University of

Esfandiar Maasoumi, Ph.D. (FRS) Professor, Department of Economics, Southern

Douglas Miller, Ph.D. Assistant Professor, Department of Economics, Iowa State

Harry J. Paarsch, Ph.D. Associate Professor, Department of Economics, University

Martin Ravallion, Ph.D. Lead Economist, Development Research Group, World

Nilanjana Roy, Ph.D. Assistant Professor, Department of Economics, University of

R. Robert Russell, Ph.D. Professor, Department of Economics, University of Cali-

Mun S. Son, Ph.D. Associate Professor, Department of Mathematics and Statistics,

Timo Teriisvirta, Ph.D. Professor, Department of Economic Statistics, Stockholm

Olivier Torres, Ph.D. Maitre de Conferences, U.F.R. MathCmatiques, Sciences

Aman Ullah, Ph.D. Professor, Department of Economics, University of California

Michael R. Veall, Ph.D. Professor, Department of Economics, McMaster University,

partial-ordering approach using disaggregated dominance criteria is more attrac-

II. DIFFERENT ECONOMIC MEASURES OF

I. Per Capita Income

t(0.5) = F-' (0.5)

c(0.25) = F-'(0.25) and c(0.75) = F ' ( 0 . 7 5 )

B. Scale Measures and Inequality Indices

If a # 0 then these distributions differ in location, and if b # 1 they differ in scale.

I. Quantile-Based Measures of Income

Quantile differences or quantile ratios have been employed to provide rough-and-

An important characteristic of the variance, which carries through to the related

3. Standard Deviation of Income

A drawback of the variance is that it is in the units squared of income. Another

4. Coefficient of Variation of Income

5. Variance ofthe Logarithm of Income

we define the variance of the logarithm of Y by

where ~ D isE“equally distributed equivalent income”-the income such that if all

8. Generalized Entropy Family of Indices