Spatial Econometrics

Spatial Econometrics
James P. LeSage
Department of Economics
University of Toledo
CIRCULATED FOR REVIEW
December, 1998
Preface
This text provides an introduction to spatial econometrics as well as a set of
MATLAB functions that implement a host of spatial econometric estimation
methods. The intended audience is faculty and students involved in modeling spatial data sets using spatial econometric methods. The MATLAB
functions described in this book have been used in my own research as well
as teaching both undergraduate and graduate econometrics courses.
Toolboxes are the name given by the MathWorks to related sets of MATLAB functions aimed at solving a particular class of problems. Toolboxes
of functions useful in signal processing, optimization, statistics, finance
and a host of other areas are available from the MathWorks as add-ons
to the standard MATLAB software distribution. I use the term Econometrics Toolbox to refer to my collection of function libraries described
in a manual entitled Applied Econometrics using MATLAB available at
http://www.econ.utoledo.edu.
The MATLAB spatial econometrics functions used to apply the spatial
econometric models discussed in this text rely on many of the functions in
the Econometrics Toolbox. The spatial econometric functions constitute a
library within the broader set of econometric functions. To use the spatial
econometrics functions library you need to install the entire set of Econometrics Toolbox functions in MATLAB. The spatial econometrics functions
library is part of the Econometrics Toolbox and will be installed and available for use as well as the econometrics functions.
Researchers currently using Gauss, RATS, TSP, or SAS for econometric
programming might find switching to MATLAB advantageous. MATLAB
software has always had excellent numerical algorithms, and has recently
been extended to include: sparse matrix algorithms and very good graphical
capabilities. MATLAB software is available on a wide variety of computing
platforms including mainframe, Intel, Apple, and Linux or Unix workstations. A Student Version of MATLAB is available for less than $100. This
version is limited in the size of problems it can solve, but many of the exi
ii
amples in this text rely on a small data sample with 49 observations that
can be used with the Student Version of MATLAB.
The collection of around 450 functions and demonstration programs are
organized into libraries, with approximately 30 spatial econometrics library
functions described in this text. For those interested in other econometric
functions or in adding programs to the spatial econometrics library, see the
manual for the Econometrics Toolbox. The 350 page manual provides many
details regarding programming techniques used to construct the functions
and examples of adding new functions to the Econometrics Toolbox. This
text does not focus on programming methods. The emphasis here is on
applying the existing spatial econometric estimation functions to modeling
spatial data sets.
A consistent design was implemented that provides documentation, example programs, and functions to produce printed as well as graphical presentation of estimation results for all of the econometric functions. This
was accomplished using the structure variables introduced in MATLAB
Version 5. Information from econometric estimation is encapsulated into a
single variable that contains fields for individual parameters and statistics
related to the econometric results. A thoughtful design by the MathWorks
allows these structure variables to contain scalar, vector, matrix, string,
and even multi-dimensional matrices as fields. This allows the econometric
functions to return a single structure that contains all estimation results.
These structures can be passed to other functions that can intelligently decipher the information and provide a printed or graphical presentation of
the results.
The Econometrics Toolbox along with the spatial econometrics library
functions should allow faculty to use MATLAB in undergraduate and graduate level courses with absolutely no programming on the part of students
or faculty.
In addition to providing a set of spatial econometric estimation routines
and documentation, the book has another goal, applied modeling strategies
and data analysis. Given the ability to easily implement a host of alternative
models and produce estimates rapidly, attention naturally turns to which
models and estimates work best to summary a spatial data sample. Much
of the discussion in this text is on these issues.
This text is provided in Adobe PDF format for online use. It attempts
to draw on the unique aspects of a computer presentation platform. The
ability to present program code, data sets and applied examples in an online
fashion is a relatively recent phenomena, so issues of how to best accomplish
a useful online presentation are numerous. For the online text the following
iii
features were included in the PDF document.
1. A detailed set of bookmarks that allow the reader to jump to any
section or subsection in the text.
2. A detailed set of bookmarks that allow the reader to jump to an
examples or figures in the text.
3. A set of bookmarks that allow the reader to view the spatial datasets
and documentation for the datasets using a Web browser.
4. A set of bookmarks that allow the reader to view all of the example
programs using a Web browser.
All of the examples in the text as well as the datasets are available offline
as well on my Web site: http://www.econ.utoledo.edu under the MATLAB
gallery icon.
Finally, there are obviously omissions, bugs and perhaps programming
errors in the Econometrics Toolbox and the spatial econometrics library
functions. This would likely be the case with any such endeavor. I would be
grateful if users would notify me when they encounter problems. It would
also be helpful if users who produce generally useful functions that extend
the toolbox would submit them for inclusion. Much of the econometric code
I encounter on the internet is simply too specific to a single research problem
to be generally useful in other applications. If econometric researchers are
serious about their newly proposed estimation methods, they should take
the time to craft a generally useful MATLAB function that others could use
in applied research. Inclusion in the spatial econometrics function library
would have the added benefit of introducing new research methods to faculty
and their students.
The latest version of the Econometrics Toolbox functions can be found on
the Internet at: http://www.econ.utoledo.edu under the MATLAB gallery
icon. Instructions for installing these functions are in an Appendix to this
text along with a listing of the functions in the library and a brief description
of each.
Contents
1 Introduction
1.1 Spatial econometrics . . . . . . . . . . . . .
1.2 Spatial dependence . . . . . . . . . . . . . .
1.3 Spatial heterogeneity . . . . . . . . . . . . .
1.4 Quantifying location in our models . . . . .
1.4.1 Quantifying spatial contiguity . . . .
1.4.2 Quantifying spatial position . . . . .
1.5 The MATLAB spatial econometrics library
1.5.1 Estimation functions . . . . . . . . .
1.5.2 Using the results structure . . . . .
1.6 Chapter Summary . . . . . . . . . . . . . .
1.7 References . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
3
6
7
8
13
18
21
23
27
28
2 Spatial autoregressive models

2.1 The first-order spatial AR model . . . . .
2.1.1 The far() function . . . . . . . . .
2.1.2 Applied examples . . . . . . . . . .
2.2 The mixed autoregressive-regressive model
2.2.1 The sar() function . . . . . . . . .
2.3 The spatial errors model . . . . . . . . . .
2.3.1 The sem() function . . . . . . . . .
2.4 The general spatial model . . . . . . . . .
2.4.1 The sac() function . . . . . . . . .
2.5 An applied exercise . . . . . . . . . . . . .
2.6 Chapter Summary . . . . . . . . . . . . .
2.7 References . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
30
31
34
36
42
43
44
50
56
57
61
62
64
70
81
81
iv
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
3 Bayesian Spatial autoregressive models

3.1 The Bayesian regression model . . . . .
3.2 The Bayesian FAR model . . . . . . . .
3.2.1 The far g() function . . . . . . .
3.2.2 Applied examples . . . . . . . . .
3.3 Other spatial autoregressive models . . .
3.4 Applied examples . . . . . . . . . . . . .
3.5 An applied exercise . . . . . . . . . . . .
3.6 Chapter Summary . . . . . . . . . . . .
3.7 References . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
84
85
91
99
101
107
110
115
124
125
4 Locally linear spatial models

4.1 Spatial expansion . . . . . . . . . . . . .
4.1.1 Implementing spatial expansion .
4.1.2 Applied examples . . . . . . . . .
4.2 DARP models . . . . . . . . . . . . . . .
4.3 Non-parametric locally linear models . .
4.3.1 Implementing GWR . . . . . . .
4.4 A Bayesian Approach to GWR . . . . .
4.4.1 Estimation of the BGWR model
4.4.2 Informative priors . . . . . . . .
4.4.3 Implementation details . . . . . .
4.5 An applied exercise . . . . . . . . . . . .
4.6 Chapter Summary . . . . . . . . . . . .
4.7 References . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
127
127
129
133
142
154
156
158
163
166
167
170
184
185
5 Limited dependent variable models

5.1 Introduction . . . . . . . . . . . . .
5.2 The Gibbs sampler . . . . . . . . .
5.3 Heteroscedastic models . . . . . . .
5.4 Implementing these models . . . .
5.5 An applied example . . . . . . . .
5.6 Chapter Summary . . . . . . . . .
5.7 References . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
187
188
192
193
195
204
210
211
6 VAR and Error Correction Models

6.1 VAR models . . . . . . . . . . . . .
6.2 Error correction models . . . . . .
6.3 Bayesian variants . . . . . . . . . .
6.4 Forecasting the models . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
213
213
220
231
243
CONTENTS
6.5
6.6
6.7
vi
An applied exercise . . . . . . . . . . . . . . . . . . . . . . . . 252

Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . 257
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Appendix
260
List of Examples
1.1
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
4.1
4.2
4.3
4.4
4.5
4.6
4.7
Demonstrate regression using the ols() function

Using the fars() function . . . . . . . . . . . . .
Using sparse matrix functions . . . . . . . . . .
Solving for rho using the far() function . . . . .
Using the sar() function . . . . . . . . . . . . .
Using the xy2cont() function . . . . . . . . . .
Least-squares bias . . . . . . . . . . . . . . . .
Testing for spatial correlation . . . . . . . . . .
Using the sem() function . . . . . . . . . . . . .
Using the sac() function . . . . . . . . . . . . .
A large sample SAC model . . . . . . . . . . .
Least-squares on the Boston dataset . . . . . .
Testing for spatial correlation . . . . . . . . . .
Spatial model estimation . . . . . . . . . . . .
Heteroscedastic Gibbs sampler . . . . . . . . .
Metropolis within Gibbs sampling . . . . . . .
Using the far g() function . . . . . . . . . . . .
Using far g() with a large data set . . . . . . .
Using sem g() in a Monte Carlo setting . . . .
An outlier example . . . . . . . . . . . . . . . .
Robust Boston model estimation . . . . . . . .
Imposing restrictions . . . . . . . . . . . . . . .
The probability of negative lambda . . . . . . .
Using the casetti() function . . . . . . . . . . .
Using the darp() function . . . . . . . . . . . .
Using darp() over space . . . . . . . . . . . . .
Using the gwr() function . . . . . . . . . . . . .
Using the bgwr() function . . . . . . . . . . . .
Producing robust BGWR estimates . . . . . . .
Posterior probabilities for models . . . . . . . .
vii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
36
39
41
44
46
48
57
59
64
68
71
74
75
88
95
101
104
110
112
115
121
123
133
147
151
157
170
171
177
LIST OF EXAMPLES
5.1
5.2
5.3
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
6.10
6.11
6.12
6.13
6.14
6.15
Using the sarp g function . . . . . . . . .

Using the sart g function . . . . . . . . . .
Right-censored Tobit Boston data . . . . .
Using the var() function . . . . . . . . . .
Using the pgranger() function . . . . . . .
VAR with deterministic variables . . . . .
Using the lrratio() function . . . . . . . .
Using the adf() and cadf() functions . . .
Using the johansen() function . . . . . . .
Estimating error correction models . . . .
Estimating BVAR models . . . . . . . . .
Using bvar() with general weights . . . . .
Estimating RECM models . . . . . . . . .
Forecasting VAR models . . . . . . . . . .
Forecasting multiple related models . . .
A forecast accuracy experiment . . . . . .
Linking national and regional models . .
Sequential forecasting of regional models .
viii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
196
202
205
215
217
218
219
222
226
229
233
235
243
245
246
248
252
255
List of Figures
1.1
1.2
1.3
1.4
1.5
Measurement error and spatial dependence

Distributions of house age versus distance .
An illustration of contiguity . . . . . . . . .
Distance expansion estimates . . . . . . . .
GWR estimates . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
6
9
16
17
2.1
2.2
2.3
2.4
2.5
Spatial autoregressive fit and residuals

Sparsity structure of W from Pace and
Generated contiguity structure results
Spatial contiguity structures . . . . . .
Actual vs. Predicted housing values .
. . . .
Barry
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
40
47
65
73
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
Vi estimates from the Gibbs sampler . . . . . .

Conditional distribution of . . . . . . . . . .
First 100 Gibbs draws for and . . . . . . .
Mean of the vi draws . . . . . . . . . . . . . . .
Mean of the vi draws for Pace and Barry data .
Posterior distribution for . . . . . . . . . . .
Vi estimates for the Boston data . . . . . . . .
Posterior distribution of . . . . . . . . . . . .
Truncated posterior distribution of . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
91
94
98
104
106
108
121
122
124
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
Spatial x-y expansion estimates . . . . . . . . . . .

Spatial x-y total impact estimates . . . . . . . . .
Distance expansion estimates . . . . . . . . . . . .
Actual versus Predicted and residuals . . . . . . .
Comparison of GWR distance weighting schemes .
GWR and BGWR diffuse prior estimates . . . . .
GWR and Robust BGWR estimates . . . . . . . .
GWR and BGWR Distance-based weights adjusted
. .
. .
. .
. .
. .
. .
. .
Vi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
139
140
141
142
158
172
173
174
ix
. .
. .
. .
. .
. .
. .
. .
by
LIST OF FIGURES
4.9
4.10
4.11
4.12
Average Vi estimates over all draws and observations

Alternative smoothing BGWR estimates . . . . . . .
Posterior model probabilities . . . . . . . . . . . . .
Smoothed parameter estimates . . . . . . . . . . . .
5.1
5.2
Results of plt() function . . . . . . . . . . . . . . . . . . . . . 199

Actual vs. simulated censored y-values . . . . . . . . . . . . . 205
6.1
6.2
Prior means and precision for important variables . . . . . . . 241

Prior means and precision for unimportant variables . . . . . 242
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
175
177
180
181
Chapter 1
Introduction
This chapter provides an overview of the nature of spatial econometrics.
An applied model-based approach is taken where various spatial econometric methods are introduced in the context of spatial data sets and associated models based on the data. The remaining chapters of the text are
organized along the lines of alternative spatial econometric estimation procedures. Each chapter illustrates applications of a different econometric
estimation method and provides references to the literature regarding these
methods.
Section 1.1 sets forth the nature of spatial econometrics and discusses differences with traditional econometrics. We will see that spatial econometrics
is characterized by: 1) spatial dependence between sample data observations
at various points in the Cartesian plane, and 2) spatial heterogeneity that
arises from relationships or model parameters that vary with our sample
data as we move over the Cartesian plane.
The nature of spatially dependent or spatially correlated data is taken
up in Section 1.2 and spatial heterogeneity is discussed in Section 1.3. Section 1.4 takes up the subject of how we formally incorporate the locational
information from spatial data in econometric models. In addition to the
theoretical discussion of incorporating locational information in econometric models, Section 1.4 provides a preview of alternative spatial econometric
estimation methods that will be covered in Chapters 2 through 4.
Finally, Section 1.5 describes software design issues related to a spatial
econometric function library based on MATLAB software from the MathWorks Inc. Functions are described throughout the text that implement the
spatial econometric estimation methods discussed. These functions provide
a consistent user-interface in terms of documentation and related functions
CHAPTER 1. INTRODUCTION
that provided printed as well as graphical presentation of the estimation results. Section 1.5 introduces the spatial econometrics function library which
is part of a broader collection of econometric estimation functions available
in my public domain Econometrics Toolbox.
1.1
Spatial econometrics
Applied work in regional science relies heavily on sample data that is collected with reference to location measured as points in space. The subject
of how we incorporate the locational aspect of sample data is deferred until Section 1.4. What distinguishes spatial econometrics from traditional
econometrics? Two problems arise when sample data has a locational component: 1) spatial dependence exists between the observations and 2) spatial
heterogeneity occurs in the relationships we are modeling.
Traditional econometrics has largely ignored these two issues that violate the traditional Gauss-Markov assumptions used in regression modeling.
With regard to spatial dependence between the observations, recall that
Gauss-Markov assumes the explanatory variables are fixed in repeated sampling. Spatial dependence violates this assumption, a point that will be
made clear in the next section. This gives rise to the need for alternative
estimation approaches. Similarly, spatial heterogeneity violates the GaussMarkov assumption that a single linear relationship exists across the sample
data observations. If the relationship varies as we move across the spatial
data sample, alternative estimation procedures are needed to successfully
model this type of variation and draw appropriate inferences.
The subject of this text is alternative estimation approaches that can be
used when dealing with spatial data samples. A subject that is seldom discussed in traditional econometrics textbooks. For example, no discussion of
issues and models related to spatial data samples can be found in Amemiya
(1985), Chow (1983), Dhrymes (1978), Fomby et al. (1984), Green (1997),
Intrilligator (1978), Kelijian and Oates (1989), Kmenta (1986), Maddala
(1977), Pindyck and Rubinfeld (1981), Schmidt (1976), and Vinod and Ullah (1981).
Anselin (1988) provides a complete treatment of many facets of spatial econometrics which this text draws upon. In addition to introducing
ideas set forth in Anselin (1988), this presentation includes some more recent approaches based on Bayesian methods applied to spatial econometric
models. In terms of focus, the materials presented here are more applied
than Anselin (1988), providing program functions and illustrations of hands-
on approaches to implementing the estimation methods described. Another

departure from Anselin (1988) is in the use of sparse matrix algorithms available in the MATLAB software to implement spatial econometric estimation
procedures. These implementation details represent previously unpublished
material that describes a set of programs (freely available) for solving largescale spatial econometric problems involving thousands of observations in a
few minutes on a modest desktop computer. Students as well as researchers
can use these programs with absolutely no programming to implement some
of the latest estimation procedures on large-scale spatial data sets.
Of course another distinction of the presentation here is the interactive
aspect of a Web-based format that allows a hands-on approach that provides
links to code, sample data and examples.
1.2
Spatial dependence
Spatial dependence in a collection of sample data observations refers to the

fact that one observation associated with a location which we might label i
depends on other observations at locations j 6= i. Formally, we might state:
yi = f (yj ), i = 1, . . ., n j 6= i
(1.1)
Note that we allow the dependence to be among several observations,

as the index i can take on any value from i = 1, . . . , n. Why would we
expect sample data observed at one point in space to be dependent on values observed at other locations? There are two reasons commonly given.
First, data collection of observations associated with spatial units such as
zip-codes, counties, states, census tracts and so on, might reflect measurement error. This would occur if the administrative boundaries for collecting
information do not accurately reflect the nature of the underlying process
generating the sample data. As an example, consider the case of unemployment rates and labor force measures. Because laborers are mobile and can
cross county or state lines to find employment in neighboring areas, labor
force or unemployment rates measured on the basis of where people live
could exhibit spatial dependence.
Consider the following illustration that demonstrates how measurement
problems could produce a set of spatial unemployment rate observations that
appear to be related to location. Recall that the labor force is measured as
persons actively seeking employment and the unemployment rate represents
the percentage of those actively seeking employment who are not employed.
Suppose that labor force is measured by place of residence and employment

is measured by place of work.
Consider six regions with differing levels of labor force measured by place
of residence and employment measured by place of work shown in Figure 1.1.
Assume the true unemployment rate is 25% in all six regions, exhibiting a
uniform distribution across space. We also assume that regions (3) and (4)
are at the center with (2) and (5) being neighboring regions and (1) and (6)
farthest from the center. In other words, the regions are assumed to occupy
a map configuration identical to that shown in Figure 1.1.
(1)
(2)
(3)
(4)
(5)
(6)
1,000
2,000
3,000
3,000
2,000
1,000
Labor force
by residence
1,000
3,000
3,000
1,000
Employment by place
100%
50%
0%
0%
50%
100%
Measured rate of
Unemployment
25%
25%
25%
25%
25%
25%
Acutal rate of
Unemployment
Figure 1.1: Measurement error and spatial dependence

If all employment were in the four central regions (2,3,4,5) as shown in
Figure 1.1, the measured unemployment rates would take on the pattern
shown. One might infer from the measured unemployment rates that a distinctive spatial pattern of unemployment exists, and that distance from the
central regions (3) and (4) is an important determinant of unemployment.
Yet, as indicated, the true pattern of unemployment exhibits a uniform dis-
tribution across the spatial dimension of the data sample.

A second and perhaps more important reason we would expect spatial
dependence is that the spatial dimension of economic activity may truly be
an important aspect of a modeling problem. Regional science is based on the
premise that location and distance are important forces at work in human
geography and market activity. All of these notions have been formalized
in regional science theory that relies on notions of spatial interaction and
diffusion effects, hierarchies of place and spatial spillovers.
As a concrete example of this type of spatial dependence, we took a
sample of 35,000 homes that sold within the last 5 years in Lucas county,
Ohio. Using the latitude-longitude coordinates, the distance of each house
was calculated from the center of the city of Toledo (which is in Lucas
county). The 35,000 homes were then sorted by distance from the center and
grouped into seven batches of 5,000 homes. Distributions were constructed
for a variable representing the year each house was built for all seven groups.
We would expect to see newer homes farther away from the city and older
homes near the center.
Figure 1.2 shows the seven house age distributions where it is clear that
the newest homes are in the group of 5,000 homes farthest from the center
of the city. There are some very old muncipalities surrounding the city of
Toledo that contain older historic homes and these appear in some of the
distributions. We also see a distinct pattern where all house construction
was adversely affected by the great depression in the 1930s and the war in
the 1940s.
The fact that these distributions are very different as we move over space
away from the center of the city, with similarity within the seven samples of
5,000 homes grouped by distance needs to be taken into account when we
attempt to model a spatial sample such as this. We also need to be aware of
the common influence on all of the distributions exerted by the depression
and the war.
Of course, this pattern of older homes in the central city and newer
homes in developments farther from the center has numerous regional science
explanations. Nonetheless, this fact leads to the distinctive spatial pattern
that would be indicative of spatial dependence in this variable vector.
Spatial dependence arising from underlying regional interactions often
observed in regional science data samples suggests the need to quantify and
model the nature of the unspecified functional dependence f , set forth in
(1.1). Before turning attention to this task, the next section discusses the
other underlying condition leading to a need for spatial econometrics
spatial heterogeneity.
0.035
mean distance = 0.021

0.03
0.025
0.02
0.015
0.01
0.005
-0.005
1860
1880
1900
1920
1940
1960
1980
2000
Figure 1.2: Distributions of house age versus distance
1.3
Spatial heterogeneity
The term spatial heterogeneity refers to variation in relationships over space.

In the most general case consider that we might expect a different relationship to hold for every point in space. Formally, we write a linear relationship
depicting this as:
yi = Xii + i
(1.2)
Where i indexes observations collected at i = 1, . . . , n points in space, Xi

represents a (1 x k) vector of explanatory variables with an associated set
of parameters i , yi is the dependent variable at observation (or location) i
and i denotes a stochastic disturbance in the linear relationship.
Given a sample of n data observations, we could not hope to estimate
a set of n parameter vectors i because there are degrees of freedom problems. We simply do not have enough sample data information with which
to produce estimates for every point in space. To proceed with analysis of
the relationship, we must provide a specification for variation over space.

This specification must be parsimonious, that is, only a handful of parameters can be used in the specification. A large amount of spatial econometric
research centers on alternative parsimonious specifications for modeling variation over space. Questions arise regarding: 1) how sensitive the inferences
are to a particular specification regarding spatial variation?, 2) is the specification consistent with the sample data information?, 3) how do competing
specifications perform and what inferences do they provide?, and a host of
other issues that will be explored in this text.
One can also view the specification task as one of placing restrictions on
the nature of variation in the relationship over space. For example, suppose
we classified our spatial observations into urban and rural regions. We could
then restrict our analysis to two relationships, one homogeneous across all
urban observational units and another for the rural units. This raises a
number of questions: 1) are two relations consistent with the data, or is
there evidence to suggest more than two?, 2) is there a trade-off between
efficiency in the estimates and the number of restrictions we use?, 3) are
the estimates biased if the restrictions are inconsistent with the sample data
information?, and other issues we will explore.
1.4
Quantifying location in our models
A first task we must undertake before we can ask questions about spatial
dependence and heterogeneity is quantification of the locational aspects of
our sample data. Given that we can always map a set of spatial data observations, we have two sources of information on which we can draw.
The location in Cartesian space represented by latitude and longitude is
one source of information. This information would also allow us to calculate
distances from any point in space, or the distance of observations located at
distinct points in space to observations at other locations. Spatial dependence should conform to the fundamental theorem of regional science, distance matters. Observations that are near each other should reflect a greater
degree of spatial dependence than those more distant from each other. In
other words, the strength of spatial dependence between observations should
decline with the distance between observations.
The second source of locational information is contiguity, reflecting the
relative position in space of one regional unit of observation to other such
units. Measures of contiguity rely on a knowledge of the size and shape of the
observational units depicted on a map. From this, we can determine which
units are neighbors (have borders that touch) or represent observational

units in reasonable proximity to each other. Regarding spatial dependence,
neighboring units should exhibit a higher degree of spatial dependence than
units located far apart.
I note in passing that these two types of information are not necessarily different. Given the latitude-longitude coordinates of an observation,
we could construct a contiguity structure by defining a neighboring observation as one that lies within a certain distance. Consider also, given the
centroid coordinates of a set of observations associated with contiguous map
regions, we can calculate distances between the regions (or observations).
We will illustrate how both types of locational information can be used
in spatial econometric modeling. We first take up the issue of quantifying
spatial contiguity, which is used in the models presented in Chapter 2.
1.4.1
Quantifying spatial contiguity
Figure 1.2 shows a hypothetical example of five regions as they would appear
on a map. We wish to construct a 5 by 5 binary matrix W containing 25
elements taking values of 0 or 1 that captures the notion of connectiveness
between the five entities depicted in the map configuration. We record in
each row of the matrix W a set of contiguity relations associated with one
of the five regions. For example the matrix element in row 1, column 2
would record the presence (represented by a 1) or absence (denoted by 0) of
a contiguity relationship between regions 1 and 2. As another example, the
row 3, column 4 element would reflect the presence or absence of contiguity
between regions 3 and 4. Of course, a matrix constructed in such fashion
must be symmetric if regions 3 and 4 are contiguous, so are regions 4 and
3.
It turns out there are an embarrassingly large number of ways to accomplish our task. Below, we enumerate some of the alternative ways we
might define a binary matrix W that represent alternative definitions of the
contiguity relationships between the five entities in Figure 1.3. For the
enumeration below, start with a matrix filled with zeros, then consider the
following alternative ways to define the presence of a contiguity relationship.
Linear contiguity: Define Wij = 1 for entities that share a common
edge to the immediate right or left of the region of interest. For row
1, where we record the relations associated with region 1, we would
have all W1j = 0, j = 1, . . ., 5. On the other hand, for row 5, where we
record relationships involving region 5, we would have W53 = 1 and
(4)
(3)
(5)
(2)
(1)
Figure 1.3: An illustration of contiguity

all other row-elements equal to zero.
Rook contiguity: Define Wij = 1 for regions that share a common side
with the region of interest. For row 1, reflecting region 1s relations
we would have W12 = 1 with all other row elements equal to zero. As
another example, row 3 would record W34 = 1, W35 = 1 and all other
row elements equal to zero.
Bishop contiguity: Define Wij = 1 for entities that share a common
vertex with the region of interest. For region 2 we would have W23 = 1
and all other row elements equal to zero.
10
Double linear contiguity: For two entities to the immediate right or left
of the region of interest, define Wij = 1. This definition would produce
the same results as linear contiguity for the regions in Figure 1.3.
Double rook contiguity: For two entities to the right, left, north and
south of the region of interest define Wij = 1. This would result in the
same matrix W as rook contiguity for the regions shown in Figure 1.3.
Queen contiguity: For entities that share a common side or vertex
with the region of interest define Wij = 1. For region 3 we would
have: W32 = 1, W34 = 1, W35 = 1 and all other row elements zero.
Believe it or not, there are even more ways one could proceed. For a
good discussion of these issues, see Appendix 1 of Kelejian and Robinson
(1995). Note also that the double linear and double rook definitions are
sometimes referred to as second order contiguity, whereas the other definitions are termed first order. More elaborate definitions sometimes rely
on the distance of shared borders. This might impact whether we considered regions (4) and (5) in Figure 1.3 as contiguous or not. They have a
common border, but it is very short. Note that in the case of a vertex, the
rook definition rules out a contiguity relation, whereas the bishop and queen
definitions would record a relationship.
The guiding principle in selecting a definition should be the nature of
the problem being modeled, and perhaps additional non-sample information
that is available. For example, suppose that a major highway connecting
regions (2) and (3) existed and we knew that region (2) was a bedroom
community for persons who work in region (3). Given this non-sample
information, we would not want to rely on the rook definition that would
rule out a contiguity relationship, as there is quite reasonably a large amount
of spatial interaction between these two regions.
We will use the rook definition to define a first-order contiguity matrix
for the five regions in Figure 1.3 as a concrete illustration. This is a definition
that is often used in applied work. Perhaps the motivation for this is that
we simply need to locate all regions on the map that have common borders
with some positive length.
The matrix W reflecting first-order rooks contiguity relations for the
five regions in Figure 1.3 is:
W =
0
1
0
0
0
11
1
0
0
0
0
0
0
0
1
1
0
0
1
0
1
0
0
1
1
0
(1.3)
Note that the matrix W is symmetric as indicated above, and by convention the matrix always has zeros on the main diagonal. A transformation
often used in applied work is to convert the matrix W to have row-sums of
unity. This is referred to as a standardized first-order contiguity matrix,
which we denote as C:
C =
0
1
0
0
0
1 0
0
0
0 0
0
0
0 0 1/2 1/2
0 1/2 0 1/2
0 1/2 1/2 0
(1.4)
The motivation for the standardization can be seen by considering what

happens if we use matrix multiplication of C and a vector of observations
on some variable associated with the five regions which we label y. This
matrix product y ? = Cy represents a new variable equal to the mean of
observations from contiguous regions:
y1?
y2?
y3?
y4?
y5?
y1?
y2?
y3?
y4?
y5?
0
1
0
0
0
1 0
0
0
y1
0 0
0
0 y2
0 0 0.5 0.5 y3
0 0.5 0 0.5 y4
0 0.5 0.5 0
y5
y2
y1
1/2y4 + 1/2y5
1/2y3 + 1/2y5
1/2y3 + 1/2y4
(1.5)
This is one way of quantifying the notion that yi = f (yj ), j 6= i, expressed in (1.1). Consider now a linear relationship that uses the variable
y ? we constructed in (1.5) as an explanatory variable in a linear regression
relationship to explain variation in y across the spatial sample of observations.
y = Cy +
12
(1.6)
Where represents a regression parameter to be estimated and denotes the

stochastic disturbance in the relationship. The parameter would reflect
the spatial dependence inherent in our sample data, measuring the average
influence of neighboring or contiguous observations on observations in the
vector y. If we posit spatial dependence between the individual observations
in the data sample y, some part of the total variation in y across the spatial
sample would be explained by each observations dependence on its neighbors. The parameter would reflect this in the typical sense of regression.
In addition, we could calculate the proportion of the total variation in y
that is explained by spatial dependence. This would be represented by Cy,
where represents the estimated value of .
We will examine spatial econometric models that rely on this type of
formulation in great detail in Chapter 2, where we set forth maximum likelihood estimation procedures for a taxonomy of these models known as spatial
autoregressive models.
One point to note is that traditional explanatory variables of the type
encountered in regression can be added to the model in (1.6). We can
represent these with the traditional matrix notation: X, allowing us to
modify (1.6) to take the form shown in (1.7).
y = Cy + X +
(1.7)
As an illustration, consider the following example which is intended to

serve as a preview of material covered in the next two chapters. We provide
a set of regression estimates based on maximum likelihood procedures for a
spatial data set consisting of 49 neighborhoods in Columbus, Ohio set forth
in Anselin (1988). The data set consists of observations on three variables:
neighborhood crime incidents, household income, and house values for all
49 neighborhoods. The model uses the income and house values to explain
variation in neighborhood crime incidents. That is, y = neighborhood crime,
X =(a constant, household income, house values). The estimates are shown
below, printed in the usual regression format with associated statistics for
precision of the estimates, fit of the model and an estimate of the disturbance
variance,
2 .
Spatial autoregressive Model Estimates
Dependent Variable =
Crime
R-squared
=
0.6518
Rbar-squared =
0.6366
13
sigma^2
=
95.5032
log-likelihood =
-165.41269
Nobs, Nvars
=
49,
3
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
45.056251
6.231261
0.000000
income
-1.030641
-3.373768
0.001534
house value
-0.265970
-3.004945
0.004331
rho
0.431381
3.625340
0.000732
For this example, we can calculate the proportion of total variation ex2
plained by spatial dependence with a comparison of the fit measured by R
from this model to the fit of a least-squares model that excludes the spatial dependence variable Cy. The least-squares regression for comparison is
shown below:
Ordinary Least-squares Estimates
Crime
R-squared
=
0.5521
Rbar-squared =
0.5327
sigma^2
= 130.8386
Durbin-Watson =
1.1934
Nobs, Nvars
=
49,
3
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
68.609759
14.484270
0.000000
income
-1.596072
-4.776038
0.000019
house value
-0.274079
-2.655006
0.010858
We see that around 10 percent of the variation in the crime incidents

2 is roughly 0.63 in the
is explained by spatial dependence, because the R
model that takes spatial dependence into account and 0.53 in the leastsquares model that ignores this aspect of the spatial data sample. Note also
that the tstatistic on the parameter for the spatial dependence variable Cy
is 3.62, indicating that this explanatory variable has a coefficient estimate
that is significantly different from zero. We will pursue more examples in
Chapters 2 and 3, with this example provided as a concrete demonstration
of some of the ideas we have discussed.
1.4.2
Quantifying spatial position
Associating location in space with observations is essential to modeling relationships that exhibit spatial heterogeneity. Recall this means there is
variation in the relationship being modeled over space. We illustrate two
14
approaches to using location that allow locally linear regressions to be fit

over sub-regions of space. These form the basis for models we will discuss
in Chapter 4.
Casetti (1972, 1992) introduced our first approach that involves a method
he labels spatial expansion. The model is shown in (1.8), where y denotes
an nx1 dependent variable vector associated with spatial observations and
X is an nxnk matrix consisting of terms xi representing kx1 explanatory
variable vectors, as shown in (1.9). The locational information is recorded
in the matrix Z which has elements Zxi , Zyi , i = 1, . . ., n, that represent
latitude and longitude coordinates of each observation as shown in (1.9).
The model posits that the parameters vary as a function of the latitude
and longitude coordinates. The only parameters that need be estimated
are the parameters in 0 that we denote, x , y . These represent a set of
2k parameters. Recall our discussion about spatial heterogeneity and the
need to utilize a parsimonious specification for variation over space. This
represents one approach to this type of specification.
We note that the parameter vector in (1.8) represents an nkx1 matrix in this model that contains parameter estimates for all k explanatory
variables at every observation. The parameter vector 0 contains the 2k
parameters to be estimated.
y = X +
= ZJ0
(1.8)
Where:
y =
y1
y2
..
.
X=
yn
x01 0 . . . 0
0 x02
..
..
.
.
0
x0n
1
2
..
.
n
1
2
..
.
n
Ik 0
Zx1 Ik Zy1 Ik
0
...
0 Ik
..
..
.
.
0
Z =
J = ..
.
..
.
Zxn Ik Zyn Ik
0 Ik
0 =
x
y
(1.9)
15
This model can be estimated using least-squares to produce estimates of

the 2k parameters x , y . Given these estimates, the remaining estimates
for individual points in space can be derived using the second equation in
(1.8). This process is referred to as the expansion process. To see this,
substitute the second equation in (1.8) into the first, producing:
y = XZJ0 +
(1.10)
Here it is clear that X, Z and J represent available information or data

observations and only 0 represent parameters in the model that need be
estimated.
The model would capture spatial heterogeneity by allowing variation
in the underlying relationship such that clusters of nearby or neighboring
observations measured by latitude-longitude coordinates take on similar parameter values. As the location varies, the regression relationship changes
to accommodate a locally linear fit through clusters of observations in close
proximity to one another.
To provide a preview of this model which is discussed in Chapter 4, Figure 1.4 presents the spatially expanded parameter estimates for the Columbus neighborhood crime data set. The observations have been ordered by
distance from the central city, so as we move to the far right along the
x-axis, we are examining parameter values associated with more suburban
locations. The estimates are based on a slightly altered version of the spatial expansion model that relies on distance from a central city neighborhood
rather than the latitude-longitude location of observations defined in (1.9).
(We will discuss this variant on the spatial expansion model more fully in
Chapter 4.)
We see from the figure that neighborhood crime is negatively influenced
by both household income and housing values. As we move away from the
central city neighborhoods, this negative influence increases.
Another approach to modeling variation over space is based on locally
weighted regressions to produce estimates for every point in space by using a
sub-sample of data information from nearby observations. McMillen (1996)
and Brundson, Fotheringham and Charlton (1996) introduce this type of
approach. It has been labeled geographically weighted regression, (GWR)
by Brundson, Fotheringham and Charlton (1996). Let y denote an nx1
vector of dependent variable observations collected at n points in space, X
an nxk matrix of explanatory variables, and an nx1 vector of normally
distributed, constant variance disturbances. Letting Wi represent an nxn
diagonal matrix containing distance-based weights for observation i that
16
Household income
0
-0.2
-0.4
-0.6
-0.8
-1
10
15
20
25
30
Distance from central city
35
40
45
50
10
15
20
25
30
35
40
45
50
House value
-0.05
-0.1
-0.15
-0.2
-0.25
Figure 1.4: Distance expansion estimates

reflects the distance between observation i and all other observations, we
can write the GWR model as:
Wi y = Wi Xi + i
(1.11)
The subscript i on i indicates that this kx1 parameter vector is associated with observation i. The GWR model produces n such vectors of
parameter estimates, one for each observation. These estimates are produced using:
i = (X 0 Wi2 X)1 (X 0Wi2 y)
(1.12)
One confusing aspect of this notation is that Wi y denotes an n-vector

of distance-weighted observations used to produce estimates for observation
i. The notation is confusing because we usually rely on subscripts to index scalar magnitudes representing individual elements of a vector. Note
17
also, that Wi X represents a distance-weighted data matrix, not a single observation and i represents a n-vector. The precise nature of the distance
weighting is taken up in Chapter 4.
As an applied illustration, we provide a graphical presentation in Figure 1.5 of the estimates produced by the GWR method sorted by distance
from the central city, making these comparable to those already presented
for the Casetti distance expansion model. The two sets of estimates are
quite different, raising the question of which approach provides a better
set of inferences regarding the relationship between the neighborhood crime
rates and the explanatory variables in the model. This topic will be taken
up when we discuss alternative modeling approaches to deal with spatial
dependence and heterogeneity.
2
Household income
1
0
-1
-2
-3
-4
10
15
20
25
30
35
40
45
50
10
15
20
25
30
35
40
45
50
House value
0.5
0
-0.5
-1
-1.5
Figure 1.5: GWR estimates

It may have occurred to the reader that a homogeneous model fit to a
spatial data sample that exhibits heterogeneity will produce residuals that
exhibit spatial dependence. The residuals or errors made by the homo-
18
geneous model should reflect unexplained variation attributable to heterogeneity in the underlying relationship over space. Spatial clustering of the
residuals would occur with positive and negative residuals appearing in distinct regions and patterns on the map. This of course was our motivation
and illustration of spatial dependence as illustrated in Figure 1.2. You might
infer correctly that spatial heterogeneity and dependence are often related
in the context of modeling. An inappropriate model that fails to capture
spatial heterogeneity will result in residuals that exhibit spatial dependence.
This is another topic we discuss in the following chapters of this text.
1.5
The MATLAB spatial econometrics library
As indicated in the preface, all of the spatial econometric methods discussed

in this text have been implemented using the MATLAB software from MathWorks Inc. Toolboxes are the name given by the MathWorks to related sets
of MATLAB functions aimed at solving a particular class of problems. Toolboxes of functions useful in signal processing, optimization, statistics, finance
and a host of other areas are available from the MathWorks as add-ons to
the standard MATLAB distribution. We will reserve the term Econometrics
Toolbox to refer to my larger collection of econometric functions available
in the public domain at www.econ.utoledo.edu. The spatial econometrics library represents a smaller part of this larger collection of software functions
for econometric analysis. I have used the term library to denote subsets of
functions aimed at various categories of estimation methods. The Econometrics Toolbox contains libraries for econometric regression analysis, timeseries and vector autoregressive modeling, optimization functions to solve
general maximum likelihood estimation problems, Bayesian Gibbs sampling
diagnostics, error correction testing and estimation methods, simultaneous
equation models and a collection of utility functions that I designate as the
utility function library. Taken together, these constitute the Econometrics
Toolbox that is described in a 350 page manual available at the Web site
listed above.
The spatial econometrics library functions rely on some of the utility
functions and are implemented using a general design that provides a common user-interface for the entire toolbox of econometric estimation functions. In Chapter 2 we will use MATLAB functions to carry out spatial
econometric estimation methods. Here, we discuss the general design that
is used to implement all of the spatial econometric estimation functions.
Having some feel for the way in which these functions work and communi-
19
cate with other functions in the Econometric Toolbox should allow you to
more effectively use these functions to solve spatial econometric estimation
problems.
The entire Econometrics Toolbox has been included in the internet-based
materials provided here, as well as an online HTML interface to examine
the functions available along with their documentation. All functions have
accompanying demonstration files that illustrate the typical use of the functions with sample data. These demonstration files can be viewed using the
online HTML interface. We have also provided demonstration files for all
of the estimation functions in the spatial econometrics library that can be
viewed online along with their documentation. Examples are provided in
this text and the program files along with the datasets have been included
in the Web-based module.
In designing a spatial econometric library of functions, we need to think
about organizing our functions to present a consistent user-interface that
packages all of our MATLAB functions in a unified way. The advent of
structures in MATLAB version 5 allows us to create a host of alternative
spatial econometric functions that all return results structures.
A structure in MATLAB allows the programmer to create a variable
containing what MATLAB calls fields that can be accessed by referencing
the structure name plus a period and the field name. For example, suppose
we have a MATLAB function to perform ordinary least-squares estimation
named ols that returns a structure. The user can call the function with
input arguments (a dependent variable vector y and explanatory variables
matrix x) and provide a variable name for the structure that the ols function
will return using:
result = ols(y,x);
The structure variable result returned by our ols function might have
fields named rsqr, tstat, beta, etc. These fields might contain the R One virtue
squared statistic, tstatistics and the least-squares estimates .
of using the structure to return regression results is that the user can access
individual fields of interest as follows:
bhat = result.beta;
disp(The R-squared is:);
result.rsqr
disp(The 2nd t-statistic is:);
result.tstat(2,1)
There is nothing sacred about the name result used for the returned
structure in the above example, we could have used:
bill_clinton
result2
restricted
unrestricted
=
=
=
=
20
ols(y,x);
ols(y,x);
ols(y,x);
ols(y,x);
That is, the name of the structure to which the ols function returns its
information is assigned by the user when calling the function.
To examine the nature of the structure in the variable result, we can
simply type the structure name without a semi-colon and MATLAB will
present information about the structure variable as follows:
result =
meth:
y:
nobs:
nvar:
beta:
yhat:
resid:
sige:
tstat:
rsqr:
rbar:
dw:
ols
[100x1
100.00
3.00
[ 3x1
[100x1
[100x1
1.01
[ 3x1
0.74
0.73
1.89
double]
double]
double]
double]
double]
Each field of the structure is indicated, and for scalar components the
value of the field is displayed. In the example above, nobs, nvar, sige,
rsqr, rbar, and dw are scalar fields, so there values are displayed. Matrix
or vector fields are not displayed, but the size and type of the matrix or
vector field is indicated. Scalar string arguments are displayed as illustrated
by the meth field which contains the string ols indicating the regression
method that was used to produce the structure. The contents of vector or
matrix strings would not be displayed, just their size and type. Matrix and
vector fields of the structure can be displayed or accessed using the MATLAB
conventions of typing the matrix or vector name without a semi-colon. For
example,
result.resid
result.y
would display the residual vector and the dependent variable vector y in the
MATLAB command window.
Another virtue of using structures to return results from our regression
functions is that we can pass these structures to another related function
21
that would print or plot the regression results. These related functions can
query the structure they receive and intelligently decipher the meth field
to determine what type of regression results are being printed or plotted.
For example, we could have a function prt that prints regression results and
another plt that plots actual versus fitted and/or residuals. Both these functions take a regression structure as input arguments. Example 1.1 provides
a concrete illustration of these ideas.
% ----- Example 1.1 Demonstrate regression using the ols() function
load y.data;
load x.data;
result = ols(y,x);
prt(result);
plt(result);
The example assumes the existence of functions ols, prt, plt and data
matrices y, x in files y.data and x.data. Given these, we carry out a regression, print results and plot the actual versus predicted as well as residuals
with the MATLAB code shown in example 1.1. We will discuss the prt and
plt functions in Section 1.5.2.
1.5.1
Estimation functions
Now to put these ideas into practice, consider implementing an ols function.
The function code would be stored in a file ols.m whose first line is:
function results=ols(y,x)
The keyword function instructs MATLAB that the code in the file ols.m
represents a callable MATLAB function.
The help portion of the MATLAB ols function is presented below and
follows immediately after the first line as shown. All lines containing the
MATLAB comment symbol % will be displayed in the MATLAB command
window when the user types help ols.
function results=ols(y,x)
% PURPOSE: least-squares regression
%--------------------------------------------------% USAGE: results = ols(y,x)
% where: y = dependent variable vector (nobs x 1)
%
x = independent variables matrix (nobs x nvar)
%--------------------------------------------------% RETURNS: a structure
22
%
results.meth = ols
%
results.beta = bhat
%
results.tstat = t-stats
%
results.yhat = yhat
%
results.resid = residuals
%
results.sige = e*e/(n-k)
%
results.rsqr = rsquared
%
results.rbar = rbar-squared
%
results.dw
= Durbin-Watson Statistic
%
results.nobs = nobs
%
results.nvar = nvars
%
results.y
= y data vector
% -------------------------------------------------% SEE ALSO: prt(results), plt(results)
%---------------------------------------------------
All functions in the spatial econometrics library present a unified documentation format for the MATLAB help command by adhering to the
convention of sections entitled, PURPOSE, USAGE, RETURNS, SEE
ALSO, and perhaps a REFERENCES section, delineated by dashed lines.
The USAGE section describes how the function is used, with each input
argument enumerated along with any default values. A RETURNS section
portrays the structure that is returned by the function and each of its fields.
To keep the help information uncluttered, we assume some knowledge on
the part of the user. For example, we assume the user realizes that the
.residuals field would be an (nobs x 1) vector and the .beta field would
consist of an (nvar x 1) vector.
The SEE ALSO section points the user to related routines that may
be useful. In the case of our ols function, the user might what to rely on
the printing or plotting routines prt and plt, so these are indicated. The
REFERENCES section would be used to provide a literature reference (for
the case of our more exotic spatial estimation procedures) where the user
could read about the details of the estimation methodology.
As an illustration of the consistency in documentation, consider the function sar that provides estimates for the spatial autoregressive model that
we presented in Section 1.4.1. The documentation for this function is shown
below:
PURPOSE: computes spatial autoregressive model estimates
y = p*W*y + X*b + e, using sparse matrix algorithms
--------------------------------------------------USAGE: results = sar(y,x,W,rmin,rmax,convg,maxit)
where: y = dependent variable vector
x = explanatory variables matrix
23
W = standardized contiguity matrix

rmin = (optional) minimum value of rho to use in search
rmax = (optional) maximum value of rho to use in search
convg = (optional) convergence criterion (default = 1e-8)
maxit = (optional) maximum # of iterations (default = 500)
--------------------------------------------------RETURNS: a structure
results.meth = sar
results.beta = bhat
results.rho = rho
results.tstat = asymp t-stat (last entry is rho)
results.yhat = yhat
results.sige = sige = (y-p*W*y-x*b)*(y-p*W*y-x*b)/n
results.lik = -log likelihood
results.nobs = # of observations
results.nvar = # of explanatory variables in x
results.y
= y data vector
results.iter
= # of iterations taken
results.romax = 1/max eigenvalue of W (or rmax if input)
results.romin = 1/min eigenvalue of W (or rmin if input)
-------------------------------------------------SEE ALSO: prt(results), sac, sem, far
--------------------------------------------------REFERENCES: Anselin (1988), pages 180-182.
---------------------------------------------------
The actual execution code to produce least-squares or spatial autoregressive parameter estimates would follow the documentation in the file
discussed above. We do not discuss programming of the spatial econometric
functions in the text, but you can of course examine all of the functions
to see how they work. The manual for the Econometrics Toolbox provides
a great deal of discussion of programming in MATLAB and examples of
how to add new functions to the toolbox or change existing functions in the
toolbox.
1.5.2
Using the results structure
To illustrate the use of the results structure returned by our ols function,
consider the associated function plt reg which plots actual versus predicted
values along with the residuals. The results structure contains everything
needed by the plt reg function to carry out its task. Earlier, we referred to
functions plt and prt rather than plt reg, but prt and plt are wrapper
24
functions that call the functions prt reg and plt reg where the real work
of printing and plotting regression results is carried out. The motivation
for taking this approach is that separate smaller functions can be devised
to print and plot results from all of the spatial econometric procedures,
facilitating development. The wrapper functions eliminate the need for the
user to learn the names of different printing and plotting functions associated
with each group of spatial econometric procedures all results structures
can be printed and plotted by simply invoking the prt and plt functions.
Documentation for the plt function which plots results from all spatial
econometrics functions as well as the Econometrics Toolbox is shown below.
This function is a the wrapper function that calls an appropriate plotting
function, plt spat based on the econometric method identified in the results
structure meth field argument.
PURPOSE: Plots results structures returned by most functions
by calling the appropriate plotting function
--------------------------------------------------USAGE: plt(results,vnames)
Where: results = a structure returned by an econometric function
vnames = an optional vector of variable names
e.g. vnames = vnames = strvcat(y,const,x1,x2);
-------------------------------------------------NOTES: this is simply a wrapper function that calls another function
-------------------------------------------------RETURNS: nothing, just plots the results
-------------------------------------------------SEE ALSO: prt()
---------------------------------------------------
A decision was made not to place the pause command in the plt function, but rather let the user place this statement in the calling program or
function. An implication of this is that the user controls viewing regression
plots in for loops or in the case of multiple invocations of the plt function.
For example, only the second plot will be shown in the following code.
result1 = sar(y,x1,W);
plt(result1);
plt(result2);
If the user wishes to see the plots associated with the first spatial autoregression, the code would need to be modified as follows:
25
plt(result1);
pause;
plt(result2);
The pause statement would force a plot of the results from the first spatial autoregression and wait for the user to strike any key before proceeding
with the second regression and accompanying plot of these results.
A more detailed example of using the results structure is the prt function which produces printed output from all of the functions in the spatial
econometrics library. The printout of estimation results is similar to that
provided by most statistical packages.
The prt function allows the user an option of providing a vector of fixed
width variable name strings that will be used when printing the regression
coefficients. These can be created using the MATLAB strvcat function
that produces a vertical concatenated list of strings with fixed width equal
to the longest string in the list. We can also print results to an indicated file
rather than the MATLAB command window. Three alternative invocations
of the prt function illustrating these options for usage are shown below:
vnames = strvcat(crime,const,income,house value);
res = sar(y,x,W);
prt(res);
% print with generic variable names
prt(res,vnames);
% print with user-supplied variable names
fid = fopen(sar.out,wr); % open a file for printing
prt(res,vnames,fid);
% print results to file sar.out
The first use of prt produces a printout of results to the MATLAB

command window that uses generic variable names:
R-squared
=
0.6518
Rbar-squared
=
0.6366
sigma^2
=
95.5033
Nobs, Nvars
=
49,
3
log-likelihood =
-165.41269
# of iterations =
17
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
variable 1
45.056480
6.231281
0.000000
variable 2
-1.030647
-3.373784
0.001513
variable 3
-0.265970
-3.004944
0.004290
rho
0.431377
3.625292
0.000720
26
The second use of prt uses the user-supplied variable names. The MATLAB function strvcat carries out a vertical concatenation of strings and
pads the shorter strings in the vnames vector to have a fixed width based
on the longer strings. A fixed width string containing the variable names is
required by the prt function. Note that we could have used:
vnames = [crime
,
const
,
income
,
house value];
but, this takes up more space and is slightly less convenient as we have to
provide the padding of strings ourselves. Using the vnames input in the prt
function would result in the following printed to the MATLAB command
window.
crime
R-squared
=
0.6518
Rbar-squared
=
0.6366
sigma^2
=
95.5033
Nobs, Nvars
=
49,
3
log-likelihood =
-165.41269
# of iterations =
12
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
45.056481
6.231281
0.000000
income
-1.030647
-3.373784
0.001513
house value
-0.265970
-3.004944
0.004290
rho
0.431377
3.625292
0.000720
The third case specifies an output file opened with the command:
fid = fopen(sar.out,wr);
The file sar.out would contain output identical to that from the second use
of prt. It is the users responsibility to close the file that was opened using
the MATLAB command:
fclose(fid);
In the following chapters that present various spatial estimation methods we will provide the documentation but not the details concerning implementation of the estimation procedures in MATLAB. A function has been
27
devised and incorporated in the spatial econometrics library for each of the
estimation procedures that we discuss and illustrate. These functions carry
out estimation and provide printed as well as graphical presentation of the
results using the design framework that was set forth in this section.
1.6
Chapter Summary
This chapter introduced two main features of spatial data sets, spatial dependence and spatial heterogeneity. Spatial dependence refers to the fact
that sample data observations exhibit correlation with reference to points or
location in space. We often observe spatial clustering of sample data observations with respect to map regions. An intuitive motivation for this type of
result is the existence of spatial hierarchical relationships, spatial spillovers
and other types of spatial interactivity studied in regional science.
Spatial heterogeneity refers to the fact that underlying relationships we
wish to study may vary systematically over space. This creates problems
for regression and other econometric methods that do not accommodate
spatial variation in the relationships being modeled. A host of methods
have arisen in spatial econometrics that allow the estimated relationship to
vary systematically over space.
A large part of the chapter was devoted to introducing how locational
information regarding sample data observations is formally incorporated in
spatial econometric models. After introducing the concept of a spatial contiguity matrix, we provided a preview of the spatial autoregressive model
that relies on the contiguity concept. Chapters 2 and 3 cover this spatial
econometric method in detail.
In addition to spatial contiguity, other spatial econometric methods rely
on the latitude-longitude information available for spatial data samples to
allow variation in the relationship being studied over space. Two approaches
to this were introduced, the spatial expansion model and geographically
weighted regression, which are the subject of Chapter 4.
A preview that provided an applied illustration of the methods that will
be detailed in Chapters 2 through 4 was provided using a spatial data set
from Anselin (1988) on crime incidents in 49 neighborhoods in Columbus,
Ohio.
Finally, a software design for implementing the spatial econometric estimation methods discussed in this text was set forth. Our estimation methods
will be implemented using MATLAB software from the MathWorks Inc. A
design based on MATLAB structure variables was set forth. This approach
28
to developing a set of spatial econometric estimation functions can provide

a consistent user-interface for the function documentation and help information as well as encapsulation of the estimation results in a MATLAB
structure variable. This construct can be accessed by related functions to
provide printed and graphical presentation of the estimation results.
1.7
References
Amemiya, T. 1985. Advanced Econometrics, (Cambridge, MA: Harvard University Press).
Anselin, L. 1988. Spatial Econometrics: Methods and Models, (Dorddrecht: Kluwer Academic Publishers).
Brundson, C. A. S. Fotheringham, and M. Charlton. 1996. Geographically weighted regression: a method for exploring spatial nonstationarity, Geographical Analysis, Vol. 28, pp. 281-298.
Casetti, E., 1972. Generating Models by the Expansion Method:
Applications to Geographic Research, Geographical Analysis, Vol. 4,
pp. 81-91.
Casetti, E. 1992. Bayesian Regression and the Expansion Method,
Geographical Analysis, Vol. 24, pp. 58-74.
Chow, G. 1983. Econometrics, (New York: McGraw-Hill);
Dhrymes, P. 1981. Distributed Lags: Problems of Estimation and
Formulation, (Amsterdam: North-Holland).
Fomby, T., R. Hill, and S. Johnson. 1984. Advanced Econometric
Methods, (New York: Springer).
Intrilligator, M. 1978. Econometric Models, Techniques, and Applications, (Englewood Cliffs: Prentice-Hall).
Green, W. H. 1997. Econometric Analysis, third edition, (Upper Saddle River, N.J: Prentice Hall).
Kelejian, H. and W. Oates. 1989. Introduction to Econometrics: Principles and Applications, (New York: Harper and Row).
29
Kelejian, H. H. and D. P. Robinson. 1995. Spatial Correlation:

A suggested alternative to the autoregressive model, in New Directions in Spatial Econometrics, L. Anselin and R.J.G.M. Florax (eds.).
(Berlin: Springer).
Kmenta, J. 1971. Elements of Econometrics, (New York: Macmillan).
Maddala, G.S. 1977. Econometrics, (New York: McGraw-Hill).
McMillen, D.P. 1996. One hundred fifty years of land values in
Chicago: a nonparameteric approach, Journal of Urban Economics,
Vol. 40, pp. 100-124.
Pindyck, R. and D. Rubinfeld. 1981. Econometric Models and Economic Forecasts, (New York: McGraw-Hill).
Schmidt, P. 1976. Econometrics, (New York: Marcel Dekker).
Vinod, H. and A. Ullah. 1981. Recent Advances in Regression Methods, (New York: Marcel Dekker).
Chapter 2
Spatial autoregressive models

This chapter discusses in detail the spatial autoregressive models introduced
in Chapter 1. A class of spatial autoregressive models have been introduced
to model cross-sectional spatial data samples taking the form shown in (2.1)
(Anselin, 1988).
y = W1 y + X + u
(2.1)
u = W2 u +
N (0, 2In )
Where y contains an nx1 vector of cross-sectional dependent variables and
X represents an nxk matrix of explanatory variables. W1 and W2 are known
nxn spatial weight matrices, usually containing first-order contiguity relations or functions of distance. As explained in Section 1.4.1, a first-order
contiguity matrix has zeros on the main diagonal, rows that contain zeros
in positions associated with non-contiguous observational units and ones in
positions reflecting neighboring units that are (first-order) contiguous based
on one of the contiguity definitions.
From the general model in (2.1) we can derive special models by imposing
restrictions. For example, setting X = 0 and W2 = 0 produces a first-order
spatial autoregressive model shown in (2.2).
y = W1 y +
(2.2)
N (0, 2In )
This model attempts to explain variation in y as a linear combination of
contiguous or neighboring units with no other explanatory variables. The
30
CHAPTER 2. SPATIAL AUTOREGRESSIVE MODELS
31
model is termed a first order spatial autoregression because its represents a

spatial analogy to the first order autoregressive model from time series analysis, yt = yt1 + t , where total reliance is on the past period observations
to explain variation in yt .
Setting W2 = 0 produces a mixed regressive-spatial autoregressive model
shown in (2.3). This model is analogous to the lagged dependent variable
model in time series. Here we have additional explanatory variables in the
matrix X that serve to explain variation in y over the spatial sample of
observations.
y = W1 y + X +
(2.3)
N (0, In )
2
Letting W1 = 0 results in a regression model with spatial autocorrelation

in the disturbances as shown in (2.4).
y = X + u
(2.4)
u = W2 u +
N (0, 2In )
This chapter is organized into sections that discuss and illustrate each
of these special cases of the spatial autoregressive model as well as the most
general model form in (2.1). Section 2.1 deals with the first-order spatial autoregressive model presented in (2.2). The mixed regressive-spatial autoregressive model is taken up in Section 2.2. Section 2.3 takes up the regression
model containing spatial autocorrelation in the disturbances and illustrates
various tests for spatial dependence using regression residuals. The most
general model is the focus of Section 2.4. Applied illustrations of all the
models are provided using a variety of spatial data sets. Spatial econometrics library functions that utilize MATLAB sparse matrix algorithms allow
us to estimate models with over 3,000 observations in around 100 seconds
on an inexpensive desktop computer.
2.1
The first-order spatial AR model
This model is seldom used in applied work, but it serves to motivate some
of the ideas that we draw on in later sections of the chapter. The model
which we label FAR, takes the form:
y = W y +
32
(2.5)
N (0, In )
2
Where the spatial contiguity matrix W has been standardized to have row
sums of unity and the variable vector y is expressed in deviations from the
means form to eliminate the constant term in the model.
To illustrate the problem with least-squares estimation of spatial autoregressive models, consider applying least-squares to the model in (2.5) which
would produce an estimate for the single parameter in the model:
= (y 0 W 0 W y)1 y 0 W 0 y
(2.6)
Can we show that this estimate is unbiased? If not, is it consistent?

Taking the same approach as in least-squares, we substitute the expression
for y from the model statement and attempt to show that E(
) = to prove
unbiasedness.
E(
) = (y 0 W 0 W y)1 y 0 W 0 (W y + )
= + (y 0 W 0 W y)1 y 0 W 0
(2.7)
Note that the least-squares estimate is biased, since we cannot show that
E(
) = . The usual argument that the explanatory variables matrix X in
least-squares is fixed in repeated sampling allows one to pass the expectation operator over terms like (y 0 W 0 W y)1 y 0 W 0 and argue that E() = 0,
eliminating the bias term. Here however, because of spatial dependence, we
cannot make the case that W y is fixed in repeated sampling. This also rules
out making a case for consistency of the least-squares estimate of , because
the probability limit (plim) for the term y 0 W 0 is not zero. In fact, Anselin
(1988) establishes that:
plimN 1 (y 0 W 0 ) = plimN 1 0 W (I W )1
(2.8)
This is equal to zero only in the trivial case where equals zero and we have
no spatial dependence in the data sample.
Given that least-squares will produce biased and inconsistent estimates
of the spatial autoregressive parameter in this model, how do we proceed
to estimate ? The maximum likelihood estimator for requires that we
find a value of that maximizes the likelihood function shown in (2.9).
L(y|, 2) =
1
2
2 (n/2)
|In W | exp{
33
1
(y W y)0 (y W y)} (2.9)
2 2
In order to simplify the maximization problem, we obtain a concentrated log likelihood function based on eliminating the parameter 2 for
the variance of the disturbances. This is accomplished by substituting
2 = (1/n)(y W y)0 (y W y) in the likelihood (2.9) and taking logs

which yields:
n
Ln(L) ln(y W y)0 (y W y) + ln|In W |
(2.10)
2
This expression can be maximized with respect to using a simplex univariate optimization routine. The estimate for the parameter 2 can be obtained
using the value of that maximizes the log-likelihood function (say, ) in:
2 = (1/n)(y W y)0 (y W y). In the next section, we discuss a sparse

matrix algorithm approach to evaluating this likelihood function that allows
us to solve problems involving thousands of observations quickly with small
amounts of computer memory.
Two implementation details arise with this approach to solving for maximum likelihood estimates. First, there is a constraint that we need to impose
on the parameter . This parameter can take on feasible values in the range
(Anselin and Florax, 1994):
1/min < < 1/max
where min represents the minimum eigenvalue of the standardized spatial
contiguity matrix W and max denotes the largest eigenvalue of this matrix.
This suggests that we need to constrain our optimization procedure search
over values of within this range.
The second implementation issue is that the numerical hessian matrix
that would result from a gradient-based optimization procedure and provide estimates of dispersion for the parameters is not available with simplex
optimization. We can overcome this problem in two ways. For problems
involving a small number of observations, we can use our knowledge of
the theoretical information matrix to produce estimates of dispersion. An
asymptotic variance matrix based on the Fisher information matrix shown
below for the parameters = (, 2) can be used to provide measures of
dispersion for the estimates of and 2 . (see Anselin, 1980 page 50):
[I()]1 = E[
2 L 1
]
0
(2.11)
34
This approach is computationally impossible when dealing with large

scale problems involving thousands of observations. In these cases we can
evaluate the numerical hessian matrix using the maximum likelihood estimates of and 2 as well as our sparse matrix function to compute the
likelihood. We will demonstrate results from using both of these approaches
in the next section.
2.1.1
The far() function
Building on the software design set forth in section 1.5 for our spatial
econometrics function library, we have implemented a function far to produce maximum likelihood estimates for the first-order spatial autoregressive
model. We rely on on the sparse matrix functionality of MATLAB so largescale problems can be solved using a minimum of time and computer RAM
memory. We demonstrate this function in action on a data set involving
3,107 U.S. counties.
Estimating the FAR model requires that we find eigenvalues for the large
n by n matrix W , as well as the determinant of the related n by n matrix
(In W ). In addition, matrix multiplications involving W and (In W )
are required to compute the information matrix used to produce estimates
of dispersion.
We constructed a function far that can produce estimates for the firstorder spatial autoregressive model in a case involving 3,107 observations in
95 seconds on a moderately fast inexpensive desktop computer. The MATLAB algorithms for dealing with sparse matrices make it ideally suited for
spatial modeling because spatial weight matrices are almost always sparse.
Another issue we need to address is computing measures of dispersion
for the estimates and 2 in large estimation problems. As already noted,
we cannot rely on the information matrix approach because this involves
matrix operations on very large matrices. An approach that we take to
produce measures of dispersion is to numerically evaluate the hessian matrix
using the maximum likelihood estimates of and 2 . The approach basically
produces a numerical approximation to the expression in (2.11). A key to
using this approach is the ability to evaluate the log likelihood function using
the sparse algorithms to handle large matrices.
It should be noted that Pace and Barry (1997) when confronted with the
task of providing measures of dispersion for spatial autoregressive estimates
based on sparse algorithms suggest using likelihood ratio tests to determine
the significance of the parameters. The approach taken here may suffer from
some numerical inaccuracy relative to measures of dispersion based on the
35
theoretical information matrix, but has the advantage that users will be
presented with traditional tstatistics on which they can base inferences.
We will have more to say about how our approach to solving large spatial autoregressive estimation problems using sparse matrix algorithms in
MATLAB compares to one proposed by Pace and Barry (1997), when we
apply the function far to a large data set in the next section.
Documentation for the function far is presented below. This function
was written to perform on both large and small problems. If the problem
is small (involving less than 500 observations), the function far computes
measures of dispersion using the theoretical information matrix. If more observations are involved, the function determines these measures by computing a numerical hessian matrix. (Users may need to decrease the number of
observations to less than 500 if they have computers without a large amount
of RAM memory.)
PURPOSE: computes 1st-order spatial autoregressive estimates
y = p*W*y + e, using sparse matrix algorithms
--------------------------------------------------USAGE: results = far(y,W,rmin,rmax,convg,maxit)
results.meth = far
results.rho = rho
results.tstat = asymptotic t-stat
results.yhat = yhat
results.sige = sige = (y-p*W*y)*(y-p*W*y)/n
results.nobs = nobs
results.nvar = nvar = 1
results.y
= y data vector
results.iter
--------------------------------------------------
One option we provide allows the user to supply minimum and maximum
values of rather than rely on the eigenvalues of W . This might be used if
36
we wished to constrain the estimation results to a range of say 0 < < 1.

Note also that this would save the time needed to compute the maximum
and minimum eigenvalues of the large W matrix.
2.1.2
Applied examples
Given our function far that implements maximum likelihood estimation of

small and large first-order spatial autoregressive models, we turn attention to
illustrating the use of the function with some spatial data sets. In addition
to the estimation functions, we have functions prt and plt that provide
printed and graphical presentation of the estimation results.
Example 2.1 provides an illustration of using these functions to estimate
a first-order spatial autoregressive model for neighborhood crime from the
Anselin (1988) spatial data sample. Note that we convert the variable vector
containing crime incidents to deviations from the means form.
% ----- Example 2.1 Using the far() function
load wmat.dat;
% standardized 1st-order contiguity matrix
load anselin.dat; % load Anselin (1988) Columbus neighborhood crime data
y = anselin(:,1);
ydev = y - mean(y);
W = wmat;
vnames = strvcat(crime,rho);
res = far(ydev,W); % do 1st-order spatial autoregression
prt(res,vnames);
% print the output
plt(res,vnames);
% plot actual vs predicted and residuals
This example produced the following printed output with the graphical
output presented in Figure 2.1. From the output we would infer that a
distinct spatial dependence among the crime incidents for the sample of 49
neighborhoods exists since the parameter estimate for has a tstatistic of
4.259. We would interpret this statistic in the typical regression fashion to
indicate that the estimated lies 4.2 standard deviations away from zero.
We also see that this model explains nearly 44% of the variation in crime
incidents in deviations from the means form.
First-order spatial autoregressive model Estimates
crime
R-squared
=
0.4390
sigma^2
= 153.8452
Nobs, Nvars
=
49,
1
log-likelihood =
-373.44669
# of iterations =
17
min and max rho =
-1.5362,
1.0000
37
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.669775
4.259172
0.000095
FAR Actual vs. Predicted

40
20
-20
-40
10
15
20
25
30
35
40
45
50
30
35
40
45
50
Residuals
40
20
0
-20
-40
-60
10
15
20
25
Figure 2.1: Spatial autoregressive fit and residuals

Another more challenging example involves a large sample of 3,107 observations representing counties in the continental U.S. from Pace and Barry
(1997). They examine presidential election results for this large sample of
observations covering the U.S. presidential election of 1980 between Carter
and Reagan. The variable we wish to explain using the first-order spatial autoregressive model is the proportion of total possible votes cast for
both candidates. Only persons 18 years and older are eligible to vote, so
the proportion is based on those voting for both candidates divided by the
population over 18 years of age.
Pace and Barry (1997) suggest an alternative approach to that implemented here in the function far. They propose overcoming the difficulty we
38
face in evaluating the determinant (I W ) by computing this determinant

once over a grid of values for the parameter ranging from 1/min to 1/max
prior to estimation. They suggest a grid based on 0.01 increments for over
the feasible range. Given these pre-determined values for the determinant
(I W ), they point out that one can quickly evaluate the log-likelihood
function for all values of in the grid and determine the optimal value of as
that which maximizes the likelihood function value over this grid. Note that
their proposed approach would involve evaluating the determinant around
200 times if the feasible range of was -1 to 1. In many cases the range
is even greater than this and would require even more evaluations of the
determinant. In contrast, our function far reports that only 17 iterations
requiring log likelihood function evaluations involving the determinant were
needed to solve for the estimates in the case of the Columbus neighborhood
crime data set.
In addition, consider that one might need to construct a finer grid around
the approximate maximum likelihood value of determined from the initial
grid search, whereas our use of the MATLAB simplex algorithm produces
an estimate that is accurate to a number of decimal digits.
After some discussion of the computational savings associated with the
use of sparse matrices, we illustrate the use of our function far and compare
it to the approach suggested by Pace and Barry.
A first point to note regarding sparsity is that large problems such as
this will inevitably involve a sparse spatial contiguity weighting matrix. This
becomes obvious when you consider the contiguity structure of our sample
of 3,107 U.S. counties. At most, individual counties exhibited only 8 firstorder (rook definition) contiguity relations, so the remaining 2,999 entries in
this row are zero. The average number of contiguity relationships is 4, so a
great many of the elements in the matrix W are zero, which is the definition
of a sparse matrix.
To understand how sparse matrix algorithms conserve on storage space
and computer memory, consider that we need only record the non-zero elements of a sparse matrix for storage. Since these represent a small fraction
of the total 3107x3107 = 9,653,449 elements in the weight matrix, we save
a tremendous amount of computer memory. In fact for our example of the
3,107 U.S. counties, only 12,429 non-zero elements were found in the firstorder spatial contiguity matrix, representing a very small fraction (far less
than 1 percent) of the total elements.
MATLAB provides a function sparse that can be used to construct a
large sparse matrix by simply indicating the row and column positions of
non-zero elements and the value of the matrix element for these non-zero
39
row and column elements. Continuing with our example, we can store the
first-order contiguity matrix in a single data file containing 12,429 rows with
3 columns that take the form:
row column value
This represents a considerable savings in computational space when compared to storing a matrix containing 9,653,449 elements. A handy utility
function in MATLAB is spy which allows one to produce a specially formatted graph showing the sparsity structure associated with sparse matrices.
We demonstrate by executing spy(W) on our weight matrix W from the
Pace and Barry data set, which produced the graph shown in Figure 2.2.
As we can see from the figure, most of the non-zero elements reside near the
diagonal.
As an example of storing a sparse first-order contiguity matrix, consider
example 2.2 below that reads data from the file ford.dat in sparse format and uses the function sparse to construct a working spatial contiguity
matrix W . The example also produces a graphical display of the sparsity
structure using the MATLAB function spy.
% ----- Example 2.2 Using sparse matrix functions
load ford.dat; % 1st order contiguity matrix
% stored in sparse matrix form
ii = ford(:,1);
jj = ford(:,2);
ss = ford(:,3);
clear ford;
% clear out the matrix to save RAM memory
W = sparse(ii,jj,ss,3107,3107);
clear ii; clear jj; clear ss; % clear out these vectors to save memory
spy(W);
To compare our function far with the approach proposed by Pace and
Barry, we implemented their approach and provide timing results. We take
a more efficient approach to the grid search over values of the parameter
than suggested by Pace and Barry. Rather than search over a large number
of values for , we based our search on a large increment of 0.1 for an initial
grid of values covering from 1/min to 1/max. Given the determinant
of (I W ) calculated using sparse matrix algorithms in MATLAB, we
evaluated the negative log likelihood function values for this grid of values
to find the value that minimizes the likelihood function. We then make
a second pass based on a tighter grid with increments of 0.01 around the
optimal value found in the first pass. A third and final pass is based on an
40
500
1000
1500
2000
2500
3000
0
500
1000
1500
nz = 12428
2000
2500
3000
Figure 2.2: Sparsity structure of W from Pace and Barry

even finer grid with increments of 0.001 around the optimal estimate from
the second pass.
Note that we used a MATLAB sparse function eigs to solve for the
eigenvalues of the contiguity matrix W , which requires 60 seconds to solve
this part of the problem as shown in the output below. The time necessary
to perform each pass over the grid of 21 values for was around 10 seconds.
With a total of 3 passes to produce an estimate of accurate to 3 decimal
digits, we have a total elapsed time of 1 minute and 30 seconds to solve
for the maximum likelihood estimate of . This is certainly a reasonable
amount of computational time for such a large problem on a reasonably in-
41
expensive desktop computing platform. Of course, there is still the problem

of producing measures of dispersion for the estimates that Pace and Barry
address by suggesting the use of likelihood ratio statistics.
elapsed_time = 59.8226 % computing min,max eigenvalues
elapsed_time = 10.5280 % carrying out 1st 21-point grid over rho
elapsed_time = 10.3791 % carrying out 2nd 21-point grid over rho
elapsed_time = 10.3747 % carrying out 3rd 21-point grid over rho
estimate of rho =
0.7220
estimate of sigma =
0.0054
How does our approach compare to that of Pace and Barry? Example 2.3
shows a program to estimate the same FAR model using our far function.
% ----- Example 2.3 Using the far() function
%
with very large data set from Pace and Barry
load elect.dat;
% load data on votes
y = elect(:,7)./elect(:,8); % proportion of voters casting votes
ydev = y - mean(y);
% deviations from the means form
clear y;
% conserve on RAM memory
clear elect; % conserve on RAM memory
load ford.dat; % 1st order contiguity matrix stored in sparse matrix form
ii = ford(:,1); jj = ford(:,2); ss = ford(:,3);
n = 3107;
clear ford; % clear ford matrix to save RAM memory
W = sparse(ii,jj,ss,n,n);
clear ii; clear jj; clear ss; % conserve on RAM memory
tic; res = far(ydev,W); toc;
prt(res);
In terms of time needed to solve the problem, our use of the simplex
optimization algorithm takes only 10.6 seconds to produce a more accurate
estimate than that based on the grid approach of Pace and Barry. Their
approach which we modified took 30 seconds to solve for a value accurate
to 3 decimal digits. Note also in contrast to Pace and Barry, we compute
a conventional measure of dispersion using the numerical hessian estimates
which takes only 1.76 seconds. The total time required to compute not only
the estimates and measures of dispersion for and , but the Rsquared
statistics and log likelihood function was around 100 seconds.
elapsed_time
elapsed_time
elapsed_time
elapsed_time
total time
= 59.8226 % computing min,max eigenvalues

= 10.6622 % time required for simplex solution of rho
= 1.7681 % time required for hessian evaluation
= 1.7743 % time required for likelihood evaluation
= 74.01
% comparable time to Pace and Barry
42

R-squared
=
0.5375
sigma^2
=
0.0054
Nobs, Nvars
=
3107,
1
log-likelihood =
1727.9824
# of iterations =
13
min and max rho =
-1.0710,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.721474
59.495159
0.000000
Many of the ideas developed in this section regarding the use of MATLAB sparse matrix algorithms will apply equally to the estimation procedures we develop in the next three sections for the other members of the
spatial autoregressive model family.
2.2
The mixed autoregressive-regressive model
This model extends the first-order spatial autoregressive model to include

a matrix X of explanatory variables such as those used in traditional regression models. Anselin (1988) provides a maximum likelihood method for
estimating the parameters of this model that he labels a mixed regressive
- spatial autoregressive model. We will refer to this model as the spatial
autoregressive model (SAR). The SAR model takes the form:
y = W y + X +
(2.12)
N (0, In )
2
Where y contains an nx1 vector of dependent variables, X represents the

usual nxk data matrix containing explanatory variables and W is a known
spatial weight matrix, usually a first-order contiguity matrix. The parameter
is a coefficient on the spatially lagged dependent variable, W y, and the
parameters reflect the influence of the explanatory variables on variation
in the dependent variable y. The model is termed a mixed regressive spatial autoregressive model because it combines the standard regression
model with a spatially lagged dependent variable, reminiscent of the lagged
dependent variable model from time-series analysis.
Maximum likelihood estimation of this model is based on a concentrated
likelihood function as was the case with the FAR model. A few regressions
43
are carried out along with a univariate parameter optimization of the concentrated likelihood function over values of the autoregressive parameter .
The steps are enumerated in Anselin (1988) as:
1. perform OLS for the model: y = X0 + 0

2. perform OLS for the model W y = XL + L
3. compute residuals e0 = y X 0 and eL = W y X L
4. given e0 and eL find that maximizes the concentrated likelihood
function: LC = C (n/2)ln(1/n)(e0 eL )0 (e0 eL ) + ln|I W |
5. given that maximizes LC , compute = (0 L) and
2 =
0
(1/n)(e0 eL ) (e0 eL )
Again we face the problem that using a univariate simplex optimization
algorithm to find a maximum likelihood estimate of based on the concentrated log likelihood function leaves us with no estimates of the dispersion
associated with the parameters. We can overcome this using the theoretical
information matrix for small problems and the numerical hessian approach
introduced for the FAR model in the case of large problems.
Since this model is quite similar to the FAR model which we already
presented, we will turn immediately to describing the function.
2.2.1
The sar() function
The function sar is fairly similar to our far function, with the documentation
presented below.
PURPOSE: computes spatial autoregressive model estimates
y = p*W*y + X*b + e, using sparse matrix algorithms
--------------------------------------------------USAGE: results = sar(y,x,W,rmin,rmax,convg,maxit)
44
results.meth = sar
results.beta = bhat
results.rho = rho
results.tstat = asymp t-stat (last entry is rho)
results.yhat = yhat
results.sige = sige = (y-p*W*y-x*b)*(y-p*W*y-x*b)/n
results.nvar = # of explanatory variables in x
results.y
= y data vector
results.iter
--------------------------------------------------
As in the case of far, we allow the user to provide minimum and maximum values of to use in the search. This may save time in cases where we
wish to restrict our estimate of to the positive range.
The other point to note is that this function also uses the numerical
hessian approach to compute measures of dispersion for large problems involving more than 500 observations.
2.2.2
Applied examples
As an illustration of using the sar function, consider the program in example

2.4, where we estimate a model to explain variation in votes casts on a
per capita basis in the 3,107 counties. The explanatory variables in the
model were: the proportion of population with high school level education
or higher, the proportion of the population that are homeowners and the
income per capita. Note that the population deflater used to convert the
variables to per capita terms was the population 18 years or older in the
county.
% ----- Example 2.4 Using the sar() function with a very large data set
load elect.dat;
% load data on votes in 3,107 counties
y = (elect(:,7)./elect(:,8));
% convert to per capita variables
x1 = log(elect(:,9)./elect(:,8)); % education
x2 = log(elect(:,10)./elect(:,8));% homeownership
x3 = log(elect(:,11)./elect(:,8));% income
n = length(y); x = [ones(n,1) x1 x2 x3];
clear x1; clear x2; clear x3;
clear elect;
45
n = 3107;
vnames = strvcat(voters,const,educ,homeowners,income);
to = clock;
res = sar(y,x,W);
etime(clock,to)
prt(res,vnames);
We use the MATLAB clock function as well as etime to determine the

overall execution time needed to solve this problem, which was 130 seconds.
The estimation results are presented below:
voters
R-squared
=
0.6419
Rbar-squared
=
0.6416
sigma^2
=
0.0042
Nobs, Nvars
=
3107,
4
log-likelihood =
5053.5179
# of iterations =
13
min and max rho =
-1.0710,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
0.753169
25.963031
0.000000
educ
0.148553
17.341543
0.000000
homeowners
0.208960
26.142340
0.000000
income
-0.085462
-9.413244
0.000000
rho
0.563764
39.797104
0.000000
We see from the results that all of the explanatory variables exhibit a
significant effect on the variable we wished to explain. The results also
indicate that the dependent variable y exhibits strong spatial dependence
even after taking the effect of these variables into account as the estimate
of on the spatial lagged variable is large and significant.
As an illustration of the bias associated with least-squares estimation
of spatial autoregressive models, we present an example based on a spatial
sample of 88 observations for counties in the state of Ohio. A sample of
average housing values for each of 88 counties in Ohio will be related to
population per square mile, the housing density and unemployment rates in
each county. This regression relationship can be written as:
HOU SEi = + P OPi + HDEN SIT Yi + U N EM P LOYi + i (2.13)
46
The motivation for the regression relationship is that population and

household density as well as unemployment rates work to determine the
house values in each county. Consider that the advent of suburban sprawl
and the notion of urban rent gradients suggests that housing values in contiguous counties should be related. The least-squares relationship in (2.13)
ignores the spatial contiguity information whereas the SAR model would
allow for this type of variation in the model.
The first task is to construct a spatial contiguity matrix for use with
our spatial autoregressive model. This could be accomplished by examining a map of the 88 counties and recording neighboring tracts for every
observation, a very tedious task. An alternative is to use the latitude and
longitude coordinates to construct a contiguity matrix. We rely on a function xy2cont that carries out this task. This function is part of Pace and
Barrys Spatial Statistics Toolbox for MATLAB, but has been modified to
fit the documentation conventions of the spatial econometrics library. The
function documentation is shown below:
PURPOSE: uses x,y coord to produce spatial contiguity weight matrices
with delaunay routine from MATLAB version 5.2
-----------------------------------------------------USAGE: [w1 w2 w3] = xy2cont(xcoord,ycoord)
where:
xcoord = x-direction coordinate vector (nobs x 1)
ycoord = y-direction coordinate vector (nobs x 1)
-----------------------------------------------------RETURNS: w1 = W*W*S, a row-stochastic spatial weight matrix
w2 = W*S*W, a symmetric spatial weight matrix (max(eig)=1)
w3 = diagonal matrix with i,i equal to 1/sqrt(sum of ith row)
-----------------------------------------------------References: Kelley Pace, Spatial Statistics Toolbox 1.0
------------------------------------------------------
This function essentially uses triangles connecting the x-y coordinates in

space to deduce contiguous entities. As an example of using the function,
consider constructing a spatial contiguity matrix for the Columbus neighborhood crime data set where we know both the first-order contiguity structure
taken from a map of the neighborhoods as well as the x-y coordinates. Here
is a program to generate the first-order contiguity matrix from the latitude
and longitude coordinates and produce a graphical comparison of the two
contiguity structures shown in Figure 2.3.
% ----- Example 2.5
load anselin.data;
xc = anselin(:,5);
yc = anselin(:,4);
Using the xy2cont() function

% Columbus neighborhood crime
% longitude coordinate
% latitude coordinate
47
load Wmat.data;
% load standardized contiguity matrix
% create contiguity matrix from x-y coordinates
[W1 W2 W3] = xy2cont(xc,yc);
% graphically compare the two
spy(W2,ok); hold on; spy(Wmat,+k);
legend(generated,actual);
10
generated
actual
15
20
25
30
35
40
45
50
10
15
20
25
nz = 232
30
35
40
45
50
Figure 2.3: Generated contiguity structure results

Example 2.6 reads in the data from two files containing a database for the
88 Ohio counties as well as data vectors containing the latitude and longitude
information needed to construct a contiguity matrix. We rely on a log transformation of the dependent variable house values to provide better scaling for
the data. Note the use of the MATLAB construct: ohio2(:,5)./ohio1(:,2),
48
which divides every element in the column vector ohio(:,5) containing total
households in each county by every element in the column vector ohio1(:,2),
which contains the population for every county. This produces the number of
households per capita for each county as an explanatory variable measuring
household density.
% ----- Example 2.6 Least-squares bias
%
demonstrated with Ohio county data base
load ohio1.dat; % 88 counties (observations)
% 10 columns
% col1 area in square miles
% col2 total population
% col3 population per square mile
% col4 black population
% col5 blacks as a percentage of population
% col6 number of hospitals
% col7 total crimes
% col8 crime rate per capita
% col9 population that are high school graduates
% col10 population that are college graduates
load ohio2.dat; % 88 counties
% 10 columns
% col1 income per capita
% col2 average family income
% col3 families in poverty
% col4 percent of families in poverty
% col5 total number of households
% col6 average housing value
% col7 unemployment rate
% col8 total manufacturing employment
% col9 manufacturing employment as a percent of total
% col10 total employment
load ohio.xy; % latitude-longitude coordinates of county centroids
[junk W junk2] = xy2cont(ohio(:,1),ohio(:,2)); % make W-matrix
y = log(ohio2(:,6)); n = length(y);
x = [ ones(n,1) ohio1(:,3) ohio2(:,5)./ohio1(:,2) ohio2(:,7) ];
vnames = strvcat(hvalue,constant,popsqm,housedensity,urate);
res = ols(y,x); prt(res,vnames);
res = sar(y,x,W); prt(res,vnames);
The results from these two regressions are shown below. The first point
to note is that the spatial autocorrelation coefficient estimate for the SAR
model is statistically significant, indicating the presence of spatial autocorrelation in the regression relationship. Least-squares ignores this type of
variation producing estimates that lead us to conclude that all three explanatory variables are significant in explaining housing values across the 88
county sample. In contrast, the SAR model leads us to conclude that the
49
population density (popsqm) is not statistically significant at conventional

levels. Keep in mind that the OLS estimates are biased and inconsistent,
so the inference of significance from OLS we would draw is likely to be
incorrect.
hvalue
R-squared
=
0.6292
Rbar-squared =
0.6160
sigma^2
=
0.0219
Durbin-Watson =
2.0992
Nobs, Nvars
=
88,
4
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
11.996858
71.173358
0.000000
popsqm
0.000110
2.983046
0.003735
housedensity
-1.597930
-3.344910
0.001232
urate
-0.067693
-7.525022
0.000000
hvalue
R-squared
=
0.7298
Rbar-squared
=
0.7201
sigma^2
=
0.0153
Nobs, Nvars
=
88,
4
log-likelihood =
87.284225
# of iterations =
13
min and max rho =
-2.0158,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
6.300144
35.621170
0.000000
popsqm
0.000037
1.196689
0.234794
housedensity
-1.251435
-3.140028
0.002332
urate
-0.055474
-7.387845
0.000000
rho
0.504131
53.749348
0.000000
A second point is that taking the spatial variation into account improves
the fit of the model, raising the R-squared statistic for the SAR model.
Finally, the magnitudes of the OLS parameter estimates indicate that house
values are more sensitive to the household density and the unemployment
rate variables than the SAR model. For example, the OLS estimates imply
that a one percentage point increase in the unemployment rate leads to
a decrease of 6.8 percent in house values whereas the SAR model places
this at 5.5 percent. Similarly, the OLS estimates for household density is
considerably larger in magnitude than that from the SAR model.
50
The point of this illustration is that ignoring information regarding the

spatial configuration of the data observations will produce different inferences that may lead to an inappropriate model specification. Anselin and
Griffith (1988) also provide examples and show that traditional specification tests are plagued by the presence of spatial autocorrelation, so that we
should not rely on these tests in the presence of significant spatial autocorrelation.
2.3
The spatial errors model
Here we turn attention to the spatial errors model shown in (2.14), where
the disturbances exhibit spatial dependence. Anselin (1988) provides a maximum likelihood method for this model which we label SEM here.
y = X + u
(2.14)
u = W u +
N (0, 2In )
y contains an nx1 vector of dependent variables and X represents the usual
nxk data matrix containing explanatory variables. W is a known spatial
weight matrix and the parameter is a coefficient on the spatially correlated
errors analogous to the serial correlation problem in time series models. The
parameters reflect the influence of the explanatory variables on variation
in the dependent variable y.
We introduce a number of statistical tests that can be used to detect
the presence of spatial autocorrelation in the residuals from a least-squares
model. Use of these tests will be illustrated in the next section.
The first test for spatial dependence in the disturbances of a regression
model is called Morans Istatistic. Given that this test indicates spatial
correlation in the least-squares residuals, the SEM model would be an appropriate way to proceed.
Morans Istatistic takes two forms depending on whether the spatial
weight matrix W is standardized or not.
1. W not standardized
2. W standardized
I = (n/s)[e0 W e]/e0 e
(2.15)
I = e0 W e/e0 e
(2.16)
51
Where e represent regression residuals. Cliff and Ord (1972, 1973, 1981)
show that the asymptotic distribution for Morans I based on least-squares
residuals corresponds to a standard normal distribution after adjusting the
Istatistic by subtracting the mean and dividing by the standard deviation
of the statistic. The adjustment takes two forms depending on whether W
is standardized or not. (Anselin, 1988, page 102).
1. W not standardized, let M = (I X(X 0X)1 X 0 ) and tr denotes the
trace operator.
E(I) = (n/s)tr(M W )/(n k)
V (i) = (n/s)2 [tr(M W M W 0 ) + tr(M W )2 + (tr(M W ))2 ]/d E(I)2
d = (n k)(n k + 2)
ZI
= [I E(I)]/V (I)1/2
(2.17)
2. W standardized
E(I) = tr(M W )/(n k)
V (i) = [tr(M W M W 0 ) + tr(M W )2 + (tr(M W ))2 ]/d E(I)2
d = (n k)(n k + 2)
ZI
= [I E(I)]/V (I)1/2
(2.18)
We implement this test in the MATLAB function moran, which takes a

regression model and spatial weight matrix W as input and returns a structure variable containing the results from a Moran test for spatial correlation
in the residuals. The prt function can be used to provide a nicely formatted
print out of the test results. The help documentation for the function is
shown below.
PURPOSE: computes Morans I-statistic for spatial correlation
in the residuals of a regression model
--------------------------------------------------USAGE: result = moran(y,x,W)
x = independent variables matrix
W = contiguity matrix (standardized or unstandardized)
--------------------------------------------------RETURNS: a structure variable
result.morani = e*W*e/e*e (I-statistic)
result.istat = [i - E(i)]/std(i), standardized version
result.imean = E(i),
expectation
52
result.ivar = var(i), variance

result.prob = std normal marginal probability
result.nobs = # of observations
result.nvar = # of variables in x-matrix
--------------------------------------------------NOTE: istat > 1.96, => small prob,
=> reject HO: of no spatial correlation
--------------------------------------------------See also: prt(), lmerrors, walds, lratios
---------------------------------------------------
A number of other asymptotic approaches exist for testing whether spatial correlation is present in the residuals from a least-squares regression
model. Some of these are the likelihood ratio test, the Wald test and a
lagrange multiplier test, all of which are based on maximum likelihood estimation of the SEM model.
The likelihood ratio test is based on the difference between the log likelihood from the SEM model and the log likelihood from a least-squares
regression. This quantity represents a statistic that is distributed 2 (1). A
function lratios carries out this test and returns a results structure which
can be passed to the prt function for presentation of the results. Documentation for the function is:
PURPOSE: computes likelihood ratio test for spatial
correlation in the errors of a regression model
--------------------------------------------------USAGE: result = lratios(y,x,W)
or: result = lratios(y,x,W,sem_result);
where:
y = dependent variable vector
W = contiguity matrix (standardized or unstandardized)
sem_result = a results structure from sem()
result.meth = lratios
result.lratio = likelihood ratio statistic
result.chi1 = 6.635
result.prob = marginal probability
result.nvar = # of variables in x-matrix
--------------------------------------------------NOTES: lratio > 6.635, => small prob,
calling the function with a results structure from sem()
can save time for large models that have already been estimated
---------------------------------------------------
53
Note that we allow the user to supply a results structure variable from
the sem estimation function, which would save the time needed to estimate the SEM model if the model has already been estimated. This could
represent a considerable savings for large problems.
Another approach is based on a Wald test for residual spatial autocorrelation. This test statistic (shown in (2.19)) is distributed 2 (1). (Anselin,
1988, page 104).
= 2 [t2 + t3 (1/n)(t21 )] 2 (1)
t1 = tr(W. B
t2 = tr(W B
t3 = tr(W B
(2.19)
1 2
1 0
) (W B 1 )
Where B = (In W ), with the maximum likelihood estimate of used,

and . denotes element-by-element matrix multiplication.
We have implemented a MATLAB function walds, that carries out this
test. The function documentation is shown below:
PURPOSE: Wald statistic for spatial autocorrelation in
the residuals of a regression model
--------------------------------------------------USAGE: result = walds(y,x,W)
W = contiguity matrix (standardized)
result.meth = walds
result.wald = Wald statistic
result.chi1 = 6.635
result.nvar = # of variables
--------------------------------------------------NOTE: wald > 6.635, => small prob,
--------------------------------------------------See also: lmerror, lratios, moran
---------------------------------------------------
A fourth approach is the Lagrange Multiplier (LM) test which is based

on the least-squares residuals and calculations involving the spatial weight
matrix W . The LM statistic takes the form: (Anselin, 1988, page 104).
LM
T
= (1/T )[(e0W e)/ 2 ]2 2 (1)
54
(2.20)
= tr(W + W ). W
Where e denote least-squares residuals and again we use . to denote element

by element matrix multiplication.
This test is implemented in a MATLAB function lmerrors with the
documentation for the function shown below.
PURPOSE: LM error statistic for spatial correlation in
the residuals of a regression model
--------------------------------------------------USAGE: result = lmerror(y,x,W)
result.meth = lmerror
result.lm = LM statistic
result.chi1 = 6.635
--------------------------------------------------NOTE: lm > 6.635, => small prob,
--------------------------------------------------See also: walds, lratios, moran
---------------------------------------------------
Finally, a test based on the residuals from the SAR model can be used
to examine whether inclusion of the spatial lag term eliminates spatial dependence in the residuals of the model. This test differs from the four tests
outlined above in that we allow for the presence of the spatial lagged variable Cy in the model. The test for spatial dependence is conditional on
having a parameter not equal to zero in the model, rather than relying on
least-squares residuals as in the case of the other four tests.
One could view this test as based on the following model:
y = Cy + X + u
u = W u +
N (0, 2In )
(2.21)
55
Where the focus of the test is on whether the parameter = 0. This test
statistic is also a Lagrange Multiplier statistic based on: (Anselin, 1988,
page 106).
(e0 W e/ 2 )[T22 (T21)2 var()]1 2 (1)
(2.22)
T22 = tr(W. W + W W )
T21 = tr(W. CA1 + W 0 CA1 )
Where W is the spatial weight matrix shown in (2.21), A = (In C) and
var() is the maximum likelihood estimate of the variance of the parameter
in the model.
We have implemented this test in a MATLAB function lmsar with the
documentation for the function shown below.
PURPOSE: LM statistic for spatial correlation in the
residuals of a spatial autoregressive model
--------------------------------------------------USAGE: result = lmsar(y,x,W1,W2)
W1 = contiguity matrix for rho
W2 = contiguity matrix for lambda
result.meth = lmsar
result.lm = LM statistic
result.chi1 = 6.635
--------------------------------------------------NOTE: lm > 6.635, => small prob,
--------------------------------------------------See also: walds, lratios, moran, lmerrors
---------------------------------------------------
It should be noted that a host of other methods to test for spatial dependence in various modeling situations have been proposed. In addition,
the small sample properties of many alternative tests have been compared
in Anselin and Florax (1994) and Anselin and Rey (1991). One point to
consider is that many of the matrix computations required for these tests
cannot be carried out with very large data samples. We discuss this issue
and suggest alternative approaches in the applied examples of Section 2.3.2.
2.3.1
56
The sem() function
To implement estimation of the spatial error model (SEM) we can draw on

the sparse matrix approach we used for the FAR and SAR models. One approach to estimating this model is based on an iterative approach that: 1)
constructs least-squares estimates and associated residuals, 2) finds a value
of that maximizes the log likelihood conditional on the least-squares values, 3) updates the least-squares values of using the value of determined
in step 2). This process is continued until convergence in the residuals.
Next, we present documentation for the function sem that carries out
the iterative estimation process. This is quite similar in approach to the
functions far and sar already described.
PURPOSE: computes spatial error model estimates
y = XB + u, u = L*W*u + e, using sparse algorithms
--------------------------------------------------USAGE: results = sem(y,x,W,lmin,lmax,convg,maxit)
lmin = (optional) minimum value of lambda to use in search
lmax = (optional) maximum value of lambda to use in search
results.meth = sem
results.beta = bhat
results.lam = L (lambda)
results.tstat = asymp t-stats (last entry is lam)
results.yhat = yhat
results.sige = sige = e(I-L*W)*(I-L*W)*e/n
results.lik = log likelihood
results.nobs = nobs
results.nvar = nvars (includes lam)
results.y
= y data vector
results.iter
results.lmax = 1/max eigenvalue of W (or lmax if input)
results.lmin = 1/min eigenvalue of W (or lmin if input)
--------------------------------------------------
It should be noted that an alternative approach to estimating this model

would be to directly maximize the log likelihood function for this model
57
using a general optimization algorithm. It might produce an improvement

in speed, depending on how many likelihood function evaluations are needed
when solving large problems. We provide an option for doing this in a
function semo that relies on a MATLAB optimization function maxlik
that is part of my Econometrics Toolbox software.
2.3.2
Applied examples
We provide examples of using the functions moran, lmerror, walds and

lratios that test for spatial correlation in the least-squares residuals as well
as lmsar to test for spatial correlation in the residuals of an SAR model.
These examples are based on the Anselin neighborhood crime data set. It
should be noted that computation of the Moran Istatistic, the LM error
statistic, and the Wald test require matrix multiplications involving the large
spatial weight matrices C and W . This is not true of the likelihood ratio
statistic implemented in the function lratios. This test only requires that
we compare the likelihood from a least-squares model to that from a spatial
error model. As we can produce SEM estimates using our sparse matrix
algorithms, this test can be implemented for large models.
Example 2.7 shows a program that carries out all of the test for spatial
correlation as well as estimating an SEM model.
% ----- Example 2.7 Testing for spatial correlation
load wmat.dat;
y = anselin(:,1); nobs = length(y);
x = [ones(nobs,1) anselin(:,2:3)];
W = wmat;
res1 = moran(y,x,W);
prt(res1);
res2 = lmerror(y,x,W);
prt(res2);
res3 = lratios(y,x,W);
prt(res3);
res4 = walds(y,x,W);
prt(res4);
res5 = lmsar(y,x,W,W);
prt(res5);
res = sem(y,x,W);% do 1st-order spatial autoregression
prt(res,vnames); % print the output
Note that we have provided code in the prt function to provide a nicely
formatted print out of the test results from our spatial correlation testing
58
functions. From the results printed below, we see that the least-squares
residuals exhibit spatial correlation. We infer this from the small marginal
probabilities that indicate significance at the 99% level of confidence. With
regard to the LM error test for spatial correlation in the residuals of the
SAR model implemented in the function lmsar,we see from the marginal
probability of 0.565 that here we can reject any spatial dependence in the
residuals from this model.
Moran I-test for spatial correlation in residuals
Moran I
0.23610178
Moran I-statistic
2.95890622
Marginal Probability
0.00500909
mean
-0.03329718
standard deviation
0.09104680
LM error tests for spatial correlation in residuals
LM value
5.74566426
0.01652940
chi(1) .01 value
17.61100000
LR tests for spatial correlation in residuals
LR value
8.01911539
0.00462862
chi-squared(1) value
6.63500000
Wald test for spatial correlation in residuals
Wald value
14.72873758
0.00012414
chi(1) .01 value
6.63500000
LM error tests for spatial
LM value
chi(1) .01 value
correlation in SAR model residuals

0.33002340
0.56564531
6.63500000
Spatial error Model Estimates

crime
R-squared
=
0.6515
Rbar-squared
=
0.6364
sigma^2
=
95.5675
log-likelihood =
-166.40057
Nobs, Nvars
=
49,
3
# iterations
=
12
min and max lam =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
59.878750
11.157027
0.000000
income
-0.940247
-2.845229
0.006605

house value
lambda
-0.302236
0.562233
-3.340320
4.351068
59
0.001667
0.000075
As an example of estimating an SEM model on a large data set, we use

the Pace and Barry data set with the same model used to demonstrate the
SAR estimation procedure.
% ----- Example 2.8 Using the sem() function with a very large data set
load elect.dat;
y = (elect(:,7)./elect(:,8));
clear elect;
n = 3107;
to = clock;
res = sem(y,x,W);
etime(clock,to)
prt(res,vnames);
We computed estimates using both the iterative procedure implemented

in the function sem and the optimization procedure implemented in the
function semo. The time required for the optimization procedure was 338
seconds, which compared to 311 seconds for the iterative procedure. The
optimization approach required only 6 function evaluations whereas the iterative procedure required 10 function evaluations. Needless to say, almost
all of the time is spent in the log likelihood function evaluations. We present
the estimates from both approaches to demonstrate that they produce estimates that are identical to 3 decimal places. Both of these functions are
part of the spatial econometrics library, as it may be the case that the optimization approach would produce estimates in less time than the iterative
approach in other applications. This would likely be the case if very good
initial estimates were available as starting values.
% estimates from iterative approach using sem() function
voters
60
R-squared
=
0.6570
Rbar-squared
=
0.6566
sigma^2
=
0.0040
log-likelihood =
5063.9904
Nobs, Nvars
=
3107,
4
# iterations
=
10
min and max lam =
-1.0710,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
1.216656
35.780274
0.000000
educ
0.192118
14.564908
0.000000
homeowners
0.250041
29.083714
0.000000
income
-0.117625
-9.560005
0.000000
lambda
0.659193
43.121763
0.000000
% estimates from optimization approach using semo() function
voters
R-squared
=
0.6570
Rbar-squared
=
0.6566
sigma^2
=
0.0040
log-likelihood =
5063.9904
Nobs, Nvars
=
3107,
4
# iterations
=
6
min and max lam =
-1.0710,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
1.216673
35.780840
0.000000
educ
0.192125
14.565561
0.000000
homeowners
0.250036
29.083079
0.000000
income
-0.117633
-9.560617
0.000000
lambda
0.659176
43.119452
0.000000
The estimates from this model indicate that after taking into account
the influence of the explanatory variables, we still have spatial correlation in
the residuals of the model that can be modeled successfully with the SEM
model. As a confirmation of this, consider that the LR test implemented
with the function lratios produced the results shown below:
LR tests for spatial correlation in residuals
LR value
1163.01773404
0.00000000
chi-squared(1) value
6.63500000
Recall that this is a test of spatial autocorrelation in the residuals from

a least-squares model, and the test results provide a strong indication of
spatial dependence in the least-squares residuals. Note also that this is the
only test that can be implemented successfully with large data sets.
61
A reasonable alternative would be to simply estimate a FAR model using

the least-squares residuals to test for the presence of spatial dependence in
the errors. We illustrate this approach in Section 2.5.
2.4
The general spatial model
A general version of the spatial model includes both the spatial lagged term
as well as a spatially correlated error structure as shown in (2.23).
y = W1 y + X + u
(2.23)
u = W2 u +
N (0, 2In )
One point to note about this model is that W1 can equal W2 , but there
may be identification problems in this case. The log likelihood for this
model can be maximized using our general optimization algorithm on a
concentrated version of the likelihood function. The parameters and 2
are concentrated out of the likelihood function, leaving the parameters
and . This eliminates the ability to use the univariate simplex optimization
algorithm fmin that we used with the other spatial autoregressive models.
We can still produce a sparse matrix algorithm for the log likelihood
function and proceed in a similar fashion to that used for the other spatial
autoregressive models. One difference is that we cannot easily impose restrictions on the parameters and to force them to lie within the ranges
defined by the maximum and minimum eigenvalues from their associated
weight matrices W1 and W2 .
When might one rely on this model? If there were evidence that spatial dependence existed in the error structure from a spatial autoregressive
(SAR) model, the SAC model is an appropriate approach to modeling this
type of dependence in the errors. Recall, we can use the LM-test implemented in the function lmsars to see if spatial dependence exists in the
residuals of an SAR model.
Another place where one might rely on this model is a case where a
second-order spatial contiguity matrix was used for W2 that corresponds to
a first-order contiguity matrix W1 . This type of model would express the belief that the disturbance structure involved higher-order spatial dependence,
perhaps due to second-round effects of a spatial phenomena being modeled.
62
A third example of using matrices W1 and W2 might be where W1 represented a first-order contiguity matrix and W2 was constructed as a diagonal
matrix measuring the distance from the central city. This type of configuration of the spatial weight matrices would indicate a belief that contiguity
alone does not suffice to capture the spatial effects at work. The distance
from the central city might also represent an important factor in the phenomena we are modeling. This raises the identification issue, should we use
the distance weighting matrix in place of W1 and the first-order contiguity
matrix for W2 , or rely on the opposite configuration? Of course, comparing likelihood function values along with the statistical significance of the
parameters and from models estimated using both configurations might
point to a clear answer.
The log likelihood function for this model is:
L = C (n/2) ln( 2 ) + ln(|A|) + ln(|B|) (1/2 2)(e0 B 0 Be)
e = (Ay X)
A = (In W1 )
(2.24)
B = (In W2 )
We concentrate the function using the following expressions for and
2:
= (X 0A0 AX)1 (X 0 A0 ABy)
e = By x
(2.25)
2 = (e0 e)/n
Given the expressions in (2.25), we can evaluate the log likelihood given
values of and . The values of the other parameters and 2 can be
calculated as a function of the , parameters and the sample data in y, X.
2.4.1
The sac() function
Documentation for the MATLAB function sac that carries out the nonlinear optimization of the log likelihood function for this model is shown
below. There are a number of things to note about this function. First, we
provide optimization options for the user in the form of a structure variable
info. These options allow the user to control some aspects of the maxlik
63
optimization algorithm and to print intermediate results while optimization

is proceeding.
This is the first example of a function that uses the MATLAB structure
variable as an input argument. This allows us to provide a large number
of input arguments using a single structure variable. Note that you can
name the structure variable used to input the options anything it is the
fieldnames that the function sac parses to find the options.
PURPOSE: computes general Spatial Model
model: y = p*W1*y + X*b + u, u = lam*W2*u + e
--------------------------------------------------USAGE: results = sac(y,x,W1,W2)
W1 = spatial weight matrix (standardized)
W2 = spatial weight matrix
info
= a structure variable with optimization options
info.parm = (optional) 2x1 starting values for rho, lambda
info.convg = (optional) convergence criterion (default = 1e-7)
info.maxit = (optional) maximum # of iterations (default = 500)
info.method = bfgs, dfp (default bfgs)
info.pflag = flag for printing of intermediate results
results.meth = sac
results.beta = bhat
results.rho = p (rho)
results.lam = L (lambda)
results.tstat = asymptotic t-stats (last 2 are rho,lam)
results.yhat = yhat
results.sige = sige = e(I-L*W)*(I-L*W)*e/n
results.lik = likelihood function value
results.nobs = nobs
results.y
= y data vector
results.iter = # of iterations taken
--------------------------------------------------
We take the same approach to optimization failure as we did with the

sem function. A message is printed to warn the user that optimization
failed, but we let the function continue to process and return a results structure consisting of failed parameter estimates. This decision was made to allow the user to examine the failed estimates and attempt estimation based
64
on alternative optimization options. For example, the user might elect to

attempt a Davidson-Fletcher-Powell (dfp) algorithm in place of the default
Broyden-Fletcher-Goldfarb-Smith (bfgs) routine or supply starting values
for the parameters and .
With regard to optimization algorithm failures, it should be noted that
the Econometrics Toolbox contains alternative optimization functions that
can be used in place of maxlik. Any of these functions could be substituted
for maxlik in the function sac. Chapter 10 in the Econometrics Toolbox
provides examples of using these functions as well as their documentation.
The next section turns to illustrating the use of the estimation functions
we have constructed for the general spatial autoregressive model.
2.4.2
Applied examples
Our first example illustrates the use of the general spatial model with the
Anselin Columbus neighborhood crime data set. We construct a matrix W2
for use in the model based on W2 = W 0 W . To provide an illustration of
what this matrix inner product involving the first-order contiguity matrix
W represents in terms of spatial structure, we use the spy function to produce comparative graphs of the two contiguity matrices that are shown in
Figure 2.4. As we can see from the graphs, the inner product allows for a
more wide-ranging set of spatial influences reflecting secondary influences
not captured by the first-order contiguity matrix. The sparse matrix literature refers to this phenomena where matrix products produce less sparsity as
fill-in. The motivation for this terminology should be clear from the figure.
For another approach to producing spatial lag matrices, see the function
slag which represents a more appropriate way to create higher-order spatial
contiguity matrices.
Our example illustrates the point discussed earlier regarding model specification with respect to the use of W and W2 by producing estimates for
two models based on alternative configurations of these two spatial weight
matrices. We also produce estimates based on a model that uses the W
matrix for both the autoregressive lag and the autoregressive error terms.
% ----- Example 2.9 Using the sac function
load Wmat.dat;
W = Wmat;
W2 = W*W;
0
10
20
30
40
50
10
20
30
40
50
First-order contiguity structure
0
10
20
30
40
50
10
20
30
40
contiguity structure of W2
50
Figure 2.4: Spatial contiguity structures
65
66
subplot(2,1,1), spy(W);
xlabel(First-order contiguity structure);
subplot(2,1,2), spy(W2);
xlabel(contiguity structure of W2);
res1 = sac(y,x,W2,W);% general spatial model W2,W
prt(res1,vnames);
% print the output
res2 = sac(y,x,W,W2);% general spatial model W,W2
prt(res2,vnames);
% print the output
res3 = sac(y,x,W,W); % general spatial model W,W
prt(res3,vnames);
% print the output
The estimation results are shown below for all three versions of the
model. The first model produced the worst fit indicated by the lower R2
value, but a slightly higher log likelihood function value. In addition, we see
that both the coefficients on and are statistically significant, as indicated
by the tstatistics and associated marginal probabilities. The negative spatial lag parameter is problematical as this type of influence is difficult to
explain intuitively. Negative spatial lags would indicate that higher crime
rates in neighboring regions exerted a negative influence on crime on average
across the entire Columbus neighborhood sample. When considered from
an overall and limiting case viewpoint, this seems a logical impossibility.
General Spatial Model Estimates
crime
R-squared
=
0.5591
Rbar-squared =
0.5400
sigma^2
= 120.9027
log-likelihood =
-164.88367
Nobs, Nvars
=
49,
3
# iterations =
8
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
37.261470
5.472898
0.000002
income
-0.958005
-2.573394
0.013360
house value
-0.219046
-2.125609
0.038937
rho
-0.713070
-2.070543
0.044041
lambda
0.571698
18.759773
0.000000
crime
R-squared
=
0.6528
Rbar-squared =
0.6377
sigma^2
=
95.2216
log-likelihood =
-166.12487
Nobs, Nvars
=
49,
3
# iterations =
6
***************************************************************

Variable
const
income
house value
rho
lambda
Coefficient
55.343773
-0.672762
-0.320741
0.512647
0.064988
t-statistic
4.628119
-2.163164
-3.552861
3.747903
0.829892
67
t-probability
0.000030
0.035760
0.000894
0.000497
0.410886

crime
R-squared
=
0.6512
Rbar-squared =
0.6360
sigma^2
=
95.6698
log-likelihood =
-165.25612
Nobs, Nvars
=
49,
3
# iterations =
7
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
47.275680
4.138342
0.000147
income
-0.977532
-2.971857
0.004696
house value
-0.284369
-3.137873
0.002968
rho
0.167197
0.476045
0.636296
lambda
0.368187
1.348037
0.184248
The second model produced the typical positive coefficient on the spatial lag term along with an insignificant estimate for . As noted above,
it also produced a better fit to the sample data. Confirmation of these results from the second model are provided by the test for spatial dependence
in the disturbances of the SAR model that we carried out in example 2.7,
which are repeated below. From these results, we see that there is no evidence of spatial dependence in the disturbances of the SAR model. This
explains why the parameter in the general spatial version of this model is
not significantly different from zero.
The third model that relies on the same W matrix for both the spatial
lag and error correlation terms produced estimates for both and that
were insignificant. Since we have statistical evidence of spatial correlation
in the residuals from least-squares as well as the statistical significance of
the parameter in the SAR model, we would reject this model as inferior
to the SAR model results.
By way of summary, I would reject the specification associated with the
first model since it produces counter-intuitive results. An important point to
note regarding any non-linear optimization problem such as that involved
in the SAC model is that the estimates may not reflect global optimum.
A few different solutions of the optimization problem based on alternative
starting values is usually undertaken to confirm that the estimates do indeed
68
represent global solutions to the problem. The function sac allows the user
to input alternative starting values, making this relatively easy to do.
Given that we accept the second model as best, because the parameter
is not statistically significant, we would conclude that an SAR model would
suffice for this sample data.
LM error tests for spatial
LM value
chi(1) .01 value
correlation in SAR model residuals

0.33002340
0.56564531
6.63500000
A final example uses the large Pace and Barry data set to illustrate the
sac function in operation on large problems. Example 2.10 turns on the
printing flag so we can observe intermediate results from the optimization
algorithm as it proceeds.
% ----- Example 2.10 Using the sac() function on a large data set
load elect.dat;
y = (elect(:,7)./elect(:,8));
clear elect;
n = 3107;
to = clock; info.pflag = 1;
res = sac(y,x,W,W*W,info);
etime(clock,to)
prt(res,vnames);
The results including intermediate results that were printed are shown
below. It took 602 seconds to solve this problem involving only 6 iterations
by the maxlik function. This function tends to be faster than the alternative
optimization algorithms available in the Econometrics Toolbox.
==== Iteration ==== 2
log-likelihood
bconvergence
fconvergence
-5052.2920
1.1386
0.4438
Parameter
Estimates
Gradient

Parameter 1
Parameter 2
69
0.6329 -479.9615
0.0636 -4927.5720

log-likelihood
bconvergence
-5068.0928
0.7587
fconvergence
0.0031
Parameter
Estimates
Gradient
Parameter 1
0.6644
-18.8294
Parameter 2
0.0185 -793.3583
log-likelihood
bconvergence
fconvergence
-5068.2786
0.2197
0.0000
Parameter
Estimates
Gradient
Parameter 1
0.6383
-54.3318
Parameter 2
0.0219
44.6226
log-likelihood
bconvergence
fconvergence
-5068.5622
0.0405
0.0001
Parameter
Estimates
Gradient
Parameter 1
0.6497
47.7892
Parameter 2
0.0223
27.5841
log-likelihood
bconvergence
fconvergence
-5068.5627
0.0107
0.0000
Parameter
Estimates
Gradient
Parameter 1
0.6500
0.6407
Parameter 2
0.0221
-3.6765
voters
R-squared
=
0.6598
Rbar-squared =
0.6594
sigma^2
=
0.0040
log-likelihood =
5068.5627
Nobs, Nvars
=
3107,
4
# iterations =
6
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
1.193076
34.732852
0.000000
educ
0.180548
13.639337
0.000000
homeowners
0.267665
30.881364
0.000000
income
-0.107153
-8.596555
0.000000
rho
0.649976
39.916262
0.000000
lambda
0.022120
3.011273
0.002623
From the estimation results we see some evidence that a general spatial
70
model might be appropriate for this modeling problem. The parameters

and are both statistically significant, but the estimated value of the
parameter is close to zero. This is consistent with the idea that any
second order effects are small. Nonetheless, the fit of this model is slightly
better than the SAR and SEM models (see example 2.4 and example 2.8) as
indicated by a slightly higher adjusted Rsquared value and the likelihood
function value is also slightly higher.
2.5
An applied exercise
In a well-known paper, Harrison and Rubinfeld (1978) used a housing data

set for the Boston SMSA with 506 observations (one observation per census
tract) containing 14 variables. They were interested in the housing demand
and clean air issues. We will use this spatial data set to illustrate specification and testing for the spatial autoregressive models presented in this
chapter. Our dependent variable will be median housing prices for each of
the 506 census tracts. The explanatory variables in the dataset are shown
below:
The Boston house-price data of Harrison, D. and Rubinfeld, D.L. Hedonic
prices and the demand for clean air, J. Environ. Economics & Management,
vol.5, 81-102, 1978.
Variables in order:
CRIM
per capita crime rate by town
ZN
proportion of residential land zoned for lots over 25,000 sq.ft.
INDUS
proportion of non-retail business acres per town
CHAS
Charles River dummy (= 1 if tract bounds river; 0 otherwise)
NOX
nitric oxides concentration (parts per 10 million)
RM
average number of rooms per dwelling
AGE
proportion of owner-occupied units built prior to 1940
DIS
weighted distances to five Boston employment centres
RAD
index of accessibility to radial highways
TAX
full-value property-tax rate per $10,000
PTRATIO pupil-teacher ratio by town
B
1000(Bk - 0.63)^2 where Bk is the proportion of blacks by town
LSTAT
percent lower status of the population
MEDV
Median value of owner-occupied homes in $1000s
Belsley, Kuh, and Welch (1980) used the data to examine the effects of
robust estimation and published the observations in an appendix on pages
244-261. It should be noted that they published data that included various
transformations, so the data in their appendix does not match our data
71
which is in the raw untransformed format. Pace (1993), Gilley and Pace
(1996), and Pace and Gilley (1997) have used this data set with spatial
econometric models and longitude-latitude coordinates for the census tracts
have been added to the dataset.
Our regression model will simply relate the median house values to all of
the explanatory variables, simplifying our specification task. We will focus
on alternative spatial specifications and models.
The next task involves some scaling and standardization issues surrounding the data set. Belsley, Kuh and Welsch (1980) used this data set to illustrate numerically ill-conditioned data that contained outliers and influential observations. Poor scaling will adversely impact our numerical hessian
approach to determining the variance-covariance structure of the spatial
autoregressive parameter estimates. Intuitively, the hessian function attempts to compute a numerical derivative by perturbing each parameter in
turn and examining the impact on the likelihood function. If the parameters
vary widely in terms of magnitudes because the data is poorly scaled, this
task will be more difficult and we may calculate negative variances.
Example 2.11 demonstrates the nature of these scaling problems, carrying out a least-squares regression. We see that the coefficient estimates
varying widely in magnitude, so we scale the variables in the model using
a function studentize from the Econometric Toolbox that subtracts the
means and divides by the standard deviations. Another least-squares regression is then carried out to illustrate the impact of scaling on the model
coefficients.
% ----- Example 2.11 Least-squares on the Boston dataset
load boston.raw; % Harrison-Rubinfeld data
[n k] = size(boston);y = boston(:,k); % median house values
x = [ones(n,1) boston(:,1:k-1)];
% other variables
vnames = strvcat(hprice,constant,crime,zoning,industry, ...
charlesr,noxsq,rooms2,houseage,distance, ...
access,taxrate,pupil/teacher,blackpop,lowclass);
res = ols(y,x); prt(res,vnames);
ys = studentize(y); xs = studentize(x(:,2:k));
res2 = ols(ys,xs);
vnames2 = strvcat(hprice,crime,zoning,industry,charlesr, ...
noxsq,rooms2,houseage,distance,access,taxrate, ...
pupil/teacher,blackpop,lowclass);
prt(res2,vnames2);
% sort actual and predicted by housing values from low to high
yhat = res2.yhat; [ysort yi] = sort(ys); yhats = yhat(yi,1);
tt=1:n; % plot actual vs. predicted
plot(tt,ysort,ok,tt,yhats,+k);
ylabel(housing values);
72
xlabel(census tract observations);
The results indicate that the coefficient estimates based on the unscaled
data vary widely in magnitude from 0.000692 to 36.459, whereas the scaled
variables produce coefficients ranging from 0.0021 to -0.407.
Ordinary Least-squares Estimates (non-scaled variables)
hprice
R-squared
=
0.7406
Rbar-squared =
0.7338
sigma^2
=
22.5179
Durbin-Watson =
1.2354
Nobs, Nvars
=
506,
14
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
36.459488
7.144074
0.000000
crime
-0.108011
-3.286517
0.001087
zoning
0.046420
3.381576
0.000778
industry
0.020559
0.334310
0.738288
charlesr
2.686734
3.118381
0.001925
noxsq
-17.766611
-4.651257
0.000004
rooms2
3.809865
9.116140
0.000000
houseage
0.000692
0.052402
0.958229
distance
-1.475567
-7.398004
0.000000
access
0.306049
4.612900
0.000005
taxrate
-0.012335
-3.280009
0.001112
pupil/teacher
-0.952747
-7.282511
0.000000
blackpop
0.009312
3.466793
0.000573
lowclass
-0.524758
-10.347146
0.000000
Ordinary Least-squares Estimates (scaled variables)
hprice
R-squared
=
0.7406
Rbar-squared =
0.7343
sigma^2
=
0.2657
Durbin-Watson =
1.2354
Nobs, Nvars
=
506,
13
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.101017
-3.289855
0.001074
zoning
0.117715
3.385011
0.000769
industry
0.015335
0.334650
0.738032
charlesr
0.074199
3.121548
0.001905
noxsq
-0.223848
-4.655982
0.000004
rooms2
0.291056
9.125400
0.000000
houseage
0.002119
0.052456
0.958187
distance
-0.337836
-7.405518
0.000000
access
0.289749
4.617585
0.000005

taxrate
pupil/teacher
blackpop
lowclass
-0.226032
-0.224271
0.092432
-0.407447
-3.283341
-7.289908
3.470314
-10.357656
73
0.001099
0.000000
0.000565
0.000000
The program in example 2.11 also produces a plot of the actual versus
predicted values from the model sorted by housing values from low to high.
From this plot (shown in Figure 2.5), we see large predicted errors for the
highest housing values. This suggests a log transformation on the dependent
variable y in the model would be appropriate. The figure also illustrates that
housing values above $50,000 have been censored to a value of $50,000, a
subject we take up in Chapter 5.
3
housing values
-1
-2
-3
100
200
300
400
census tract observations
500
600
Figure 2.5: Actual vs. Predicted housing values

We adopt a model based on the scaled data and a log transformation
for the dependent variable and carry out least-squares estimation again in
example 2.12. As a test for spatial autocorrelation in the least-squares residuals, we employ a first-order spatial autoregressive model on the residuals.
74
We also carry out a Morans I test for spatial autocorrelation, which may
or may not work depending on how much RAM memory you have in your
computer. Note that we can directly use the prt function without first
returning the results from moran to a results structure. We rely on our
function xy2cont to generate a spatial contiguity matrix needed by far and
moran to test for spatial autocorrelation.
% ----- Example 2.12 Testing for spatial correlation
load latitude.data; load longitude.data;
[W1 W W3] = xy2cont(latitude,longitude); % create W-matrix
[n k] = size(boston);y = boston(:,k);
% median house values
x = boston(:,1:k-1);
% other variables
vnames = strvcat(hprice,crime,zoning,industry,charlesr, ...
ys = studentize(log(y)); xs = studentize(x);
res = ols(ys,xs); prt(res,vnames);
resid = res.resid; % recover residuals
rmin = 0; rmax = 1;
res2 = far(resid,W,rmin,rmax); prt(res2);
prt(moran(ys,xs,W));
The results shown below indicate that we have strong evidence of spatial
autocorrelation in the residuals from the least-squares model. Our FAR
model produced a spatial correlation coefficient estimate of 0.647 with a large
tstatistic and this indication of spatial autocorrelation in the residuals is
confirmed by the Moran test results.
hprice
R-squared
=
0.7896
Rbar-squared =
0.7845
sigma^2
=
0.2155
Durbin-Watson =
1.0926
Nobs, Nvars
=
506,
13
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.216146
-7.816230
0.000000
zoning
0.066897
2.136015
0.033170
industry
0.041401
1.003186
0.316263
charlesr
0.062690
2.928450
0.003564
noxsq
-0.220667
-5.096402
0.000000
rooms2
0.156134
5.435516
0.000000
houseage
0.014503
0.398725
0.690269
distance
-0.252873
-6.154893
0.000000
access
0.303919
5.377976
0.000000

taxrate
pupil/teacher
blackpop
lowclass
-0.258015
-0.202702
0.092369
-0.507256
-4.161602
-7.316010
3.850718
-14.318127
75
0.000037
0.000000
0.000133
0.000000
(Test for spatial autocorrelation using FAR model)

R-squared
=
0.3085
sigma^2
=
0.1452
Nobs, Nvars
=
506,
1
log-likelihood =
-911.39591
# of iterations =
9
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.647624
14.377912
0.000000
Moran I-test for spatial correlation in residuals
Moran I
0.35833634
Moran I-statistic
14.70009315
0.00000000
mean
-0.01311412
standard deviation
0.02526858
Example 2.13 carries out a series of alternative spatial autoregressive

models in an attempt to specify the most appropriate model.
% ----- Example 2.13 Spatial autoregressive model estimation
% other variables
rmin = 0; rmax = 1;
tic; res1 = sar(ys,xs,W,rmin,rmax); prt(res1,vnames); toc;
tic; res2 = sem(ys,xs,W,rmin,rmax); prt(res2,vnames); toc;
tic; res3 = sac(ys,xs,W,W);
prt(res3,vnames); toc;
The results from example 2.13 are presented below. We see that all
three models produced estimates indicating significant spatial autocorrelation. For example, the SAR model produced a coefficient estimate for
equal to 0.4508 with a large tstatistic, and the SEM model produced an
estimate for of 0.7576 that was also significant. The SAC model produced
estimates of and that were both significant at the 99% level.
76
Which model is best? The log-likelihood function values are much higher
for the SEM and SAC models, so this would be evidence against the SAR
model. A further test of the SAR model would be to use the function lmsar
that tests for spatial autocorrelation in the residuals of the SAR model. If
we find evidence of residual spatial autocorrelation, it suggests that the SAC
model might be most appropriate. Note that the SAC model exhibits the
best log-likelihood function value.
hprice
R-squared
=
0.8421
Rbar-squared
=
0.8383
sigma^2
=
0.1576
Nobs, Nvars
=
506,
13
log-likelihood =
-85.099051
# of iterations =
9
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.165349
-6.888522
0.000000
zoning
0.080662
3.009110
0.002754
industry
0.044302
1.255260
0.209979
charlesr
0.017156
0.918665
0.358720
noxsq
-0.129635
-3.433659
0.000646
rooms2
0.160858
6.547560
0.000000
houseage
0.018530
0.595675
0.551666
distance
-0.215249
-6.103520
0.000000
access
0.272237
5.625288
0.000000
taxrate
-0.221229
-4.165999
0.000037
pupil/teacher
-0.102405
-4.088484
0.000051
blackpop
0.077511
3.772044
0.000182
lowclass
-0.337633
-10.149809
0.000000
rho
0.450871
12.348363
0.000000
hprice
R-squared
=
0.8708
Rbar-squared
=
0.8676
sigma^2
=
0.1290
log-likelihood =
-58.604971
Nobs, Nvars
=
506,
13
# iterations
=
10
min and max lam =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.186710
-8.439402
0.000000
zoning
0.056418
1.820113
0.069348
industry
-0.000172
-0.003579
0.997146

charlesr
noxsq
rooms2
houseage
distance
access
taxrate
pupil/teacher
blackpop
lowclass
lambda
-0.014515
-0.220228
0.198585
-0.065056
-0.224595
0.352244
-0.257567
-0.122363
0.129036
-0.380295
0.757669
-0.678562
-3.683553
8.325187
-1.744224
-3.421361
5.448380
-4.527055
-3.839952
4.802657
-10.625978
19.133467
77
0.497734
0.000255
0.000000
0.081743
0.000675
0.000000
0.000008
0.000139
0.000002
0.000000
0.000000

hprice
R-squared
=
0.8662
Rbar-squared =
0.8630
sigma^2
=
0.1335
log-likelihood =
-55.200525
Nobs, Nvars
=
506,
13
# iterations =
7
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.198184
-8.766862
0.000000
zoning
0.086579
2.824768
0.004923
industry
0.026961
0.585884
0.558222
charlesr
-0.004154
-0.194727
0.845687
noxsq
-0.184557
-3.322769
0.000958
rooms2
0.208631
8.573808
0.000000
houseage
-0.049980
-1.337513
0.181672
distance
-0.283474
-5.147088
0.000000
access
0.335479
5.502331
0.000000
taxrate
-0.257478
-4.533481
0.000007
pupil/teacher
-0.120775
-3.974717
0.000081
blackpop
0.126116
4.768082
0.000002
lowclass
-0.374514
-10.707764
0.000000
rho
0.625963
9.519920
0.000000
lambda
0.188257
3.059010
0.002342
The results from the lmsar test shown below indicate the presence of
spatial autocorrelation in the residuals of the SAR model, suggesting that
the SAC model would be appropriate.
LM error tests for spatial correlation in SAR model residuals
LM value
60.37309581
0.00000000
chi(1) .01 value
6.63500000
We would conclude that the SAC model is most appropriate here. Regarding inferences, an interesting point is that the Charles River location
78
dummy variable was statistically significant in the least-squares version of

the model, but not in any of the three spatial autoregressive models. Intuitively, taking explicit account of the spatial nature of the data eliminates
the need for this locational dummy variable. Other differences in the inferences that would be made from least-squares versus the SAC model center
on the magnitudes of pupil/teacher ratio and lower class population variables. The least-squares estimates for these two variables are roughly twice
the magnitude of those from the SAC model. Since the other two spatial
autoregressive models produce similar estimates for these two variables, we
would infer that the least-squares estimates for these coefficients are exhibiting upward bias.
To a lesser extent, we would draw a different inference regarding the magnitude of impact for the rooms2 variable from the two spatial autoregressive
models that we think most appropriate (SEM and SAC) than least-squares.
The SEM and SAC models produce estimates around 0.2 compared to a
value of 0.156 from least-squares. Interestingly, the SAR model estimate for
this variable is 0.160, close to that from least-squares.
We used this applied data set to explore the accuracy of the numerical
hessian versus information matrix approach to determining the variancecovariance matrix for the estimates. Since this is a reasonably large dataset
with a large number of explanatory variables, it should provide a good indication of how well the numerical hessian approach does. The tstatistics
from the SAR, SEM and SAC models estimated with both the information
matrix approach and the numerical hessian are presented below.
11.2 sec.
SAR(info)
-6.763498
3.007078
1.255180
0.923109
-3.397977
6.521229
0.593747
-6.037894
5.577032
-4.147236
-4.025972
3.744277
-9.900943
10.351829
46.1 sec.
79.7 sec.
58.8 sec. 170.1 sec. 114.6 sec.
SAR(hess)
SEM(info)
SEM(hess) SAC(info)
SAC(hess)
-6.888522
-8.474183
-8.439402 -8.718561
-8.766862
3.009110
1.820262
1.820113
3.266408
2.824768
1.255260
-0.003584
-0.003579 0.750766
0.585884
0.918665
-0.690453
-0.678562 -0.227615
-0.194727
-3.433659
-3.687834
-3.683553 -4.598758
-3.322769
6.547560
8.345316
8.325187
8.941594
8.573808
0.595675
-1.756358
-1.744224 -1.597997
-1.337513
-6.103520
-3.435146
-3.421361 -7.690499
-5.147088
5.625288
5.478342
5.448380
6.906236
5.502331
-4.165999
-4.527097
-4.527055 -4.985237
-4.533481
-4.088484
-3.886484
-3.839952 -4.765931
-3.974717
3.772044
4.808276
4.802657
5.971414
4.768082
-10.149809 -10.954441 -10.625978 -11.650367 -10.707764
12.348363
20.890074
19.133467 24.724209
9.519920
3.747132
3.059010
79
The numerical approach appears to work very well, producing identical

inferences for all coefficients. The SAC model produced the greatest divergence between the tstatistics and unfortunately a large discrepancy exists
for the spatial parameters and in this model. Nonetheless, we would draw
the same inferences regarding the significance of these parameters from both
approaches to computing measures of dispersion. Pace and Barry (1998) express the belief that the information matrix approach (whether computed
numerically or analytically) may not work well for spatial autoregressive
models because of the asymmetry of the profile likelihood for these models
that arises from the restricted range of the parameter . Additionally, they
argue that the necessary smoothness needed to compute well-behaved
second derivatives may not always occur as Ripley (1988) documents for a
particular case. Given this, we should perhaps qualify our evaluation of success regarding the variance-covariance calculations. On the other hand, for
this particular data set and model, the variance-covariance structure for the
spatial autoregressive models is remarkably similar to that from the leastsquares model as indicated by the similar magnitudes for the tstatistics.
Further, for the single case of the Charles River dummy variable where the
least-squares and spatial tstatistics differed, we have a plausible explanation for this difference.
The times (in seconds) required by both information matrix and numerical hessian approaches to estimating the variance-covariance structure of the
parameters are reported, and we see that the information matrix approach
was quite a bit faster for the SAR model but slower for the other two models.
One point to consider regarding the comparative execution times is that the
spatial contiguity weight matrix is not exceptionally sparse for this problem.
Of the (506x506)=256,036 elements, there are 3,006 non-zero entries which
is 1.17 percent of the elements. This would bias upward the times reported
for the numerical hessian approach since the sparse algorithms cant work
to their fullest.
By way of concluding this applied illustration, we might estimate an
SAC model based on a second order spatial contiguity matrix in place of the
first-order matrix used in example 2.13. We can also produce FAR model
estimates using the residuals from the SAC model in example 2.13 as well
as the new model based on a second-order spatial weight matrix to check
for second order spatial effects in the residuals of our model. Example 2.14
implements these final checks.
% ----- Example 2.14 Final model checking
80

% other variables
rmin = 0; rmax = 1;
W2 = W*W;
res = sac(ys,xs,W,W2); prt(res,vnames);
res2 = sac(ys,xs,W,W);
resid = res2.resid; % recover SAC residuals
prt(far(resid,W2,rmin,rmax));
The results shown below indicate that the SAC model using a secondorder spatial weight matrix produces a slightly lower likelihood function
value than the SAC model in example 2.13, but estimates that are reasonably
similar in all other regards.
hprice
R-squared
=
0.8766
Rbar-squared =
0.8736
sigma^2
=
0.1231
log-likelihood =
-56.71359
Nobs, Nvars
=
506,
13
# iterations =
7
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.195223
-8.939992
0.000000
zoning
0.085436
2.863017
0.004375
industry
0.018200
0.401430
0.688278
charlesr
-0.008227
-0.399172
0.689939
noxsq
-0.190621
-3.405129
0.000715
rooms2
0.204352
8.727857
0.000000
houseage
-0.056388
-1.551836
0.121343
distance
-0.287894
-5.007506
0.000001
access
0.332325
5.486948
0.000000
taxrate
-0.245156
-4.439223
0.000011
pupil/teacher
-0.109542
-3.609570
0.000338
blackpop
0.127546
4.896422
0.000001
lowclass
-0.363506
-10.368125
0.000000
rho
0.684424
12.793448
0.000000
lambda
0.208597
2.343469
0.019502
R-squared
=
0.0670
sigma^2
=
0.1245
81
Nobs, Nvars
=
506,
1
log-likelihood =
-827.5289
# of iterations =
9
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.188278
1.420560
0.156062
The residuals from the SAC model in example 2.13 show no significant
second-order spatial autocorrelation, since the FAR model estimate is not
statistically significant.
To further explore this model as an exercise, you might consider replacing
the W2 weight matrix in example 2.14 with a matrix based on distances.
See if this variant of the SAC model is successful. Another exercise would be
to estimate a model using: sac(ys,xs,W2,W), and compare it to the model
in example 2.14.
2.6
Chapter Summary
We have seen that spatial autoregressive models can be estimated using

univariate and bivariate optimization algorithms to solve for estimates by
maximizing the likelihood function. The sparse matrix routines in MATLAB
allow us to write functions that evaluate the log likelihood function for
large models rapidly and with a minimum of computer RAM memory. This
approach was used to construct a library of estimation functions that were
illustrated on a problem involving all 3,107 counties in the continental U.S.
on an inexpensive desktop computer.
In addition to providing functions that estimate these models, the use of
a general software design allowed us to provide both printed and graphical
presentation of the estimation results.
Another place where we produced functions that can be used in spatial
econometric analysis was in the area of testing for spatial dependence in
the residuals from least-squares models and SAR models. Functions were
devised to implement Morans Istatistic as well as likelihood ratio and
Lagrange multiplier tests for spatial autocorrelation in the residuals from
least-squares and SAR models. These tests are a bit more hampered by
large-scale data sets, but alternative approaches based on using the FAR
model on residuals or likelihood ratio tests can be used.
2.7
References
82
Anselin, L. 1980. Estimation Methods for Spatial Autoregressive Structures, (New York: Regional Science Dissertation and Monograph Series 8).
Anselin, L. and D.A. Griffith. 1988. Do spatial effects really matter
in regression analysis? Papers of the Regional Science Association, 65,
pp. 11-34.
Anselin, L. and R.J.G. Florax. 1994. Small Sample Properties of
Tests for Spatial Dependence in Regression Models: Some Further
Results, Research paper 9414, Regional Research Institute, West Virginia University, Morgantown, West Virginia.
Anselin, L. and S. Rey. 1991. Properties of tests for spatial dependence in linear regression models, Geographical Analysis, Volume 23,
pages 112-31.
Belsley, D. A., E. Kuh, and R. E. Welsch. 1980. Regression Diagnostics: Identifying Influential Data and Source of Collinearity, (John
Wiley: New York).
Cliff, A. and J. Ord, 1972. Testing for spatial autocorrelation among
regression residuals, Geographical Analysis, Vol. 4, pp. 267-84.
Cliff, A. and J. Ord, 1973. Spatial Autocorrelation, (London: Pion)
Cliff, A. and J. Ord, 1981. Spatial Processes, Models and Applications,
(London: Pion)
Gilley, O.W., and R. Kelley Pace. 1996. On the Harrison and Rubinfeld Data, Journal of Environmental Economics and Management,
Vol. 31 pp. 403-405.
Pace, R. K. and R. Barry. 1997. Quick Computation of Spatial
Autoregressive Estimators, forthcoming in Geographical Analysis.
Pace, R. Kelley. 1993. Nonparametric Methods with Applications to
Hedonic Models, Journal of Real Estate Finance and Economics Vol.
7, pp. 185-204.
83
Pace, R. Kelley, and O.W. Gilley. 1997. Using the Spatial Configuration of the Data to Improve Estimation, Journal of the Real Estate
Finance and Economics Vol. 14 pp. 333-340.
Pace, R. Kelley, and R. Barry. 1998. Simulating mixed regressive
spatially autoregressive estimators, Computational Statistics, Vol. 13
pp. 397-418.
Ripley, Brian D. 1988. Statistical Inference for Spatial Processes,
(Cambridge University Press: Cambridge, U.K.).
Chapter 3
Bayesian Spatial
autoregressive models
This chapter discusses spatial autoregressive models from a Bayesian perspective. It is well-known that Bayesian regression methods implemented
with diffuse prior information can replicate maximum likelihood estimation
results. We demonstrate this type of application, but focus on some extensions that are available with the Bayesian approach. The maximum likelihood estimation methods set forth in the previous chapter are based on the
presumption that the underlying disturbance process involved in generating
the model is normally distributed.
Further, many of the formal tests for spatial dependence and heterogeneity including those introduced in the previous chapter rely on characteristics
of quadratic forms for normal variates in order to derive the asymptotic distribution of the test statistics.
There is a history of Bayesian literature that deals with heteroscedastic and leptokurtic disturbances, treating these two phenomena in a similar
fashion. Geweke (1993) points out that a non-Bayesian regression methodology introduced by Lange, Little and Taylor (1989), which assume an independent Student-t distribution for the regression disturbances, is identical
to a Bayesian heteroscedastic linear regression model he introduces. Lange,
Little and Taylor (1989) show that their regression model provides robust
results in a wide range of applied data settings.
Geweke (1993) argues the same for his method and makes a connection to
the Bayesian treatment of symmetric leptokurtic disturbance distributions
through the use of scale mixtures of normal distributions. We adopt the
approach of Geweke (1993) in order to extend the spatial autoregressive
84
CHAPTER 3. BAYESIAN SPATIAL AUTOREGRESSIVE MODELS 85

models introduced in Chapter 2.
The extended version of the model is:
y = W1 y + X + u
(3.1)
u = W2 u +
N (0, 2V )
V
= diag(v1 , v2 , . . . , vn )
Where the change made to the basic model is in the assumption regarding
the disturbances . We assume that they exhibit non-constant variance,
taking on different values for every observation. The magnitudes vi , i =
1, .. . . . , n represent parameters to be estimated. This assumption of inherent
spatial heterogeneity seems more appropriate than the traditional GaussMarkov assumption that the variance of the disturbance terms is constant
over space.
The first section of this chapter introduces a Bayesian heteroscedastic
regression model and the topic of Gibbs sampling estimation without complications introduced by the spatial autoregressive model. The next section
applies these ideas to the simple FAR model and implements a Gibbs sampling estimation procedure for this model. Following sections deal with the
other spatial autoregressive models that we introduced in the previous chapter.
3.1
The Bayesian regression model
We consider the case of a heteroscedastic linear regression model with an

informative prior that can be written as in (3.2).
y = X +
N (0, V )
2
= diag(v1 , v2 , . . . , vn )
N (c, T )
(1/)
r/vi ID 2 (r)/r
r
(m, k)
(3.2)

Where y is an nx1 vector of dependent variables and X represents the nxk
matrix of explanatory variables. We assume that is an nx1 vector of normally distributed random variates with non-constant variance. We place
a normal prior on the parameters and a diffuse prior on . The relative variance terms (v1 , v2 , . . . , vn), are assumed fixed but unknown parameters that need to be estimated. The thought of estimating n parameters,
v1 , v2 , . . . , vn, in addition to the k+1 parameters, , using n data observations seems problematical. Bayesian methods dont encounter the same
degrees of freedom constraints, because we can rely on an informative prior
for these parameters. This prior distribution for the vi terms will take the
form of an independent 2 (r)/r distribution. Recall that the 2 distribution
is a single parameter distribution, where we have represented this parameter
as r. This allows us to estimate the additional n vi parameters in the model
by adding the single parameter r to our estimation procedure.
This type of prior has been used by Lindley (1971) for cell variances
in an analysis of variance problem, and Geweke (1993) in modeling heteroscedasticity and outliers. The specifics regarding the prior assigned to
the vi terms can be motivated by considering that the prior mean equals
unity and the variance of the prior is 2/r. This implies that as r becomes
very large, the terms vi will all approach unity, resulting in V = In , the
traditional Gauss-Markov assumption. We will see that the role of V 6= In
is to robustify against outliers and observations containing large variances
by downweighting these observations. Large r values are associated with a
prior belief that outliers and non-constant variances do not exist.
Now consider the posterior distribution from which we would derive our
estimates. Following the usual Bayesian methodology, we would combine
the likelihood function for our simple model with the prior distributions
for , and V to arrive at the posterior. There is little use in doing this
as we produce a complicated function that is not amenable to analysis. As
an alternative, consider the conditional distributions for the parameters ,
and V . These distributions are those that would arise from assuming each of
the other parameters were known. For example, the conditional distribution
for assuming that we knew and V would looks as follows:
|(, V ) N [H(X 0V 1 y + 2 R0 T 1 c) , 2 H].
0
H = (X V
X +RT
R)
(3.3)
Note that this is quite analogous to a generalized least-squares (GLS) version of the Theil and Goldberger (1961) estimation formulas, known as the

mixed estimator. Consider also that this would be fast and easy to compute.
Next consider the conditional distribution for the parameter assuming
that we knew the parameters and V in the problem. This distribution
would be:
[
n
X
(e2i /vi )/ 2 ]|(, V ) 2 (n)
(3.4)
i=1
yi x0i .
Where we let ei =
This result parallels the simple regression case
where we know that the residuals are 2 distributed. A difference from the
standard case is that we adjust the ei using the relative variance terms vi .
Finally, the conditional distribution for the parameters V represent a 2
distribution with r + 1 degrees of freedom (see Geweke, 1993):
[( 2 e2i + r)/vi]|(, ) 2 (r + 1)
(3.5)
The Gibbs sampler provides a way to sample from a multivariate posterior probability density of the type we encounter in this estimation problem
based only on the densities of subsets of vectors conditional on all others. In
other words, we can use the conditional distributions set forth above to produce estimates for our model, despite the fact that the posterior distribution
was not tractable.
The Gibbs sampling approach that we will use throughout this chapter
to estimate Bayesian variants of the spatial autoregressive models is based
on this simple idea. We specify the conditional distributions for all of the
parameters in the model and proceed to carry out random draws from these
distributions until we collect a large sample of parameter draws. Gelfand
and Smith (1990) demonstrate that Gibbs sampling from the sequence of
complete conditional distributions for all parameters in the model produces
a set of draws that converge in the limit to the true (joint) posterior distribution of the parameters. That is, despite the use of conditional distributions
in our sampling scheme, a large sample of the draws can be used to produce
valid posterior inferences about the mean and moments of the multivariate
posterior parameter distribution.
The method is most easily described by developing and implementing a
Gibbs sampler for our heteroscedastic Bayesian regression model. Given the
three conditional posterior densities in (3.3) through (3.5), we can formulate
a Gibbs sampler for this model using the following steps:
1. Begin with arbitrary values for the parameters 0 , 0 and vi0 which we
designate with the superscript 0.

2. Compute the mean and variance of using (3.3) conditional on the
initial values 0 and vi0 .
3. Use the computed mean and variance of to draw a multivariate
normal random vector, which we label 1 .
4. Calculate expression (3.4) using 1 determined in step 3 and use this
value along with a random 2 (n) draw to determine 1 .
5. Using 1 and 1 , calculate expression (3.5) and use the value along
with an nvector of random 2 (r + 1) draws to determine vi , i =
1, . . ., n.
These steps constitute a single pass of the Gibbs sampler. We wish
to make a large number of passes to build up a sample ( j , j , vij ) of j
values from which we can approximate the posterior distributions for our
parameters.
To illustrate this approach in practice, consider example 3.1 which shows
the MATLAB code for carry out estimation using the Gibbs sampler set
forth above. In this example, we generate a regression model data set that
contains a heteroscedastic set of disturbances based on a time trend variable.
Only the last 50 observations in the generated data sample contain nonconstant variances. This allows us to see if the estimated vi parameters
detect this pattern of non-constant variance over the last half of the sample.
The generated data set used values of unity for 0 , the intercept term,
and the two slope parameters, 1 and 2 . The prior means for the parameters were set to the true values of unity with prior variances of unity,
reflecting a fair amount of uncertainty. The following MATLAB program
implements the Gibbs sampler for this model. Note how easy it is to implement the mathematical equations in MATLAB code.
% ----- Example 3.1 Heteroscedastic Gibbs sampler
n=100; k=3; % set number of observations and variables
x = randn(n,k); b = ones(k,1); % generate data set
tt = ones(n,1); tt(51:100,1) = [1:50];
y = x*b + randn(n,1).*sqrt(tt); % heteroscedastic disturbances
ndraw = 1100; nomit = 100;
% set the number of draws
bsave = zeros(ndraw,k);
% allocate storage for results
ssave = zeros(ndraw,1);
vsave = zeros(ndraw,n);
c = [1.0 1.0 1.0];
% prior b means
R = eye(k); T = eye(k);
% prior b variance
Q = chol(inv(T)); q = Q*c;
b0 = x\y;
% use ols starting values

sige = (y-x*b0)*(y-x*b0)/(n-k);
V = ones(n,1); in = ones(n,1); % initial value for V
rval = 4;
% initial value for rval
qpq = Q*Q; qpv = Q*q;
% calculate QQ, Qq only once
tic;
% start timing
for i=1:ndraw;
% Start the sampling
ys = y.*sqrt(V); xs = matmul(x,sqrt(V));
xpxi = inv(xs*xs + sige*qpq);
b = xpxi*(xs*ys + sige*qpv); % update b
b = norm_rnd(sige*xpxi) + b; % draw MV normal mean(b), var(b)
bsave(i,:) = b;
% save b draws
e = ys - xs*b; ssr = e*e;
% update sige
chi = chis_rnd(1,n);
% do chisquared(n) draw
sige = ssr/chi; ssave(i,1) = sige; % save sige draws
chiv = chis_rnd(n,rval+1);
% update vi
vi = ((e.*e./sige) + in*rval)./chiv;
V = in./vi; vsave(i,:) = vi; % save the draw
end;
% End the sampling
toc;
% stop timing
bhat = mean(bsave(nomit+1:ndraw,:)); % calculate means and std deviations
bstd = std(bsave(nomit+1:ndraw,:)); tstat = bhat./bstd;
smean = mean(ssave(nomit+1:ndraw,1));
vmean = mean(vsave(nomit+1:ndraw,:));
tout = tdis_prb(tstat,n); % compute t-stat significance levels
% set up for printing results
in.cnames = strvcat(Coefficient,t-statistic,t-probability);
in.rnames = strvcat(Variable,variable 1,variable 2,variable 3);
in.fmt = %16.6f; tmp = [bhat tstat tout];
fprintf(1,Gibbs estimates \n); % print results
mprint(tmp,in);
fprintf(1,Sigma estimate = %16.8f \n,smean);
result = theil(y,x,c,R,T);
% compare to Theil-Goldberger estimates
prt(result); plot(vmean);
% plot vi-estimates
title(mean of vi-estimates);
We rely on MATLAB functions norm rnd and chis rnd to provide the
multivariate normal and chi-squared random draws. These functions are
part of the Econometrics Toolbox and are discussed in Chapter 8 of the
manual. Note also, we omit the first 100 draws at start-up to allow the
Gibbs sampler to achieve a steady state before we begin sampling for the
parameter distributions.
The results are shown below, where we find that it took only 11.5 seconds
to carry out the 1100 draws and produce a sample of 1000 draws on which
we can base our posterior inferences regarding the parameters and .
For comparison purposes, we produced estimates using the theil function
from the Econometrics Toolbox that implements mixed estimation. These

estimates are similar, but the tstatistics are lower because they suffer from
the heteroscedasticity. Our Gibbs sampled estimates take this into account
increasing the precision of the estimates.
elapsed_time =
11.5534
Gibbs estimates
Variable
Coefficient
t-statistic
variable 1
1.071951
2.784286
variable 2
1.238357
3.319429
variable 3
1.254292
3.770474
Sigma estimate =
10.58238737
t-probability
0.006417
0.001259
0.000276
Theil-Goldberger Regression Estimates

R-squared
=
0.2316
Rbar-squared =
0.2158
sigma^2
=
14.3266
Durbin-Watson =
1.7493
Nobs, Nvars
=
100,
3
***************************************************************
Variable
Prior Mean
Std Deviation
variable 1
1.000000
1.000000
variable 2
1.000000
1.000000
variable 3
1.000000
1.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
variable 1
1.056404
0.696458
0.487808
variable 2
1.325063
0.944376
0.347324
variable 3
1.293844
0.928027
0.355697
Figure 3.1 shows the mean of the 1,000 draws for the parameters vi
plotted for the 100 observation sample. Recall that the last 50 observations
contained a time-trend generated pattern of non-constant variance. This
pattern was detected quite accurately by the estimated vi terms.
One point that should be noted about Gibbs sampling estimation is
that convergence of the sampler needs to be diagnosed by the user. The
Econometrics Toolbox provides a set of convergence diagnostic functions
along with illustrations of their use in Chapter 5 of the manual. Fortunately,
for simple regression models (and spatial autoregressive models) convergence
of the sampler is usually a certainty, and convergence occurs quite rapidly.
A simple approach to testing for convergence is to run the sampler once
to carry out a small number of draws, say 300 to 500, and a second time
to carry out a larger number of draws, say 1000 to 2000. If the means and
variances for the posterior estimates are similar from both runs, convergence
seems assured.

mean of vi-estimates
4
3.5
2.5
1.5
10
20
30
40
50
60
70
80
90
100
Figure 3.1: Vi estimates from the Gibbs sampler
3.2
The Bayesian FAR model
In this section we turn attention to a Gibbs sampling approach for the

FAR model that can accommodate heteroscedastic disturbances and outliers. Note that the presence of a few spatial outliers due to enclave effects
or other aberrations in the spatial sample will produce a violation of normality in small samples. The distribution of disturbances will take on a
fat-tailed or leptokurtic shape. This is precisely the type of problem that
the heteroscedastic modeling approach of Geweke (1993) based on Gibbs
sampling estimation was designed to address.
The Bayesian extension of the FAR model takes the form:
y = W y +
N (0, V )
2
(3.6)

V
= diag(v1 , v2 , . . . , vn )
N (c, T )
r/vi ID 2 (r)/r
r
(m, k)
(0 , d0)
Where as in Chapter 2, the spatial contiguity matrix W has been standardized to have row sums of unity and the variable vector y is expressed
in deviations from the means to eliminate the constant term in the model.
We allow for an informative prior on the spatial autoregressive parameter
, the heteroscedastic control parameter r and the disturbance variance .
This is the most general Bayesian model, but practitioners would probably
implement diffuse priors for and .
A diffuse prior for would be implemented by setting the prior mean
c to zero and using a large prior variance for T , say 1e+12. To implement
a diffuse prior for we would set 0 = 0, d0 = 0. The prior for r is based
on a (m, k) distribution which has a mean equal to m/k and a variance
equal to m/k2 . Recall our discussion of the role of the prior hyperparameter
r in allowing the vi estimates to deviate from their prior means of unity.
Small values for r around 2 to 7 allow for non-constant variance and are
associated with a prior belief that outliers or non-constant variance exist.
Large values such as r = 20 or r = 50 would produce vi estimates that are all
close to unity, forcing the model to take on a homoscedastic character and
produce estimates equivalent to those from the maximum likelihood FAR
model discussed in Chapter 2. This would make little sense if we wished
to produce maximum likelihood estimates, it would be much quicker to use
the far function from Chapter 2. In the heteroscedastic regression model
demonstrated in example 3.1, we set r = 4 to allow ample opportunity for
the vi parameters to deviate from unity. Note that in example 3.1, the vi
estimates for the first 50 observations were all close to unity despite our
prior setting of r = 4. We will provide examples that suggests an optimal
strategy for setting r is to use small values in the range from 2 to 7. If the
sample data exhibits homoscedastic disturbances that are free from outliers,
the vi estimates will reflect this fact. On the other hand, if there is evidence
of heterogeneity in the errors, these settings for the hyperparameter r will
allow the vi estimates to deviate substantially from unity. Estimates for the
vi parameters that deviate from unity are needed to produce an adjustment
in the estimated and that take non-constant variance into account or
robustify our estimates in the presence of outliers.

Econometric estimation problems amenable to Gibbs sampling can take
one of two forms. The simplest case is where all of the conditional distributions are from well-known distributions allowing us to sample random
deviates using standard computational algorithms. This was the case with
our heteroscedastic Bayesian regression model.
A second more complicated case that one sometimes encounters in Gibbs
sampling is where one or more of the conditional distributions can be expressed mathematically, but they take an unknown form. It is still possible
to implement a Gibbs sampler for these models using a host of alternative methods that are available to produce draws from distributions taking
non-standard forms.
One of the more commonly used ways to deal with this situation is
known as the Metropolis algorithm. It turns out that the FAR model
falls into this latter category requiring us to rely on what is known as a
Metropolis-within-Gibbs sampler. To see how this problem arises, consider
the conditional distributions for the FAR model parameters where we rely
on diffuse priors, () and () for the parameters (, ) shown in (3.7).
() constant
(3.7)
() (1/), 0 < < +

These priors can be combined with the likelihood for this model producing
a joint posterior distribution for the parameters, p(, |y).
p(, |y) |In W | (n+1) exp{
1
(y W y)0 (y W y)}
2 2
(3.8)
If we treat as known, the kernel for the conditional posterior (that part
of the distribution that ignores inessential constants) for given takes the
form:
1 0
}
(3.9)
2 2
where = y W y. It is important to note that by conditioning on
(treating it as known) we can subsume the determinant, |In W |, as part
of the constant of proportionality, leaving us with one of the standard distributional forms. From (3.9) we conclude that 2 2 (n).
Unfortunately, the conditional distribution of given takes the following non-standard form:
p(|, y) (n+1) exp{
p(|, y) n/2 |In W |{(y W y)0(y W y)}n/2
(3.10)
To sample from (3.10) we can rely on a method called Metropolis sampling,

within the Gibbs sampling sequence, hence it is often labeled Metropoliswithin-Gibbs.
Metroplis sampling is described here for the case of a symmetric normal
candidate generating density. This should work well for the conditional
distribution of because, as Figure 3.2 shows, a conditional distribution of
is similar to a normal distribution with the same mean value. The figure
also shows a tdistribution with 3 degrees of freedom, which would also
work well in this application.
-150
x 10
0.5
0.1
0.2
0.3
0.4
0.5
0.6
conditional distribution of
0.7
0.8
0.9
3
2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
normal distribution
0.7
0.8
0.9
3
2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
t-distribution with 3 dof
0.7
0.8
0.9
Figure 3.2: Conditional distribution of

To describe Metropolis sampling in general, suppose we are interested in
sampling from a density f () and x0 denotes the current draw from f . Let
the candidate value be generated by y = x0 + cZ, where Z is a draw from
a standard normal distribution and c is a known constant. (If we wished

to rely on a tdistribution, we could simply replace Z with a random draw
from the tdistribution.)
An acceptance probability is computed using: p = min{1, f (y)/f (x0)}.
We then draw a uniform random deviate we label U , and if U < p, the next
draw from f is given by x1 = y. If on the other hand, U p, the draw is
taken to be the current value, x1 = x0 .
A MATLAB program to implement this approach for the case of the
homoscedastic first-order spatial autoregressive (FAR) model is shown in
example 3.2. We use this simple case as an introduction before turning to
the heteroscedastic case. Note also that we do not rely on the sparse matrix
algorithms which we will turn attention to after this introductory example.
An implementation issue is that we need to impose the restriction:
1/min < < 1/max
where min and max are the minimum and maximum eigenvalues of the
standardized spatial weight matrix W . We impose this restriction using an
approach that has been labeled rejection sampling. Restrictions such as
this, as well as non-linear restrictions, can be imposed on the parameters
during Gibbs sampling by simply rejecting values that do not meet the
restrictions (see Gelfand, Hills, Racine-Poon and Smith, 1990).
% ----- Example 3.2 Metropolis within Gibbs sampling FAR model
n=49; ndraw = 1100; nomit = 100; nadj = ndraw-nomit;
% generate data based on a given W-matrix
load wmat.dat; W = wmat; IN = eye(n); in = ones(n,1); weig = eig(W);
lmin = 1/min(weig); lmax = 1/max(weig); % bounds on rho
rho = 0.7;
% true value of rho
y = inv(IN-rho*W)*randn(n,1); ydev = y - mean(y); Wy = W*ydev;
% set starting values
rho = 0.5;
% starting value for the sampler
sige = 10.0; % starting value for the sampler
c = 0.5;
% for the Metropolis step (adjusted during sampling)
rsave = zeros(nadj,1); % storage for results
ssave = zeros(nadj,1); rtmp = zeros(nomit,1);
iter = 1; cnt = 0;
while (iter <= ndraw);
% start sampling;
e = ydev - rho*Wy; ssr = (e*e);
% update sige;
chi = chis_rnd(1,n); sige = (ssr/chi);
% metropolis step to get rho update
rhox = c_rho(rho,sige,ydev,W);
% c_rho evaluates conditional
rho2 = rho + c*randn(1); accept = 0;
while accept == 0;
% rejection bounds on rho
if ((rho2 > lmin) & (rho2 < lmax)); accept = 1; end;
rho2 = rho + c*randn(1); cnt = cnt+1;

end;
% end of rejection for rho
rhoy = c_rho(rho2,sige,ydev,W);
% c_rho evaluates conditional
ru = unif_rnd(1,0,1); ratio = rhoy/rhox; p = min(1,ratio);
if (ru < p)
rho = rho2;
end;
rtmp(iter,1) = rho;
if (iter >= nomit);
if iter == nomit
% update c based on initial draws
c = 2*std(rtmp(1:nomit,1));
end;
ssave(iter-nomit+1,1) = sige; rsave(iter-nomit+1,1) = rho;
end; % end of if iter > nomit
iter = iter+1;
end; % end of sampling loop
% print-out results
fprintf(1,hit rate = %6.4f \n,ndraw/cnt);
fprintf(1,mean and std of rho %6.3f %6.3f \n,mean(rsave),std(rsave));
fprintf(1,mean and std of sig %6.3f %6.3f \n,mean(ssave),std(ssave));
% maximum likelihood estimation for comparison
res = far(ydev,W);
prt(res);
Rejection sampling is implemented in the example with the following

code fragment that examines the candidate draws in rho2 to see if they
are in the feasible range. If rho2 is not in the feasible range, another
candidate value rho2 is drawn and we increment a counter variable cnt
to keep track of how many candidate values are found outside the feasible
range. The while loop continues to draw new candidate values and examine
whether they are in the feasible range until we find a candidate value within
the limits. Finding this value terminates the while loop. This approach
ensures than any values of that are ultimately accepted as draws will meet
the constraints.
% metropolis step to get rho update
rho2 = rho + c*randn(1); accept = 0;
while accept == 0;
% rejection bounds on rho
if ((rho2 > lmin) & (rho2 < lmax)); accept = 1; end;
rho2 = rho + c*randn(1); cnt = cnt+1;
end;
% end of rejection for rho
Another point to note about the example is that we adjust the parameter
c used to produce the random normal candidate values. This is done by
using the initial nomit=100 values of rho to compute a new value for c
based on two standard deviations of the initial draws. The following code

fragment carries this out, where the initial rho draws have been stored in
a vector rtmp.
if iter == nomit
% update c based on initial draws
c = 2*std(rtmp(1:nomit,1));
end;
Consider also, that we delay collecting our sample of draws for the parameters and until we have executed nomit burn-in draws, which is
100 in this case. This allows the sampler to settle into a steady state, which
might be required if poor values of and were used to initialize the sampler. In theory, any arbitrary values can be used, but a choice of good
values will speed up convergence of the sampler. A plot of the first 100
values drawn from this example is shown in Figure 3.3. We used = 0.5
and 2 = 100 as starting values despite our knowledge that the true values
were = 0.7 and 2 = 1. The plots of the first 100 values indicate that even
if we start with very poor values, far from the true values used to generate
the data, only a few iterations are required to reach a steady state. This is
usually true for regression-based Gibbs samplers.
The function c rho evaluates the conditional distribution for given 2
at any value of . Of course, we could use sparse matrix algorithms in this
function to handle large data sample problems, which is the way we approach
this task when constructing our function far g for the spatial econometrics
library.
function yout = c_rho(rho,sige,y,W)
% evaluates conditional distribution of rho
% given sige for the spatial autoregressive model
n = length(y);
IN = eye(n); B = IN - rho*W;
detval = log(det(B));
epe = y*B*B*y;
yout = (n/2)*log(sige) + (n/2)*log(epe) - detval;
Finally, we present results from executing the code shown in example 3.2,
where both Gibbs estimates based on the mean of the 1,000 draws for and
as well as the standard deviations are shown. For contrast, we present maximum likelihood estimates, which for the case of the homoscedastic Gibbs
sampler implemented here with a diffuse prior on and should produce
similar estimates.
Gibbs sampling estimates
hit rate = 0.3561

1.2
1
0.8
0.6
0.4
0.2
0
-0.2
10
20
30
40
50
60
first 100 draws for rho
70
80
90
100
10
20
30
40
50
60
first 100 draws for sige
70
80
90
100
4
3.5
3
2.5
2
1.5
1
0.5
Figure 3.3: First 100 Gibbs draws for and

mean and std of rho
mean and std of sig
0.649
1.348
0.128
0.312

R-squared
=
0.4067
sigma^2
=
1.2575
Nobs, Nvars
=
49,
1
log-likelihood =
-137.94653
# of iterations =
13
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.672436
4.298918
0.000084
The time needed to generate 1,100 draws was around 10 seconds, which
represents 100 draws per second. We will see that similar speed can be
achieved even for large data samples.
From the results we see that the mean and standard deviations from

the Gibbs sampler produce estimates very close to the maximum likelihood
estimates, and to the true values used to generate the model data. The
mean estimate for = 0.649 divided by the standard deviation of 0.128
implies a tstatistic of 5.070, which is very close to the maximum likelihood
tstatistic. The estimate of 2 = 1.348 based on the mean from 1,000 draws
is also very close to the true value of unity used to generate the model.
The reader should keep in mind that we do not advocate using the Gibbs
sampler in place of maximum likelihood estimation. That is, we dont really wish to implement a homoscedastic version of the FAR model Gibbs
sampler that relies on diffuse priors. We turn attention to the more general
heteroscedastic case that allows for either diffuse or informative priors in the
next section.
3.2.1
The far g() function
We discuss some of the implementation details concerned with constructing a MATLAB function far g to produce estimates for the Bayesian FAR
model. This function will rely on a sparse matrix algorithm approach to
handle problems involving large data samples. It will also allow for diffuse
or informative priors and handle the case of heterogeneity in the disturbance
variance.
The first thing we need to consider is that to produce a large number of
draws, say 1,000, we would need to evaluate the conditional distribution of
2,000 times. (Note that we called this function twice in example 3.2). Each
evaluation would require that we compute the determinant of the matrix
(In W ), which we have already seen is a non-trivial task for large data
samples. To avoid this, we rely on the Pace and Barry (1997) approach
discussed in the previous chapter. Recall that they suggested evaluating
this determinant over a grid of values in the feasible range of once at the
outset. Given that we have carried out this evaluation and stored the determinant values along with associated values, we can simply look-up
the appropriate determinant in our function that evaluates the conditional
distribution. That is, the call to the conditional distribution function will
provide a value of for which we need to evaluate the conditional distribution. If we already know the determinant for a grid of all feasible values,
we can simply look up the determinant value closest to the value and use
it during evaluation of the conditional distribution. This saves us the time
involved in computing the determinant twice for each draw of .
Since we need to carry out a large number of draws, this approach works
better than computing determinants for every draw. Note that in the case

of maximum likelihood estimation from Chapter 2, the opposite was true.
There we only needed 10 to 20 evaluations of the likelihood function, making
the initial grid calculation approach of Pace and Barry much slower.
Now we turn attention to the function far g that implements the Gibbs
sampler for the FAR model. The documentation for the function is shown
below. You can of course examine the code in the function, but it essentially
carries out the approach set forth in example 3.2, with modifications to allow
for the relative variance parameters vi and informative priors for , and r
in this extended version of the model.
PURPOSE: Gibbs sampling estimates of the 1st-order Spatial
model: y = rho*W*y + e,
e = N(0,sige*V),
V = diag(v1,v2,...vn), r/vi = ID chi(r)/r, r = Gamma(m,k)
rho = N(c,T), sige = gamma(nu,d0)
---------------------------------------------------------------USAGE: result = far_g(y,W,ndraw,nomit,prior,start)
where: y = nobs x 1 independent variable vector
W = nobs x nobs 1st-order contiguity matrix (standardized)
ndraw = # of draws
nomit = # of initial draws omitted for burn-in
prior = a structure variable for prior information input
prior.rho, prior mean for rho, c above, default = 0 (diffuse)
prior.rcov, prior rho variance, T above, default = 1e+12 (diffuse)
prior.nu,
informative Gamma(nu,d0) prior on sige
prior.d0
default: nu=0,d0=0 (diffuse prior)
prior.rval, r prior hyperparameter, default=4
prior.m,
informative Gamma(m,k) prior on r
prior.k,
default: not used
prior.rmin, (optional) min value of rho to use in sampling
prior.rmax, (optional) max value of rho to use in sampling
start = (optional) (2x1) vector of rho, sige starting values
(defaults, rho = 0.5, sige = 1.0)
--------------------------------------------------------------RETURNS: a structure:
results.meth
= far_g
results.pdraw = rho draws (ndraw-nomit x 1)
results.sdraw = sige draws (ndraw-nomit x 1)
results.vmean = mean of vi draws (1 x nobs)
results.yhat
= predicted values of y
results.rdraw = r-value draws (ndraw-nomit x 1)
results.pmean = rho prior mean
(if prior input)
results.pstd
= rho prior std dev (if prior input)
results.nu
= prior nu-value for sige (if prior input)
results.d0
= prior d0-value for sige (if prior input)
results.r
= value of hyperparameter r (if input)
results.m
= m prior parameter (if input)
results.k
= k prior parameter (if input)

results.nobs
= # of observations
results.ndraw = # of draws
results.nomit = # of initial draws omitted
results.y
= actual observations
results.yhat
= predicted values for y
results.time
= time taken for sampling
results.accept = acceptance rate
results.pflag = 1 for prior, 0 for no prior
results.rmax
= 1/max eigenvalue of W (or rmax if input)
results.rmin
= 1/min eigenvalue of W (or rmin if input)
---------------------------------------------------------------NOTE: use either improper prior.rval
or informative Gamma prior.m, prior.k, not both of them
----------------------------------------------------------------
3.2.2
Applied examples
As the documentation makes clear, there are a number of user options to

facilitate different models. Example 3.3 illustrates using the function with
various input options. We generate a FAR model vector y based on the standardized W weight matrix from the Columbus neighborhood crime data set.
The program then produces maximum likelihood estimates for comparison
to our Gibbs sampled estimates. The first set of Gibbs estimates are produced with a homoscedastic prior based on r = 30 and diffuse priors for
and . Diffuse priors for and are the defaults used by far g, and
the default for r equals 4. My experience indicates this represents a good
rule-of-thumb value.
After producing the first estimates, we add two outliers to the data set
at observations 10 and 39. We then compare maximum likelihood estimates
to the Gibbs sampled estimates based on a heteroscedastic prior with r = 4.
% ----- Example 3.3 Using the far_g function
load wmat.dat; % standardized 1st-order spatial weight matrix
W = wmat;
% from the Columbus neighborhood data set
[n junk] = size(W); IN = eye(n);
rho = 0.75;
% true value of rho
y = inv(IN-rho*W)*randn(n,1)*5; % generate data
ydev = y - mean(y);
vnames = strvcat(y-simulated,y-spatial lag);
rmin = 0; rmax = 1;
resml = far(ydev,W,rmin,rmax); % do maximum likelihood for comparison
prt(resml,vnames);
ndraw = 1100; nomit = 100;
prior.rval = 30; % homoscedastic prior diffuse rho,sigma (the default)
prior.rmin = 0; prior.rmax = 1;
result = far_g(ydev,W,ndraw,nomit,prior); % call Gibbs sampling function

prt(result,vnames);
% add outliers to the generated data
ydev(20,1) = ydev(20,1)*10; ydev(39,1) = ydev(39,1)*10;
prior.rval = 4; % heteroscedastic model, diffuse rho,sigma (the default)
resml2 = far(ydev,W); % do maximum likelihood for comparison
prt(resml2,vnames);
result2 = far_g(ydev,W,ndraw,nomit,prior); % call Gibbs sampling function
prt(result2,vnames);
% plot the mean of the vi-draws, which represent vi-estimates
plot(result2.vmean);
The program produced the following output. Note that our printing
function computes means and standard deviations using the draws returned
in the results structure of far g. We also compute tstatistics and evaluate
the marginal probabilities. This allows us to provide printed output in the
form of a traditional regression model.
% homoscedastic models
y-simulated
R-squared
=
0.5908
sigma^2
=
24.9531
Nobs, Nvars
=
49,
1
log-likelihood =
-285.86289
# of iterations =
9
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.771980
6.319186
0.000000
Gibbs sampling First-order spatial autoregressive model
y-simulated
R-squared
=
0.5805
sigma^2
=
25.2766
r-value
=
30
Nobs, Nvars
=
49,
1
ndraws,nomit
=
1100,
100
acceptance rate =
0.8886
time in secs
=
13.7984
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.739867
8.211548
0.000000
% outlier models
y-simulated

R-squared
=
0.0999
sigma^2
= 267.3453
Nobs, Nvars
=
49,
1
log-likelihood =
-398.06025
# of iterations =
14
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.368404
1.591690
0.118019
y-simulated
R-squared
=
0.1190
sigma^2
= 107.8131
r-value
=
3
Nobs, Nvars
=
49,
1
ndraws,nomit
=
1100,
100
acceptance rate =
0.8568
time in secs
=
8.2693
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.503992
3.190913
0.002501
The first two sets of output illustrate the point we made regarding
Bayesian analysis implemented with a diffuse prior. The results from the
Gibbs sampling approach are very close to those from maximum likelihood
estimation. Note that the printed output shows the time required to carry
out 1,100 draws along with the acceptance rate. It took only 13.7 seconds
to produce 1,100 draws.
After introducing two outliers, we see that the maximum likelihood estimates produce a poor fit to the data as well as an inflated estimate of 2 .
The coefficient estimate for is also affected adversely, deviating from the
true value of 0.75, and the precision of the estimate is degraded. In contrast,
the Gibbs sampled estimate of was closer to the true value and exhibits
greater precision as indicated by the larger tstatistic. The estimate for 2
is much smaller than that from maximum likelihood. Robust estimates will
generally exhibit a smaller R2 statistic as the estimates downweight outliers
rather than try to fit these data observations.
We also produced a plot of the mean of the vi draws which serve as an
estimate of these relative variance terms. This graph is shown in Figure 3.4,
where we see that the two outliers were identified.
Example 3.4 illustrates the use of the far g function on the large Pace
and Barry data set. We set an r value of 4 which will capture heterogeneity

12
10
mean of V draws
10
15
20
25
30
Observations
35
40
45
50
Figure 3.4: Mean of the vi draws

if it exists. We rely on the input options to set a minimum and maximum
value of between 0 and 1 over which to search, to speed up computation.
This avoids computation of the eigenvalues for the large matrix W which
would provide the range over which to search. If you find an estimate for
near zero, this restriction to the (0,1) interval is unwise.
% ----- Example 3.4 Using far_g with a large data set
load elect.dat;
y = (elect(:,7)./elect(:,8));
ydev = y - mean(y);
clear elect;
n = 3107;
prior.rval = 4; prior.rmin = 0; prior.rmax = 1;

ndraw = 1100; nomit = 100;
res = far_g(ydev,W,ndraw,nomit,prior);
prt(res);
plot(res.vmean);
xlabel(Observations);
ylabel(V_i estimates);
pause;
pltdens(res.pdraw,0.1,0,1);
We present maximum likelihood results for comparison with the Gibbs

sampling results. If there is no substantial heterogeneity in the disturbance,
the two sets of estimates should be similar, as we saw from example 3.3.
% Maximum likelihood results
R-squared
=
0.5375
sigma^2
=
0.0054
Nobs, Nvars
=
3107,
1
log-likelihood =
3506.3203
# of iterations =
13
min and max rho =
-1.0710,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.721474
59.567710
0.000000
% Gibbs sampling estimates
R-squared
=
0.5337
sigma^2
=
0.0052
r-value
=
4
Nobs, Nvars
=
3107,
1
ndraws,nomit
=
1100,
100
acceptance rate =
0.7131
time in secs
= 262.4728
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
rho
0.706526
47.180554
0.000000
From the results we see that the maximum likelihood and Bayesian robust estimates are very similar, suggesting a lack of heterogeneity. We can
further explore this issue by examining a plot of the mean vi draws, which
serve as estimates for these parameters in the model. Provided we use a
small value of r, the presence of heterogeneity and outliers will be indicated
by large vi estimates that deviate substantially from unity. Figure 3.5 shows
a plot of the mean of the vi draws, confirming that a handful of large vi values exist. Close inspection reveals that only 58 vi values greater than 3 exist

in a sample of 3107 observations. Apparently this amount of heterogeneity
does not affect the estimates for this model.
8
Mean of vi draws
500
1000
1500
2000
observations
2500
3000
3500
Figure 3.5: Mean of the vi draws for Pace and Barry data
An advantage of Gibbs sampling is that valid estimates of dispersion for
the parameters as well as the entire posterior distribution associated with
the estimated parameters are available. Recall that this presents a problem for large data sets estimated using maximum likelihood methods, which
we solved using a numerical hessian calculation. In the presence of outliers
or non-constant variance the numerical hessian approach may not be valid
because normality in the disturbance generating process might be violated.
In the case of Gibbs sampling, the law of large numbers suggests that we
can compute valid means and measures of dispersion from the sample of
draws. As an illustration, we use a function pltdens from the Econometrics
Toolbox to produce a non-parametric density estimate of the posterior distribution for . Figure 3.6 shows the posterior density that is plotted using
the command:

pltdens(res.pdraw,0.1,0,1);
An optional argument to the function allows a kernel smoothing parameter

to be input as indicated in the documentation for the function pltdens
shown below. The default kernel bandwidth produces fairly non-smooth
densities that tend to overfit the data, so we supply our own value.
PURPOSE: Draw a nonparametric density estimate.
--------------------------------------------------USAGE: [h f y] = pltdens(x,h,p,kernel)
or pltdens(x) which uses gaussian kernel default
where:
x is a vector
h is the kernel bandwidth
default=1.06 * std(x) * n^(-1/5); Silverman page 45
p is 1 if the density is 0 for negative values
k is the kernel type:
=1 Gaussian (default)
=2 Epanechnikov
=3 Biweight
=4 Triangular
A jittered plot of the
observations is shown below the density.
--------------------------------------------------RETURNS:
h = the interval used
f = the density
y = the domain of support
plot(y,f) will produce a plot of the density
--------------------------------------------------
The disadvantage of the Gibbs sampling estimation approach is the time

required. This is reported in the printed output which indicates that it took
262 seconds to produce 1,100 draws. This is relatively competitive with the
maximum likelihood estimation method that took around 100 seconds to
produce estimates.
3.3
Other spatial autoregressive models
It should perhaps be clear that implementation of Gibbs samplers for the

other spatial autoregressive models is quite straightforward. We need simply to determine the complete sequence of conditional distributions for the
parameters in the model and code a loop to carry out the draws. LeSage
(1997) sets forth an alternative approach to the Metropolis within Gibbs
25
20
15
10
0.66
0.68
0.7
0.72
0.74
0.76
0.78
Figure 3.6: Posterior distribution for

sampling in that article. There are many ways to generate samples from an
unknown conditional distribution and the ratio of uniforms approach set
forth in LeSage (1997) is another approach. My experience has convinced
me that the Metropolis approach set forth here is superior as it requires far
less time.
All of the spatial autoregressive models will have in common the need
to produce a Metropolis-within Gibbs estimate for based on a conditional
distribution involving the determinant (In W ). In the case of the SAC
model, we need two determinants, one for (In W1 ) and another for (In
W2 ). Of course we will carry this out initially over a grid of values and
store the results. These will be passed to the functions that perform the
conditional distribution calculations.
There are functions sar g, sem g and sac g that implement Gibbs sampling estimation for the Bayesian variants of the spatial autoregressive models. The documentation for sem g (which is similar to that for the other

models) is shown below:
PURPOSE: Gibbs sampling estimates of the heteroscedastic
spatial error model:
y = XB + u, u = lam*W + e
e is N(0,sige*V)
B = N(c,T), sige = gamma(nu,d0), lam = diffuse prior
--------------------------------------------------USAGE: results = sem_g(y,x,W,ndraw,nomit,prior,start)
where: y = dependent variable vector (nobs x 1)
W = 1st order contiguity matrix (standardized, row-sums = 1)
prior = a structure for: B = N(c,T), sige = gamma(nu,d0)
prior.beta, prior means for beta,
c above (default 0)
prior.bcov, prior beta covariance , T above (default 1e+12)
prior.m,
prior.k,
(default: not used)
prior.nu,
a prior parameter for sige
prior.d0,
(default: diffuse prior for sige)
prior.lmin, (optional) min value of lambda to use in sampling
prior.lmax, (optional) max value of lambda to use in sampling
ndraw = # of draws
start = (optional) structure containing starting values:
defaults: beta=ones(k,1),sige=1,rho=0.5, V=ones(n,1)
start.b
= beta starting values (nvar x 1)
start.lam = lam starting value
(scalar)
start.sig = sige starting value (scalar)
start.V
= V starting values (n x 1)
--------------------------------------------------RETURNS: a structure:
results.meth = sem_g
results.bdraw = bhat draws (ndraw-nomit x nvar)
results.pdraw = lam draws (ndraw-nomit x 1)
results.rdraw = r draws (ndraw-nomit x 1) (if m,k input)
results.bmean = b prior means, prior.beta from input
results.bstd = b prior std deviations sqrt(diag(prior.bcov))
results.r
results.nvar = # of variables in x-matrix
results.y
= actual observations (nobs x 1)
results.yhat = predicted values
results.nu
= nu prior parameter

results.d0
= d0 prior parameter
results.time = time taken for sampling
results.accept= acceptance rate
results.lmax = 1/max eigenvalue of W (or lmax if input)
results.lmin = 1/min eigenvalue of W (or lmin if input)
As the other functions are quite similar, we leave it to the reader to examine the documentation and demonstration files for these functions. One
point to note regarding use of these functions is that two options exist for
specifying a value for the hyperparameter r. The default is to rely on an improper prior based on r = 4. The other option allows a proper Gamma(m,k)
prior to be assigned for r.
The first option has the virtue that convergence will be quicker and less
draws are required to produce estimates. The drawback is that the estimates
are conditional on the single value of r set for the hyperparameter. The
second approach produces draws from a Gamma(m,k) distribution for r on
each pass through the sampler. This produces estimates that average over
alternative r values, in essence integrating over this parameter, resulting in
unconditional estimates.
My experience is that estimates produced with a Gamma(8,2) prior having a mean of r = 4 and variance of 2 are quite similar to those based on an
improper prior with r = 4. Use of the Gamma prior tends to require a larger
number of draws based on convergence diagnostics routines implemented by
the function coda from the Econometrics Toolbox described in Chapter 5
of the manual.
3.4
Applied examples
We turn attention to some applications involving the use of these models.

First we present example 3.5 that generates SEM models for a set of
parameters ranging from 0.1 to 0.9, based on the spatial weight matrix from
the Columbus neighborhood crime data set. Both maximum likelihood and
Gibbs estimates are produced by the program and a table is printed out to
compare the estimation results. The hyperparameter r was set to 30 in this
example, which should produce estimates similar to the maximum likelihood
results.
During the loop over alternative data sets, we recover the estimates and
other information we are interested in from the results structures. These are
stored in a matrix that we print using the mprint function to add column
and row labels.

% ----- Example 3.5 Using the sem_g function
load wmat.dat;
y = anselin(:,1); n = length(y);
x = [ones(n,1) anselin(:,2:3)];
W = wmat; IN = eye(n);
tt = ones(n,1); tt(25:n,1) = [1:25];
rvec = 0.1:.1:.9; b = ones(3,1);
nr = length(rvec); results = zeros(nr,6);
ndraw = 1100; nomit = 100; prior.rval = 30; bsave = zeros(nr,6);
for i=1:nr, rho = rvec(i);
u = (inv(IN-rho*W))*randn(n,1);
y = x*b + u;
% do maximum likelihood for comparison
resml = sem(y,x,W); prt(resml);
results(i,1) = resml.lam;
results(i,2) = resml.tstat(4,1);
bsave(i,1:3) = resml.beta;
% call Gibbs sampling function
result = sem_g(y,x,W,ndraw,nomit,prior); prt(result);
results(i,3) = mean(result.pdraw);
results(i,4) = results(i,3)/std(result.pdraw);
results(i,5) = result.time;
results(i,6) = result.accept;
bsave(i,4:6) = mean(result.bdraw);
end;
in.rnames = strvcat(True lam,0.1,0.2,0.3, ...
0.4,0.5,0.6,0.7,0.8,0.9);
in.cnames = strvcat(ML lam,lam t,Gibbs lam,lam t, ...
time,accept);
mprint(results,in);
in2.cnames = strvcat(b1 ML,b2 ML,b3 ML,...
b1 Gibbs,b2 Gibbs,b3 Gibbs);
mprint(bsave,in2);
From the results we see that 1100 draws took around 13 seconds. The
acceptance rate falls slightly for values of near 0.9, which we would expect.
Since this is close to the upper limit of unity, we will see an increase in
rejections of candidate values for that lie outside the feasible range.
The estimates are reasonably similar to the maximum likelihood results
even for the relatively small number of draws used. In addition to presenting estimates for , we also provide estimates for the parameters in
the problem.
% sem model demonstration
True lam
ML lam
lam t Gibbs lam
lam t
time
accept

0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
-0.6357
0.0598
0.3195
0.2691
0.5399
0.7914
0.5471
0.7457
0.8829
b1 ML
1.0570
1.2913
0.7910
1.5863
1.2081
0.3893
1.2487
1.8094
-0.9454
-3.1334
0.3057
1.8975
1.5397
4.0460
10.2477
4.1417
8.3707
17.5062
b2 ML
1.0085
1.0100
1.0298
0.9343
0.9966
1.0005
0.9584
1.0319
1.0164
-0.5049
0.0881
0.3108
0.2091
0.5141
0.7466
0.5303
0.7093
0.8539
b3 ML
0.9988
1.0010
0.9948
1.0141
1.0065
1.0155
1.0021
0.9920
0.9925
-2.5119
0.4296
1.9403
1.1509
3.6204
7.4634
3.8507
6.7513
14.6300
b1 Gibbs
1.0645
1.2684
0.7876
1.5941
1.2250
0.3529
1.3191
1.8366
-0.9783
12.9968
12.8967
13.0212
12.9347
13.1345
13.3044
13.2014
13.5251
13.7529
b2 Gibbs
1.0112
1.0121
1.0310
0.9283
0.9980
1.0005
0.9544
1.0348
1.0158
0.4475
0.4898
0.4855
0.4827
0.4770
0.4616
0.4827
0.4609
0.4349
b3 Gibbs
0.9980
1.0005
0.9936
1.0157
1.0058
1.0171
1.0029
0.9918
0.9923
As another example, we use a generated data set where we insert 2

outliers. The program generates a vector y based on the Columbus neighborhood crime data set and then adjusts two of the generated y values for
observations 10 and 40 to create outliers.
Example 3.6 produces maximum likelihood and two sets of Bayesian
SAR model estimates. One Bayesian model uses a homoscedastic prior and
the other sets r = 4, creating a a heteroscedastic prior. This is to illustrate
that the differences in the parameter estimates are due to robustification,
not the Gibbs sampling approach to estimation. A point to consider is that
maximum likelihood estimates of precision based on the information matrix
rely on normality, which is violated by the existence of outliers. These
create a disturbance distribution that contains fatter tails than the normal
distribution, not unlike the tdistribution. In fact, this is the motivation for
Gewekes approach to robustifying against outliers. Gibbs estimates based
on a heteroscedastic prior dont rely on normality. If you find a difference
between the estimates of precision from maximum likelihood estimates and
Bayesian Gibbs estimates, it is a good indication that outliers may exist.
% ----- Example 3.6 An outlier example
load anselin.dat; load wmat.dat;
x = [anselin(:,2:3)]; [n k] = size(x); x = [ones(n,1) x];
W = wmat; IN = eye(n);
rho = 0.5;
% true value of rho

b = ones(k+1,1); % true value of beta
Winv = inv(IN-rho*W);
y = Winv*x*b + Winv*randn(n,1);
vnames = strvcat(y-simulated,constant,income,house value);
% insert outliers
y(10,1) = y(10,1)*2; y(40,1) = y(40,1)*2;
% do maximum likelihood for comparison
resml = sar(y,x,W); prt(resml,vnames);
ndraw = 1100; nomit = 100;
prior.rval = 100; % homoscedastic model,
resg = sar_g(y,x,W,ndraw,nomit,prior);
prt(resg,vnames);
prior.rval = 4; % heteroscedastic model,
resg2 = sar_g(y,x,W,ndraw,nomit,prior);
prt(resg2,vnames);
% plot the vi-estimates
plot(resg2.vmean);
ylabel(mean of V_i draws);
The maximum likelihood SAR estimates along with the two sets of
Bayesian model estimates are shown below. We see that the homoscedastic
Gibbs estimates are similar to the maximum likelihood estimates, demonstrating that the Gibbs sampling estimation procedure is not responsible
for the difference in estimates we see between maximum likelihood and the
heteroscedastic Bayesian model. For the case of the heteroscedastic prior
we see much better estimates for both and . Note that the Rsquared
statistic is lower for the robust estimates which will be the case because
robustification requires that we not attempt to fit the outlying observations. This will generally lead to a worse fit for models that produce robust
estimates.
y-simulated
R-squared
=
0.6779
Rbar-squared
=
0.6639
sigma^2
= 680.8004
Nobs, Nvars
=
49,
3
log-likelihood =
-212.59468
# of iterations =
14
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
19.062669
1.219474
0.228880
income
-0.279572
-0.364364
0.717256
house value
1.966962
8.214406
0.000000

rho
0.200210
1.595514
0.117446
Gibbs sampling spatial autoregressive model

y-simulated
R-squared
=
0.6770
sigma^2
= 1077.9188
r-value
=
100
Nobs, Nvars
=
49,
3
ndraws,nomit
=
1100,
100
acceptance rate =
0.9982
time in secs
=
28.9612
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000
1000000.000000
income
0.000000
1000000.000000
house value
0.000000
1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
18.496993
1.079346
0.286060
income
-0.123720
-0.142982
0.886929
house value
1.853332
10.940066
0.000000
rho
0.219656
1.589213
0.118863
y-simulated
R-squared
=
0.6050
sigma^2
= 1374.8421
r-value
=
3
Nobs, Nvars
=
49,
3
ndraws,nomit
=
1100,
100
acceptance rate =
0.9735
time in secs
=
17.2292
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000
1000000.000000
income
0.000000
1000000.000000
house value
0.000000
1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
13.673728
0.778067
0.440513
income
0.988163
0.846761
0.401512
house value
1.133790
3.976679
0.000245
rho
0.388923
2.844965
0.006610
One point that may be of practical importance is that using large values

for the hyperparameter r slows down the Gibbs sampling process. This is
because the chi-squared random draws take longer for large r values. This
shows up in this example where the 1100 draws for the homoscedastic prior
based on r = 100 took close to 29 seconds and those for the model based on
r = 4 took only 17.2 seconds. This suggests that a good operational strategy
would be not to rely on values of r greater than 30 or 40. These values
may produce vi estimates that deviate from unity somewhat, but should in
most cases replicate the maximum likelihood estimates when there are no
outliers.
A better strategy is to always rely on a small r value between 2 and 8,
in which case a divergence between the maximum likelihood estimates and
those from the Bayesian model reflect the existence of non-constant variance
or outliers.
3.5
An applied exercise
We applied the series of spatial autoregressive models from Section 2.5 as

well as corresponding Bayesian spatial autoregressive models to the Boston
data set. Recall that Belsley, Kuh and Welsch (1980) used this data set to
illustrate the impact of outliers and influential observations on least-squares
estimation results. Here we have an opportunity to see the how the Bayesian
spatial autoregressive models deal with the outliers.
Example 3.7 shows the program code needed to implement both maximum likelihood and Bayesian models for this data set.
% ----- Example 3.7 Robust Boston model estimation
% other variables
rmin = 0; rmax = 1;
tic; res1 = sar(ys,xs,W,rmin,rmax); prt(res1,vnames); toc;
prior.rval = 4;
ndraw = 1100; nomit=100;
tic; resg1 = sar_g(ys,xs,W,ndraw,nomit,prior);
prt(resg1,vnames); toc;
tic; res2 = sem(ys,xs,W,rmin,rmax); prt(res2,vnames); toc;

tic; resg2 = sem_g(ys,xs,W,ndraw,nomit,prior);
tic; res3 = sac(ys,xs,W,W);
prt(res3,vnames); toc;
tic; resg3 = sac_g(ys,xs,W,W,ndraw,nomit,prior);
An interesting aspect is the timing results which were produced using

the MATLAB tic and toc commands. Maximum likelihood estimation
of the SAR model took 44 seconds while Gibbs sampling using 1100 draws
and omitting the first 100 took 124 seconds. For the SEM model the corresponding times were 59 and 164 seconds and for the SAC model 114 and
265 seconds. These times seem quite reasonable for this moderately sized
problem.
The results are shown below. (We eliminated the printed output showing
the prior means and standard deviations because all of the Bayesian models
were implemented with diffuse priors for the parameters in the model.) A
prior value of r = 4 was used to produce robustification against outliers and
non-constant variance.
What do we learn from this exercise? First, the parameters and for
the SAR and SEM models from maximum likelihood and Bayesian models
are in agreement regarding both the magnitude and significance. For the
SAC model, the Bayesian estimate for is in agreement with the maximum
likelihood estimate, but that for is not. The Bayesian estimate is 0.107
versus 0.188 for maximum likelihood. The maximum likelihood estimate is
significant whereas the Bayesian estimate is not. This would impact on our
decision regarding which model represents the best specification.
Most of the estimates are remarkably similar with one notable exception, that for the noxsq pollution variable. In all three Bayesian models this
estimate is smaller than the maximum likelihood estimate and insignificant
at the 95% level. All maximum likelihood estimates indicate significance.
This would represent an important policy difference in the inference made
regarding the impact of air pollution on housing values.
Spatial autoregressive Model Estimates (elapsed_time = 44.2218)
hprice
R-squared
=
0.8421
Rbar-squared
=
0.8383
sigma^2
=
0.1576
Nobs, Nvars
=
506,
13
log-likelihood =
-85.099051
# of iterations =
9
min and max rho =
0.0000,
1.0000
***************************************************************

Variable
crime
zoning
industry
charlesr
noxsq
rooms2
houseage
distance
access
taxrate
pupil/teacher
blackpop
lowclass
rho
Coefficient
-0.165349
0.080662
0.044302
0.017156
-0.129635
0.160858
0.018530
-0.215249
0.272237
-0.221229
-0.102405
0.077511
-0.337633
0.450871
t-statistic
-6.888522
3.009110
1.255260
0.918665
-3.433659
6.547560
0.595675
-6.103520
5.625288
-4.165999
-4.088484
3.772044
-10.149809
12.348363
t-probability
0.000000
0.002754
0.209979
0.358720
0.000646
0.000000
0.551666
0.000000
0.000000
0.000037
0.000051
0.000182
0.000000
0.000000
Gibbs sampling spatial autoregressive model (elapsed_time = 126.4168)

hprice
R-squared
=
0.8338
sigma^2
=
0.1812
r-value
=
4
Nobs, Nvars
=
506,
13
ndraws,nomit
=
1100,
100
acceptance rate =
0.9910
time in secs
= 110.2646
min and max rho =
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
crime
-0.127092
-3.308035
0.001008
zoning
0.057234
1.467007
0.143012
industry
0.045240
0.950932
0.342105
charlesr
0.006076
0.249110
0.803379
noxsq
-0.071410
-1.512866
0.130954
rooms2
0.257551
5.794703
0.000000
houseage
-0.031992
-0.748441
0.454551
distance
-0.171671
-5.806800
0.000000
access
0.173901
2.600740
0.009582
taxrate
-0.202977
-5.364128
0.000000
pupil/teacher
-0.086710
-3.081886
0.002172
blackpop
0.094987
3.658802
0.000281
lowclass
-0.257394
-5.944543
0.000000
rho
0.479126
5.697191
0.000000
(elapsed_time = 59.0996)
hprice
R-squared
=
0.8708
Rbar-squared
=
0.8676
sigma^2
=
0.1290

log-likelihood =
-58.604971
Nobs, Nvars
=
506,
13
# iterations
=
10
min and max lam =
0.0000,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.186710
-8.439402
0.000000
zoning
0.056418
1.820113
0.069348
industry
-0.000172
-0.003579
0.997146
charlesr
-0.014515
-0.678562
0.497734
noxsq
-0.220228
-3.683553
0.000255
rooms2
0.198585
8.325187
0.000000
houseage
-0.065056
-1.744224
0.081743
distance
-0.224595
-3.421361
0.000675
access
0.352244
5.448380
0.000000
taxrate
-0.257567
-4.527055
0.000008
pupil/teacher
-0.122363
-3.839952
0.000139
blackpop
0.129036
4.802657
0.000002
lowclass
-0.380295
-10.625978
0.000000
lambda
0.757669
19.133467
0.000000
Gibbs sampling spatial error model (elapsed_time = 164.8779)
hprice
R-squared
=
0.7313
sigma^2
=
0.1442
r-value
=
4
Nobs, Nvars
=
506,
13
ndraws,nomit
=
1100,
100
acceptance rate
=
0.4715
time in secs
= 116.2418
min and max lambda =
-1.9826,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
crime
-0.165360
-3.967705
0.000083
zoning
0.048894
1.226830
0.220472
industry
-0.002985
-0.051465
0.958976
charlesr
-0.014862
-0.538184
0.590693
noxsq
-0.145616
-1.879196
0.060807
rooms2
0.339991
7.962844
0.000000
houseage
-0.130692
-2.765320
0.005900
distance
-0.175513
-2.398220
0.016846
access
0.276588
3.121642
0.001904
taxrate
-0.234511
-4.791976
0.000002
pupil/teacher
-0.085891
-2.899236
0.003908
blackpop
0.144119
4.773623
0.000002
lowclass
-0.241751
-5.931583
0.000000
lambda
0.788149
19.750640
0.000000

General Spatial Model Estimates (elapsed_time = 114.5267)
hprice
R-squared
=
0.8662
Rbar-squared =
0.8630
sigma^2
=
0.1335
log-likelihood =
-55.200525
Nobs, Nvars
=
506,
13
# iterations =
7
***************************************************************
Variable
Coefficient
t-statistic
t-probability
crime
-0.198184
-8.766862
0.000000
zoning
0.086579
2.824768
0.004923
industry
0.026961
0.585884
0.558222
charlesr
-0.004154
-0.194727
0.845687
noxsq
-0.184557
-3.322769
0.000958
rooms2
0.208631
8.573808
0.000000
houseage
-0.049980
-1.337513
0.181672
distance
-0.283474
-5.147088
0.000000
access
0.335479
5.502331
0.000000
taxrate
-0.257478
-4.533481
0.000007
pupil/teacher
-0.120775
-3.974717
0.000081
blackpop
0.126116
4.768082
0.000002
lowclass
-0.374514
-10.707764
0.000000
rho
0.625963
9.519920
0.000000
lambda
0.188257
3.059010
0.002342
Gibbs sampling general spatial model (elapsed_time = 270.8657)
hprice
R-squared
=
0.7836
sigma^2
=
0.1487
r-value
=
4
Nobs, Nvars
=
506,
13
ndraws,nomit
=
1100,
100
accept rho rate
=
0.8054
accept lam rate
=
0.9985
time in secs
= 205.6773
min and max rho
=
0.0000,
1.0000
-1.9826,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
crime
-0.161673
-3.905439
0.000107
zoning
0.047727
1.274708
0.203013
industry
0.024687
0.433283
0.664999
charlesr
0.008255
0.319987
0.749114
noxsq
-0.121782
-1.814226
0.070251
rooms2
0.333823
7.332684
0.000000
houseage
-0.098357
-2.080260
0.038018
distance
-0.193060
-3.798247
0.000164

access
taxrate
pupil/teacher
blackpop
lowclass
rho
lambda
0.227007
-0.231393
-0.110537
0.137065
-0.293952
0.689346
0.107479
2.592018
-4.713901
-3.852098
4.555835
-7.201166
7.938580
1.188044
0.009825
0.000003
0.000133
0.000007
0.000000
0.000000
0.235388
One other variable where we would draw a different inference is the

houseage variable in both the SEM and SAC models. The Bayesian estimates indicate significance for this variable whereas maximum likelihood
estimates do not.
Another interesting difference between the Bayesian models and maximum likelihood is the lower R2 statistics for the Bayesian versus corresponding maximum likelihood estimates. The Bayesian SEM and SAC models
both show a dramatic reduction in fit indicative that a substantial amount
of robustification is taking place. The Bayesian SAR model shows only a
modest reduction in fit when compared to the maximum likelihood estimates. Recall that we rejected the SAR model in Chapter 2 for a number
of reasons.
How do we decide between the maximum likelihood and Bayesian estimates? One issue we should explore is that of outliers and non-constant
variance. A plot of the vi estimates based on the mean of the draws from
the Bayesian SAC model is shown in Figure 3.7. All three Bayesian models
produced very similar estimates for the vi terms as shown for the SAC model
in Figure 3.7.
Given these estimates for the vi terms, it would be hard to maintain the
hypothesis of a constant variance normally distributed disturbance process
for this model and data set.
Choosing between the Bayesian SEM and SAC models is not necessary
as a similar set of inferences regarding the significance of the noxsq air
pollution variable and houseage are produced by both models. Note that
these are different inferences than one would draw from maximum likelihood
estimates for either the SEM or SAC models.
If we wished to pursue this point we might examine the posterior distribution of from the SAC model, which is shown in Figure 3.8.
Suppose we felt comfortable imposing the restriction that must be
positive? We can do this with the Bayesian SAC model, whereas this is not
possible (given our optimization procedures) for the maximum likelihood
estimation procedure. This can be accomplished using an input option to
the sac g function as shown in example 3.8. Note that we increased the

12
10
Vi estimates
100
200
300
400
observations --- census tracts
500
600
Figure 3.7: Vi estimates for the Boston data

number of draws to 2100 in example 3.8 to increase the precision regarding
this models estimates.
% ----- Example 3.8 Imposing restrictions
% other variables
prior.lmin = 0; prior.lmax = 1;
prior.rval = 4; ndraw = 2100; nomit=100;
resg3 = sac_g(ys,xs,W,W,ndraw,nomit,prior);
prt(resg3,vnames);
posterior density
-0.3
-0.2
-0.1
0.1
0.2
lambda values
0.3
0.4
0.5
Figure 3.8: Posterior distribution of

Carrying out estimation of this version of the model, we found the following results. The estimate for the parameter is now perhaps different
from zero as indicated by a tstatistic that is significant at the 0.10 level.
Gibbs sampling general spatial model
hprice
R-squared
=
0.7891
sigma^2
=
0.1491
r-value
=
4
Nobs, Nvars
=
506,
13
ndraws,nomit
=
2100,
100
accept rho rate
=
0.9464
accept lam rate
=
0.6243
time in secs
= 403.2714
min and max rho
=
-1.9826,
1.0000
0.0000,
1.0000
***************************************************************

Posterior Estimates
Variable
Coefficient
crime
-0.164682
zoning
0.048255
industry
0.018722
charlesr
0.009160
noxsq
-0.123344
rooms2
0.328860
houseage
-0.096837
distance
-0.192698
access
0.232155
taxrate
-0.233451
pupil/teacher
-0.109920
blackpop
0.136174
lowclass
-0.297580
rho
0.671345
lambda
0.134279
t-statistic
-4.096886
1.252200
0.334302
0.360944
-1.915048
7.352177
-2.045083
-3.829066
2.771251
-5.075241
-3.730565
4.462714
-7.136230
7.891604
1.700944
t-probability
0.000049
0.211091
0.738294
0.718296
0.056065
0.000000
0.041377
0.000145
0.005795
0.000001
0.000213
0.000010
0.000000
0.000000
0.089584
A graphical examination of the posterior density for this parameter

shown in Figure 3.9 might make us a bit concerned about the nature of
the restriction we imposed on .
A virtue of Gibbs sampling is that we can compute the probability of
< 0 by simply counting the Gibbs draws for this parameter that are less
than zero in the unrestricted model. This can be done easily as shown in
example 3.9. To improve the accuracy of our calculation, we increased the
number of draws to 2100 in this program.
% ----- Example 3.9 The probability of negative lambda
% other variables
prior.rval = 4; ndraw = 2100; nomit=100;
resg3 = sac_g(ys,xs,W,W,ndraw,nomit,prior);
prt(resg3,vnames);
% find the number of lambda draws < 0
nlam = find(resg3.ldraw < 0); numl = length(nlam);
fprintf(1,The # of negative lambda values is: %5d \n,length(nlam));
fprintf(1,Probability lambda < 0 is: %6.2f \n,numl/(ndraw-nomit));
The results indicate a probability of 6%, which should not make us nervous about imposing the restriction on .
Posterior density
-0.1
0.1
0.2
lambda values
0.3
0.4
Figure 3.9: Truncated posterior distribution of

The # of negative lambda values is:
Probability lambda < 0 is:
0.06
126
Summarizing, I would use the Gibbs SAC model based on the restriction
that > 0. With the exception of , this would produce the same inferences
regarding all parameters as the model without this restriction. It would also
produce the same inferences regarding the parameters as the SEM model,
which is comforting.
3.6
Chapter Summary
We have seen that spatial autoregressive models can be extended to allow

for Bayesian prior information as well as non-constant variances over space.
These models require a Gibbs sampling estimation approach, making them
take more time than the maximum likelihood estimation methods. The time

is quite reasonable, even for large sample problems because we rely on the
MATLAB sparse matrix algorithms and a grid-based approach to compute
determinants that we use in the sampling process. In fact, as the sample
size gets larger, the time difference between maximum likelihood and Gibbs
sampling methods diminishes.
An advantage of these models is that they can serve as a check on the assumption of homogeneity that is inherent in the maximum likelihood models.
It should be noted that some work has been done on accommodating nonconstant variance in the case of maximum likelihood methods (see Anselin,
1988). Unfortunately, these approaches require that the investigator add a
specification for the changing variance over space. This adds to the specification problems facing the practitioner, whereas the Bayesian approach set
forth here requires no such specification. Outliers and non-constant variance
are automatically detected during estimation and the estimates are adjusted
for these problems.
We saw in an application from Section 3.5 an example where the existence of outliers produced different inferences from maximum likelihood and
Bayesian robust estimates. These differences would have important policy
implications for the conclusions one were to draw regarding the impact of
air quality on housing values. These Bayesian models would produce different inferences than maximum likelihood regarding two of the explanatory
variables noxsq and houseage and for the SAC model there is a difference
regarding the significance of the parameter .
3.7
References
Casella, G. and E.I. George. 1992. Explaining the Gibbs Sampler,
American Statistician, Vol. 46, pp. 167-174.
Gelfand, Alan E., and A.F.M Smith. 1990. Sampling-Based Approaches to Calculating Marginal Densities, Journal of the American
Statistical Association, Vol. 85, pp. 398-409.
Gelfand, Alan E., Susan E. Hills, Amy Racine-Poon and Adrian F.M.
Smith. 1990. Illustration of Bayesian Inference in Normal Data Models Using Gibbs Sampling, Journal of the American Statistical Association, Vol. 85, pp. 972-985.

Gelman, Andrew, John B. Carlin, Hal S. Stern, and Donald B. Rubin.
1995. Bayesian Data Analysis, (London: Chapman & Hall).
Geweke, John. 1993. Bayesian Treatment of the Independent Student t Linear Model, Journal of Applied Econometrics, Vol. 8, s19s40.
Gilks, W.R., S. Richardson and D.J. Spiegelhalter. 1996. Markov
Chain Monte Carlo in Practice, (London: Chapman & Hall).
Lange, K.L., R.J.A. Little, and J.M.G. Taylor (1989). Robust Statistical Modeling Using the t Distribution, Journal of the American
Statistical Association, 84, pp. 881-896.
LeSage, James P. 1997. Bayesian Estimation of Spatial Autoregressive Models, International Regional Science Review, 1997 Vol. 20,
number 1&2, pp. 113-129. Also available at www.econ.utoledo.edu.
Lindley, David V. 1971. The estimation of many parameters, in
Foundations of Statistical Science, V.P. Godambe and D.A. Sprout
(eds.) (Toronto: Holt, Rinehart, and Winston).
Theil, Henri and Arthur S. Goldberger. 1961. On Pure and Mixed
Statistical Estimation in Economics, International Economic Review,
Vol. 2, pp. 65-78.
Zellner, Arnold. 1984. Basic Issues in Econometrics, (Chicago: University of Chicago Press), Chapter 3.
Chapter 4
Locally linear spatial models

This chapter discusses in detail a set of estimation methods that attempt
to accommodate spatial heterogeneity by allowing the parameters of the
model to vary with the spatial location of the sample data. The first section deals with spatial and distance expansion models introduced by Casetti
(1972,1992). A more recent variant of this model presented in Casetti (1982)
and Casetti and Can (1998) called a DARP model is the subject of Section 4.2.
Non-parametric locally linear regression models introduced in McMillen
(1996), McMillen and McDonald (1997) and Brunsdon, Fotheringham and
Charlton (1997) (sometimes labeled geographically weighted regression) represent another way to deal with spatial heterogeneity. These models are covered in Section 4.3. Finally, a Bayesian approach to geographically weighted
regressions is presented in Section 4.4.
4.1
Spatial expansion
The first model of this type was introduced by Casetti (1972) and labeled
a spatial expansion model. The model is shown in (4.1), where y denotes
an nx1 dependent variable vector associated with spatial observations and
X is an nxnk matrix consisting of terms xi representing kx1 explanatory
variable vectors, as shown in (4.2). The locational information is recorded
in the matrix Z which has elements Zxi , Zyi , i = 1, . . ., n, that represent
latitude and longitude coordinates of each observation as shown in (4.2).
The model posits that the parameters vary as a function of the latitude
and longitude coordinates. The only parameters that need be estimated
are the parameters in 0 that we denote, x , y . These represent a set of
127
CHAPTER 4. LOCALLY LINEAR SPATIAL MODELS
128
2k parameters. Recall our discussion about spatial heterogeneity and the

need to utilize a parsimonious specification for variation over space. This
represents one approach to this type of specification.
We note that the parameter vector in (4.1) represents an nkx1 matrix in this model that contains parameter estimates for all k explanatory
variables at every observation. The parameter vector 0 contains the 2k
parameters to be estimated.
y = X +
= ZJ0
(4.1)
Where:
y =
y1
y2
..
.
X=
yn
x01 0 . . . 0
0 x02
..
..
.
.
0
x0n
1
2
..
.
n
1
2
..
.
n
Ik 0
Zx1 Ik Zy1 Ik
0
...
0 Ik
..
..
.
.
0
Z =
J = ..
.
..
.
Zxn Ik Zyn Ik
0 Ik
0 =
x
y
(4.2)
This model can be estimated using least-squares to produce estimates of

the 2k parameters x , y . Given these estimates, the remaining estimates
for individual points in space can be derived using the second equation in
(4.1). This process is referred to as the expansion process. To see this,
substitute the second equation in (4.1) into the first, producing:
y = XZJ0 +
(4.3)
Here it is clear that X, Z and J represent available information or data

observations and only 0 represent parameters in the model that need be
estimated.
The model would capture spatial heterogeneity by allowing variation
in the underlying relationship such that clusters of nearby or neighboring
129
observations measured by latitude-longitude coordinates take on similar parameter values. As the location varies, the regression relationship changes
to accommodate a locally linear fit through clusters of observations in close
proximity to one another.
Another way to implement this model is to rely on a vector of distances
rather than the latitude-longitude coordinates. This implementation defines
the distance from a central observation,
q
di =
(Zxi Zxc )2 + (Zyi Zyc )2
(4.4)
Where Zxc , Zyc denote the latitude-longitude coordinates of the centrally

located observation and Zxi , Zyi denote the latitude-longitude coordinates
for observations i = 1, . . . , n in the data sample.
This approach allows one to ascribe different weights to observations
based on their distance from the central place origin. The formulation discussed above would result in a distance vector that increased with distance
from the central observation. This would be suitable if one were modeling
a phenomena reflecting a hollowing out of the central city or a decay of
influence with distance from the central point.
The distance expansion model can be written as:
y = X +
= DJ0
(4.5)
Where D = diag(d1 , d2 , . . ., dn ), represents the distance of each observation

from the central place and 0 represents a kx1 vector of parameters for the
central place. The matrix J in (4.5) is an nxk matrix, J = (Ik , Ik , . . ., Ik )0 .
4.1.1
Implementing spatial expansion
Estimating this model is relatively straightforward as we can rely on leastsquares. One issue is that there are a number of alternative expansion
specifications. For example, one approach would be to construct a model
that includes the base k explanatory variables in the matrix X estimated
with fixed parameters, plus an additional 2k expansion variables based on
the latitude-longitude expansion. Another approach would be to include the
base k variables in the matrix X and only 2(k1) variables in expansion form
by excluding the constant term from the expansion process. Yet another
approach would be to rely on a simple expansion of all variables as was
illustrated in (4.1).
130
The second approach was taken in implementing the MATLAB function

casetti that carries out spatial expansion estimation. This choice was made
because it seems unwise to include the constant term in the expansion as one
can overfit the sample data when the intercept is allowed to very over space.
A motivation for not relying on a simple expansion of all variables is that
we would like our model to partition the influence of explanatory variables
into fixed plus spatial effects. A simple expansion assigns all influence to
spatial effects and also falls prey to the overfitting problem by allowing the
intercept term to vary.
The expansion implemented by our function casetti can be written as:
y = + X + XZxx + XZy y +
(4.6)
The function allows the user to specify an option for distance expansion based on a particular point in the spatial data sample or the latitudelongitude expansion. In the case of the distance expansion, the k explanatory
variables in the matrix X are used as non-expansion variables estimated with
fixed parameters and the k 1 variables excluding the constant are included
as distance-expanded variables. This version of the model can be written
as:
y = + X + XD0 +
(4.7)
q For the case of distance expansion, a distance vector is calculated as: di =
(Zxi Zxc )2 + (Zyi Zyc )2 , where Zxc , Zyc denote the latitude-longitude
coordinates of the centrally located observation and Zxi , Zyi denote the coordinates for observation i in the data sample. The distance of the central
point is zero of course.
An optional input is provided to carry out isotropic normalization of the
x-y coordinates which essentially puts the coordinates in deviations from
the means form then standardizes by dividing by the square root of the sum
of the variances in the x-y directions. That is:
q
x? = (x x
)/ (x2 + y2 )
q
y ? = (y y)/ (x2 + y2 )
(4.8)
This normalization is carried out by a function normxy in the spatial econometrics library.
This normalization should make the center points xc, yc close to zero and
produces a situation where the coefficients for the base model represent a
131
central observation. The distance-expanded estimates provide information

about variation in the model parameters with reference to the central point.
The documentation for casetti is:
PURPOSE: computes Casettis spatial expansion regression
--------------------------------------------------USAGE: results = casetti(y,x,xc,yc,option)
where:
xc = latitude (or longitude) coordinate
yc = longitude (or latitude) coordinate
option = a structure variable containing options
option.exp = 0 for x-y expansion (default)
= 1 for distance from ctr expansion
option.ctr = central point observation # for distance expansion
option.norm = 1 for isotropic x-y normalization (default=0)
--------------------------------------------------RETURNS:
results.meth = casetti
results.b0
= bhat (underlying b0x, b0y)
results.t0
= t-stats (associated with b0x, b0y)
results.beta = spatially expanded estimates (nobs x nvar)
results.yhat = yhat
results.nobs = nobs
results.nvar = # of variables in x
results.y
= y data vector
results.xc
= xc
results.yc
= yc
results.ctr
= ctr (if input)
results.dist = distance vector (if ctr used)
results.exp
= exp input option
results.norm = norm input option
-------------------------------------------------NOTE: assumes x(:,1) contains a constant term
--------------------------------------------------
Given the exclusion of the constant term from the spatial expansion
formulation, we need to impose that the user place the constant term vector
in the first column of the explanatory variables matrix X used as an input
argument to the function.
Of course, we have an associated function to print the results structure and another to provide graphical presentation of the estimation results.
Printing these estimation results is a bit challenging because of the large
132
number of parameter estimates that we produce using this method. Graphical presentation may provide a clearer picture of the variation in coefficients
over space. A call to plt using the results structure variable will produce
plots of the coefficients in both the x- and y-directions for the case of the
latitude-longitude expansion, where we sort the x-direction from left to right
and the y-direction from left to right. This provides a visual picture of how
the coefficients vary over space. If the x-coordinates are largest for the east
and smallest for the west, the plot will show coefficient variation from west
to east as in map space. Similarly, if the y-coordinates are smallest for the
south and largest in the north, the plot will present coefficient variation
from south to north. (Note that if you enter Western hemisphere latitudelongitude coordinates, the x-direction plots will be from east to west, but
the y-direction plots will be south to north.)
For the case of distance expansion estimates, the plots present coefficients sorted by distance from the central point, provided by the user in the
structure field option.ctr. The central observation (smallest distance) will
be on the left of the graph and the largest distance on the right.
Another point to note regarding the graphical presentation of the estimates relates to the fact that we present the coefficients in terms of the
individual variables total impact on the dependent variable y. It was felt
that users would usually be concerned with the total impact of a particular
variable on the dependent variable as well as the decomposition of impacts
into spatial and non-spatial effects. The printed output provides the coefficient estimates for the base model as well as the expansion coefficients that
can be used to analyze the marginal effects from the spatial and non-spatial
decomposition. To provide another view of the impact of the explanatory
variables in the model on the dependent variable, the graphical presentation
plots the coefficient estimates in a form representing their total impact on
the dependent variable. That is we graph:
xi = i + Zx xi
yi = i + Zy yi
di = i + D0i
(4.9)
Where x, y are plotted for the x-y expansion and d is graphed for the
distance expansion. This should provide a feel for the total impact of variable
i on the dependent variable since it takes into account the non-spatial impact
attributed to i , as well as the spatially varying impacts in the x-y direction
or with respect to distance.
133
An illustration in the next section will pursue this point in more detail.
4.1.2
Applied examples
Example 4.1 provides an illustration of using the function casetti based

on the Columbus neighborhood crime data set. Both types of expansion
models are estimated by changing the structure variable option field .exp.
For the case of distance expansion, we rely on a central observation number
20 which lies near the center of the spatial sample of neighborhoods. One
point to note is that the x-coordinate in Anselins data set represents the
south-north direction and the y-coordinate reflects the west-east direction.
% ----- example 4.1 Using the casetti() function
% load Anselin (1988) Columbus neighborhood crime data
load anselin.dat; y = anselin(:,1); n = length(y);
% Anselin (1988) x-y coordinates
xc0 = anselin(:,4); yc0 = anselin(:,5);
vnames = strvcat(crime,const,income,hse value);
% do Casetti regression using x-y expansion (default)
res1 = casetti(y,x,xc0,yc0);
prt(res1,vnames); % print the output
plt(res1,vnames); % graph the output
pause;
% do Casetti regression using distance expansion
option.exp = 1; option.ctr = 20; % Obs # of a central neighborhood
res2 = casetti(y,x,xc0,yc0,option);
plt(res2,vnames); % graph the output
The default option is to implement an x-y expansion, which produces the

result structure variable res1. The next case relies on a structure variable
option to select distance expansion. The printed output is shown below.
Both the base estimates as well as the expansion estimates are presented
in the printed output. If you are working with a large model containing
numerous observations, you can rely on the printing option that places the
output in a file. Recall from Section 1.5, we need simply open an output file
and input the file-id as an option to the prt function.
Another point to note regarding the printed output is that in the case of
a large number of explanatory variables, the printed estimates will wrap. A
set of estimates that take up 80 columns will be printed for all observations,
and remaining estimates will be printed below for all observations. This
wrapping will continue until all of the parameter estimates are printed.

Casetti X-Y Spatial Expansion Estimates
crime
R-squared
=
0.6330
Rbar-squared =
0.5806
sige
= 117.4233
Nobs, Nvars
=
49,
3
***************************************************************
Base x-y estimates
Variable
Coefficient
t-statistic
t-probability
const
69.496160
15.105146
0.000000
income
-4.085918
-1.951941
0.057048
hse value
0.403956
0.517966
0.606965
x-income
-0.046062
-1.349658
0.183731
x-hse value
0.026732
2.027587
0.048419
y-income
0.121440
2.213107
0.031891
y-hse value
-0.048606
-2.341896
0.023571
***************************************************************
Expansion estimates
Obs#
x-income x-hse value
y-income y-hse value
1
-1.6407
0.9522
5.1466
-2.0599
2
-1.6813
0.9757
4.9208
-1.9695
3
-1.6909
0.9813
4.7009
-1.8815
4
-1.5366
0.8918
4.6645
-1.8670
5
-1.7872
1.0372
5.3519
-2.1421
6
-1.8342
1.0645
5.0009
-2.0016
7
-1.8429
1.0695
4.6147
-1.8470
8
-2.0152
1.1695
4.7702
-1.9092
9
-1.8245
1.0588
4.2395
-1.6968
10
-2.1930
1.2727
4.4228
-1.7702
11
-2.2377
1.2986
4.1848
-1.6750
12
-2.2851
1.3262
3.9650
-1.5870
13
-2.3082
1.3395
3.6323
-1.4538
14
-2.3602
1.3697
3.3760
-1.3512
15
-2.3441
1.3604
3.0651
-1.2268
16
-2.2312
1.2949
3.3918
-1.3576
17
-2.1525
1.2492
3.8752
-1.5510
18
-2.0009
1.1612
4.3621
-1.7459
19
-1.9977
1.1594
4.0634
-1.6264
20
-1.8945
1.0995
4.0245
-1.6108
21
-2.0244
1.1749
3.8387
-1.5364
22
-2.0313
1.1789
3.6918
-1.4776
23
-2.0129
1.1682
3.5436
-1.4183
24
-1.8904
1.0971
3.4950
-1.3989
25
-1.9913
1.1556
3.3165
-1.3274
26
-1.9655
1.1406
3.0311
-1.2132
27
-1.8982
1.1016
3.1453
-1.2589
28
-1.8112
1.0511
3.1392
-1.2565
29
-1.8927
1.0984
3.3384
-1.3362
30
-1.7651
1.0244
3.4999
-1.4008
134

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
-1.9028
-1.8130
-1.8296
-1.7637
-1.6859
-1.7319
-1.7103
-1.7434
-1.6559
-1.6453
-1.6472
-1.6651
-1.5698
-1.3966
-1.2870
-1.2561
-1.1170
-1.1732
-1.3367
1.1043
1.0522
1.0618
1.0236
0.9784
1.0051
0.9926
1.0118
0.9610
0.9549
0.9559
0.9664
0.9110
0.8105
0.7469
0.7290
0.6482
0.6809
0.7758
3.7525
3.9929
3.7209
3.6857
3.8970
4.1387
4.3864
4.4083
4.4204
4.3233
4.2091
4.1192
3.6942
3.4319
3.6250
3.4258
3.2412
3.1222
3.2279
-1.5019
-1.5982
-1.4893
-1.4752
-1.5598
-1.6565
-1.7556
-1.7644
-1.7693
-1.7304
-1.6847
-1.6487
-1.4786
-1.3736
-1.4509
-1.3712
-1.2973
-1.2497
-1.2919
Casetti Distance Spatial Expansion Estimates

crime
R-squared
=
0.6307
Rbar-squared =
0.5878
sige
= 112.7770
Nobs, Nvars
=
49,
3
central obs
=
20
***************************************************************
Base centroid estimates
Variable
Coefficient
t-statistic
t-probability
const
62.349645
12.794160
0.000000
income
-0.855052
-1.048703
0.299794
hse value
-0.138951
-0.520305
0.605346
d-income
-0.048056
-0.613545
0.542538
d-hse value
-0.013384
-0.473999
0.637743
***************************************************************
Expansion estimates
Obs#
income
hse value
1
-0.5170
-0.1440
2
-0.4187
-0.1166
3
-0.3417
-0.0952
4
-0.4512
-0.1257
5
-0.5371
-0.1496
6
-0.3915
-0.1090
7
-0.2397
-0.0668
8
-0.3208
-0.0893
9
-0.1121
-0.0312
10
-0.3490
-0.0972
11
-0.3636
-0.1013
135

12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
-0.4082
-0.4586
-0.5495
-0.6034
-0.4314
-0.2755
-0.1737
-0.1087
-0.0000
-0.1542
-0.1942
-0.2269
-0.2096
-0.2978
-0.4000
-0.3479
-0.3610
-0.2715
-0.2477
-0.1080
-0.0860
-0.1379
-0.1913
-0.2235
-0.1755
-0.2397
-0.2189
-0.2941
-0.2856
-0.2682
-0.2422
-0.3631
-0.5700
-0.6533
-0.7069
-0.8684
-0.8330
-0.6619
136
-0.1137
-0.1277
-0.1530
-0.1681
-0.1201
-0.0767
-0.0484
-0.0303
-0.0000
-0.0429
-0.0541
-0.0632
-0.0584
-0.0829
-0.1114
-0.0969
-0.1005
-0.0756
-0.0690
-0.0301
-0.0239
-0.0384
-0.0533
-0.0622
-0.0489
-0.0668
-0.0610
-0.0819
-0.0795
-0.0747
-0.0675
-0.1011
-0.1587
-0.1820
-0.1969
-0.2419
-0.2320
-0.1843
We turn attention to interpreting the output from this example. For

this purpose we compare the base spatial expansion estimates to those
from least-squares which are presented below. The addition of the four x-y
expansion variables increased the fit of the model slightly as indicated by the
higher adjusted R2 statistic. We see that the intercept estimate is relatively
unaffected by the inclusion of expansion variables, but the coefficients on
income and house value take on very different values. The significance of
the income variable falls as indicated by the lower tstatistic, and the house
137
value variable becomes insignificant.

Three of the four x-y expansion variables are significant at the 0.05 level,
providing evidence that the influence of these variables on neighborhood
crime varies over space. Keep in mind that depending on the amount of
inherent variation in the x-y coordinates, we may introduce a substantial
amount of collinearity into the model when we add expansion variables.
These are likely highly correlated with the base variables in the model, and
this may account for the lack of significance of the house value variable in
the base model. A strict interpretation of these base estimates would be
that income expansion in the x-direction (south-north) is not significant,
whereas it is in the y-direction (west-east).
crime
R-squared
=
0.5521
Rbar-squared =
0.5327
sigma^2
= 130.8386
Durbin-Watson =
1.1934
Nobs, Nvars
=
49,
3
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
68.609759
14.484270
0.000000
income
-1.596072
-4.776038
0.000019
house value
-0.274079
-2.655006
0.010858
Casetti X-Y Spatial Expansion Estimates
crime
R-squared
=
0.6330
Rbar-squared =
0.5806
sige
= 117.4233
Nobs, Nvars
=
49,
3
***************************************************************
Base x-y estimates
Variable
Coefficient
t-statistic
t-probability
const
69.496160
15.105146
0.000000
income
-4.085918
-1.951941
0.057048
hse value
0.403956
0.517966
0.606965
x-income
-0.046062
-1.349658
0.183731
x-hse value
0.026732
2.027587
0.048419
y-income
0.121440
2.213107
0.031891
y-hse value
-0.048606
-2.341896
0.023571
Casetti Distance Spatial Expansion Estimates
crime
R-squared
=
0.6307
Rbar-squared =
0.5878
sige
= 112.7770
138
Nobs, Nvars
=
49,
3
central obs
=
20
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
62.349645
12.794160
0.000000
income
-0.855052
-1.048703
0.299794
hse value
-0.138951
-0.520305
0.605346
d-income
-0.048056
-0.613545
0.542538
d-hse value
-0.013384
-0.473999
0.637743
In order to interpret the x-y expansion estimates, we need to keep in mind

that the x-direction reflects the south-north direction with larger values of
xc indicating northward movement. Similarly, larger values for yc reflect
west-east movement. Using these interpretations, the base model estimates
indicate that income exerts an insignificant negative influence on crime as
we move from south to north. Considering the y-direction representing
west-east movement, we find that income exerts a positive influence as we
move in the easterly direction. One problem with interpreting the expansion
estimates is that the base model coefficient is -4.085, indicating that income
exerts a negative influence on crime. It is difficult to assess the total impact
on neighborhood crime from both the base model coefficient representing
non-spatial impact plus the small 0.121 value for the expanded coefficient
reflecting the impact of spatial variation.
If we simply plotted the expansion coefficients, they would suggest that
income in the y-direction has a positive influence on crime, a counterintuitive
result. This is shown in the plot of the expanded coefficients sorted by the
south-north and east-west directions shown in Figure 4.1. We are viewing a
positive coefficient on the y-income variable in the graph.
Figure 4.2 shows a graph of the total impact of income on the dependent
variable crime that takes into account both the base model non-spatial impact plus the spatial impact indicated by the expansion coefficient. Here we
see that the total impact of income on crime is negative, except for neighborhoods in the extreme east at the right of the graph in the lower left-hand
corner of Figure 4.2. The coefficient graphs produced using the plt function on the results structure from casetti are identical to those shown in
Figure 4.2. You can of course recover the spatial expansion estimates from
the results structure returned by casetti, sort the estimates in the x-y directions and produce your own plots of just the expansion estimates if this
is of interest. As an example, the following code would produce this type of
graph, where we are assuming the existence of a structure result returned
by casetti.
139
[xcs xci] = sort(result.xc);

[ycs yci] = sort(result.yc);
beta = result.beta;
[nobs nvar] = size(beta);
nvar = nvar/2;
betax = beta(xci,1:nvar); % sort estimates
betay = beta(yci,nvar+1:2*nvar);
tt=1:nobs;
for j=1:nvar
plot(tt,betax(:,j)); pause;
end;
for j=1:nvar
plot(tt,betay(:,j)); pause;
end;
-1
1.4
x-house value
1.2
x-income
-1.5
-2
0.8
0.6
10
20
30
40
50
sorted by x-direction, left=smallest x
5.5
-1.2
-1.4
y-house value
y-income
-2.5
4.5
4
3.5
3
10
20
30
40
50
10
20
30
40
50
sorted by y-direction, left=smallest x
-1.6
-1.8
-2
10
20
30
40
50
-2.2
Figure 4.1: Spatial x-y expansion estimates

The distance expansion method produced a similar fit to the data as
indicated by the adjusted R2 statistic. This is true despite the fact that
only the constant term is statistically significant at conventional levels.
-5
1.8
-5.5
1.6
x-house value
x-income
-6
-6.5
1.4
1.2
10
20
30
40
50
1.5
-0.8
-1
y-house value
y-income
-7
0.5
0
-0.5
10
20
30
40
50
10
20
30
40
50
-1.2
-1.4
-1.6
-1
-1.5
140
10
20
30
40
50
-1.8
Figure 4.2: Spatial x-y total impact estimates

These estimates take on a value of zero for the central observation since
the distance is zero at that point. This makes them somewhat easier to
interpret than the estimates for the case of x-y coordinates. We know that
the expansion coefficients take on values of zero at the central point, producing estimates based on the non-spatial base coefficients for this point.
As we move outward from the center, the expansion estimates take over
and adjust the constant coefficients to account for variation over space. The
printed distance expansion estimates reported for the base model reflect values near the distance-weighted average of all points in space. Given this,
we would interpret the coefficients for the model at the central point to be
(62.349, -0.855, -0.138 ) for the intercept, income and house value variables
respectively. In Figure 4.3 we see the total impact of income and house values on neighborhood crime as we move away from the central point. Both
income and house values have a negative effect on neighborhood crime as
we move away from the central city. Note that this is quite different from
141
the pattern shown for the x-y expansion. Anselin (1988) in analyzing the
x-y model shows that heteroscedastic disturbances produce problems that
plague inferences for the model. Adjusting for the heteroscedastic disturbances dramatically alters the inferences. We turn attention to this issue
when we discuss the DARP version of this model in Section 4.2.
The plots of the coefficient estimates provide an important source of
information about the nature of coefficient variation over space, but you
should keep in mind that they do not indicate levels of significance, simply
point estimates.
Our plt wrapper function works to call the appropriate function plt cas
that provides individual graphs of each coefficient in the model as well as
a two-part graph showing actual versus predicted and residuals. Figure 4.4
shows the actual versus predicted and residuals from the distance expansion
model. This plot is produced by the plt function when given a results
structure from the casetti function.
-0.8
d-income
-1
-1.2
-1.4
-1.6
-1.8
10
15
20
25
30
35
distance from the center, left=center
40
45
50
10
15
20
25
30
35
distance from the center, left=center
40
45
50
-0.05
d-house value
-0.1
-0.15
-0.2
-0.25
-0.3
-0.35
-0.4
Figure 4.3: Distance expansion estimates
142
CASETTI Actual vs. Predicted

70
60
50
40
30
20
10
0
10
15
20
25
30
35
40
45
50
30
35
40
45
50
Residuals
30
20
10
0
-10
-20
-30
-40
10
15
20
25
Figure 4.4: Actual versus Predicted and residuals
4.2
DARP models
A problem with the spatial expansion model is that heteroscedasticity is inherent in the way the model is constructed. To see this, consider the slightly
altered version of the distance expansion model shown in (4.10), where we
have added a stochastic term u to reflect some error in the expansion relationship.
y = X + e
= DJ0 + u
(4.10)
Now consider substituting the second equation from (4.10) into the first,
producing:
y = XDJ0 + Xu + e
(4.11)
143
It should be clear that the new composite disturbance term Xu+e will reflect
heteroscedasticity unless the expansion relationship is exact and u = 0.
Casetti (1982) and Casetti and Can (1998) propose a model they label
DARP, an acronym for Drift Analysis of Regression Parameters, that aims at
solving this problem. This model case be viewed as an extended expansion
model taking the form:
y = X + e
= f (Z, ) + u
(4.12)
Where f (Z, ) represents the expansion relationship based on a function f ,

variables Z and parameters . Estimation of this model attempts to take
into account that the expanded model will have a heteroscedastic error as
shown in (4.11).
To keep our discussion concrete, we will rely on f (Z, ) = DJ0 , the
distance expansion relationship in discussing this model. In order to take
account of the spatial heteroscedasticity an explicit model of the composite
disturbance term: = Xu + e, is incorporated during estimation. This
disturbance is assumed to have a variance structure that can be represented
in alternative ways shown below. We will rely on these alternative scalar and
matrix representations in the mathematical development and explanation of
the model.
E(0 ) = = 2
= exp(diag(d1 , d2, . . . , dn))
= diag(12 , 22 , . . . , n2 )
i2 = exp(0 + idi )
(4.13)
Where di denotes the squared distance between the ith observation and the
central point, and 2 , 0, 1 are parameters to be estimated. Of course, a
more general statement of the model would be that i2 = g(hi , ), indicating
that any functional form g involving some known variable hi and associated
parameters could be employed to specify the non-constant variance over
space.
Note that the alternative specifications in (4.13) imply: 2 = exp(0 ),
the constant scalar component of the variance structure, and exp(1 di) reflects the non-constant component which is modeled as a function of distance
from the central point.
144
An intuitive motivation for this type of variance structure is based on

considering the nature of the composite disturbance: Xu + e. The constant
scalar 2 reflects the constant component e, while the role of the scalar parameter 1 associated with distance is to measure the impact of Xu, the
non-constant variance component on average. Somewhat loosely, consider
a linear regression involving the residuals on a constant term plus a vector
of distances from the central place. The constant term estimate would reflect 2 = exp(0 ), while the distance coefficient is intended to capture the
influence of the non-constant Xu component: = exp(1 d).
If 1 = 0, we have that = 2 In , a constant scalar value across all observations in space. This would be indicative of a situation where u, the error
made in the expansion specification is small. This homoscedastic case would
be indicative that a simple deterministic spatial expansion specification for
spatial coefficient variation is performing well.
On the other hand, if 1 > 0, we find that moving away from the central
point produces a positive increase in the variance. This is interpreted as
evidence that parameter drift is present in the relationship being modeled.
The motivation is that increasing variance would be indicative of larger
errors (u) in the expansion relationship as we move away from the central
point.
Note that one can assume a value for 1 rather than estimate this parameter. If you impose a positive value, you are assuming a DARP model
that will generate locally linear estimates since movement in the parameters
will increase with movement away from the central point. This is because
allowing increasing variance in the stochastic component of the expansion
relation brings about more rapid change or adjustment in the parameters.
Another way to view this is that the change in parameters need not adhere
as strictly to the deterministic expansion specification. We will argue in
Section 4.4 that a Bayesian approach to this type of specification is more
intuitively appealing.
Negative values for 1 suggest that the errors made by the deterministic
expansion specification are smaller as we move away from the central point.
This indicates that the expansion relation works well for points farther from
the center, but not as well for the central area observations. Casetti and
Can (1998) interpret this as suggesting differential performance of the
base model with movement over space, and they label this phenomena as
performance drift. The intuition here is most likely based on the fact that
the expansion relationship is of a locally linear nature. Given this, better
performance with distance is indicative of a need to change the deterministic
expansion relationship to improve performance. Again, I will argue that a
145
Bayesian model represents a more intuitively appealing way to deal with

these issues in Section 4.4.
Estimation of the parameters of the model require either feasible generalized least squares (FGLS) or maximum likelihood (ML) methods. Feasible
generalized least squares obtains a consistent estimate of the unknown parameter 1 and then proceeds to estimate the remaining parameters in the
model conditional on this estimate.
As an example, consider using least-squares to estimate the expansion
model and associated residuals, e = y XDJ0 . We could then carry out
a regression of the form:
log(
e2) = 0 + 1 d +
(4.14)
Casetti and Can (1998) argue that the estimate 1 from this procedure
would be consistent. Given this estimate and our knowledge of the distances
Using ,
in the vector d, we construct and estimate of that we label .

generalized least-squares can produce:
F GLS
F2 GLS
1 X)1 X 0
1 y
= (X 0
1 (y X F GLS )/(n k)
= (y X F GLS )
Of course, the usual GLS variance-covariance matrix for the estimates

applies:
1 X)1
var cov(F GLS ) =
2 (X 0
(4.15)
P
Casetti and Can (1998) also suggest using a statistic: 12 /4.9348 di

which is chi-squared distributed with one degree of freedom to test the null
hypothesis that 1 = 0.
Maximum likelihood estimation involves using optimization routines to
solve for a minimum of the negative of the log-likelihood function. We have
already seen how to solve optimization problems in the context of spatial
autoregressive models in Chapter 2. We will take the same approach for this
model. The log-likelihood is:
L(, 1 |.) = c (1/2)ln| 2| (1/2)(y X)01 (y X)
(4.16)
As in Chapter 2 we can construct a MATLAB function to evaluate the

negative of this log-likelihood function and rely on our function maxlik.
The asymptotic variance-covariance matrix for the estimates is equal to
146
that for the FGLS estimates shown in (4.15). The asymptotic variancecovariance matrix for the parameters ( 2 , 1) is given by:
var cov( 2, 1) = 2(D0 D)1
D = (, d)
(4.17)
In the case of maximum likelihood estimates, a Wald statistic based on

that has a chi-squared distribution with one degree of freedom can
be used to test the null hypothesis that 1 = 0. Note that the maximum
likelihood estimate of 1 is more efficient than the FGLS estimate. This can
P
be seen by comparing the ML estimates asymptotic variance of 2 di , to
P
that for the FGLS which equals 4.9348 di . Bear in mind, tests regarding
the parameter 1 are quite often the focus of this methodology as it provides
exploratory evidence regarding performance drift versus parameter drift,
so increased precision regarding this parameter may be important.
The function darp implements this estimation procedure using the maxlik function from the Econometrics Toolbox to find estimates via maximum
likelihood. The documentation for darp is:
P
12 /2 di
PURPOSE: computes Casettis DARP model

--------------------------------------------------USAGE: results = darp(y,x,xc,yc,option)
where:
xc = latitude (or longitude) coordinate
yc = longitude (or latitude) coordinate
option = a structure variable containing options
option.exp = 0 for x-y expansion (default)
= 1 for distance from ctr expansion
option.ctr = central point observation # for distance expansion
option.iter = # of iterations for maximum likelihood routine
option.norm = 1 for isotropic x-y normalization (default=0)
--------------------------------------------------RETURNS:
results.meth = darp
results.b0
= bhat (underlying b0x, b0y)
results.t0
= t-stats (associated with b0x, b0y)
results.beta = spatially expanded estimates (nobs x nvar)
results.yhat = yhat
results.nobs = nobs
147
results.nvar = # of variables in x
results.y
= y data vector
results.xc
= xc
results.yc
= yc
results.ctr
= ctr (if input)
results.dist = distance vector
results.exp
= exp input option
results.norm = norm input option
results.iter = # of maximum likelihood iterations
-------------------------------------------------NOTE: assumes x(:,1) contains a constant term
--------------------------------------------------
Because we are relying on maximum likelihood estimation which may

not converge, we provide FGLS estimates as output in the event of failure.
A message is printed to the MATLAB command window indicating that this
has occurred. We rely on the FGLS estimates to provide starting values for
the maxlik routine, which should speed up the optimization process.
The DARP model can be invoked with either the x-y or distance expansion as in the case of the spatial expansion model. Specifically, for x-y
expansion the variance specification is based on:
log(
e2) = 0 + 1 xc + 2 yc +
(4.18)
This generalizes the distance expansion approach presented previously.

Of course, we have an accompanying prt and plt function to provide
printed and graphical presentation of the estimation results.
Example 4.2 shows how to use the function darp for both x-y and distance expansion using the Columbus neighborhood crime data set.
% ----- example 4.2 Using the darp() function
xc = anselin(:,4); yc = anselin(:,5); % Anselin x-y coordinates
% do Casetti darp using x-y expansion
res1 = darp(y,x,xc,yc);
plt(res1,vnames); % plot the output
pause;
% do Casetti darp using distance expansion from observation #20
option.exp = 1; option.ctr = 20;
res2 = darp(y,x,xc,yc,option);
plt(res2,vnames); % plot the output
148
The printed results are shown below, where we report not only the estimates for 0 and the expansion estimates, but estimates for the parameters
as well. A chi-squared statistic to test the null hypothesis that 1 = 0
is provided as well as a marginal probability level. For the case of the x-y
expansion, we see that 1 parameter is negative and significant by virtue
of the large chi-squared statistic and associated marginal probability level
of 0.0121. The inference we would draw is that performance drift occurs in
the south-north direction. For the 2 parameter, we find a positive value
that is not significantly different from zero because of the marginal probability level of 0.8974. This indicates that the simple deterministic expansion
relationship is working well in the west-east direction. Note that these results conform to those found with the spatial expansion model, where we
indicated that parameter variation in the west-east direction was significant,
but not for the south-north.
DARP X-Y Spatial Expansion Estimates
crime
R-squared
=
0.6180
Rbar-squared
=
0.5634
sige
= 122.2255
gamma1,gamma2 =
-0.0807,
0.0046
gamma1, prob
=
6.2924,
0.0121
gamma2, prob
=
0.0166,
0.8974
# of iterations =
16
log-likelihood =
-181.3901
Nobs, Nvars
=
49,
3
***************************************************************
Base x-y estimates
Variable
Coefficient
t-statistic
t-probability
const
66.783527
6.024676
0.000000
income
-2.639184
-0.399136
0.691640
hse value
0.249214
0.095822
0.924078
x-income
-0.048337
-0.537889
0.593247
x-hse value
0.021506
0.640820
0.524819
y-income
0.084877
0.564810
0.574947
y-hse value
-0.037460
-0.619817
0.538436
***************************************************************
Expansion estimates
Obs#
x-income x-hse value
y-income y-hse value
1
-1.3454
0.6595
4.0747
-1.6828
2
-1.3786
0.6758
3.8959
-1.6089
3
-1.3865
0.6796
3.7218
-1.5370
4
-1.2600
0.6176
3.6930
-1.5251
5
-1.4655
0.7183
4.2372
-1.7499
6
-1.5040
0.7372
3.9593
-1.6351
7
-1.5112
0.7407
3.6536
-1.5088

8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
-1.6524
-1.4961
-1.7982
-1.8349
-1.8738
-1.8926
-1.9353
-1.9221
-1.8296
-1.7650
-1.6407
-1.6381
-1.5535
-1.6600
-1.6657
-1.6505
-1.5501
-1.6328
-1.6116
-1.5565
-1.4851
-1.5520
-1.4473
-1.5603
-1.4866
-1.5002
-1.4462
-1.3824
-1.4201
-1.4024
-1.4296
-1.3578
-1.3491
-1.3507
-1.3654
-1.2872
-1.1452
-1.0553
-1.0300
-0.9159
-0.9620
-1.0961
0.8100
0.7333
0.8814
0.8994
0.9185
0.9277
0.9487
0.9422
0.8968
0.8652
0.8042
0.8029
0.7615
0.8137
0.8165
0.8091
0.7598
0.8004
0.7900
0.7630
0.7280
0.7607
0.7095
0.7648
0.7287
0.7354
0.7089
0.6776
0.6961
0.6874
0.7008
0.6656
0.6613
0.6621
0.6693
0.6310
0.5613
0.5173
0.5049
0.4490
0.4715
0.5373
DARP Distance Expansion Estimates

crime
R-squared
=
0.6083
Rbar-squared =
0.5628
sige
= 119.6188
gamma
=
-0.0053
3.7766
3.3565
3.5017
3.3132
3.1392
2.8757
2.6729
2.4267
2.6854
3.0680
3.4536
3.2171
3.1863
3.0392
2.9229
2.8056
2.7671
2.6258
2.3998
2.4902
2.4854
2.6431
2.7709
2.9709
3.1613
2.9459
2.9181
3.0853
3.2767
3.4728
3.4901
3.4997
3.4228
3.3324
3.2613
2.9248
2.7171
2.8700
2.7123
2.5662
2.4719
2.5556
-1.5597
-1.3862
-1.4461
-1.3683
-1.2964
-1.1876
-1.1038
-1.0022
-1.1090
-1.2670
-1.4263
-1.3286
-1.3159
-1.2551
-1.2071
-1.1586
-1.1428
-1.0844
-0.9911
-1.0284
-1.0264
-1.0915
-1.1443
-1.2269
-1.3055
-1.2166
-1.2051
-1.2742
-1.3532
-1.4342
-1.4413
-1.4453
-1.4136
-1.3762
-1.3468
-1.2079
-1.1221
-1.1852
-1.1201
-1.0598
-1.0209
-1.0554
149

gamma, prob
=
0.0467,
0.8289
# of iterations =
10
log-likelihood =
-138.64471
Nobs, Nvars
=
49,
3
central obs
=
20
***************************************************************
Variable
Coefficient
t-statistic
t-probability
const
62.323508
5.926825
0.000000
income
-0.889528
-0.925670
0.359448
hse value
-0.312813
-1.015500
0.315179
d-income
-0.004348
-0.790909
0.433056
d-hse value
0.000659
0.349488
0.728318
***************************************************************
Expansion estimates
Obs#
income
hse value
1
-0.4032
0.0055
2
-0.2644
0.0036
3
-0.1761
0.0024
4
-0.3070
0.0042
5
-0.4350
0.0060
6
-0.2311
0.0032
7
-0.0866
0.0012
8
-0.1552
0.0021
9
-0.0190
0.0003
10
-0.1837
0.0025
11
-0.1994
0.0027
12
-0.2513
0.0034
13
-0.3172
0.0043
14
-0.4554
0.0062
15
-0.5492
0.0075
16
-0.2807
0.0038
17
-0.1145
0.0016
18
-0.0455
0.0006
19
-0.0178
0.0002
20
-0.0000
0.0000
21
-0.0359
0.0005
22
-0.0569
0.0008
23
-0.0776
0.0011
24
-0.0662
0.0009
25
-0.1338
0.0018
26
-0.2413
0.0033
27
-0.1826
0.0025
28
-0.1965
0.0027
29
-0.1112
0.0015
30
-0.0925
0.0013
31
-0.0176
0.0002
32
-0.0111
0.0002
33
-0.0287
0.0004
150

34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
-0.0552
-0.0753
-0.0465
-0.0867
-0.0723
-0.1305
-0.1230
-0.1085
-0.0885
-0.1989
-0.4900
-0.6437
-0.7538
-1.1374
-1.0465
-0.6607
151
0.0008
0.0010
0.0006
0.0012
0.0010
0.0018
0.0017
0.0015
0.0012
0.0027
0.0067
0.0088
0.0103
0.0156
0.0143
0.0090
For the case of the distance expansion we find a single parameter that
is negative but not significant. This would be interpreted to mean that the
deterministic expansion relationship is performing adequately over space.
A comparison of the base model estimates from the x-y darp model
versus those from casetti show relatively similar coefficient estimates as we
would expect. In the case of the x-y expansion all of the signs are the same
for spatial expansion and darp models. The distance expansion version of
the model exhibits a sign change for the coefficient on income, which goes
from positive in the expansion model to negative in the darp model. Correcting for the heteroscedastic character of the estimation problem produces
dramatic changes in the statistical significance found for the base model
estimates. They all become insignificant, a finding consistent with results
reported by Anselin (1988) based on a jacknife approach to correcting for
heteroscedasticity in this model. (Anselin finds that the income coefficient
is still marginally significant after the correction.)
One approach to using this model is to expand around every point in
space and examine the parameters for evidence indicating where the model
is suffering from performance or parameter drift. Example 4.3 shows how
this might be accomplished using a for loop over all observations. For this
purpose, we wish only to recover the estimated values for the parameter
along with the marginal probability level.
% ----- example 4.3 Using darp() over space
xc = anselin(:,4); yc = anselin(:,5); % Anselin x-y coordinates
152

% do Casetti darp using distance expansion from all
% observations in the data sample
option.exp = 1;
output = zeros(n,2);
tic;
for i=1:n % loop over all observations
option.ctr = i;
res = darp(y,x,xc,yc,option);
output(i,1) = res.gamma(1);
output(i,2) = res.cprob(1);
end;
toc;
in.cnames = strvcat(gamma estimate,marginal probability);
in.rflag = 1;
mprint(output,in)
We use the MATLAB tic and toc commands to time the operation
of producing these maximum likelihood estimates across the entire sample.
The results are shown below, where we find that it took around 70 seconds
to solve the maximum likelihood estimation problem, calculate expansion
estimates and produce all of the ancillary statistics 49 times, once for each
observation in the data set.
elapsed_time = 69.2673
Obs#
gamma estimate marginal probability
1
-0.2198
0.0714
2
-0.2718
0.0494
3
-0.3449
0.0255
4
-0.4091
0.0033
5
-0.2223
0.0532
6
-0.3040
0.0266
7
-0.4154
0.0126
8
-0.2071
0.1477
9
-0.5773
0.0030
10
0.1788
0.1843
11
0.1896
0.1526
12
0.1765
0.1621
13
0.1544
0.1999
14
0.1334
0.2214
15
0.1147
0.2708
16
0.1429
0.2615
17
0.1924
0.2023
18
-0.1720
0.3112
19
0.1589
0.3825
20
-0.3471
0.0810
21
0.2020
0.2546
22
0.1862
0.2809

23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
0.1645
0.0904
0.1341
0.1142
0.1150
0.0925
0.1070
-0.2765
0.0453
-0.6580
-0.3293
-0.5949
-0.8133
-0.5931
-0.4853
-0.4523
-0.5355
-0.6050
-0.6804
-0.7257
-0.7701
-0.5150
-0.3997
-0.3923
-0.3214
-0.3586
-0.4668
153
0.3334
0.6219
0.4026
0.4264
0.4618
0.5584
0.5329
0.1349
0.8168
0.0012
0.0987
0.0024
0.0000
0.0023
0.0066
0.0121
0.0016
0.0005
0.0001
0.0001
0.0000
0.0001
0.0005
0.0003
0.0004
0.0001
0.0000
From the estimated values of and the associated marginal probabilities, we infer that the model suffers from performance drift over the initial 9
observations and sample observations from 32 to 49. We draw this inference
from the negative estimates that are statistically significant for these observations. (Note that observation #8 is only marginally significant.) Over
the middle range of the sample, from observations 10 to 31 we find that
the deterministic distance expansion relationship works well. This inference
arises from the fact that none of the estimated parameters are significantly
different from zero.
In Section 4.4 we provide evidence that observations 2 and 4 represent
outliers that impact on estimates for neighboring observations 1 to 9. We
also show that this is true of observation 34, which influences observations
31 to 44. This suggests that the DARP model is working correctly to spot
places where the model encounters problems.
4.3
154
Non-parametric locally linear models
These models represent an attempt to draw on the flexibility and tractability

of non-parametric estimators. Note that the use of the spatial expansion and
DARP methods in the previous section did not involve matrix manipulations
or inversion of large sparse matrices. The models presented in this section
share that advantage over the spatial autoregressive models.
Locally linear regression methods introduced in McMillen (1996,1997)
and labeled geographically weighted regressions (GWR) in Brunsdon, Fotheringham and Charlton (1996) (BFC hereafter) are discussed in this section.
The main contribution of the GWR methodology is the use of distanceweighted sub-samples of the data to produce locally linear regression estimates for every point in space. Each set of parameter estimates is based on
a distance-weighted sub-sample of neighboring observations, which has a
great deal of intuitive appeal in spatial econometrics. While this approach
has a definite appeal, it also presents some problems that are discussed in
Section 4.4, where a Bayesian approach is used to overcome the problems.
The distance-based weights used by BFC for data at observation i take
the form of a vector Wi which can be determined based on a vector of
distances di between observation i and all other observations in the sample.
This distance vector along with a distance decay parameter are used to
construct a weighting function that places relatively more weight on sample
observations from neighboring observations in the spatial data sample.
A host of alternative approach have been suggested for constructing the
weight function. One approach suggested by BFC is:
Wi2 = exp(di /)
(4.19)
The parameter is a decay parameter that BFC label bandwidth.

Changing the bandwidth results in a different exponential decay profile,
which in turn produces estimates that vary more or less rapidly over space.
Another weighting scheme is the tri-cube function proposed by McMillen
(1998):
Wi2 = (1 (/di )3 )3
(4.20)
Still another approach is to rely on a Gaussian function :

Wi2 = (di/)
(4.21)
Where denotes the standard normal density and represents the standard
deviation of the distance vector di .
155
The distance vector is calculated for observation i as:

q
di =
(Zxi Zxj )2 + (Zyi Zyj )2
(4.22)
Where Zxj , Zyj denote the latitude-longitude coordinates of the observations j = 1, . . ., n.

Note that the notation used here may be confusing as we usually rely on
subscripted variables to denote scalar elements of a vector. In the notation
used here, the subscripted variable di represents a vector of distances between observation i and all other sample data observations. Similarly, the
Wi is a vector of distance-based weights associated with observation i.
BFC use a single value of , the bandwidth parameter for all observations
determined using a cross-validation procedure that is often used in locally
linear regression methods. A score function taking the form:
n
X
[yi y6=i ()]2
(4.23)
i=1
is used, where y6=i () denotes the fitted value of yi with the observations for
point i omitted from the calibration process. A value of that minimizes
this score function is used as the distance-weighting bandwidth to produce
GWR estimates.
The non-parametric GWR model relies on a sequence of locally linear
regressions to produce estimates for every point in space by using a subsample of data information from nearby observations. Let y denote an nx1
vector of dependent variable observations collected at n points in space, X
an nxk matrix of explanatory variables, and an nx1 vector of normally
distributed, constant variance disturbances. Letting Wi represent an nxn
diagonal matrix containing distance-based weights for observation i that
reflect the distance between observation i and all other observations, we can
write the GWR model as:
1/2
Wi
1/2
y = Wi
Xi + i
(4.24)
The subscript i on i indicates that this kx1 parameter vector is associated with observation i. The GWR model produces n such vectors of
parameter estimates, one for each observation. These estimates are produced using:
i = (X 0 Wi X)1 (X 0Wi y)
(4.25)
156
Keep in mind the notation, Wi y denotes an n-vector of distance-weighted

observations used to produce estimates for observation i. Note also, that
Wi X represents a distance-weighted data matrix, not a single observation
and i represents an n-vector.
Note that these GWR estimates for i are conditional on the parameter
. That is, changing , the distance decay parameter will produce a different
set of GWR estimates.
The best way to understand this approach to dealing with spatial heterogeneity is to apply the method, a subject to which we turn in the next
section.
4.3.1
Implementing GWR
We have an optimization problem to solve regarding minimizing the score

function to find the cross-validation bandwidth parameter . We first construct a MATLAB function to compute the scores associated with different
bandwidths. This univariate function of the scalar bandwidth parameter
can then be minimized using the simplex algorithm fmin.
Given the optimal bandwidth, estimation of the GWR parameters
and associated statistics can proceed via generalized least-squares. A function gwr whose documentation is shown below implements the estimation
procedure.
PURPOSE: compute geographically weighted regression
---------------------------------------------------USAGE: results = gwr(y,x,east,north,info)
where:
x = explanatory variable matrix
east = x-coordinates in space
north = y-coordinates in space
info = a structure variable with fields:
info.bwidth = scalar bandwidth to use or zero
for cross-validation estimation (default)
info.dtype = gaussian
for Gaussian weighting (default)
= exponential for exponential weighting
NOTE: res = gwr(y,x,east,north) does CV estimation of bandwidth
--------------------------------------------------RETURNS: a results structure
results.meth = gwr
results.beta = bhat matrix
(nobs x nvar)
results.tstat = t-stats matrix (nobs x nvar)
results.yhat = yhat
results.sige = ee/(n-dof) (nobs x 1)
157
results.nobs = nobs
results.bwidth = bandwidth
results.dtype
= input string for Gaussian, exponential weights
results.iter
= # of simplex iterations for cv
results.north = north (y-coordinates)
results.east = east (x-coordinates)
results.y
= y data vector
--------------------------------------------------NOTES: uses auxiliary function scoref for cross-validation
---------------------------------------------------
The following program illustrates using the gwr function on the Anselin
(1988) neighborhood crime data set to produce estimates based on both
Gaussian and exponential weighting functions. Figure 4.5 shows a graph of
these two sets of estimates, indicating that they are not very different.
% ----- example 4.4 Using the gwr() function
% load the Anselin data set
load anselin.dat; y = anselin(:,1); nobs = length(y);
x = [ones(nobs,1) anselin(:,2:3)]; tt=1:nobs;
north = anselin(:,4); east = anselin(:,5);
info.dtype = gaussian;
% Gaussian weighting function
res1 = gwr(y,x,east,north,info);
info.dtype = exponential; % Exponential weighting function
res2 = gwr(y,x,east,north,info);
subplot(3,1,1), plot(tt,res1.beta(:,1),tt,res2.beta(:,1),--);
legend(Gaussian,Exponential); ylabel(Constant term);
legend(Gaussian,Exponential); ylabel(Household income);
legend(Gaussian,Exponential); ylabel(House value);
The printed output for these models is voluminous as illustrated below,

where we only print estimates associated with two observations.
Geometrically weighted regression estimates
crime
R-squared
=
0.9418
Rbar-squared =
0.9393
Bandwidth
=
0.6518
Decay type
=
gaussian
# iterations =
17
Nobs, Nvars
=
49,
3
***************************************
Obs =
1, x-coordinate= 42.3800, y-coordinate= 35.6200, sige=
Variable
Coefficient
t-statistic
t-probability
1.1144

const
income
hse value
51.198618
-0.461074
-0.434240
16.121809
-2.938009
-6.463775
158
0.000000
0.005024
0.000000
Obs =
2, x-coordinate= 40.5200, y-coordinate= 36.5000, sige=
Variable
Coefficient
t-statistic
t-probability
const
63.563830
15.583144
0.000000
income
-0.369869
-1.551568
0.127201
hse value
-0.683562
-7.288304
0.000000
Constant term
100
2.7690
Gaussian
Exponential
80
60
40
20
10
15
20
25
30
35
40
45
50
Household income
2
0
-2
-4
-6
Gaussian
Exponential
0
10
15
20
25
30
35
40
45
50
House value
1
0
-1
Gaussian
Exponential
-2
10
15
20
25
30
35
40
45
50
Figure 4.5: Comparison of GWR distance weighting schemes
4.4
A Bayesian Approach to GWR
A Bayesian treatment of locally linear geographically weighted regressions

is set forth in this section. While the use of locally linear regression seems
appealing, it is plagued by some problems. A Bayesian treatment can resolve
159
these problems and has some advantages over the non-parametric approach
discussed in Section 4.3.
One problem with the non-parametric approach is that valid inferences
cannot be drawn for the GWR regression parameters. To see this, consider
that the locally linear estimates use the same sample data observations (with
different weights) to produce a sequence of estimates for all points in space.
Given a lack of independence between estimates for each location, conventional measures of dispersion for the estimates will likely be incorrect. These
(invalid) conventional measures are what we report in the results structure
from gwr, as this is the approach taken by Brunsdon, Fotheringham and
Charlton (1996).
Another problem is that non-constant variance over space, aberrant observations due to spatial enclave effects, or shifts in regime can exert undue
influence on locally linear estimates. Consider that all nearby observations
in a sub-sequence of the series of locally linear estimates may be contaminated by an outlier at a single point in space. The Bayesian approach
introduced here solves this problem by robustifying against aberrant observations. Aberrant observations are automatically detected and downweighted to lessen their influence on the estimates. The non-parametric
implementation of the GWR model assumed no outliers.
A third problem is that the non-parametric estimates may suffer from
weak data problems because they are based on a distance weighted subsample of observations. The effective number of observations used to produce estimates for some points in space may be very small. This problem
can be solved with the Bayesian approach by incorporating subjective prior
information during estimation. Use of subjective prior information is a wellknown approach for overcoming weak data problems.
In addition to overcoming these problems, the Bayesian formulation introduced here specifies the relationship that is used to smooth parameters
over space. This allows us to subsume the non-parametric GWR method as
part of a much broader class of spatial econometric models. For example,
the Bayesian GWR can be implemented with parameter smoothing relationships that result in: 1) a locally linear variant of the spatial expansion
methods discussed in section 4.1, 2) a parameter smoothing relation appropriate for monocentric city models where parameters vary systematically
with distance from the center of the city, 3) a parameter smoothing scheme
based on contiguity relationships, and 4) a parameter smoothing scheme
based on distance decay.
The Bayesian approach, which we label BGWR is best described using
matrix expressions shown in (4.26) and (4.27). First, note that (4.26) is the
160
same as the non-parametric GWR relationship, but the addition of (4.27)

provides an explicit statement of the parameter smoothing that takes place
across space. The parameter smoothing involves a locally linear combination
of neighboring areas, where neighbors are defined in terms of the distance
weighting function that decays over space.
1/2
Wi
1/2
y = Wi

i =
Xi + i
wi1 Ik . . . win Ik
..
. + ui
(4.26)
(4.27)
n
The terms wij represent a normalized distance-based weight so the rowvector (wi1 , . . ., win ) sums to unity, and we set wii = 0. That is, wij =
P
exp(dij /)/ nj=1 exp(dij /).
To complete our model specification, we add distributions for the terms
i and ui :
i N [0, 2Vi ],
Vi = diag(v1 , v2 , . . . , vn )
0
ui N [0, (X Wi X)
2 2
)]
(4.28)
(4.29)
The Vi = diag(v1 , v2 , . . . , vn ), represent our n variance scaling parameters from Chapter 3. These allow for non-constant variance as we move
across space. One point to keep in mind is that here we have n2 terms to estimate, reflecting n Vi vectors, one vector for each of the n observations. We
will use the same assumption as in Chapter 3 regarding the Vi parameters.
All n2 parameters are assumed to be i.i.d. 2 (r) distributed, where r is our
hyperparameter that controls the amount of dispersion in the Vi estimates
across observations. As in Chapter 3, we introduce a single hyperparameter
r to the estimation problem and receive in return n2 parameter estimates.
Consider that as r becomes very large, the prior imposes homoscedasticity
on the BGWR model and the disturbance variance becomes 2 In for all
observations i.
The distribution for ui in the parameter smoothing relationship is normally distributed with mean zero and a variance based on Zellners (1971)
gprior. This prior variance is proportional to the parameter variancecovariance matrix, 2 (X 0 Wi X)1 ) with 2 acting as the scale factor. The
use of this prior specification allows individual parameters i to vary by
different amounts depending on their magnitude.
161
The parameter 2 acts as a scale factor to impose tight or loose adherence to the parameter smoothing specification. Consider a case where
is very small, then the smoothing restriction would force i to look like
a distance-weighted linear combination of other i from neighboring observations. On the other hand, as (and Vi = In ) we produce the
non-parametric GWR estimates. To see this, we rewrite the BGWR model
in a more compact form:
i i + i
yi = X
(4.30)
i = Ji + ui
(4.31)
Where the definitions of the matrix expressions are:
1/2
yi = Wi y
i = W 1/2X
X
i

Ji =
wi1 Ik . . . win Ik
1
..
= .
n
As indicated earlier, the notation is somewhat confusing in that yi denotes an nvector, not a scalar magnitude. Similarly, i is an nvector and
i is an n by k matrix. Note that (4.30) can be written in the form of a
X
Theil and Goldberger (1961) estimation problem as shown in (4.32).
yi
Ji
i
X
Ik
i +
i
ui
(4.32)
Assuming Vi = In , the estimates i take the form:

2
0yi + X
0X
i = R(X
i
i i Ji / )
0X
i + X
i / 2 )1
0X
R = (X
i
As approaches , the terms associated with the Theil and Goldberger

2
0X
0 2
stochastic restriction, X
i iJi / and Xi Xi/ become zero, and we have
the GWR estimates:
i )1 (X
i0X
i0yi )
i = (X
162
(4.33)
In practice, we can use a diffuse prior for which allows the amount of
parameter smoothing to be estimated from sample data information, rather
than by subjective prior information.
Details concerning estimation of the parameters in the BGWR model
are taken up in the next section. Before turning to these issues, we consider
some alternative spatial parameter smoothing relationships that might be
used in lieu of (4.27) in the BGWR model.
One alternative smoothing specification would be the monocentric city
smoothing set forth in (4.34). This relation assumes that the data observations have been ordered by distance from the center of the spatial sample.
i = i1 + ui
(4.34)
0
ui N [0, (X Wi X)
2 2
Given that the observations are ordered by distance from the center, the
smoothing relation indicates that i should be similar to the coefficient i1
from a neighboring concentric ring. Note that we rely on the same GWR
distance-weighted data sub-samples, created by transforming the data using:
Wi y, WiX. This means that the estimates still have a locally linear interpretation as in the GWR. We rely on the same distributional assumption
for the term ui from the BGWR which allows us to estimate the parameters from this model by making minor changes to the approach used for the
BGWR.
Another alternative is a spatial expansion smoothing based on the
ideas introduced by Casetti (1972). This is shown in (4.35), where Zxi , Zyi
denote latitude-longitude coordinates associated with observation i.

i =
Zxi Ik Zyi Ik
x
y
+ ui
(4.35)
ui N [0, 2 2 (X 0 Wi X)1 )]
This parameter smoothing relation creates a locally linear combination
based on the latitude-longitude coordinates of each observation. As in the
case of the monocentric city specification, we retain the same assumptions
regarding the stochastic term ui , making this model simple to estimate with
minor changes to the BGWR methodology.
163
Finally, we could adopt a contiguity smoothing relationship based on

a first-order spatial contiguity matrix as shown in (4.36). The terms cij represent the ith row of a row-standardized first-order contiguity matrix. This
creates a parameter smoothing relationship that averages over the parameters from observations that neighbor observation i.

i =
ci1 Ik . . . cin Ik
1
..
. + ui
n
(4.36)
ui N [0, 2(X 0 Wi2 X)1 )]

Alternative approaches to specifying geographically weighted regression
models suggest that researchers need to think about which type of spatial
parameter smoothing relationship is most appropriate for their application.
Additionally, where the nature of the problem does not clearly favor one
approach over another, statistical tests of alternative models based on different smoothing relations might be carried out. Posterior odds ratios can
be constructed that will shed light on which smoothing relationship is most
consistent with the sample data. We illustrate model specification issues in
an applied example in Section 4.5.
4.4.1
Estimation of the BGWR model
We use Gibbs sampling to estimate the BGWR model. This approach is

particularly attractive in this application because the conditional densities
all represent know distributions that are easy to obtain. In Chapter 3 we
saw an example of Gibbs sampling where the conditional distribution for the
spatial autoregressive parameters were from an unknown distribution and
we had to rely on the more complicated case of Metropolis within-Gibbs
sampling.
To implement the Gibbs sampler we need to derive and draw samples
from the conditional posterior distributions for each group of parameters,
i , , , and Vi in the model. Let P (i |, , Vi, ) denote the conditional
density of i , where represents the values of other j for observations
j 6= i. Using similar notation for the the other conditional densities, the
Gibbs sampling process can be viewed as follows:
1. start with arbitrary values for the parameters i0 , 0 , 0 , Vi0 , 0
2. for each observation i = 1, . . . , n,
164
(a) sample a value, i1 from P (i | 0 , 0 , Vi0 , 0)

(b) sample a value, Vi1 from P (Vi |i1 , 0, 0 , 0)
3. use the sampled values i1 , i = 1, . . . , n from each of the n draws above
to update 0 to 1.
4. sample a value, 1 from P (| 0 , Vi1 , 1)
5. sample a value, 1 from P (| 1 , Vi1 , 1)
6. go to step 1 using i1 , 1 , 1 , Vi1 , 1 in place of the arbitrary starting
values.
The sequence of draws outlined above represents a single pass through
the sampler, and we make a large number of passes to collect a large sample
of parameter values from which we construct our posterior distributions.
We rely on the compact statement of the BGWR model in (4.30) to
facilitate presentation of the conditional distributions that we rely on during
the sampling.
The conditional posterior distribution of i given , , and Vi is a multivariate normal shown in (4.37).
p(i| . . .) N (i, 2R)
(4.37)
Where:
2
0V 1 yi + X
0X
i = R(X
i i
i i Ji / )
i0V 1 X
i + X
i / 2 )1
i0X
R = (X
i
(4.38)
This result follows from the assumed variance-covariance structures for
i , ui and the Theil-Goldberger (1961) representation shown in (4.32).
The conditional posterior distribution for is a 2 (m = n2 ) distribution
shown in (4.39).
p(| . . .) (m+1) exp{

i i
i = yi X
n
1 X
(0 V 1 i )}
2 2 i=1 i i
(4.39)
165
The sum in (4.39) extends over the subscript i to indicate that the n
vector of the squared residuals (deflated by the n individual Vi terms) from
each sub-sample of n observations are summed, and then these n sums are
summed as well.
The conditional posterior distribution for Vi is shown in (4.40), which indicates that we draw an n-vector based on a 2 (r+1) distribution. Note that
the individual elements of the matrix Vi act on the spatial weighting scheme
0 V 1 X
i = X 0 Wi V 1 Wi X. The
because the estimates
involve terms like: X
i i
i
p
terms Wi = exp(di /) from the weighting scheme will be adjusted by
the Vi estimates, which are large for aberrant observations or outliers. In
the event of an outlier, observation i will receive less weight when the spatial
distance-based weight is divided by a large Vi value.
p{[(e2i / 2 ) + r]/Vi | . . .} 2 (r + 1)
(4.40)
Finally, the conditional distribution for is a 2 (nk) distribution based

on (4.41).
p(| . . .) nk exp{
n
X
2 2
1
0X
(i Ji )0 (X
i i ) (i Ji )/2 }
(4.41)
i=1
Now consider the modifications needed to the conditional distributions

to implement the alternative spatial smoothing relationships set forth in
Section 4.4.1. Since we maintained the assumptions regarding the disturbance terms i , ui, we need only alter the conditional distributions for i and
. First, consider the case of the monocentric city smoothing relationship.
The conditional distribution for i is multivariate normal with mean i and
variance-covariance 2 R as shown in (4.42).
2
0 V 1 yi + X
0X
i = R(X
i i
i i i1 / )
0 V 1 X
i + X
i/ 2 )1
0X
R = (X
i i
(4.42)
The conditional distribution for is a 2 (nk) based on the expression in

(4.43).
p(| . . .) nk exp{
n
X
i )1 (i i1 )/ 2 2 }
i0X
(i i1 )0 (X
(4.43)
i=1
For the case of the spatial expansion and contiguity smoothing relationships, we can maintain the conditional expressions for i and from the case
166
of the BGWR, and simply modify the definition of J, to be consistent with

these smoothing relations. One additional change needs to be made for the
case of the spatial expansion smoothing relationship, we need to add a conditional distribution for the parameters x, y in the model. This distribution
is a multivariate normal with mean = (x y )0 and variance-covariance
0 Q1 X
i Ji )1 as defined in (4.44).
matrix 2 (Ji0 X
i
0Q1 X
iJi )1 (J 0 X
0 1 i )
= (Ji0 X
i
i iQ y
i )1 X
0 / 2 )
i (X
0X
Q = (Vi + X
i
4.4.2
(4.44)
Informative priors
Implementing the BGWR model with diffuse priors on may lead to large
values that essentially eliminate the parameter smoothing relationship from
the model. The BGWR estimates will then collapse on the GWR estimates
(in the case of a large value for the hyperparameter r that leads to Vi = In ).
In cases where the sample data is weak or objective prior information suggests spatial parameter smoothing should follow a particular specification,
we can use an informative prior for the parameter . A Gamma(a, b) prior
distribution which has a mean of a/b and variance of a/b2 seems appropriate. Given this prior, we could eliminate the conditional density for and
replace it with a random draw from the Gamma(a, b) distribution.
In order to devise an appropriate prior setting for , consider that the
1 , so setting values for > 1
0 X)
GWR variance-covariance matrix is: 2 (X
would represent a relatively loose imposition of the parameter smoothing
relationship. Values of < 1 would impose the parameter smoothing prior
more tightly.
A similar approach can be taken for the hyperparameter r. Using a
Gamma prior distribution with a = 8, b = 2 that indicates small values of r
around 4, should provide a fair amount of robustification if there is spatial
heterogeneity. In the absence of heterogeneity, the resulting Vi estimates
will be near unity so the BGWR distance weights will be similar to those
from GWR, even with a small value of r.
Additionally, a 2 (c, d) natural conjugate prior for the parameter could
be used in place of the diffuse prior set forth here. This would affect the
conditional distribution used during Gibbs sampling in only a minor way.
Some other alternatives offer additional flexibility when implementing
the BGWR model. For example, one can restrict specific parameters to
exhibit no variation over the spatial sample observations. This might be
167
useful if we wish to restrict the constant term to be constant over space.

Or, it may be that the constant term is the only parameter that would be
allowed to vary over space.
These alternatives can be implemented by adjusting the prior variances
in the parameter smoothing relationship:
1
0X
var cov(i) = 2 2 (X
i i)
(4.45)
For example, assuming the constant term is in the first column of the matrix
i , setting the first row and column elements of (X
1 to zero would
0X
X
i i)
restrict the intercept term to remain constant over all observations.
4.4.3
Implementation details
We have devised a function bgwr to carry out Gibbs sampling estimation of

the Bayesian GWR model. The documentation for the function is shown below, where a great many user-supplied options are available. These options
are input using a structure variable named prior with which alternative
types of parameter smoothing relationships can be indicated. Note that only
three of the four parameter smoothing relationships discussed in Section 4.4
are implemented. The Casetti spatial expansion parameter smoothing relationship is not yet implemented. Another point to note is that you can
implement a contiguity smoothing relationship by either specifying a spatial
weight matrix or relying on the function to calculate this matrix based on
the x-y coordinates using the function xy2cont discussed in Chapter 2.
PURPOSE: compute Bayesian geographically weighted regression
model: y = Xb(i) + e,
e = N(0,sige*V),
b(i) = f[b(j)] + u, u = delta*sige*inv(xx)
V = diag(v1,v2,...vn),
r/vi = ID chi(r)/r,
delta = gamma(s,t),
r = Gamma(m,k)
f[b(j)] = b(i-1) for concentric city prior
f[b(j)] = W(i) b for contiguity prior
f[b(j)] = [exp(-d/c)/sum(exp(-d/c)] b for distance prior
c = GWR c.v. determined bandwidth
---------------------------------------------------USAGE: results = bgwr(y,x,xcoord,ycoord,ndraw,nomit,prior)
x = explanatory variable matrix
xcoord = x-coordinates in space
ycoord = y-coordinates in space
prior = a structure variable with fields:
prior.rval, improper r value, default=4
prior.m,
168
prior.k,
(default: not used)
prior.dval, improper delta value (default=diffuse)
prior.s,
informative Gamma(s,t) prior on delta
prior.t,
(default: not used)
prior.ptype, concentric for concentric city smoothing
distance for distance based smoothing (default)
contiguity for contiguity smoothing
casetti
for casetti smoothing (not implemented)
prior.ctr, observation # of central point (for concentric prior)
prior.W,
(optional) prior weight matrix (for contiguity prior)
ndraw = # of draws
results.meth
= bgwr
results.bdraw
= beta draws (ndraw-nomitxnobsxnvar) (3-d matrix)
results.sdraw
= sige draws (ndraw-nomit x 1)
results.vmean
= mean of vi draws (1 x nobs)
results.rdraw
= r-value draws (ndraw-nomit x 1)
results.ddraw
= delta draws (if diffuse prior used)
results.r
results.d
= value of hyperparameter delta (if input)
results.m
= m prior parameter (if input)
results.k
= k prior parameter (if input)
results.s
= s prior parameter (if input)
results.t
= t prior parameter (if input)
results.nobs
= nobs
results.nvar
= nvars
results.ptype
= input string for parameter smoothing relation
results.xcoord = x-coordinates
results.ycoord = y-coordinates
results.ctr
= central point observation # (if concentric prior)
results.dist
= distance vector (if ptype = 0)
results.y
= y data vector
results.logpost = (nobs x 1) vector of approximate log posterior
results.time
= time taken for sampling
--------------------------------------------------NOTE: use either improper prior.rval
or informative Gamma prior.m, prior.k, not both of them
uses exponential distance weighting function exp(-d/bwdith)
where d=distance and bwidth = gwr c.v. determined bandwidth
The user also has control over options for assigning a prior to the hyperparameter r that robustifies with respect to outliers and accommodates
non-constant variance. Either an improper prior value can be set (as a ruleof-thumb I recommend r = 4), or a proper prior based on a Gamma(m,k)
distribution can be used. Here, one would try to rely on a prior in the range
of 4 to 10, because larger values produce estimates that are not robust to
169
heteroscedasticity or outliers. As an example, m = 8, k = 2 would implement a prior with the mean of r = 4 and the variance of r = 2, since the
mean of the Gamma distribution is m/k, and the variance is (m/k2 ).
The hyperparameter can be handled in three ways: 1) we can simply assign an improper prior value using say, prior.dval=20 as an input
option, 2) we can input nothing about this parameter producing a default
implementation based on a diffuse prior where will be estimated, and 3)
we can assign a Gamma(s,t) prior as in the case of the hyperparameter r.
Implementation with a diffuse prior for and a large value for the hyperparameter r will most likely reproduce the non-parameter GWR estimates,
and this approach to producing those estimates requires more computing
time. It is possible (but not likely) that a model and the sample data are
very consistent with the parameter smoothing relationship. If this occurs,
a diffuse prior for will produce a relatively small value as the posterior
estimate. In the most likely cases encountered in practice, small deviations
of the parameters from the smoothing relationship will lead to very large
estimates for , producing BGWR parameter estimates that come very close
to those from the non-parametric GWR model.
The value of the Bayesian approach outlined here lies in the ability to robustifying against outliers, so a default value of r = 4 has been implemented
if the user enters no information regarding the hyperparameter r.
Consider how the following alternative implementations of the various
prior settings could be used to shed light on the nature of parameter variation over space. We can compare the results from a GWR model to a
BGWR model implemented with r = 4 and either a diffuse prior for or
an improper prior based on large value. This comparison should show the
impact of robustification on the estimates, and a plot of the Vi estimates
can be used to detect outliers. Another model based on r = 4 along with
an informative prior for 1 that places some weight on the parameter
smoothing restrictions can be used to see how this alters the estimates when
compared to the robust BGWR estimates. A dramatic difference between
the robust BGWR estimates and those based on the informative prior for
1 indicates that the parameter smoothing relation is inconsistent with
the sample data.
It may be necessary to experiment with alternative values of because
the scale is unclear in any given problem. One way to deal with the scale issue is to calibrate based on the diffuse estimate. As an example, if 2 = 10,
then a value of < 10 will impose the parameter smoothing restriction more
tightly. To see this, consider that the GWR variance-covariance matrix is
1 , so using < 1 moves in the direction of tightening the pa 0 X)
2 (X
170
rameter smoothing prior. We will provide an example of this in the next

section.
The function bgwr returns the entire sequence of draws made for the
parameters in the problem, allowing one to: check for convergence of the
Gibbs sampler, plot posterior density functions using a function pltdens
from the Econometrics Toolbox or compute statistics from the posterior
distributions of the parameters. Most users will rely on the prt function
that calls a related function to produce printed output very similar to the
results printed for the gwr function. If you do want to access the estimates,
note that they are stored in a MATLAB 3-dimensional matrix structure. We
illustrate how to access these in Example 4.5 of the next section.
4.5
An applied exercise
The program in example 4.5 shows how to use the bgwr function to produce
estimates for the Anselin neighborhood crime data set. We begin by using
an improper prior for = 1000000 and setting r = 30 to demonstrate that
the BGWR model can replicate the estimates from the GWR model. The
program plots these estimates for comparison. We produce estimates for
all three parameter smoothing priors, but given the large = 1000000,
these relationships should not effectively enter the model. This will result in
all three sets of estimates identical to those from the GWR. We did this to
illustrate that the BGWR can replicate the GWR estimates with appropriate
settings for the hyperparameters.
% ----- example 4.5 Using the bgwr() function
load anselin.data; % load the Anselin data set
y = anselin(:,1); nobs = length(y); x = [ones(nobs,1) anselin(:,2:3)];
east = anselin(:,4); north = anselin(:,5); tt=1:nobs;
ndraw = 250; nomit = 50;
prior.ptype = contiguity; prior.rval = 30; prior.dval = 1000000;
tic; r1 = bgwr(y,x,east,north,ndraw,nomit,prior); toc;
prior.ptype = concentric; prior.ctr = 20;
dist = res2.dist; [dists di] = sort(dist); % recover distance vector
prior.ptype = distance;
vnames = strvcat(crime,constant,income,hvalue);
% compare gwr estimates with posterior means
info2.dtype = exponential;
result = gwr(y,x,east,north,info2);
bgwr = result.beta(di,:);
b1 = r1.bdraw(:,di,1); b2 = r1.bdraw(:,di,2); b3 = r1.bdraw(:,di,3);
171
b1m = mean(b1); b2m = mean(b2); b3m = mean(b3);

c1 = r2.bdraw(:,:,1); c2 = r2.bdraw(:,:,2); c3 = r2.bdraw(:,:,3);
c1m = mean(c1); c2m = mean(c2); c3m = mean(c3);
d1 = r3.bdraw(:,di,1); d2 = r3.bdraw(:,di,2); d3 = r3.bdraw(:,di,3);
d1m = mean(d1); d2m = mean(d2); d3m = mean(d3);
% plot mean of vi draws (sorted by distance from #20)
plot(tt,r1.vmean(1,di),-b,tt,r2.vmean,--r,tt,r3.vmean(1,di),-.k);
title(vi means); legend(contiguity,concentric,distance);
pause;
% plot beta estimates (sorted by distance from #20)
subplot(3,1,1),
plot(tt,bgwr(:,1),-k,tt,b1m,--k,tt,c1m,-.k,tt,d1m,:k);
legend(gwr,contiguity,concentric,distance);
xlabel(b1 parameter);
subplot(3,1,2),
subplot(3,1,3),
As we can see from the graph of the GWR and BGWR estimates shown
in Figure 4.6, the three sets of BGWR estimates are nearly identical to the
GWR. Keep in mind that we dont recommend using the BGWR model
to replicate GWR estimates, as this problem involving 250 draws took 193
seconds for the contiguity prior, 185 for the concentric city prior and 190
seconds for the distance prior.
Example 4.6 produces estimates based on the contiguity smoothing relationship with a value of r = 4 for this hyperparameter to indicate a prior
belief in heteroscedasticity or outliers. We keep the parameter smoothing
relationship from entering the model by setting an improper prior based
on = 1000000, so there is no need to run more than a single parameter
smoothing model. Estimates based on any of the three parameter smoothing
relationships would produce the same results because the large value for
keeps this relation from entering the model. The focus here is on the impact
of outliers in the sample data and how BGWR robust estimates compare
to the GWR estimates based on the assumption of homoscedasticity. We
specify 250 draws with the first 50 to be discarded for burn-in of the Gibbs
sampler. Figure 4.7 shows a graph of the mean of the 200 draws which represent the posterior parameter estimates, compared to the non-parametric
estimates.
% ----- example 4.6 Producing robust BGWR estimates
172
load anselin.data;
east = anselin(:,4); north = anselin(:,5);
prior.ptype = contiguity; prior.rval = 4; prior.dval = 1000000;
% use diffuse prior for distance decay smoothing of parameters
result = bgwr(y,x,east,north,ndraw,nomit,prior);
info.dtype = exponential;
result2 = gwr(y,x,east,north,info);
% compare gwr and bgwr estimates
b1 = result.bdraw(:,:,1); b1mean = mean(b1);
betagwr = result2.beta;
tt=1:nobs;
subplot(3,1,1),
plot(tt,betagwr(:,1),-k,tt,b1mean,--k);
100
50
gwr
contiguity
concentric
distance
0
10
15
20
25
30
b1 parameter
35
40
45
50
10
15
20
25
30
b2 parameter
35
40
45
50
10
15
20
25
30
b3 parameter
35
40
45
50
2
0
-2
-4
1
0
-1
-2
Figure 4.6: GWR and BGWR diffuse prior estimates
173
legend(gwr,bgwr);
subplot(3,1,2),
subplot(3,1,3),
pause;
plot(result.vmean);
ylabel(V_{i} estimates);
100
80
60
40
20
gwr
bgwr
0
10
15
20
25
30
b1 parameter
35
40
45
50
10
15
20
25
30
b2 parameter
35
40
45
50
10
15
20
25
30
b3 parameter
35
40
45
50
2
0
-2
-4
1
0
-1
-2
Figure 4.7: GWR and Robust BGWR estimates

We see a departure of the two sets of estimates around the first 10
observations and around observations 30 to 45. To understand why, we
need to consider the Vi terms that represent the only difference between the
BGWR and GWR models given that we used a large value. Figure 4.9
shows the mean of the draws for the Vi parameters. Large estimates for the
174
Vi terms at observations 2, 4 and 34 indicate aberrant observations. This

accounts for the difference in trajectory taken by the non-parametric GWR
estimates and the Bayesian estimates that robustify against these aberrant
observations.
To illustrate how robustification takes place, Figure 4.8 shows the weighting terms Wi from the GWR model plotted alongside the weights adjusted
1/2
1/2
by the Vi terms, Wi Vi1 Wi from the BGWR model. A sequence of six
observations from 30 to 35 are plotted, with a symbol o placed at observation #34 on the BGWR weights to help distinguish this observation in the
figure.
Beginning with observation #30, the aberrant observation #34 is downweighted when estimates are produced for observations #30 to #35. This
downweighting of the distance-based weight for observation #34 occurs during estimation of i for observations #30 through #35, all of which are near
#34 in terms of the GWR distance measure. This alternative weighting
produces the divergence between the GWR and BGWR estimates that we
observe in Figure 4.7 starting around observation #30.
1.5
BGWR
GWR
Observation 31
Observation 30
0.5
1
0.5
O
0
10
20
30
40
50
1
0.5
0
O
0
10
20
30
40
30
10
20
30
10
20
30
40
50
40
50
40
50
0.5
O
1.5
Observation 35
Observation 34
20
50
0.5
10
1.5
Observation 33
Observation 32
1.5
O
0
O
0
10
20
30
40
50
1
0.5
0
Figure 4.8: GWR and BGWR Distance-based weights adjusted by Vi
175
Ultimately, the role of the parameters Vi in the model and the prior
assigned to these parameters reflects our prior knowledge that distance alone
may not be reliable as the basis for spatial relationships between variables.
If distance-based weights are used in the presence of aberrant observations,
inferences will be contaminated for whole neighborhoods and regions in our
analysis. Incorporating this prior knowledge turns out to be relatively simple
in the Bayesian framework, and it appears to effectively robustify estimates
against the presence of spatial outliers.
2.6
2.4
2.2
1.8
1.6
Vi
-1
estimates
1.4
1.2
0.8
10
15
20
25
30
neighborhoods
35
40
45
50
Figure 4.9: Average Vi estimates over all draws and observations

The function bgwr has associated prt and plt methods to produced
printed and graphical presentation of the results. Some of the printed output
is shown below. Note that the time needed to carry out 550 draws was 289
seconds, making this estimation approach quite competitive to DARP or
GWR. The plt function produces the same output as for the GWR model.
Gibbs sampling geographically weighted regression model
crime
R-squared
=
0.5650
sigma^2
=
0.7296
Nobs, Nvars
=
49,
3
ndraws,nomit
=
550,
50
176
r-value
=
4.0000
delta-value
=
25.0776
gam(m,k) d-prior =
50,
2
time in secs
= 289.2755
prior type
=
distance
***************************************************************
Obs =
1, x-coordinate= 35.6200, y-coordinate= 42.3800
Variable
Coefficient
t-statistic
t-probability
constant
74.130145
15.143596
0.000000
income
-2.208382
-9.378908
0.000000
hvalue
-0.197050
-5.166565
0.000004
Obs =
2, x-coordinate= 36.5000, y-coordinate= 40.5200
Variable
Coefficient
t-statistic
t-probability
constant
82.308344
20.600005
0.000000
income
-2.559334
-11.983369
0.000000
hvalue
-0.208478
-4.682454
0.000023
The next task was to implement the BGWR model with a diffuse prior
on the parameter. The results indicated that the mean of the draws for
was: around 32 for the contiguity prior, 43 for the concentric prior and 33
for the distance prior. The BGWR estimates were almost identical to those
from the improper prior = 1000000, so we do not present these graphically.
Given these estimates for with a diffuse prior, we can impose the parameter restrictions by setting smaller values for , say in the range of 1
to 10. This should produce differing estimates that rely on the alternative
parameter smoothing relationships. We used = 1 to impose the parameter
smoothing relationships fairly tightly. The resulting parameter estimates
are shown in Figure 4.10.
Here we see some departure between the estimates based on alternative smoothing relationships. This raises the question of which smoothing
relationship is most consistent with the data.
We can compute posterior probabilities for each of the three models
based on alternative parameter smoothing relationships using an approximation from Leamer (1983). Since this is a generally useful way of comparing
alternative Bayesian models, the bgwr() function returns a vector of the
approximate log posterior for each observation. The nature of this approximation as well as the computation is beyond the scope of our discussion
here. The program in example 4.7 demonstrates how to use the vector of
log posterior magnitudes to compute posterior probabilities for each model.
We set the parameter = 0.5, to impose the three alternative parameter
smoothing relationships even tighter than the hyperparameter value of = 1
used to generate the estimates in Figure 4.10. A successive tightening of the
177
parameter smoothing relationships will show which relationship is most consistent with the sample data and which relationship is rejected by the sample
data. A fourth model based on = 1000 is also estimated to test whether
the sample data rejects all three parameter smoothing relationships.
% ----- example 4.7 Posterior probabilities for models
load anselin.data; y = anselin(:,1); nobs = length(y);
x = [ones(nobs,1) anselin(:,2:3)]; [junk nvar] = size(x);
east = anselin(:,4); north = anselin(:,5);
ndraw = 550; nomit = 50; % estimate all three models
prior.ptype = contiguity;
prior.rval = 4; prior.dval = 0.5;
res1 = bgwr(y,x,east,north,ndraw,nomit,prior);
prior2.ptype = concentric;
prior2.ctr = 20; prior2.rval = 4; prior2.dval = 0.5;
res2 = bgwr(y,x,east,north,ndraw,nomit,prior2);
prior3.ptype = distance;
prior3.rval = 4; prior3.dval = 0.5;
100
80
60
contiguity
concentric
distance
40
20
10
15
20
25
30
b1 parameter
35
40
45
50
10
15
20
25
30
b2 parameter
35
40
45
50
10
15
20
25
30
b3 parameter
35
40
45
50
0
-1
-2
-3
-4
1
0.5
0
-0.5
Figure 4.10: Alternative smoothing BGWR estimates

prior4.ptype = distance;
prior4.rval = 4; prior4.dval = 1000;
% compute posterior model probabilities
nmodels = 4;
pp = zeros(nobs,nmodels); lpost = zeros(nobs,nmodels);
lpost(:,1) = res1.logpost;
psum = sum(lpost);
for j=1:nmodels
pp(:,j) = lpost(:,j)./psum;
end;
% compute posterior means for beta
bb = zeros(nobs,nvar*nmodels);
b1 = res1.bdraw(:,:,1); bb(:,1) = mean(b1);
c1 = res2.bdraw(:,:,1); bb(:,4) = mean(c1);
d1 = res3.bdraw(:,:,1); bb(:,7) = mean(d1);
e1 = res4.bdraw(:,:,1); bb(:,10) = mean(e1);
tt=1:nobs;
plot(tt,pp(:,1),ok,tt,pp(:,2),*k,tt,pp(:,3),+k, tt,pp(:,4),-);
legend(contiguity,concentric,distance,diffuse);
xlabel(observations); ylabel(probabilities);
pause;
subplot(3,1,1),
plot(tt,bb(:,1),-k,tt,bb(:,4),--k,tt,bb(:,7),-.,tt,bb(:,10),+);
legend(contiguity,concentric,distance,diffuse);
subplot(3,1,2),
subplot(3,1,3),
% produce a Bayesian model averaging set of estimates
bavg = zeros(nobs,nvar); cnt = 1;
for j=1:nmodels
bavg = bavg + matmul(pp(:,j),bb(:,cnt:cnt+nvar-1)); cnt = cnt+nvar;
end;
ttp = tt;
b1out = [ttp bavg(:,1) bb(:,1) bb(:,4) bb(:,7) bb(:,10)];
178
179
in.fmt = strvcat(%4d,%8.2f,%8.2f,%8.2f,%8.2f,%8.2f);
in.cnames = strvcat(Obs,avg,contiguity,concentric,...
distance,diffuse);
fprintf(1,constant term parameter \n);
mprint(b1out,in);
fprintf(1,household income parameter \n);
mprint(b2out,in);
in.fmt = strvcat(%4d,%8.3f,%8.3f,%8.3f,%8.3f,%8.3f);
fprintf(1,house value parameter \n);
mprint(b3out,in);
The graphical display of the posterior probabilities produced by the program in example 4.7 are shown in Figure 4.11 and the parameter estimates
are shown in Figure 4.12. We see some divergence between the parameter estimates produced by the alternative spatial smoothing priors and the model
with no smoothing prior whose estimates are graphed using + symbols,
but not a dramatic amount. Given the similarity of the parameter estimates, we would expect relatively uniform posterior probabilities for the
four models.
In Figure 4.11 the posterior probabilities for the model with no parameter smoothing is graphed as a line to make comparison with the three models
that impose parameter smoothing easy. We see certain sample observations
where the parameter smoothing relationships produce lower model probabilities than the model without parameter smoothing. For most observations
however, the parameter smoothing relationships are relatively consistent
with the sample data, producing posterior probabilities above the model
with no smoothing relationship. As we would expect, none of the models
dominates.
A Bayesian solution to the problem of model specification and choice is
to produce a mixed or averaged model that relies on the posterior probabilities as weights. We would simply multiply the four sets of coefficient
estimates by the four probability vectors to produce a Bayesian model averaging solution to the problem of which estimates are best.
A tabular presentation of this type of result is shown below. The averaged results have the virtue that a single set of estimates are available from
which to draw inferences and one can feel comfortable that the inferences
are valid for a wide variety of model specifications.
constant term parameter
Obs
average contiguity concentric
1
65.92
62.72
63.53
distance
77.64
diffuse
57.16

2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
73.88
78.40
72.94
62.61
73.14
78.72
72.28
75.88
63.02
61.74
59.85
57.39
54.34
52.68
61.30
65.12
71.63
71.78
75.95
71.09
70.93
75.45
68.92
64.05
72.44
75.96
70.77
74.12
60.14
57.78
56.30
55.26
53.01
53.36
58.86
61.18
70.19
66.98
72.70
67.98
69.09
76.31
69.34
59.64
69.78
79.39
72.37
77.12
67.73
65.78
63.28
61.57
55.23
48.64
64.67
70.64
73.53
76.34
78.37
75.85
79.02
79.41
78.35
74.42
78.08
78.53
73.27
75.90
65.27
64.69
63.03
60.29
58.39
57.87
62.91
66.82
72.92
72.72
75.42
71.74
180
76.78
82.78
75.52
49.72
72.32
81.11
72.71
76.40
55.47
55.75
55.18
52.13
51.30
51.69
58.67
60.95
69.68
70.83
77.34
68.54
0.34
contiguity
concentric
distance
diffuse
0.32
probabilities
0.3
0.28
0.26
0.24
0.22
0.2
10
15
20
25
observations
30
35
40
Figure 4.11: Posterior model probabilities
45
50

22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
69.67
70.48
72.66
68.20
63.00
68.06
66.06
70.67
74.97
74.99
76.53
76.54
75.99
74.54
73.08
75.10
75.99
73.69
71.52
65.65
67.08
70.60
64.11
55.68
61.08
58.06
66.11
71.80
72.23
75.21
75.11
75.03
73.76
72.79
72.84
74.56
72.20
70.39
73.99
74.32
74.26
72.25
65.25
73.99
67.35
74.73
76.25
77.88
78.22
77.90
75.24
76.34
73.09
78.38
76.65
74.50
72.53
71.23
71.02
73.61
69.78
68.75
71.04
72.25
72.47
75.23
74.64
75.86
75.72
75.90
75.27
75.27
76.24
76.88
76.46
75.59
181
67.62
69.39
72.16
66.48
62.21
65.97
66.56
69.26
76.81
75.23
76.86
77.53
77.92
72.63
71.00
72.66
75.82
71.30
67.05
100
80
60
40
20
10
15
20
25
30
35
40
b1 parameter, + = diffuse, - = contiguity, -- = concentric -. = distance
45
50
10
15
20
25
30
35
40
b2 parameter + = diffuse, - = contiguity, -- = concentric -. = distance
45
50
10
15
20
25
30
35
40
b3 parameter + = diffuse, - = contiguity, -- = concentric -. = distance
45
50
0
-1
-2
-3
-4
1
0.5
0
-0.5
-1
Figure 4.12: Smoothed parameter estimates

41
42
43
44
45
46
47
48
49
71.88
72.53
70.53
49.13
43.33
36.60
32.64
30.66
39.28
70.31
71.24
67.93
48.73
47.61
39.29
42.26
37.17
41.89
182
75.41
76.56
68.68
51.45
45.99
40.08
30.70
34.90
43.01
75.16
75.02
70.68
44.19
38.48
34.98
30.76
31.87
37.57
66.08
66.81
75.03
51.55
40.45
31.96
28.13
20.91
33.87
household income parameter

Obs
1
-1.62
-1.63
-1.79
2
-2.26
-1.95
-2.72
3
-2.40
-2.09
-2.72
4
-1.76
-1.77
-2.29
5
-1.64
-1.80
-1.49
6
-2.45
-2.20
-2.96
7
-2.67
-2.38
-3.00
8
-2.53
-2.36
-2.65
9
-2.40
-2.31
-2.54
10
-2.38
-1.89
-3.09
11
-2.36
-1.83
-3.10
12
-2.16
-1.74
-2.78
13
-1.72
-1.62
-1.85
14
-1.62
-1.71
-1.46
15
-1.77
-2.21
-1.44
16
-2.12
-2.14
-2.28
17
-2.16
-1.93
-2.46
18
-2.37
-2.24
-2.49
19
-2.41
-2.13
-2.67
20
-2.72
-2.59
-2.78
21
-2.57
-2.38
-2.73
22
-2.67
-2.52
-2.64
23
-3.10
-3.06
-2.99
24
-3.34
-3.45
-3.04
25
-3.19
-3.31
-2.71
26
-3.10
-2.79
-3.08
27
-3.41
-3.40
-3.24
28
-3.31
-2.91
-3.40
29
-3.27
-3.52
-2.63
30
-3.03
-3.26
-2.47
31
-3.07
-3.05
-2.95
32
-2.59
-2.64
-2.70
33
-3.03
-3.19
-2.76
34
-2.80
-2.93
-2.44
35
-2.31
-2.39
-2.66
36
-1.95
-1.97
-2.20
37
-2.05
-1.74
-2.68
distance
-2.50
-2.57
-2.60
-2.52
-2.41
-2.60
-2.67
-2.54
-2.57
-2.34
-2.32
-2.20
-2.02
-2.12
-2.32
-2.33
-2.44
-2.52
-2.62
-2.79
-2.80
-2.99
-3.25
-3.39
-3.37
-3.42
-3.59
-3.55
-3.48
-3.17
-3.15
-2.72
-3.08
-2.98
-2.61
-2.43
-2.41
diffuse
-0.27
-1.68
-2.16
-0.07
-0.68
-1.92
-2.63
-2.55
-2.16
-1.76
-1.76
-1.67
-1.37
-1.26
-1.34
-1.73
-1.76
-2.22
-2.19
-2.71
-2.34
-2.53
-3.09
-3.48
-3.42
-3.12
-3.42
-3.36
-3.50
-3.21
-3.14
-2.29
-3.10
-2.83
-1.51
-1.14
-1.30

38
39
40
41
42
43
44
45
46
47
48
49
-2.31
-1.80
-1.65
-1.63
-1.77
-2.75
-1.38
-1.12
-0.90
-0.63
-0.80
-1.04
-2.04
-1.67
-1.49
-1.67
-1.97
-2.51
-1.47
-1.25
-1.04
-1.03
-1.38
-1.39
183
-2.90
-2.22
-2.25
-2.12
-2.25
-3.22
-1.27
-1.28
-1.00
-0.57
-0.76
-1.13
-2.48
-2.40
-2.32
-2.31
-2.36
-2.61
-1.16
-0.90
-0.80
-0.62
-0.67
-0.88
-1.78
-0.83
-0.39
-0.28
-0.40
-2.62
-1.60
-0.99
-0.74
-0.40
-0.44
-0.75
house value parameter

Obs
1
-0.230
-0.183
-0.124
2
-0.141
-0.218
0.181
3
-0.152
-0.213
0.026
4
-0.229
-0.183
0.023
5
-0.163
-0.150
-0.154
6
-0.055
-0.167
0.255
7
-0.070
-0.130
0.063
8
0.005
-0.063
0.120
9
-0.037
-0.048
-0.008
10
0.142
-0.008
0.348
11
0.164
0.010
0.409
12
0.129
0.013
0.345
13
-0.005
0.001
-0.033
14
0.008
0.088
-0.097
15
0.160
0.318
0.030
16
0.098
0.157
0.099
17
0.049
0.004
0.118
18
-0.011
-0.042
0.007
19
0.021
-0.012
0.049
20
0.068
0.065
0.045
21
0.093
0.077
0.070
22
0.178
0.172
0.115
23
0.350
0.395
0.236
24
0.395
0.468
0.257
25
0.472
0.591
0.222
26
0.510
0.539
0.434
27
0.555
0.711
0.327
28
0.532
0.514
0.541
29
0.455
0.624
0.160
30
0.217
0.344
0.026
31
0.223
0.255
0.124
32
0.030
0.062
0.021
33
0.165
0.241
0.052
34
0.104
0.149
0.011
distance
-0.113
-0.110
-0.084
-0.071
-0.101
-0.094
-0.056
-0.010
0.018
0.089
0.096
0.075
0.056
0.118
0.232
0.134
0.118
0.036
0.089
0.115
0.186
0.275
0.380
0.379
0.485
0.544
0.555
0.519
0.468
0.259
0.258
0.080
0.200
0.165
diffuse
-0.566
-0.475
-0.364
-0.823
-0.265
-0.258
-0.166
-0.037
-0.113
0.020
0.017
0.004
-0.041
-0.059
0.097
0.006
-0.057
-0.048
-0.046
0.044
0.039
0.153
0.393
0.479
0.604
0.526
0.635
0.554
0.578
0.235
0.256
-0.051
0.164
0.086

35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
4.6
-0.005
-0.091
-0.122
-0.059
-0.165
-0.139
-0.139
-0.114
0.186
-0.059
-0.054
-0.032
-0.029
0.022
-0.014
0.017
-0.090
-0.183
-0.146
-0.216
-0.205
-0.133
-0.057
0.127
0.018
-0.057
0.025
-0.022
0.231
0.124
0.061
-0.043
-0.003
0.131
-0.027
0.029
-0.048
-0.063
0.420
-0.128
-0.031
-0.052
-0.036
-0.042
-0.047
0.051
0.002
-0.035
-0.030
-0.044
-0.040
-0.030
-0.013
0.116
-0.055
-0.056
-0.045
-0.037
-0.039
-0.052
184
-0.164
-0.246
-0.280
-0.209
-0.399
-0.370
-0.367
-0.341
0.068
-0.065
-0.076
-0.052
-0.023
-0.062
-0.076
Chapter Summary
We have seen that locally linear regression models can be estimated using distance weighted sub-samples of the observations to produce different
estimates for every point in space. This approach can deal with spatial heterogeneity and provide some feel for parameter variation over space in the
relationships being explored.
Some problems arise in using spatial expansion models because they tend
to produce heteroscedastic disturbances by construction. This problem is
overcome to some extent with the DARP model approach.
The non-parametric locally linear regression models produce problems
with respect to inferences about the parameters as they vary over space.
We saw how a Bayesian approach can provide valid posterior inferences
overcome these problems.
The Bayesian GWR model also solves some problems with the nonparametric implementation of the GWR regarding non-constant variance
over space or outliers. Given the locally linear nature of the GWR estimates, aberrant observations tend to contaminate whole subsequences of
the estimates. The BGWR model robustifies against these observations by
automatically detecting and downweighting their influence on the estimates.
A further advantage of this approach is that a diagnostic plot can be used
to identity observations associated with regions of non-constant variance or
spatial outliers.
Finally, an advantage of the Bayesian approach is that is subsumes the
spatial expansion, DARP and GWR models as special cases and provides
a more flexible implementation by explicitly including a relationship that
185
describes parameter smoothing over space to the model. Diffuse implementation of the parameter smoothing specification leads to the non-parametric
GWR model. In addition to replicating the GWR estimates, the Bayesian
model presented here can produce estimates based on parameter smoothing
specifications that rely on: distance decay relationships, contiguity relationships, monocentric distance from a central point, or the latitude-longitude
locations proposed by Casetti (1972).
4.7
References
Brunsdon, C., A. S. Fotheringham, and M.E. Charlton. (1996) Geographically weighted regression: A method for exploring spatial nonstationarity, Geographical Analysis, Volume 28, pp. 281-298.
Casetti, E. (1972) Generating Models by the Expansion Method: Applications to Geographic Research, Geographical Analysis, December,
Volume 4, pp. 81-91.
Casetti, E. (1992) Bayesian Regression and the Expansion Method,
Geographical Analysis, January, Volume 24, pp. 58-74.
Casetti, E. (1982) Drift Analysis of Regression Parameters: An Application to the Investigation of Fertility Development Relations, Modeling and Simulation 13, Part 3:, pp. 961-66.
Casetti, E. and A. Can (1998) The Econometric estimation and testing of DARP models. Paper presented at the RSAI meetings, Sante
Fe, New Mexico.
Leamer, Edward. 1983. Model Choice and Specification Analysis,
in Handbook of Econometrics, Volume 1, Chapter 5, Zvi Griliches and
Michael Intriligator (eds.) (Amsterdam: North-Holland).
LeSage, James P. (1997) Bayesian Estimation of Spatial Autoregressive Models, International Regional Science Review, 1997 Volume 20,
number 1&2, pp. 113-129
McMillen, Daniel P. (1996) One Hundred Fifty Years of Land Values
in Chicago: A Nonparameteric Approach, Journal of Urban Economics, Vol. 40, pp. 100-124.
186
McMillen, Daniel P. and John F. McDonald. (1997) A Nonparametric

Analysis of Employment Density in a Polycentric City, Journal of
Regional Science, Vol. 37, pp. 591-612.
Theil, Henri and Arthur S. Goldberger. (1961) On Pure and Mixed
Statistical Estimation in Economics, International Economic Review,
Vol. 2, pp. 65-78.
Zellner, Arnold. (1971) An Introduction to Bayesian Inference in
Econometrics. (New York: John Wiley & Sons.)
Chapter 5
Limited dependent variable

models
These models arise when the dependent variable y in our spatial autoregressive model takes values 0, 1, 2, . . . representing counts of some event
or a coding system for qualitative outcomes. For example, y = 0 might
represent a coding scheme indicating a lack of highways in our sample of geographical regions, and y = 1 denotes the presence of a highway. As another
example where the values taken on by y represent counts, we might have
y = 0, 1, 2, . . . denoting the number of foreign direct investment projects in
a given county where our sample of observations represent counties for a
particular state.
Spatial autoregressive modeling of these data would be interpreted in the
framework of a probability model that attempts to describe the P rob(event
i occurs) = F (X: parameters). If the outcomes represent two possibilities,
y = 0, 1, the model is said to be binary, whereas models with more than two
outcomes are referred to as multinomial or polychotomous.
Traditional spatial autoregressive models could be used to carry out a
spatial autoregression using the binary response variable y = 0, 1, but two
problems arise. First, the errors are by construction heteroscedastic. This
is because the actual y = 0, 1 minus the value W y + X = W y X
or W y X. Note also that the heteroscedastic errors are a function
of the parameter vector and . The second problem with using spatial
autoregressive models in this setting is that the predicted values can take
on values outside the (0,1) interval, which is problematic given a probability
model framework. In this setting we would like to see:
187
CHAPTER 5. LIMITED DEPENDENT VARIABLE MODELS
188
limW y+X+ P rob(y = 1) = 1
(5.1)
limW y+X P rob(y = 1) = 0
(5.2)
Two distributions that have been traditionally used to produce this type
of outcome in the case of regression models (that ensures predicted values
between zero and one) are the logisitic and normal distributions resulting
in the logit model shown in (5.3) and probit model shown in (5.4), where
denotes the cumulative normal probability function.
P rob(y = 1) = eX /(1 + eX )
(5.3)
P rob(y = 1) = (X)
(5.4)
The logistic distribution is similar to the normal except in the tails where
it is fatter resembling a Student tdistribution. Green (1997) and others
indicate that the logistic distribution resembles a tdistribution with seven
degrees of freedom.
5.1
Introduction
McMillen (1992) proposed methods for estimating SAR and SEM probit
models containing spatial heteroscedasticity that rely on the EM algorithm.
Aside from McMillen (1992), very little work has appeared regarding spatial autoregressive models that contain binary or polychotomous dependent
variables.
McMillen (1995) investigates the impact of heteroscedasticity that is
often present in spatial models on probit estimation for non-spatial autoregressive models. McMillen and McDonald (1998) propose a non-parametric
locally linear probit method for GWR models of the type discussed in Chapter 4.
Bayesian estimation of logit/probit and tobit variants of spatial autoregressive models that exhibit heteroscedasticity is developed in this chapter.
The approach taken draws on work by Chib (1992) and Albert and Chib
(1993) as well as the Bayesian estimation of spatial autoregressive models
set forth in Chapter 3.
Accounting for heteroscedasticity in logit/probit and tobit models is important because estimates based on the assumption of homoscedasticity in
189
the presence of heteroscedastic disturbances are inconsistent. The proposed

Bayesian estimation methodology overcomes several drawbacks associated
with McMillens (1992) EM approach to estimating these models in the
presence of heteroscedastic disturbances.
EM estimation methods rely on an iterative sequencing between the Estep that involves estimation and the M-step that solves a conditional maximization problem. The maximization problem is conditional on parameters
determined in the E-step, making it easier to solve than the entire problem
involving all of the parameters in the problem.
The approach proposed here extends work of Chib (1992) for the tobit
model and Albert and Chib (1993) for the probit model to the case of spatial
autoregressive and spatial error models. The basic idea exhibits a similarity
to the EM algorithm proposed by McMillen (1992), where the censored or
latent unobserved observations on the dependent variable y in the model are
replaced by estimated values. Given estimates of the missing y values, the
EM algorithm proceeds to estimate the other parameters in the model using
methods applied to non-truncated data samples. In other words, conditional
on the estimated y-values, the estimation problem is reduced to a noncensored estimation problem which can be solved using maximum likelihood
methods.
There are some drawbacks to McMillens EM estimator that we will
overcome using the Bayesian approach set forth in this chapter. One drawback to McMillens EM estimator is that the information matrix approach
to determining measures of precision for the parameter estimates cannot be
used. The likelihood function for the heteroscedastic probit model contains
a number of integrals equal to the number of observations, so evaluating the
likelihood function for these models is impossible. McMillen (1992) overcomes this problem using a non-linear weighted least-squares interpretation
of the probit estimator conditional on the spatial lag parameters in the
SAR model and in the SEM model. This rules out estimates of dispersion
for these important parameters. The use of a covariance matrix conditional
on the spatial lag parameters produces biased, but consistent confidence
intervals that may be too small.
Another problem with McMillens approach is the need to specify a functional form for the non-constant variance over space. That is, one must
specify a model for the noise vector such that, [var(i )]1/2 = g(Zi), where
g is a continuous, twice differentiable function and Zi is a vector of explanatory variables for var(i ). This approach was illustrated by McMillen (1992)
for a simple 2-variable model where both of the variables in squared form
2 + X 2 ). In larger
were used to form the Zi vector, i.e., gi = exp(1 X1i
2 2i
190
models a practitioner would need to devote considerable effort to testing

and specifying the functional form and variables involved in the model for
var(i ). Assuming success in finding a few candidate specifications, there is
still the problem of inferences that may vary across alternative specifications
for the non-constant variance.
We rely on Gibbs sampling to estimate the spatial logit/probit and tobit
models. During sampling, we introduce a conditional distribution for the
censored or latent observations conditional on all other parameters in the
model. This distribution is used to produce a random draw for each censored
value of yi in the case of tobit and for all yi in the probit model. The
conditional distribution for the latent variables takes the form of a normal
distribution centered on the predicted value truncated at the right by 0 in
the case of tobit, and truncated by 0 from the left and right in the case of
probit, for yi = 1 and yi = 0 respectively.
An important difference between the EM approach and the samplingbased approach set forth here is the proof outlined by Gelfand and Smith
(1990) that Gibbs sampling from the sequence of complete conditional distributions for all parameters in the model produces a set of draws that
converge in the limit to the true (joint) posterior distribution of the parameters. Because of this, we overcome the bias inherent in the EM algorithms
use of conditional distributions. Valid measures of dispersion for all parameters in the model can be constructed from the large sample of parameter
draws produced by the Gibbs sampler. Further, if one is interested in linear
or non-linear functions of the parameters, these can be constructed using
the underlying parameter draws in the linear or non-linear function and
then computing the mean over all draws. A valid measure of dispersion for
this linear or non-linear combination of the parameters can be based on the
distribution of the functional combination of parameters.
The Gibbs sampling approach to estimating the spatial autoregressive
models presented in Chapter 3 can be adapted to produce probit and tobit estimates by adding a single conditional distribution for the censored or
latent observations. Intuitively, once we have a sample for the unobserved latent dependent variables, the problem reduces to the Bayesian heteroscedastic spatial autoregressive models presented in Chapter 3. The conditional
distributions for all other parameters in spatial autoregressive model presented in Chapter 3 remain valid.
Another important advantage of the method proposed here is that heteroscedasticity and outliers can be easily accommodated with the use of the
methods outlined in Chapter 3. For the case of the probit model, an interesting interpretation can be given to the family of tdistributions that
191
arise from the methods in Chapter 3 to deal with spatial heterogeneity and
spatial outliers. Recall that models involving binary data can rely on any
continuous probability function as the probability rule linking fitted probabilities with the binary observations. Probit models arise from a normal
probability rule and logit models from a logistic probability rule. When
one introduces the latent variables zi in the probit model to reflect unobserved values based on the binary dependent variables yi , we have an underlying conditional regression involving z and the usual spatial regression
model variables X, W , where X represents the explanatory variables and
W denotes the row-standardized spatial weight matrix. The heteroscedastic
spatial autoregressive model introduced in Chapter 3 can be viewed in the
case of binary dependent variables as a probability rule based on a family
of tdistributions that represent a mixture of the underlying normal distribution used in the probit regression. (It is well known that the normal
distribution can be modeled as a mixture of tdistributions, see Albert and
Chib, 1993).
The most popular choice of probability rule to relate fitted probabilities
with binary data is the logit function corresponding to a logistic distribution
for the cumulative density function. Albert and Chib (1993) show that the
quantiles of the logistic distribution correspond to a t-distribution around
7 or 8 degrees of freedom. We also know that the normal probability density is similar to a tdistribution when the degrees of freedom are large.
This allows us to view both the probit and logit models as special cases of
the family of models introduced here using a chi-squared prior based on a
hyperparameter specifying alternative degrees of freedom to model spatial
heterogeneity and outliers.
By using alternative values for the prior hyperparameter that we labeled
r in Chapter 3, one can test the sensitivity of the fitted probabilities to
alternative distributional choices for the regression model. For example, if
we rely on a value of r near 7 or 8, the estimates resulting from the Bayesian
version of the heteroscedastic probit model correspond to those one would
achieve using a logit model. On the other hand, using a large degrees of
freedom parameter, say r = 50 would lead to estimates that produce fitted
probabilities based on the probit model choice of a normal probability rule.
The implication of this is that the heteroscedastic spatial probit model we
introduce here represents a more general model than either probit or logit.
The generality derives from the family of tdistributions associated with
alternative values of the hyperparameter r in the model.
A final advantage of the method described here is that estimates of the
non-constant variance for each point in space are provided and the prac-
192
titioner need not specify a functional form for the non-constant variance.
Spatial outliers or aberrant observations as well as patterns of spatial heterogeneity will be identified in the normal course of estimation. This represents a considerable improvement over the approach described by McMillen
(1992), where a separate model for the non-constant variance needs to be
specified.
5.2
The Gibbs sampler
For spatial autoregressive models with uncensored y observations where the

error process is homoscedastic and outliers are absent, the computational
intensity of the Gibbs sampler is a decided disadvantage over maximum likelihood methods. As demonstrated in Chapter 3, the estimates produced by
the Bayesian model estimated with Gibbs sampling are equivalent to those
from maximum likelihood in these cases. For the case of probit and tobit
models, the Gibbs sampler might be very competitive to the EM algorithm
presented in McMillen (1992) because numerous probit or tobit maximum
likelihood problems need to be solved to implement the EM method.
Before turning attention to the heteroscedastic version of the Bayesian
spatial autoregressive logit/probit and tobit models, consider the case of
a homoscedastic spatial autoregressive tobit model where the observed y
variable is censored. One can view this model in terms of a latent but unobservable variable z such that values of zi < 0, produce an observed variable
yi = 0. Similarly, spatial autoregressive probit models can be associated
with a latent variable zi < 0 that produces an observed variable yi = 0 and
zi 0 resulting in yi = 1. In both these models, the posterior distribution
of z conditional on all other parameters takes the form of truncated normal
distribution (see Chib, 1992 and Albert and Chib, 1993).
For the spatial tobit model, the conditional distribution of zi given all
other parameters is a truncated normal distribution constructed by trun2 ] distribution from the right by zero. Where the predicted
cating a N [
yi , ti
value for zi is denoted by yi which represents the ith row of y = B 1 X for
the SAR model and the ith row of y = X for the SEM model. The variance
2 = 2 P 2 , where denotes the ijth element of
of the prediction is ti
ij
j ij
(In W )1 for both the SAR and SEM models. The pdf of the latent
variables zi is then:
f (zi |, , ) =
193
[1 (
yi /ti)]1 exp[(zi yi )2 /2it], if zi 0
0
if zi > 0
(5.5)
Similarly, for the case of probit, the conditional distribution of zi given
all other parameters is:
(
f (zi |, , )
2 ), truncated at the left by 0 if y = 1

N (
yi , pi
i
2 ), truncated at the right by 0 if y = 0
N (
yi , pi
i
(5.6)
2
2
Where pi
= j ij
, because the probit model is unable to identify both
2
and , leading us to scale our problem so 2 equals unity. The predicted
value yi takes the same form for the SAR and SEM models as described
above for the case of tobit.
The tobit expression (5.5) indicates that we rely on the actual observed
y values for non-censored observations and use the sampled latent variables
for the unobserved values of y. For the case of the probit model, we replace
values of yi = 1 with the sampled normals truncated at the left by 0 and
values of yi = 0 with sampled normals truncated at the right by 0.
Given these sampled continuous variables from the conditional distribution of zi , we can implement the remaining steps of the Gibbs sampler
described in Chapter 3 to determine draws from the conditional distributions for , ( and in the case of tobit) using the sampled zi values in
place of the censored variables yi .
5.3
Heteroscedastic models
The models described in this section can be expressed in the form shown
in (5.5), where we relax the usual assumption of homogeneity for the disturbances used in SAR, SEM and SAC modeling. Given the discussion in
Section 5.2, we can initially assume the existence of non-censored observations on the dependent variable y because these can be replaced with
sampled values z as motivated in the previous section.
y = W1 y + X + u
u = W2 u +
N(0, 2V ),
V = diag(v1 , v2 , . . ., vn )
(5.7)
194
Where vi , i = 1, . . . , n represent a set of relative variance parameters to be

estimated. We restrict the spatial lag parameters to the interval 1/min <
, < 1/max .
It should be clear that we need only add one additional step to the
Gibbs sampler developed in Chapter 3 for Bayesian heteroscedastic spatial
autoregressive models. The additional step will provide truncated draws for
the censored or limited dependent variables.
The above reasoning suggest the following Gibbs sampler. Begin with
arbitrary values for the parameters 0 , 0 , 0 and vi0 , which we designate
with the superscript 0. (Keep in mind that we do not sample the conditional
distribution for in the case of the probit model where this parameter is
set to unity.)
1. Calculate p(|0, 0 , vi0), which we use along with a random 2 (n)
draw to determine 1 .
2. Calculate p(|0, 1 vi0 ) using 1 from the previous step. Given the
means and variance-covariance structure for , we carry out a multivariate random draw based on this mean and variance to determine
1.
3. Calculate p(vi|0 , 1 1 ), which is based on an nvector of random
2 (r + 1) draws to determine vi1 , i = 1, . . ., n.
4. Use metropolis within Gibbs sampling to determine 1 as explained in
Chapter 3, using the the values 1 , 1 and vi1 , i = 1, . . ., n determined
in the previous steps.
5. Sample the censored yi observations from a truncated normal centered
on the predictive mean and variance determined using 1 , 1 , 1 , vi1 (as
described in Section 5.2) for the probit and tobit models.
In the above development of the Gibbs sampler we assumed the hyperparameter r that determines the extent to which the disturbances take on a
leptokurtic character was known. It is unlikely in practice that investigators
would have knowledge regarding this parameter, so an issue that confronts
us when attempting to implement the heteroscedastic model is setting the
hyperparameter r. As already discussed in Chapter 3, I suggest using a small
value near r = 4, which produces estimates close to those based on a logit
probability rule. If you wish to examine the sensitivity of your inferences
to use of a logit versus probit probability rule, you can produce estimates
based on a larger value of r = 30 for comparison.
5.4
195
Implementing these models
We have functions sarp g and sart g that carry out Gibbs sampling estimation of the probit and tobit spatial autoregressive models. The documentation for sarp g is:
PURPOSE: Gibbs sampling spatial autoregressive Probit model
y = p*Wy + Xb + e, e is N(0,sige*V)
y is a 0,1 vector
B = N(c,T), sige = gamma(nu,d0), p = diffuse prior
--------------------------------------------------USAGE: results = sarp_g(y,x,W,ndraw,nomit,prior,start)
W = 1st order contiguity matrix (standardized, row-sums = 1)
ndraw = # of draws
prior = a structure for: B = N(c,T), sige = gamma(nu,d0)
c above (default 0)
prior.m,
prior.k,
(default: not used)
prior.nu,
prior.d0,
prior.rmin = (optional) min rho used in sampling
prior.rmax = (optional) max rho used in sampling
defaults: beta=1,sige=1,rho=0.5, V= ones(n,1)
start.b
start.p
= rho starting value
(scalar)
start.V
results.meth = sarp_g
results.ymean = mean of y draws (1 x nobs)
results.rdraw = r draws (ndraw-nomit x 1) (if m,k input)
results.pdraw = p draws
(ndraw-nomit x 1)
results.pmean = b prior means, prior.beta from input
results.pstd = b prior std deviations sqrt(diag(T))
results.r
results.r2mf = McFadden R-squared
results.rsqr = Estrella R-squared
196
results.zip
= # of zero y-values
results.y
results.nu
results.d0
results.rmax = 1/max eigenvalue of W (or rmax if input)
results.rmin = 1/min eigenvalue of W (or rmin if input)
This function documentation and use is very similar to the sar g function
from Chapter 3. One difference is in the measures of fit calculated. These
are Rsquared measures that are traditionally used for limited dependent
variable models.
Following an example provided by McMillen (1992) for his EM algorithm
approach to estimating SAR and SEM probit models, we employ the data
set from Anselin (1988) on crime in Columbus, Ohio. McMillen censored the
dependent variable on crime such that yi = 1 for values of crime greater than
40 and yi = 0 for values of crime less than or equal to 40. The explanatory
variables in the model are neighborhood housing values and neighborhood
income. Example 5.1 demonstrates how to implement a spatial probit model
Gibbs sampler using the probit g function.
% ----- Example 5.1 SAR Probit Model
load anselin.data;
y = anselin(:,1); [n junk] = size(y);
load Wmat.data; W = Wmat;
yc = zeros(n,1);
% now convert the data to 0,1 values
for i=1:n
if y(i,1) > 40.0
yc(i,1) = 1;
end;
end;
ndraw = 1100; nomit = 100;
prior.rval = 4; prior.rmin = 0; prior.rmax = 1;
result = sarp_g(yc,x,W,ndraw,nomit,prior);
prt(result,vnames);
plt(result,vnames);
197
The printed results are shown below and the graphical results provided
by the plt function are shown in Figure 5.1. For comparison we also present
the results from ignoring the limited dependent variable nature of the y
variable in this model and using the sar function to produce maximum
likelihood estimates.
Gibbs sampling spatial autoregressive Probit model
crime
McFadden R^2
=
0.4122
Estrella R^2
=
0.5082
sigma^2
=
2.4771
r-value
=
4
Nobs, Nvars
=
49,
3
# 0, 1 y-values =
30,
19
ndraws,nomit
=
1100,
100
acceptance rate =
0.7836
time in secs
=
69.0622
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000 1000000.000000
income
0.000000 1000000.000000
hvalue
0.000000 1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
3.616236
2.408302
0.020092
income
-0.197988
-1.698013
0.096261
hvalue
-0.036941
-1.428174
0.159996
rho
0.322851
2.200875
0.032803
R-squared
=
0.5216
Rbar-squared
=
0.5008
sigma^2
=
0.1136
Nobs, Nvars
=
49,
3
log-likelihood =
-1.1890467
# of iterations =
13
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
variable 1
0.679810
2.930130
0.005259
variable 2
-0.019912
-1.779421
0.081778
variable 3
-0.005525
-1.804313
0.077732
rho
0.539201
2.193862
0.033336
Gibbs sampling spatial autoregressive Probit model
crime
198
McFadden R^2
=
0.3706
Estrella R^2
=
0.4611
sigma^2
=
2.2429
r-value
=
40
Nobs, Nvars
=
49,
3
# 0, 1 y-values =
30,
19
ndraws,nomit
=
1100,
100
acceptance rate =
0.9616
time in secs
=
70.3423
min and max rho =
0.0000,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000 1000000.000000
income
0.000000 1000000.000000
hvalue
0.000000 1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
4.976554
2.408581
0.020078
income
-0.259836
-2.047436
0.046351
hvalue
-0.053615
-1.983686
0.053278
rho
0.411042
3.285293
0.001954
There is a remarkable difference between the maximum likelihood estimates that ignore the limited dependent nature of the variable y which we
would expect. To explore the difference between logit and probit estimates, we produced a set of Bayesian estimates based on a hyperparameter
value of r = 40, which would correspond to the Probit model. Green (1997)
states that the issue of which distributional form should be used on applied
econometric problems is unresolved. He further indicates that inferences
from either logit or probit models are often the same. This does not appear
to be the case for the data set in example 5.1, where we see a difference
in both the magnitude and significance of the parameters from the models
based on r = 4 and r = 40.
A final point with respect to interpreting the estimates is that the marginal
impacts of the variables on the fitted probabilities is usually the inference
aim of these models. The estimates would need to be converted according
to the probability rule implied by the alternative underlying tdistributions
associated with the r value employed. These marginal impacts would be
very similar (as in the case of non-spatial maximum likelihood logit versus
probit marginal impacts) despite the apparent differences in the coefficients.
For the purpose of computing marginal impacts, one needs to evaluate the
posterior density of pk , which we denote
(pk ) for k ranging over all sample
observations. (It is conventional to assess marginal impacts across all obser-

SAR Probit Gibbs Actual vs. Predicted
199
Residuals
1.2
1
0.5
0.8
0.6
0.4
-0.5
0.2
0
10
20
30
-1
40
10
Mean of Vi draws
2.5
1.9
1.8
1.5
1.7
1.6
0.5
1.5
0
0
10
20
30
30
40
50
Posterior Density for rho
1.4
20
40
50
0.5
Figure 5.1: Results of plt() function

vations in the model and then average these.) For the heteroscedastic model
(1/2)
(1/2) 0
pk will take the form: (
yk ) = (Wk vk
y + vk
xk ). To find the
posterior density estimate of
(pk ) we would employ the draws for i , vki , i
in normal densities, denoted N(, ) as indicated in (5.8)
(pk ) = (1/m)
m
X
i=1
N[
yki , (1/vki )x0k (X 0 Vi1 X)1 xk ]/N[0, 1]
(5.8)
yki = i Wk (1/ vki )y + (1/ vki )x0k i

Vi1 = diag(1/vji ), j = 1, . . . , n
Table 5.1 shows a comparison of McMillens EM algorithm estimates and
those from Gibbs sampling. The Gibbs estimates are based on 1,100 draws
with the first 100 discarded for startup or burn-in. Gibbs SAR and SEM
200
estimates are reported for both r = 4 and r = 40. Results for two values of r
were reported because the inferences are different from these two models. It
should be noted that there is a tendency of the consistent EM estimates of
dispersion to overstate the precision, producing generally larger tstatistics
for the EM versus Gibbs estimates.
Table 5.1: EM versus Gibbs estimates
CONSTANT
tvalue
INCOME
tvalue
HOUSING
tvalue
tvalue
EM
SAR
2.587
2.912
-0.128
-2.137
-0.029
-1.617
0.429
Gibbs
SAR r = 4
3.758
2.150
-0.213
-1.583
-0.037
-1.416
0.325
2.175
Gibbs
SAR r = 40
4.976
2.408
-0.259
-2.047
-0.053
-1.983
0.411
3.285
EM
SEM
2.227
3.115
-0.123
-2.422
-0.025
-1.586
0.279
Gibbs
SEM r = 4
3.925
1.936
-0.214
-1.636
-0.048
-1.439
0.311
1.766
Gibbs
r = 40
2.710
2.168
-0.143
-1.719
-0.032
-1.791
0.315
1.796
Another reason why the EM and Gibbs estimates may be different is that
McMillens approach requires that a model for the non-constant variance be
specified. The specification used by McMillen was: vi = 0.0007INCOME2 +
0.0004HOUSING2 . This is quite different from the approach taken in the
Bayesian model that relies on vi estimates from Gibbs sampling.
There are functions for carrying out Bayesian Probit and Tobit versions
of all spatial autoregressive models. SAR models are implemented by sarp g
and sart g, SEM models by semp g and semt g, SAC models by sacp g
and sact g.
Tobit models can involve either left or right truncation. That is censored
observations can be y values that fall below a limiting value or censoring can
take place above a limit value. The functions sart g, sact g and semt g
allow the user to specify the type of censoring and a limit value. This
defaults to the typical case of left censoring at zero. The documentation for
the function sart g is:
PURPOSE: Gibbs sampling spatial autoregressive Tobit model
y = p*Wy + Xb + e, e is N(0,sige*V)
y is a censored vector (assumed to be at zero)
B = N(c,T), sige = gamma(nu,d0), p = diffuse prior
--------------------------------------------------USAGE: results = sart_g(y,x,W,prior,ndraw,nomit,start)

x
W
ndraw
nomit
prior
=
=
=
=
=
independent variables matrix (nobs x nvar)

1st order contiguity matrix (standardized, row-sums = 1)
# of draws
# of initial draws omitted for burn-in
a structure for: B = N(c,T), sige = gamma(nu,d0)
c above (default 0)
prior.m,
prior.k,
(default: not used)
prior.nu,
prior.d0,
prior.trunc = left or right censoring (default = left)
prior.limit = value for censoring (default = 0)
defaults: beta=1,sige=1,rho=0.5, V= ones(n,1)
start.b
start.p
= rho starting value
(scalar)
start.V
--------------------------------------------------NOTE: 1st column of x-matrix must contain iota vector (constant term)
results.meth = sart_g
results.rdraw = sige draws (ndraw-nomit x 1)
results.pdraw = p draws
(ndraw-nomit x 1)
results.ymean = mean of y draws (1 x nobs)
results.pmean = b prior means, prior.beta from input
results.pstd = b prior std deviations sqrt(diag(T))
results.r
results.nobsc = # of censored y-values
results.y
results.nu
results.d0
results.rmax = 1/max eigenvalue of W (or rmax if input)
results.rmin = 1/min eigenvalue of W (or rmin if input)
201
202
To illustrate the case of Tobit model estimation, example 5.2 shows an

example where we generate an SAR model based on the Anselin neighborhood crime spatial contiguity matrix. We censor observations that are less
than zero. Estimates based on the uncensored data using the function sar g
are compared to those from using sar g and sart g on the censored data.
We would expect that ignoring censoring should produce poor estimates
from application of sar g on the censored data whereas use of sart g would
produce better estimates that are closer to those from application of sar g
to the uncensored data.
A vector y is generated based on the neighborhood crime independent
variables standardized. This standardization produces a relatively even
number of negative and positive values for y. Negative values are then
censored.
% ----- Example 5.2 SAR Tobit Model
load anselin.dat;
xt = [anselin(:,2:3)]; n = length(xt);
% center and scale the data so our y-values
% are evenly distributed around zero, the censoring point
x = [ones(n,1) studentize(xt)];
[n k] = size(x);
load wmat.dat;
W = wmat;
sige = 5.0;
evec = randn(n,1)*sqrt(sige);
rho = 0.75;
beta = ones(k,1);
B = eye(n) - rho*W;
BI = inv(B);
y = BI*x*beta + BI*evec;
yc = y;
% now censor neighborhoods with crime < 1
for i=1:n
if y(i,1) < 1
yc(i,1) = 1;
end;
end;
Vnames = strvcat(crime,constant,income,hvalue);
prior.rval = 30;
res1 = sar(y,x,W);
res2 = sar_g(yc,x,W,ndraw,nomit,prior);
prior.limit = 1;
prior.trunc = left;
res3 = sart_g(yc,x,W,ndraw,nomit,prior);
prt(res1,Vnames);
prt(res2,Vnames);
prt(res3,Vnames);
203
The printed results for an SAR model based on the uncensored data as
well as a set of estimates that ignore the sample censoring and the Tobit
version of the SAR model are presented below.
crime
R-squared
=
0.7344
Rbar-squared
=
0.7229
sigma^2
=
5.6564
Nobs, Nvars
=
49,
3
log-likelihood =
-99.041858
# of iterations =
13
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Coefficient
t-statistic
t-probability
constant
0.617226
1.328646
0.190519
income
1.064967
2.271643
0.027831
hvalue
0.897449
2.254232
0.028988
rho
0.724180
5.068285
0.000007
crime
R-squared
=
0.7013
sigma^2
=
3.4570
r-value
=
30
Nobs, Nvars
=
49,
3
ndraws,nomit
=
600,
100
acceptance rate =
0.9662
time in secs
=
14.4129
min and max rho =
-1.5362,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000 1000000.000000
income
0.000000 1000000.000000
hvalue
0.000000 1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
1.412658
3.037593
0.003922
income
1.253225
3.479454
0.001111
hvalue
0.364549
1.267454
0.211372
rho
0.599010
6.097006
0.000000
Gibbs sampling spatial autoregressive Tobit model
crime
R-squared
=
0.6977
sigma^2
=
6.9121
r-value
=
30
Nobs, Nvars
=
49,
3
204
# censored values =
16
ndraws,nomit
=
600,
100
acceptance rate
=
0.9120
time in secs
=
19.7692
min and max rho
=
-1.5362,
1.0000
***************************************************************
Variable
Prior Mean
Std Deviation
constant
0.000000 1000000.000000
income
0.000000 1000000.000000
hvalue
0.000000 1000000.000000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
0.936753
1.847863
0.071058
income
1.479955
2.844135
0.006624
hvalue
0.544580
1.320079
0.193340
rho
0.629394
6.446799
0.000000
A key to understanding how the Gibbs sampler works on these problems

is the generation of predicted values for the censored observations. These
values may also be useful for purposes of inference regarding the censored
observations. The tobit spatial autoregressive functions return a structure
variable field results.ymean that represents the mean of the sampled values
for the censored observations as well as the actual values of the uncensored
observations. Figure 5.2 shows a plot of this data vector against the actual y
variables. Ordinarily, we wouldnt know the values of the censored y values,
but in this case because we generated the data set and then censored the
observations we have this information.
5.5
An applied example
The Harrison and Rubinfeld Boston data set used in Chapter 3 to illustrate
Gibbs sampling Bayesian spatial autoregressive models contains censored
values. Median house values greater than $50,000 were set equal to 50,000
for 16 of the 506 sample observations (see Gilley and Pace, 1995). This
provides an opportunity to see if using tobit estimation to take the sample
truncation into account produces different parameter estimates.
Example 5.3 reads the Boston data set and sorts by median housing
values. Note that we must also sort the explanatory variables using the
index vector yind returned by the sort for the y values as well as the latitude
and longitude vectors. After carrying out Gibbs sampling estimation for the
SAR, SEM and SAC models, we add to the prior structure variable a field
for right-truncation and we supply the limit value which is the log of 50,000
205
15
Actual
Mean of Gibbs samples
10
Crime
-5
-10
10
15
20
25
30
Observations
35
40
45
50
Figure 5.2: Actual vs. simulated censored y-values

standardized. By virtue of our sorting the y vector, this transformed limit
value must equal the last 16 observations, so we use the last to define the
limit value.
% ----- Example 5.3 Right-censored Tobit for the Boston data
% sort by median house values
[ys yind] = sort(y); xs = boston(yind,1:k-1);
lats = latitude(yin,1); lons = longitude(yin,1);
[W1 W W3] = xy2cont(lats,lons); % create W-matrix
vnames = strvcat(hprice,constant,crime,zoning,industry, ...
charlesr,noxsq,rooms2,houseage,distance,access,taxrate, ...
y = studentize(log(ys)); x = [ones(n,1) studentize(xs)];
% define censoring limit
limit = y(506,1); % median values >=50,000 are censored to 50
206
ndraw = 1100; nomit = 100;

prior.rval = 4;
prior.lmin = 0; prior.lmax = 1;
% ignore censoring
res1 = sar_g(y,x,W,ndraw,nomit,prior);
prt(res1,vnames);
res2 = sem_g(y,x,W,ndraw,nomit,prior);
prt(res2,vnames);
res3 = sac_g(y,x,W,W,ndraw,nomit,prior);
prt(res3,vnames);
% use Tobit for censoring
prior.trunc = right;
prior.limit = limit;
res4 = sart_g(y,x,W,ndraw,nomit,prior);
prt(res4,vnames);
res5 = semt_g(y,x,W,ndraw,nomit,prior);
prt(res5,vnames);
res6 = sact_g(y,x,W,W,ndraw,nomit,prior);
prt(res6,vnames);
Intuitively, we might not expect a large difference in the parameter

estimates for this case where only 16 of the 506 sample observations are
censored. The results are presented below, where the information that is
typically printed regarding prior means and standard deviations has been
eliminated because we used a diffuse prior. The order of the results has been
ordered to present both SAR and the tobit SAR, then the SEM and tobit
SEM and finally SAC and tobit SAC estimates.
hprice
R-squared
=
0.8243
sigma^2
=
0.1921
r-value
=
4
Nobs, Nvars
=
506,
14
ndraws,nomit
=
1100,
100
acceptance rate =
0.8662
time in secs
= 129.8548
min and max rho =
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
-0.025169
-1.126015
0.260708
crime
-0.152583
-3.509177
0.000491
zoning
0.050736
1.304237
0.192763
industry
0.045046
0.915162
0.360555
charlesr
0.020157
0.787959
0.431100

noxsq
rooms2
houseage
distance
access
taxrate
pupil/teacher
blackpop
lowclass
rho
-0.089610
0.267168
-0.036438
-0.178140
0.188203
-0.212748
-0.117601
0.107424
-0.313225
0.314435
-1.803644
5.731838
-0.878105
-5.673916
2.713291
-5.525728
-3.980184
3.873596
-7.068474
3.212364
0.071899
0.000000
0.380315
0.000000
0.006895
0.000000
0.000079
0.000122
0.000000
0.001403
Gibbs sampling spatial autoregressive Tobit model

hprice
R-squared
=
0.8225
sigma^2
=
0.1595
r-value
=
4
Nobs, Nvars
=
506,
14
# censored values =
16
ndraws,nomit
=
1100,
100
acceptance rate
=
0.9206
time in secs
= 158.1523
min and max rho
=
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
-0.034993
-1.763936
0.078363
crime
-0.159346
-3.631451
0.000311
zoning
0.051875
1.399918
0.162168
industry
0.025951
0.561893
0.574445
charlesr
0.010791
0.446021
0.655778
noxsq
-0.084991
-1.800372
0.072414
rooms2
0.240183
5.251582
0.000000
houseage
-0.048693
-1.265013
0.206465
distance
-0.176466
-5.657902
0.000000
access
0.193611
3.125944
0.001877
taxrate
-0.219662
-6.201483
0.000000
pupil/teacher
-0.115413
-4.128876
0.000043
blackpop
0.107766
4.064385
0.000056
lowclass
-0.301859
-7.543143
0.000000
rho
0.296382
2.937506
0.003464
Gibbs sampling spatial error model
hprice
R-squared
=
0.7304
sigma^2
=
0.1445
r-value
=
4
Nobs, Nvars
=
506,
14
ndraws,nomit
=
1100,
100
acceptance rate
=
0.4870
207

time in secs
= 114.7631
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
-0.039799
-0.420595
0.674234
crime
-0.165742
-4.010247
0.000070
zoning
0.049197
1.177168
0.239698
industry
-0.005192
-0.087059
0.930660
charlesr
-0.015074
-0.579339
0.562625
noxsq
-0.147566
-1.925225
0.054777
rooms2
0.338988
8.416259
0.000000
houseage
-0.127886
-2.885708
0.004077
distance
-0.175342
-2.494614
0.012936
access
0.276185
3.014877
0.002704
taxrate
-0.237284
-4.804194
0.000002
pupil/teacher
-0.084627
-2.720154
0.006756
blackpop
0.144584
4.584164
0.000006
lowclass
-0.243651
-5.941416
0.000000
lambda
0.786425
19.084355
0.000000
Gibbs sampling spatial error Tobit model
hprice
R-squared
=
0.7313
sigma^2
=
0.1493
r-value
=
4
Nobs, Nvars
=
506,
14
# censored values =
16
ndraws,nomit
=
1100,
100
acceptance rate
=
0.4808
time in secs
= 142.1482
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
-0.039910
-0.544675
0.586224
crime
-0.155688
-4.035574
0.000063
zoning
0.049964
1.139609
0.255004
industry
-0.013470
-0.220422
0.825634
charlesr
-0.019834
-0.649249
0.516481
noxsq
-0.108960
-1.401207
0.161783
rooms2
0.277678
5.450260
0.000000
houseage
-0.116271
-2.463411
0.014103
distance
-0.161768
-2.319974
0.020751
access
0.255551
2.665245
0.007946
taxrate
-0.241885
-4.658895
0.000004
pupil/teacher
-0.088207
-2.694077
0.007300
blackpop
0.134053
4.262707
0.000024
lowclass
-0.259032
-5.544375
0.000000
208

lambda
0.728140
12.943930
0.000000
Gibbs sampling general spatial model

hprice
R-squared
=
0.8601
sigma^2
=
0.1468
r-value
=
4
Nobs, Nvars
=
506,
14
ndraws,nomit
=
1100,
100
accept rho rate
=
0.9857
accept lam rate
=
0.5040
time in secs
= 214.1628
min and max rho
=
0.0000,
1.0000
0.0000,
1.0000
***************************************************************
Posterior Estimates
Variable
Coefficient
t-statistic
t-probability
constant
-0.026721
-0.556941
0.577821
crime
-0.167938
-4.414590
0.000012
zoning
0.047543
1.179076
0.238938
industry
0.016986
0.299507
0.764680
charlesr
0.009328
0.343660
0.731249
noxsq
-0.128901
-1.944391
0.052418
rooms2
0.330110
7.369387
0.000000
houseage
-0.099118
-2.108214
0.035518
distance
-0.191117
-3.638851
0.000303
access
0.235348
2.793913
0.005411
taxrate
-0.232871
-4.977616
0.000001
pupil/teacher
-0.108696
-3.825547
0.000147
blackpop
0.133964
4.398670
0.000013
lowclass
-0.297188
-7.347890
0.000000
rho
0.717992
13.037376
0.000000
lambda
0.083388
1.815698
0.070025
Gibbs sampling general spatial Tobit model
hprice
R-squared
=
0.8602
sigma^2
=
0.1333
r-value
=
4
Nobs, Nvars
=
506,
14
# censored values =
16
ndraws,nomit
=
1100,
100
accept rho rate
=
0.9839
accept lam rate
=
0.7113
time in secs
= 249.0965
min and max rho
=
0.0000,
1.0000
0.0000,
1.0000
***************************************************************
Posterior Estimates
209

Variable
constant
crime
zoning
industry
charlesr
noxsq
rooms2
houseage
distance
access
taxrate
pupil/teacher
blackpop
lowclass
rho
lambda
Coefficient
-0.040444
-0.155084
0.045257
0.005911
-0.003574
-0.106601
0.292600
-0.104126
-0.173827
0.214266
-0.239525
-0.110112
0.131408
-0.283951
0.666444
0.100070
t-statistic
-0.967258
-4.123056
1.152237
0.111250
-0.144284
-1.739330
6.640997
-2.252888
-3.428468
2.627556
-5.153612
-4.067087
4.760042
-6.968683
9.336428
1.795139
210
t-probability
0.333890
0.000044
0.249783
0.911463
0.885335
0.082602
0.000000
0.024706
0.000658
0.008869
0.000000
0.000055
0.000003
0.000000
0.000000
0.073245
Contrary to our expectation that these two sets of estimates would produce identical inferences, an interesting and perhaps substantive conclusion
arises. In comparing the estimates that ignore sample censoring to the tobit
estimates we find further evidence regarding the noxsq air pollution variable. For all of the tobit models, this variable is less significant than for
the non-tobit models. Recall that in Chapter 3 we found that maximum
likelihood estimates produced estimates for this variable that were significantly different from zero for all three spatial autoregressive models at the
traditional 5% level. After introducing the Bayesian heteroscedastic spatial
models we found estimates that were not significantly different from zero
at the 5% level, but still significant at the 10% level. After introducing tobit variants of the Bayesian heteroscedastic spatial models we find further
movement away from significance for this variable. In the case of the SEM
model we find that noxsq has a marginal probability of 0.16. Recall that
the SEM and SAC models were judged to be most appropriate for this data
set.
It is especially interesting that none of the other variables in the models
change in significance or magnitude by very much. This is as we would
expect, given the small number (16) of censored observations in a relatively
large sample (506).
5.6
Chapter Summary
A Gibbs sampling approach to estimating heteroscedastic spatial autoregressive and spatial error probit and tobit models was presented. With the
211
exception of McMillen (1992) who set forth an EM algorithm approach to

estimating spatial autoregressive models in the presence of heteroscedastic
disturbances, no other methods exist for producing estimates under these
conditions. It was argued that the Bayesian approach set forth here has
several advantages over the EM algorithm approach suggested by McMillen
(1992). First, the method produces posterior distributions for all parameters in the model whereas McMillens approach does not provide estimates
of precision for the spatial parameters and .. The posteriors allow for
inferences regarding the mean and dispersion of all parameters, including
the important spatial lag
A second advantage is that the Gibbs sampled measures of dispersion
based on the posterior distributions are valid whereas the EM algorithm
produces consistent estimates of dispersion that are likely to overstate parameter precision. Some evidence of overstatement was in fact found in the
results in Table 5.1.
Perhaps the greatest advantage of the Bayesian approach introduced
here is that no model for the non-constant variance need be specified by the
investigator. The Gibbs sampling approach produces estimates of the nonconstant variance for every observation in space. These estimates can be
used to draw inferences regarding the presence of spatial outliers or general
patterns of non-constant variance over space.
Another point is that the EM methods introduced in McMillen do not
apply to tobit models where the likelihood function takes a more complicated
form than the probit model. The Gibbs sampling approach introduced here
applies to the tobit model as well as probit and is equally easy to implement.
In fact, the Gibbs sampling approach to estimating the heteroscedastic spatial probit model introduced here subsumes a logit version of the model as
a special case.
Finally, because the approach introduced here is quite similar to the
Gibbs sampling approach for spatial autoregressive and spatial error models
presented in Chapter 3, it provides a unified methodology for estimating
spatial models that involve continuous or dichotomous dependent variables.
5.7
References
Albert, James H. and Siddhartha Chib (1993), Bayesian Analysis of
Binary and Polychotomous Response Data, Journal of the American
Statistical Association, Volume 88, number 422, pp. 669-679.
Anselin, Luc (1988). Spatial Econometrics: Methods and Models,
212
(Dordrecht: Kluwer Academic Publishers).

Anselin, Luc and D.A. Griffith (1988). Do spatial effects really matter
in regression analysis? Papers of the Regional Science Association, 65,
pp. 11-34.
Chib, Siddhartha (1992), Bayes Inference in the Tobit Censored Regression Model, Journal of Econometrics, Volume 51, pp. 79-99.
Gelfand, Alan E., and A.F.M Smith. 1990. Sampling-Based Approaches to Calculating Marginal Densities, Journal of the American
Statistical Association, Vol. 85, pp. 398-409.
Geweke, John (1991) Evaluating the Accuracy of Sampling-Based
Approaches to the Calculation of Posterior Moments, Federal Reserve
Bank of Minneapolis, Research Department Staff Report 148.
Geweke, John, (1992a). Bayesian Treatment of the Heteroscedastic
Normal and Independent Student-t Linear Models, Federal Reserve
Bank of Minneapolis, Institute for Empirical Macroeconomics, Working paper 495.
Geweke John, (1992b). Priors for Macroeconomic Time Series and
Their Application, Federal Reserve Bank of Minneapolis, Research
Department, Discussion Paper 64.
Gilley, O.W., and R. Kelley Pace. 1996. On the Harrison and Rubinfeld Data, Journal of Environmental Economics and Management,
Vol. 31 pp. 403-405.
McMillen, Daniel P. (1992) Probit with spatial autocorrelation, Journal of Regional Science, Volume 32, number 3, pp. 335-348.
Chapter 6
VAR and Error Correction

Models
This chapter describes vector autoregressive (VAR) and error correction
(EC) models which have been used to model regional labor markets and
other types of time-series representing regional economic activity. The MATLAB functions described here provide a way to implement more appropriate
spatial prior information in Bayesian vector autoregressive models than that
found in RATS software.
Section 6.1 describes the basic VAR model and our function to estimate
and print results for this method. Section 6.2 turns attention to EC models
while Section 6.3 discusses Bayesian spatial variants on these models. Finally, we take up forecasting in Section 6.4 and Section 6.5 provides applied
examples.
6.1
VAR models
A VAR model is shown in (6.1) that contains n variables. The it denote

independent disturbances, Ci represent constants and yit , i = 1, . . . , n denote
the n variables in the model at time t. Model parameters Aij (`) take the
P
k
k
form, m
k=1 aijk ` , where ` is the lag operator defined by ` yt = ytk , and
m is the lag length specified by the modeler.
213
CHAPTER 6. VAR AND ERROR CORRECTION MODELS
y1t
y2t
..
.
ynt
. . . A1n (`)
..
..
.
.
An1 (`) . . . Ann (`)
A11 (`)

..
=
.
y1t
y2t
..
.
ynt
C1
C2
..
.
Cn
1t
2t
..
.
214
(6.1)
nt
The VAR model posits a set of relationships between past lagged values
of all variables in the model and the current value of each variable in the
model. For example, if the yit represent employment in state i at time
t, the VAR model structure allows employment variation in each state to
be explained by past employment variation in the state itself, yitk , k =
1, . . . , m as well as past employment variation in other states, yjtk , k =
1, . . . , m, j 6= i. This is attractive since regional or state differences in
business cycle activity suggest lead/lag relationships in employment of the
type set forth by the VAR model structure.
The model is estimated using ordinary least-squares, so we can draw on
our ols routine from the Econometrics Toolbox. A function var produces
estimates for the coefficients in the VAR model as well as related regression
statistics and Granger-causality test statistics.
The documentation for the var function is:
PURPOSE: performs vector autoregressive estimation
--------------------------------------------------USAGE: result = var(y,nlag,x)
where:
y
= an (nobs x neqs) matrix of y-vectors
nlag = the lag length
x
= optional matrix of variables (nobs x nx)
NOTE:
constant vector automatically included
--------------------------------------------------RETURNS a structure
results.meth = var
results.nobs = nobs, # of observations
results.neqs = neqs, # of equations
results.nlag = nlag, # of lags
results.nvar = nlag*neqs+nx+1, # of variables per equation
--- the following are referenced by equation # --results(eq).beta = bhat for equation eq
results(eq).tstat = t-statistics
results(eq).tprob = t-probabilities
results(eq).resid = residuals
results(eq).yhat = predicted values
results(eq).y
= actual values
results(eq).sige = ee/(n-k)
results(eq).rsqr = r-squared
215
results(eq).rbar = r-squared adjusted

results(eq).boxq = Box Q-statistics
results(eq).ftest = Granger F-tests
results(eq).fprob = Granger marginal probabilities
--------------------------------------------------SEE ALSO: varf, prt_var, pgranger, pftests
---------------------------------------------------
This function utilizes a new aspect of MATLAB structure variables, arrays that can store information for each equation in the VAR model. Estimates of the parameters for the first equation can be accessed from
the results structure using: result(1).beta as can other results that are
equation-specific.
In most applications, the user would simply pass the results structure
on to the prt function that will provide an organized printout of the regression results for each equation. Here is an example of a typical program to
estimate a VAR model.
% ----- Example 6.1 Using the var() function
dates = cal(1982,1,12);
load test.dat;
% monthly mining employment in 8 states
y = growthr(test(:,1:2),12); % (use only two states) convert to growth-rates
yt = trimr(y,dates.freq,0); % truncate to account for lags in growth-rates
dates = cal(1983,1,1);
% redefine calendar for truncation
vnames
= strvcat(illinos,indiana);
nlag = 2;
result = var(yt,nlag);
% estimate 2-lag VAR model
prt(result,vnames);
% printout results
It would produce the following printout of the estimation results:

***** Vector Autoregressive Model *****
illinos
R-squared
=
0.9044
Rbar-squared =
0.9019
sige
=
3.3767
Q-statistic
=
0.2335
Nobs, Nvars
=
159,
5
******************************************************************
Variable
Coefficient
t-statistic
t-probability
illinos lag1
1.042540
13.103752
0.000000
illinos lag2
-0.132170
-1.694320
0.092226
indiana lag1
0.228763
3.790802
0.000215
indiana lag2
-0.213669
-3.538905
0.000531
constant
-0.333739
-1.750984
0.081940
****** Granger Causality Tests *******

Variable
illinos
indiana
F-value
363.613553
7.422536
216
Probability
0.000000
0.000837
indiana
R-squared
=
0.8236
Rbar-squared =
0.8191
sige
=
6.5582
Q-statistic
=
0.0392
Nobs, Nvars
=
159,
5
******************************************************************
Variable
Coefficient
t-statistic
t-probability
illinos lag1
0.258853
2.334597
0.020856
illinos lag2
-0.195181
-1.795376
0.074555
indiana lag1
0.882544
10.493894
0.000000
indiana lag2
-0.029384
-0.349217
0.727403
constant
-0.129405
-0.487170
0.626830
Variable
F-value
Probability
illinos
2.988892
0.053272
indiana
170.063761
0.000000
There are two utility functions that help analyze VAR model Grangercausality output. The first is pgranger, which prints a matrix of the
marginal probabilities associated with the Granger-causality tests in a convenient format for the purpose of inference. The documentation is:
PURPOSE: prints VAR model Granger-causality results
-------------------------------------------------USAGE: pgranger(results,varargin);
where: results = a structure returned by var(), ecm()
varargin = a variable input list containing
vnames = an optional variable name vector
cutoff = probability cutoff used when printing
usage example 1: pgranger(result,0.05);
example 2: pgranger(result,vnames);
example 3: pgranger(result,vnames,0.01);
example 4: pgranger(result,0.05,vnames);
---------------------------------------------------e.g. cutoff = 0.05 would only print
marginal probabilities < 0.05
--------------------------------------------------NOTES: constant term is added automatically to vnames list
you need only enter VAR variable names plus deterministic
---------------------------------------------------
As example of using this function, consider our previous program to estimate the VAR model for monthly mining employment in eight states. Rather
217
than print out the detailed VAR estimation results, we might be interested
in drawing inferences regarding Granger-causality from the marginal probabilities. The following program would produce a printout of just these probabilities. It utilizes an option to suppress printing of probabilities greater
than 0.1, so that our inferences would be drawn on the basis of a 90% confidence level.
% ----- Example 6.2 Using the pgranger() function
dates = cal(1982,1,12);
% monthly data starts in 82,1
load test.dat;
% monthly mining employment in 8 states
y = growthr(test,12);
% convert to growth-rates
yt = trimr(y,dates.freq,0); % truncate
dates = cal(1983,1,1);
% redefine the calendar for truncation
vname = strvcat(il,in,ky,mi,oh,pa,tn,wv);
nlag = 12;
res = var(yt,nlag);
% estimate 12-lag VAR model
cutoff = 0.1;
% examine Granger-causality at 90% level
pgranger(res,vname,cutoff); % print Granger probabilities
We use the NaN symbol to replace marginal probabilities above the

cutoff point (0.1 in the example) so that patterns of causality are easier to
spot. The results from this program would look as follows:
****** Granger Causality Probabilities *******
Variable
il
in
ky
mi
oh
il
0.00
0.01
0.01
NaN
NaN
in
0.02
0.00
NaN
NaN
NaN
ky
NaN
NaN
0.00
NaN
0.10
mi
NaN
0.01
NaN
0.00
NaN
oh
NaN
0.05
0.08
NaN
0.00
pa
0.05
NaN
NaN
NaN
NaN
tn
0.02
0.05
NaN
NaN
NaN
wv
0.02
0.05
0.06
0.01
NaN
pa
0.04
NaN
NaN
NaN
0.01
0.00
0.09
0.00
tn
0.09
NaN
0.07
NaN
NaN
NaN
0.00
NaN
wv
0.02
NaN
NaN
NaN
0.01
0.06
NaN
0.03
The format of the output is such that the columns reflect the Grangercausal impact of the column-variable on the row-variable. That is, Indiana,
Kentucky, Pennsylvania, Tennessee and West Virginia exert a significant
Granger-causal impact on Illinois employment whereas Michigan and Ohio
do not. Indiana exerts the most impact, affecting Illinois, Michigan, Ohio,
Tennessee, and West Virginia.
The second utility is a function pftest that prints just the Grangercausality joint F-tests from the VAR model. Use of this function is similar
to pgranger, we simply call the function with the results structure returned
by the var function, e.g., pftest(result,vnames), where the vnames argument is an optional string-vector of variable names. This function would
218
produce the following output for each equation of a VAR model based on
all eight states:
Equation illinois
F-value
illinois
395.4534
indiana
3.3255
kentucky
0.4467
michigan
0.6740
ohio
2.9820
pennsylvania
6.6383
tennessee
0.9823
west virginia
3.0467
F-probability
0.0000
0.0386
0.6406
0.5112
0.0536
0.0017
0.3768
0.0504
For an example of using Granger causality tests in regional modeling see

LeSage and Reed (1990).
Although the illustrations so far have not involved use of deterministic
variables in the VAR model, the var function is capable of handling these
variables. As an example, we could include a set of seasonal dummy variables
in the VAR model using:
% ----- Example 6.3 VAR with deterministic variables
dates = cal(1982,1,12);
% monthly data starts in 82,1
load test.dat;
y = test;
% use levels data
[nobs neqs] = size(test);
sdum = sdummy(nobs,dates.freq); % create seasonal dummies
sdum = trimc(sdum,1,0);
% omit 1 column because we have
% a constant included by var()
vnames = strvcat(illinos,indiana,kentucky,michigan,ohio, ...
pennsylvania,tennessee,west virginia);
dnames = strvcat(dum1,dum2,dum3,dum4,dum5,dum6,dum7, ...
dum8,dum9,dum10,dum11);
vnames = strvcat(vnames,dnames);
nlag = 12;
result = var(y,nlag,sdum);
prt(result,vnames);
A handy option on the prt function is the ability to print the VAR model
estimation results to an output file. Because these results are quite large,
they can be difficult to examine in the MATLAB command window.
In addition to the prt function, there is plt that produce graphs of the
actual versus predicted and residuals for these models.
One final issue associated with specifying a VAR model is the lag length
to employ. A commonly used approach to determining the lag length is to
219
perform statistical tests of models with longer lags versus shorter lag lengths.
We view the longer lag models as an unrestricted model versus the restricted
shorter lag version of the model, and construct a likelihood ratio statistic
to test for the significance of imposing the restrictions. If the restrictions
are associated with a statistically significant degradation in model fit, we
conclude that the longer lag length model is more appropriate, rejecting the
shorter lag model.
Specifically, the chi-squared distributed test statistic which has degrees
of freedom equation to the number of restrictions imposed is:
LR = (T c)(log|r | log|u |)
(6.2)
where T is the number of observations, c is a degrees of freedom correction

factor proposed by Sims (1980), and |r |, |u| denote the determinant of
the error covariance matrices from the restricted and unrestricted models
respectively. The correction factor, c, recommended by Sims was the number
of variables in each unrestricted equation of the VAR model.
A function lrratio implements a sequence of such tests beginning at a
maximum lag (specified by the user) down to a minimum lag (also specified
by the user). The function prints results to the MATLAB command window
along with marginal probability levels. As an example, consider the following
program to determine the statistically optimal lag length to use for our
VAR model involving the eight-state sample of monthly employment data
for the mining industry.
% ----- Example 6.4 Using the lrratio() function
load test.dat;
y = test; % use all eight states
maxlag = 12;
minlag = 3;
% Turn on flag for Sims correction factor
sims = 1;
disp(LR-ratio results with Sims correction);
lrratio(y,maxlag,minlag,sims);
The output from this program is:

LR-ratio results
nlag = 12 11, LR
nlag = 11 10, LR
nlag = 10 9, LR
nlag = 9 8, LR
nlag = 8 7, LR
nlag = 7 6, LR
with Sims
statistic
statistic
statistic
statistic
statistic
statistic
correction
=
75.6240,
=
89.9364,
=
91.7983,
=
108.8114,
=
125.7240,
=
114.2624,
probability
probability
probability
probability
probability
probability
=
=
=
=
=
=
0.1517
0.01798
0.01294
0.0004052
6.573e-06
0.0001146

nlag =
nlag =
nlag =
6
5
4
5, LR statistic =
4, LR statistic =
3, LR statistic =
220
81.3528, probability = 0.07059

118.5982, probability = 4.007e-05
127.1812, probability = 4.489e-06
There exists an option flag to use the degrees of freedom correction

suggested by Sims, whereas the default behavior of lrratio is to set c = 0.
Example 4.4 turns on the correction factor by setting a flag that we named
sims equal to 1. The results suggest that the lag length of 11 cannot be
viewed as significantly degrading the fit of the model relative to a lag of
12. For the comparison of lags 11 and 10, we find that at the 0.05 level,
we might reject lag 10 in favor of lag 11 as the optimal lag length. On the
other hand, if we employ a 0.01 level of significance, we would conclude that
the optimal lag length is 9, because the likelihood ratio tests reject lag 8 as
significantly degrading the fit of the model at the 0.01 level of confidence.
Another function for use with VAR models is irf that estimates response
functions and provides a graphical presentation of the results. LeSage and
Reed (1989a, 1989b) provide examples of using impulse response functions
to examining regional wages in an urban hierarchy.
6.2
Error correction models
We provide a cursory introduction to co-integration and error correction

models and refer the reader to an excellent laymans introduction by Dickey,
Jansen and Thornton (1991) as well as a more technical work by Johansen
(1995). LeSage (1990) and Shoesmith (1995) cover co-integration and EC
models in the context of forecasting.
Focusing on the practical case of I(1), (integrated of order 1) series,
let yt be a vector of n time-series that are I(1). An I(1) series requires one
difference to transform it to a zero mean, purely non-deterministic stationary
process. The vector yt is said to be co-integrated if there exists an n x r
matrix such that:
zt = 0 yt
(6.3)
Engle and Granger (1987) provide a Representation Theorem stating

that if two or more series in yt are co-integrated, there exists an error correction representation taking the following form:
yt = A(`)yt + zt1 + t
(6.4)
where is a matrix of coefficients of dimension n x r of rank r, zt1 is of

dimension r x 1 based on r n 1 equilibrium error relationships, zt =
221
0 yt from (6.3), and t is a stationary multivariate disturbance. The error

correction (EC) model in (6.4) is simply a VAR model in first-differences
with r lagged error correction terms (zt1 ) included in each equation of the
model. If we have deterministic components in yt , we add these terms as
well as the error correction variables to each equation of the model.
With the case of only two series yt and xt in the model, a two-step
procedure proposed by Engle and Granger (1987) can be used to determine
the co-integrating variable that we add to our VAR model in first-differences
to make it an EC model. The first-step involves a regression: yt = +xt +zt
to determine estimates of and zt . The second step carries out tests on zt to
determine if it is stationary, I(0). If we find this to be the case, the condition
yt = + xt is interpreted as the equilibrium relationship between the two
series and the error correction model is estimated as:
yt = 1 zt1 + lagged(xt, yt) + c1 + 1t
xt = 2 zt1 + lagged(xt, yt) + c2 + 2t
where: zt1 = yt1 xt1 , ci are constant terms and it denote disturbances in the model.
We provide a function adf, (augmented Dickey-Fuller) to test time-series
for the I(1), I(0) property, and another routine cadf (co-integrating augmented Dickey-Fuller) to carry out the tests from step two above on zt to
determine if it is stationary, I(0). These routines as well as the function
johansen that implements a multivariate extension of the two-step Engle
and Granger procedure were designed to mimic a set of Gauss functions by
Sam Quilaris named coint.
The adf function documentation is:
PURPOSE: carry out DF tests on a time series vector
--------------------------------------------------USAGE: results = adf(x,p,nlag)
where:
x = a time-series vector
p = order of time polynomial in the null-hypothesis
p = -1, no deterministic part
p = 0, for constant term
p = 1, for constant plus time-trend
p > 1, for higher order polynomial
nlags = # of lagged changes of x included
results.meth = adf
results.alpha = estimate of the autoregressive parameter
222
results.adf = ADF t-statistic

results.crit = (6 x 1) vector of critical values
[1% 5% 10% 90% 95% 99%] quintiles
results.nlag = nlag
This would be used to test a time-series vector for I(1) or I(0) status.
Allowance is made for polynomial time trends as well as constant terms in
the function and a set of critical values are returned in a structure by the
function. A function prt coint (as well as prt) can be used to print output
from adf, cadf and johansen, saving users the work of formatting and
printing the result structure output.
The function cadf is used for the case of two variables, yt , xt, where we
wish to test whether the condition yt = xt can be interpreted as an equilibrium relationship between the two series. The function documentation
is:
PURPOSE: compute augmented Dickey-Fuller statistic for residuals
from a cointegrating regression, allowing for deterministic
polynomial trends
-----------------------------------------------------------USAGE: results = cadf(y,x,p,nlag)
where: y = dependent variable time-series vector
nlag = # of lagged changes of the residuals to include in regression
-----------------------------------------------------------RETURNS: results structure
results.meth = cadf
results.alpha = autoregressive parameter estimate
results.adf
= ADF t-statistic
results.crit = (6 x 1) vector of critical values
[1% 5% 10% 90% 95% 99%] quintiles
results.nvar = cols(x)
results.nlag = nlag
As an illustration of using these two functions, consider testing our two

monthly time-series on mining employment in Illinois and Indiana for I(1)
status and then carrying out a test of whether they exhibit an equilibrating
relationship. The program would look as follows:
% ----- Example 6.5 Using the adf() and cadf() functions
dates = cal(1982,1,12);
223
load test.dat;
y = test(:,1:2); % use only two series
vnames = strvcat(illinois,indiana);
% test Illinois for I(1) status
nlags = 6;
for i=1:nlags;
res = adf(y(:,1),0,i);
prt(res,vnames(1,:));
end;
% test Indiana for I(1) status
nlags = 6;
for i=1:nlags;
res = adf(y(:,2),0,i);
prt(res,vnames(2,:));
end;
% test if Illinois and Indiana are co-integrated
for i=1:nlags;
res = cadf(y(:,1),y(:,2),0,i);
prt(res,vnames);
end;
The program sets a lag length of 6, and loops over lags of 1 to 6 to

provide some feel for how the augmented Dickey-Fuller tests are affected
by the number of lags used. We specify p = 0 because the employment
time-series do not have zero mean, so we wish to include a constant term.
The result structures returned by the adf and cadf functions are passed on
to prt for printing. We present the output for only lag 6 to conserve on
space, but all lags produced the same inferences. One point to note is that
the adf and cadf functions return a set of 6 critical values for significance
levels 1%,5%,10%,90%,95%,99% as indicated in the documentation for these
functions. Only three are printed for purposes of clarity, but all are available
in the results structure returned by the functions.
Augmented DF test for unit root variable:
ADF t-statistic
# of lags
AR(1) estimate
-0.164599
6
0.998867
1% Crit Value
5% Crit Value
10% Crit Value
-3.464
-2.912
-2.588
illinois
Augmented DF test for unit root variable:

ADF t-statistic
# of lags
AR(1) estimate
-0.978913
6
0.987766
1% Crit Value
5% Crit Value
10% Crit Value
-3.464
-2.912
-2.588
indiana
Augmented DF test for co-integration variables:
illinois,indiana

CADF t-statistic
-1.67691570
1% Crit Value
-4.025
# of lags
6
5% Crit Value
-3.404
224
AR(1) estimate
-0.062974
10% Crit Value
-3.089
We see from the adf function results that both Illinois and Indiana are
I(1) variables. We reject the augmented Dickey-Fuller hypothesis of I(0)
because our t-statistics for both Illinois and Indiana are less than (in absolute
value terms) the critical value of -2.588 at the 90% level.
From the results of cadf we find that Illinois and Indiana mining employment are not co-integrated, again because the t-statistic of -1.67 does not
exceed the 90% critical value of -3.08 (in absolute value terms). We would
conclude that an EC model is not appropriate for these two time-series.
For most EC models, more than two variables are involved so the Engle
and Granger two-step procedure needs to be generalized. Johansen (1988)
provides this generalization which takes the form of a likelihood-ratio test.
We implement this test in the function johansen. The Johansen procedure
provides a test statistic for determining r, the number of co-integrating
relationships between the n variables in yt as well as a set of r co-integrating
vectors that can be used to construct error correction variables for the EC
model.
As a brief motivation for the work carried out by the johansen function,
we start with a reparameterization of the EC model:
yt = 1 yt1 + . . . + k1 ytk+1 ytk + t
(6.5)
where = (In A1 A2 . . . Ak ). If the matrix contains all zeros, (has

rank=0), there are no co-integrating relationships between the variables in
yt . If is of full-rank, then we have n long-run equilibrium relationships,
so all variables in the model are co-integrated. For cases where the matrix
has rank r < n, we have r co-integrating relationships. The Johansen
procedure provides two tests for the number of linearly independent cointegrating relationships among the series in yt , which we have labeled r in
our discussion. Both tests are based on an eigenvalue-eigenvector decomposition of the matrix , constructed from canonical correlations between
yt and ytk with adjustments for intervening lags, and taking into account
that the test is based on estimates that are stochastic. The test statistics
are labeled the trace statistic and the maximal eigenvalue statistic.
Given the value of r, (the number of co-integrating relationships), we
can use the eigenvectors provided by the johansen function along with the
levels of yt lagged one period to form a set of error correction variables for
225
our EC model. In practice, the function ecm does this for you, so you need
not worry about the details.
The documentation for johansen is shown below. A few points to note,
the published literature contains critical values for the trace statistic for
VAR models with up to 12 variables in Johansen (1995), and for the maximal
eigenvalue statistic, Johansen and Juselius (1988) present critical values for
VAR models containing up to 5 variables. To extend the number of variables
for which critical values are available, a procedure by MacKinnon (1996) was
used to generate critical values for both the trace and maximal eigenvalue
statistics for models with up to 12 variables. MacKinnons method is an
approximation, but it produces values close to those in Johansen (1995). The
critical values for the trace statistic have been entered in a function c sjt
and those for the maximal eigenvalue statistic are in c sja. The function
johansen calls these two functions to obtain the necessary critical values. In
cases where the VAR model has more than 12 variables, zeros are returned
as critical values in the structure field result.cvt for the trace statistic and
the result.cvm field for the maximal eigenvalue.
Another less serious limitation is that the critical values for these statistics are only available for trend transformations where 1 p 1. This
should not present a problem in most applications where p will take on
values of -1, 0 or 1.
PURPOSE: perform Johansen cointegration tests
------------------------------------------------------USAGE: result = johansen(x,p,k)
where:
x = input matrix of time-series in levels, (nobs x m)
k = number of lagged difference terms used when
computing the estimator
------------------------------------------------------RETURNS: a results structure:
result.eig = eigenvalues (m x 1)
result.evec = eigenvectors (m x m), where first
r columns are normalized coint vectors
result.lr1 = likelihood ratio trace statistic for r=0 to m-1
(m x 1) vector
result.lr2 = maximum eigenvalue statistic for r=0 to m-1
(m x 1) vector
result.cvt = critical values for trace statistic
(m x 3) vector [90% 95% 99%]
226
result.cvm = critical values for max eigen value statistic

(m x 3) vector [90% 95% 99%]
result.ind = index of co-integrating variables ordered by
size of the eigenvalues from large to small
------------------------------------------------------NOTE: c_sja(), c_sjt() provide critical values generated using
a method of MacKinnon (1994, 1996).
critical values are available for n<=12 and -1 <= p <= 1,
zeros are returned for other cases.
As an illustration of the johansen function, consider the eight-state

sample of monthly mining employment. We would test for the number of
co-integrating relationships using the following code:
% ----- Example 6.6 Using the johansen() function
vnames = strvcat(illinos,indiana,kentucky,michigan,ohio, ...
pennsylvania,tennessee,west virginia);
y = load(test.dat); % use all eight states
nlag = 9;
pterm = 0;
result = johansen(y,pterm,nlag);
prt(result,vnames);
The johansen function is called with the y matrix of time-series variables for the eight states, a value of p = 0 indicating we have a constant
term in the model, and 9 lags. (We want p = 0 because the constant term
is necessary where the levels of employment in the states differ.) The lag of
9 was determined to be optimal using the lrratio function in the previous
section.
The johansen function will return results for a sequence of tests against
alternative numbers of co-integrating relationships ranging from r 0 up
to r m 1, where m is the number of variables in the matrix y.
The function prt provides a printout of the trace and maximal eigenvalue
statistics as well as the critical values returned in the johansen results
structure.
Johansen MLE estimates
NULL:
Trace Statistic Crit 90%
r <= 0
illinos
307.689
153.634
r <= 1
indiana
205.384
120.367
r <= 2
kentucky
129.133
91.109
r <= 3
ohio
83.310
65.820
r <= 4
pennsylvania
52.520
44.493
r <= 5
tennessee
30.200
27.067
r <= 6
west virginia
13.842
13.429
Crit 95%
159.529
125.618
95.754
69.819
47.855
29.796
15.494
Crit 99%
171.090
135.982
104.964
77.820
54.681
35.463
19.935

r <= 7
michigan
NULL:
r <= 0
r <= 1
r <= 2
r <= 3
r <= 4
r <= 5
r <= 6
r <= 7
illinos
indiana
kentucky
ohio
pennsylvania
tennessee
west virginia
michigan
0.412
227
2.705
3.841
6.635
Eigen Statistic Crit 90%

102.305
49.285
76.251
43.295
45.823
37.279
30.791
31.238
22.319
25.124
16.359
18.893
13.430
12.297
0.412
2.705
Crit 95%
52.362
46.230
40.076
33.878
27.586
21.131
14.264
3.841
Crit 99%
58.663
52.307
45.866
39.369
32.717
25.865
18.520
6.635
The printout does not present the eigenvalues and eigenvectors, but they
are available in the results structure returned by johansen as they are
needed to form the co-integrating variables for the EC model. The focus of
co-integration testing would be the trace and maximal eigenvalue statistics
along with the critical values. For this example, we find: (using the 95%
level of significance) the trace statistic rejects r 0 because the statistic of
307.689 is greater than the critical value of 159.529; it also rejects r 1,
r 2, r 3, r 4, and r 5 because these trace statistics exceed the
associated critical values; for r 6 we cannot reject H0, so we conclude that
r = 6. Note that using the 99% level, we would conclude r = 4 as the trace
statistic of 52.520 associated with r 4 does not exceed the 99% critical
value of 54.681.
We find a different inference using the maximal eigenvalue statistic. This
statistic allows us to reject r 0 as well as r 1 and r 2 at the 95% level.
We cannot reject r 3, because the maximal eigenvalue statistic of 30.791
does not exceed the critical value of 33.878 associated with the 95% level.
This would lead to the inference that r = 3, in contrast to r = 6 indicated
by the trace statistic. Using similar reasoning at the 99% level, we would
infer r = 2 from the maximal eigenvalue statistics.
After the johansen test determines the number of co-integrating relationships, we can use these results along with the eigenvectors returned by
the johansen function, to form a set of error correction variables. These are
constructed using yt1 , (the levels of y lagged one period) multiplied by the
r eigenvectors associated with the co-integrating relationships to form r cointegrating variables. This is carried out by the ecm function, documented
below.
PURPOSE: performs error correction model estimation
--------------------------------------------------USAGE: result = ecm(y,nlag,r)
where:
y
= an (nobs x neqs) matrix of y-vectors in levels
228

r
= # of cointegrating relations to use
(optional: this will be determined using
Johansens trace test at 95%-level if left blank)
NOTES: constant vector automatically included
x-matrix of exogenous variables not allowed
error correction variables are automatically
constructed using output from Johansens ML-estimator
--------------------------------------------------RETURNS a structure
results.meth = ecm
results.nobs = nobs, # of observations
results.neqs = neqs, # of equations
results.nlag = nlag, # of lags
results.nvar = nlag*neqs+nx+1, # of variables per equation
results.coint= # of co-integrating relations (or r if input)
results.index= index of co-integrating variables ranked by
size of eigenvalues large to small
--- the following are referenced by equation # --results(eq).beta = bhat for equation eq (includes ec-bhats)
results(eq).yhat = predicted values (levels) (nlag+2:nobs,1)
results(eq).dyhat = predicted values (differenced) (nlag+2:nobs,1)
results(eq).y
= actual y-level values (nobs x 1)
results(eq).dy
= actual y-differenced values (nlag+2:nobs,1)
results(eq).ftest = Granger F-tests
results(eq).fprob = Granger marginal probabilities
---------------------------------------------------
The ecm function allows two options for implementing an EC model.

One option is to specify the number of co-integrating relations to use, and the
other is to let the ecm function determine this number using the johansen
function and the trace statistics along with the critical values at the 95%
level of significance. The motivation for using the trace statistic is that it
seems better suited to the task of sequential hypotheses for our particular
decision problem.
An identical approach can be taken to implement a Bayesian variant
of the EC model based on the Minnesota prior as well as a more recent
Bayesian variant based on a random-walk averaging prior. Both of these
are discussed in the next section.
The prt function will produce a printout of the results structure re-
229
turned by ecm showing the autoregressive plus error correction variable

coefficients along with Granger-causality test results as well as the trace,
maximal eigenvalue statistics, and critical values from the johansen procedure. As an example, we show a program to estimate an EC model based
on our eight-state sample of monthly mining employment, where we have
set the lag-length to 2 to conserve on the amount of printed output.
% ----- Example 6.7 Estimating error correction models
y = load(test.dat); % monthly mining employment for
% il,in,ky,mi,oh,pa,tn,wv 1982,1 to 1996,5
vnames = strvcat(il,in,ky,mi,oh,pa,tn,wv);
nlag = 2; % number of lags in var-model
% estimate the model, letting ecm determine # of co-integrating vectors
result = ecm(y,nlag);
prt(result,vnames); % print results to the command window
The printed output is shown below for a single state indicating the presence of two co-integrating relationships involving the states of Illinois and
Indiana. The estimates for the error correction variables are labeled as such
in the printout. Granger causality tests are printed, and these would form
the basis for valid causality inferences in the case where co-integrating relationships existed among the variables in the VAR model.
wv
R-squared
=
0.1975
Rbar-squared =
0.1018
sige
= 341.6896
Nobs, Nvars
=
170,
19
******************************************************************
Variable
Coefficient
t-statistic
t-probability
il lag1
0.141055
0.261353
0.794176
il lag2
0.234429
0.445400
0.656669
in lag1
1.630666
1.517740
0.131171
in lag2
-1.647557
-1.455714
0.147548
ky lag1
0.378668
1.350430
0.178899
ky lag2
0.176312
0.631297
0.528801
mi lag1
0.053280
0.142198
0.887113
mi lag2
0.273078
0.725186
0.469460
oh lag1
-0.810631
-1.449055
0.149396
oh lag2
0.464429
0.882730
0.378785
pa lag1
-0.597630
-2.158357
0.032480
pa lag2
-0.011435
-0.038014
0.969727
tn lag1
-0.049296
-0.045237
0.963978
tn lag2
0.666889
0.618039
0.537480
wv lag1
-0.004150
-0.033183
0.973572
wv lag2
-0.112727
-0.921061
0.358488

ec term il
-2.158992
-1.522859
ec term in
-2.311267
-1.630267
constant
8.312788
0.450423
Variable
F-value
Probability
il
0.115699
0.890822
in
2.700028
0.070449
ky
0.725708
0.485662
mi
0.242540
0.784938
oh
1.436085
0.241087
pa
2.042959
0.133213
tn
0.584267
0.558769
wv
1.465858
0.234146
Johansen MLE estimates
NULL:
Trace Statistic
Crit 90%
Crit 95%
r <= 0
il
214.390
153.634
159.529
r <= 1
in
141.482
120.367
125.618
r <= 2
ky
90.363
91.109
95.754
r <= 3
oh
61.555
65.820
69.819
r <= 4
tn
37.103
44.493
47.855
r <= 5
wv
21.070
27.067
29.796
r <= 6
pa
10.605
13.429
15.494
r <= 7
mi
3.192
2.705
3.841
NULL:
Eigen Statistic
Crit 90%
Crit 95%
r <= 0
il
72.908
49.285
52.362
r <= 1
in
51.118
43.295
46.230
r <= 2
ky
28.808
37.279
40.076
r <= 3
oh
24.452
31.238
33.878
r <= 4
tn
16.034
25.124
27.586
r <= 5
wv
10.465
18.893
21.131
r <= 6
pa
7.413
12.297
14.264
r <= 7
mi
3.192
2.705
3.841
230
0.129886
0.105129
0.653052
Crit 99%
171.090
135.982
104.964
77.820
54.681
35.463
19.935
6.635
Crit 99%
58.663
52.307
45.866
39.369
32.717
25.865
18.520
6.635
The results indicate that given the two lag model, two co-integrating
relationships were found leading to the inclusion of two error correction
variables in the model. The co-integrating relationships are based on the
trace statistics compared to the critical values at the 95% level. From the
trace statistics in the printed output we see that, H0: r 2 was rejected
at the 95% level because the trace statistic of 90.363 is less than the associated critical value of 95.754. Keep in mind that the user has the option
of specifying the number of co-integrating relations to be used in the ecm
function as an optional argument. If you wish to work at the 90% level of
significance, we would conclude from the johansen results that r = 4 cointegrating relationships exist. To estimate an ecm model based on r = 4
we need simply call the ecm function with:
% estimate the model, using 4 co-integrating vectors
231
result = ecm(y,nlag,4);
6.3
Bayesian variants
Despite the attractiveness of drawing on cross-sectional information from

related regional economic time series, the VAR model has empirical limitations. For example, a model with eight variables and six lags produces
49 independent variables in each of the eight equations of the model for a
total of 392 coefficients to estimate. Large samples of observations involving
time series variables that cover many years are needed to estimate the VAR
model, and these are not always available. In addition, the independent
variables represent lagged values, e.g., y1t1 , y1t2, . . . , y1t6, which tend to
produce high correlations that lead to degraded precision in the parameter
estimates. To overcome these problems, Doan, Litterman and Sims (1984)
proposed the use of Bayesian prior information. The Minnesota prior means
and variances suggested take the following form:
i N (1, 2i )
j
N (0, 2j )
(6.6)
where i denotes the coefficients associated with the lagged dependent variable in each equation of the VAR and j represents any other coefficient.
The prior means for lagged dependent variables are set to unity in belief
that these are important explanatory variables. On the other hand, a prior
mean of zero is assigned to all other coefficients in the equation, j in (6.6),
indicating that these variables are viewed as less important in the model.
The prior variances, 2i , specify uncertainty about the prior means i =
1, and 2j indicates uncertainty regarding the means j = 0. Because
the VAR model contains a large number of parameters, Doan, Litterman
and Sims (1984) suggested a formula to generate the standard deviations
as a function of a small number of hyperparameters: , and a weighting
matrix w(i, j). This approach allows a practitioner to specify individual
prior variances for a large number of coefficients in the model using only a
few parameters that are labeled hyperparameters. The specification of the
standard deviation of the prior imposed on variable j in equation i at lag k
is:
ijk = w(i, j)k
uj
ui
(6.7)
232
where
ui is the estimated standard error from a univariate autoregression
involving variable i, so that (
uj /
ui ) is a scaling factor that adjusts for varying magnitudes of the variables across equations i and j. Doan, Litterman
and Sims (1984) labeled the parameter as overall tightness, reflecting the
standard deviation of the prior on the first lag of the dependent variable.
The term k is a lag decay function with 0 1 reflecting the decay
rate, a shrinkage of the standard deviation with increasing lag length. This
has the effect of imposing the prior means of zero more tightly as the lag
length increases, based on the belief that more distant lags represent less
important variables in the model. The function w(i, j) specifies the tightness of the prior for variable j in equation i relative to the tightness of the
own-lags of variable i in equation i.
The overall tightness and lag decay hyperparameters used in the standard Minnesota prior have values = 0.1, = 1.0. The weighting matrix
used is:
W=
1 0.5 . . . 0.5
0.5 1
0.5
..
..
..
. .
.
0.5 0.5 . . . 1
(6.8)
This weighting matrix imposes i = 1 loosely, because the lagged dependent variable in each equation is felt to be an important variable. The
weighting matrix also imposes the prior mean of zero for coefficients on
other variables in each equation more tightly since the j coefficients are
associated with variables considered less important in the model.
A function bvar will provide estimates for this model. The function
documentation is:
PURPOSE: Performs a Bayesian vector autoregression of order n
--------------------------------------------------USAGE: result = bvar(y,nlag,tight,weight,decay,x)
where:
y
= an (nobs x neqs) matrix of y-vectors
tight = Littermans tightness hyperparameter
weight = Littermans weight (matrix or scalar)
decay = Littermans lag decay = lag^(-decay)
x
= an optional (nobs x nx) matrix of variables
NOTE: constant vector automatically included
results.meth
= bvar
233
results.nobs
= nobs, # of observations
results.neqs
= neqs, # of equations
results.nlag
= nlag, # of lags
results.nvar
= nlag*neqs+1+nx, # of variables per equation
results.tight
= overall tightness hyperparameter
results.weight
= weight scalar or matrix hyperparameter
results.decay
= lag decay hyperparameter
--- the following are referenced by equation # --results(eq).beta = bhat for equation eq
results(eq).yhat = predicted values
results(eq).y
= actual values
--------------------------------------------------SEE ALSO: bvarf, var, ecm, rvar, plt_var, prt_var
---------------------------------------------------
The function bvar allows us to input a scalar weight value or a more

general matrix. Scalar inputs will be used to form a symmetric prior, where
the scalar is used on the off-diagonal elements of the matrix. A matrix will
be used in the form submitted to the function.
As an example of using the bvar function, consider our case of monthly
mining employment for eight states. A program to estimate a BVAR model
based on the Minnesota prior is shown below:
% ----- Example 6.8 Estimating BVAR models
nlag = 2;
tight = 0.1; % hyperparameter values
weight = 0.5;
decay = 1.0;
result = bvar(y,nlag,tight,weight,decay);
prt(result,vnames);
The printout shows the hyperparameter values associated with the prior.
It does not provide Granger-causality test results as these are invalid given
the Bayesian prior applied to the model. Results for a single equation of the
mining employment example are shown below.
***** Bayesian Vector Autoregressive Model *****
*****
Minnesota type Prior
*****

PRIOR hyperparameters
tightness =
0.10
decay
=
1.00
Symmetric weights based on
234
0.50
il
R-squared
=
0.9942
Rbar-squared =
0.9936
sige
=
12.8634
Nobs, Nvars
=
171,
17
******************************************************************
Variable
Coefficient
t-statistic
t-probability
il lag1
1.134855
11.535932
0.000000
il lag2
-0.161258
-1.677089
0.095363
in lag1
0.390429
1.880834
0.061705
in lag2
-0.503872
-2.596937
0.010230
ky lag1
0.049429
0.898347
0.370271
ky lag2
-0.026436
-0.515639
0.606776
mi lag1
-0.037327
-0.497504
0.619476
mi lag2
-0.026391
-0.377058
0.706601
oh lag1
-0.159669
-1.673863
0.095996
oh lag2
0.191425
2.063498
0.040585
pa lag1
0.179610
3.524719
0.000545
pa lag2
-0.122678
-2.520538
0.012639
tn lag1
0.156344
0.773333
0.440399
tn lag2
-0.288358
-1.437796
0.152330
wv lag1
-0.046808
-2.072769
0.039703
wv lag2
0.014753
0.681126
0.496719
constant
9.454700
2.275103
0.024149
There exists a number of attempts to alter the fact that the Minnesota
prior treats all variables in the VAR model except the lagged dependent
variable in an identical fashion. Some of the modifications suggested have
focused entirely on alternative specifications for the prior variance. Usually,
this involves a different (non-symmetric) weight matrix W and a larger value
of 0.2 for the overall tightness hyperparameter in place of the value = 0.1
used in the Minnesota prior. The larger overall tightness hyperparameter
setting allows for more influence from other variables in the model. For
example, LeSage and Pan (1995) constructed a weight matrix based on
first-order spatial contiguity to emphasize variables from neighboring states
in a multi-state agricultural output forecasting model. LeSage and Magura
(1991) employed interindustry input-output weights to place more emphasis
on related industries in a multi-industry employment forecasting model.
These approaches can be implemented using the bvar function by constructing an appropriate weight matrix. For example, the first order conti-
235
guity structure for the eight states in our mining employment example can
be converted to a set of prior weights by placing values of unity on the main
diagonal of the weight matrix, and in positions that represent contiguous
entities. An example is shown in (6.9), where row 1 of the weight matrix
is associated with the time-series for the state of Illinois. We place a value
of unity on the main diagonal to indicate that autoregressive values from
Illinois are considered important variables. We also place values of one in
columns 2 and 3, reflecting the fact that Indiana (variable 2) and Kentucky
(variable 3) are states that have borders touching Illinois. For other states
that are not neighbors to Illinois, we use a weight of 0.1 to downweight
their influence in the BVAR model equation for Illinois. A similar scheme
is used to specify weights for the other seven states based on neighbors and
non-neighbors.
W =
1.0
1.0
1.0
0.1
0.1
0.1
0.1
0.1
1.0
1.0
1.0
1.0
1.0
0.1
0.1
0.1
1.0
1.0
1.0
0.1
1.0
0.1
1.0
1.0
0.1
1.0
0.1
1.0
1.0
0.1
0.1
0.1
0.1
1.0
1.0
1.0
1.0
1.0
0.1
1.0
0.1
0.1
0.1
0.1
1.0
1.0
0.1
1.0
0.1
0.1
1.0
0.1
0.1
0.1
1.0
0.1
0.1
0.1
1.0
0.1
1.0
1.0
0.1
1.0
(6.9)
The intuition behind this set of weights is that we really dont believe
the prior means of zero placed on the coefficients for mining employment
in neighboring states. Rather, we believe these variables should exert an
important influence. To express our lack of faith in these prior means, we
assign a large prior variance to the zero prior means for these states by
increasing the weight values. This allows the coefficients for these timeseries variables to be determined by placing more emphasis on the sample
data and less emphasis on the prior.
This could of course be implemented using bvar with a weight matrix
specified, e.g.,
% ----- Example 6.9 Using bvar() with general weights
dates = cal(1982,1,12);
nlag = 2;
tight = 0.1;
decay = 1.0;

w = [1.0
1.0
1.0
0.1
0.1
0.1
0.1
0.1
1.0
1.0
1.0
1.0
1.0
0.1
0.1
0.1
1.0
1.0
1.0
0.1
1.0
0.1
1.0
1.0
0.1
1.0
0.1
1.0
1.0
0.1
0.1
0.1
0.1
1.0
1.0
1.0
1.0
1.0
0.1
1.0
0.1
0.1
0.1
0.1
1.0
1.0
0.1
1.0
0.1
0.1
1.0
0.1
0.1
0.1
1.0
0.1
236
0.1
0.1
1.0
0.1
1.0
1.0
0.1
1.0];
result = bvar(y,nlag,tight,w,decay);
prt(result,vnames);
Another more recent approach to altering the equal treatment character

of the Minnesota prior is a random-walk averaging prior suggested by
LeSage and Krivelyova (1997, 1998).
As noted above, previous attempts to alter the fact that the Minnesota
prior treats all variables in the VAR model except the first lag of the dependent variable in an identical fashion have focused entirely on alternative
specifications for the prior variance. The prior proposed by LeSage and
Krivelyova (1998) involves both prior means and variances motivated by the
distinction between important and unimportant variables in each equation
of the VAR model. To motivate the prior means, consider the weighting
matrix for a five variable VAR model shown in (6.10). The weight matrix
contains values of unity in positions associated with important variables in
each equation of the VAR model and values of zero for unimportant variables. For example, the important variables in the first equation of the
VAR model are variables 2 and 3 whereas the important variables in the
fifth equation are variables 4 and 5.
Note that if we do not believe that autoregressive influences reflected by
lagged values of the dependent variable are important, we have a zero on
the main diagonal of the weight matrix. In fact, the weighting matrix shown
in (6.10) classifies autoregressive influences as important in only two of the
five equations in the VAR system, equations three and five. As an example
of a case where autoregressive influences are totally ignored LeSage and
Krivelyova (1997) constructed a VAR system based on spatial contiguity
that relies entirely on the influence of neighboring states and ignores the
autoregressive influence associated with past values of the variables from
the states themselves.
W =
0
1
1
0
0
1
0
1
0
0
1
1
1
1
0
0
0
0
0
1
0
0
0
1
1
237
(6.10)
The weight matrix shown in (6.10) is standardized to produce row-sums

of unity resulting in the matrix labeled C shown in (6.11).
C=
0
0.5 0.5
0
0
0.5
0
0.5
0
0
0.33 0.33 0.33 0
0
0
0
0.5
0 0.5
0
0
0
0.5 0.5
(6.11)
Using the row-standardized matrix C, we consider the random-walk with

drift that averages over the important variables in each equation i of the
VAR model as shown in (6.12).
yit = i +
n
X
Cij yjt1 + uit
(6.12)
j=1
Expanding expression (6.12) we see that multiplying yjt1 , j = 1, . . . , 5

containing 5 variables at time t1 by the row-standardized weight matrix C
shown in (6.11) produces a set of explanatory variables for each equation of
the VAR system equal to the mean of observations from important variables
in each equation at time t 1 as shown in (6.13).
y1t
y2t
y3t
y4t
y5t
1
2
3
4
5
0.5y2t1 + 0.5y3t1
0.5y1t1 + 0.5y3t1
0.33y1t1 + 0.33y2t1 + 0.33y3t1
0.5y3t1 + 0.5y5t1
0.5y4t1 + 0.5y5t1
u1t
u2t
u3t
u4t
u5t
(6.13)
This suggests a prior mean for the VAR model coefficients on variables
associated with the first own-lag of important variables equal to 1/ci, where
ci is the number of important variables in each equation i of the model. In
the example shown in (6.13), the prior means for the first own-lag of the
important variables y2t1 and y3t1 in the y1t equation of the VAR would
equal 0.5. The prior means for unimportant variables, y1t1 , y4t1 and y5t1
in this equation would be zero.
238
This prior is quite different from the Minnesota prior in that it may
downweight the lagged dependent variable using a zero prior mean to discount the autoregressive influence of past values of this variable. In contrast,
the Minnesota prior emphasizes a random-walk with drift model that relies
on prior means centered on a model: yit = i + yit1 + uit , where the intercept term reflects drift in the random-walk model and is estimated using
a diffuse prior. The random-walk averaging prior is centered on a randomwalk model that averages over important variables in each equation of the
model and allows for drift as well. As in the case of the Minnesota prior,
the drift parameters i are estimated using a diffuse prior.
Consistent with the Minnesota prior, LeSage and Krivelyova use zero
as a prior mean for coefficients on all lags other than first lags. Litterman
(1986) motivates reliance on zero prior means for many of the parameters
of the VAR model by appealing to ridge regression. Recall, ridge regression
can be interpreted as a Bayesian model that specifies prior means of zero
for all coefficients, and as we saw in Chapter 3 can be used to overcome
collinearity problems in regression models.
One point to note about the random walk averaging approach to specifying prior means is that the time series for the variables in the model need
to be scaled or transformed to have similar magnitudes. If this is not the
case, it would make little sense to indicate that the value of a time series
observation at time t was equal to the average of values from important
related variables at time t 1. This should present no problem as time series data can always be expressed in percentage change form or annualized
growth rates which meets our requirement that the time series have similar
magnitudes.
The prior variances LeSage and Krivelyova specify for the parameters in
the model differ according to whether the coefficients are associated with
variables that are classified as important or unimportant as well as the lag
length. Like the Minnesota prior, they impose lag decay to reflect a prior
belief that time series observations from the more distant past exert a smaller
influence on the current value of the time series we are modeling. Viewing
variables in the model as important versus unimportant suggests that the
prior variance (uncertainty) specification should reflect the following ideas:
1. Parameters associated with unimportant variables should be assigned
a smaller prior variance, so the zero prior means are imposed more
tightly or with more certainty.
2. First own-lags of important variables are given a smaller prior variance,
239
so the prior means force averaging over the first own-lags of important
variables.
3. Parameters associated with unimportant variables at lags greater than
one will be given a prior variance that becomes smaller as the lag length
increases to reflect our belief that influence decays with time.
4. Parameters associated with lags other than first own-lag of important
variables will have a larger prior variance, so the prior means of zero
are imposed loosely. This is motivated by the fact that we dont
really have a great deal of confidence in the zero prior mean specification for longer lags of important variables. We think they should exert
some influence, making the prior mean of zero somewhat inappropriate. We still impose lag decay on longer lags of important variables
by decreasing our prior variance with increasing lag length. This reflects the idea that influence decays over time for important as well as
unimportant variables.
It should be noted that the prior relies on inappropriate zero prior means
for the important variables at lags greater than one for two reasons. First, it
is difficult to specify a reasonable alternative prior mean for these variables
that would have universal applicability in a large number of VAR model
applications. The difficulty of assigning meaningful prior means that have
universal appeal is most likely the reason that past studies relied on the
Minnesota prior means while simply altering the prior variances. A prior
mean that averages over previous period values of the important variables
has universal appeal and widespread applicability in VAR modeling. The
second motivation for relying on inappropriate zero prior means for longer
lags of the important variables is that overparameterization and collinearity
problems that plague the VAR model are best overcome by relying on a
parsimonious representation. Zero prior means for the majority of the large
number of coefficients in the VAR model are consistent with this goal of parsimony and have been demonstrated to produce improved forecast accuracy
in a wide variety of applications of the Minnesota prior.
A flexible form with which to state prior standard deviations for variable
j in equation i at lag length k is shown in (6.14).
(aijk ) = N (1/ci, c), j C,
k = 1,
i, j = 1, . . . , n
(aijk ) = N (0, c/k), j C,
k = 2, . . ., m, i, j = 1, . . . , n
(aijk ) = N (0, c/k), j C, k = 1, . . ., m, i, j = 1, . . . , n
(6.14)
240
where:
0 < c < 1
(6.15)
> 1
(6.16)
0< < 1
(6.17)
For variables j = 1, . . ., m in equation i that are important in explaining

variation in variable i, (j C), the prior mean for lag length k = 1 is
set to the average of the number of important variables in equation i and
to zero for unimportant variables (j C). The prior standard deviation
is set to c for the first lag, and obeys the restriction set forth in (6.15),
reflecting a tight imposition of the prior mean that forces averaging over
important variables. To see this, consider that the prior means 1/ci range
between zero and unity so typical c values might be in the range of 0.1
to 0.25. We use c /k for lags greater than one which imposes a decrease
in this variance as the lag length k increases. Equation (6.16) states the
restriction necessary to ensure that the prior mean of zero is imposed on the
parameters associated with lags greater than one for important variables
loosely, relative to a tight imposition of the prior mean of 1/ci on first ownlags of important variables. We use c /k for lags on unimportant variables
whose prior means are zero, imposing a decrease in the variance as the lag
length increases. The restriction in (6.17) would impose the zero means for
unimportant variables with more confidence than the zero prior means for
important variables.
This mathematical formulation adequately captures all aspects of the
intuitive motivation for the prior variance specification enumerated above.
A quick way to see this is to examine a graphical depiction of the prior mean
and standard deviation for an important versus unimportant variable. An
artificial example was constructed for an important variable in Figure 6.1
and an unimportant variable in Figure 4.3. Figure 6.1 shows the prior
mean along with five upper and lower limits derived from the prior standard
deviations in (6.14). The five standard deviation limits shown in the figure
reflect 2 standard deviation limits resulting from alternative settings for
the prior hyperparameter ranging from 5 to 9 and a value of c = 0.25.
Larger values of generated the wider upper and lower limits.
The solid line in Figure 6.1 reflects a prior mean of 0.2 for lag 1 indicating
five important variables, and a prior mean of zero for all other lags. The prior
standard deviation at lag 1 is relatively tight producing a small band around
the averaging prior mean for this lag. This imposes the averaging prior
belief with a fair amount of certainty. Beginning at lag 2, the prior standard
241
0.5
mean
upper
lower
0.4
0.3
prior standard deviation
0.2
0.1
-0.1
-0.2
-0.3
-0.4
-0.5
6
lag length
10
12
Figure 6.1: Prior means and precision for important variables

deviation is increased to reflect relative uncertainty about the new prior
mean of zero for lags greater than unity. Recall, we believe that important
variables at lags greater than unity will exert some influence, making the
prior mean of zero not quite appropriate. Hence, we implement this prior
mean with greater uncertainty.
Figure 6.2 shows an example of the prior means and standard deviations
for an unimportant variable based on c = 0.25 and five values of ranging
from .35 to .75. Again, the larger values produce wider upper and lower
limits. The prior for unimportant variables is motivated by the Minnesota
prior that also uses zero prior means and rapid decay with increasing lag
length.
A function rvar implements the random-walk averaging prior and a related function recm carries out estimation for an EC model based on this
prior. The documentation for the rvar function is shown below, where we
have eliminated information regarding the results structure variable returned
by the function to save space.
PURPOSE: Estimates a Bayesian vector autoregressive model
242
0.2
mean
upper
lower
0.15
prior standard deviation
0.1
0.05
-0.05
-0.1
-0.15
-0.2
6
lag length
10
12
Figure 6.2: Prior means and precision for unimportant variables

using the random-walk averaging prior
--------------------------------------------------USAGE: result = rvar(y,nlag,w,freq,sig,tau,theta,x)
where:
y
= an (nobs x neqs) matrix of y-vectors (in levels)
w
= an (neqs x neqs) matrix containing prior means
(rows should sum to unity, see below)
freq = 1 for annual, 4 for quarterly, 12 for monthly
sig = prior variance hyperparameter (see below)
tau = prior variance hyperparameter (see below)
theta = prior variance hyperparameter (see below)
x
= an (nobs x nx) matrix of deterministic variables
(in any form, they are not altered during estimation)
(constant term automatically included)
priors important variables:
N(w(i,j),sig) for 1st own lag
N( 0 ,tau*sig/k) for lag k=2,...,nlag
priors unimportant variables: N(w(i,j),theta*sig/k) for lag 1
N( 0 ,theta*sig/k) for lag k=2,...,nlag
e.g., if y1, y3, y4 are important variables in eq#1, y2 unimportant
w(1,1) = 1/3, w(1,3) = 1/3, w(1,4) = 1/3, w(1,2) = 0
typical values would be: sig = .1-.3, tau = 4-8, theta = .5-1
243
--------------------------------------------------NOTES: - estimation is carried out in annualized growth terms because

the prior means rely on common (growth-rate) scaling of variables
hence the need for a freq argument input.
- constant term included automatically
---------------------------------------------------
Because this model is estimated in growth-rates form, an input argument

for the data frequency is required. As an illustration of using both the rvar
and recm functions, consider the following example based on the eight-state
mining industry data. We specify a weight matrix for the prior means using
first-order contiguity of the states.
% ----- Example 6.10 Estimating RECM models
y = load(test.dat); % a test data set
sig = 0.1;
tau = 6;
theta = 0.5;
freq = 12;
% monthly data
% this is an example of using 1st-order contiguity
% of the states as weights to produce prior means
W=[0
0.5
0.5
0
0
0
0
0
0.25
0
0.25
0.25 0.25 0
0
0
0.20
0.20
0
0
0.20 0
0.20 0.20
0
0.50
0
0
0.50 0
0
0
0
0.20
0.20
0.20 0
0.20 0.20 0.20
0
0
0
0
0.50 0
0
0.50
0
0
1
0
0
0
0
0
0
0
0.33
0
0.33 0.33 0
0];
% estimate the rvar model
results = rvar(y,nlag,W,freq,sig,tau,theta);
% print results to a file
fid = fopen(rvar.out,wr);
prt(results,vnames,fid);
% estimate the recm model letting the function
% determine the # of co-integrating relationships
results = recm(y,nlag,W,freq,sig,tau,theta);
% print results to a file
fid = fopen(recm.out,wr);
prt(results,vnames,fid);
6.4
Forecasting the models
A set of forecasting functions are available that follow the format of the
var, bvar, rvar, ecm, becm, recm functions named varf, bvarf, rvarf,
244
ecmf, becmf, recmf. These functions all produce forecasts of the timeseries levels to simplify accuracy analysis and forecast comparison from the
alternative models. They all take time-series levels arguments as inputs and
carry out the necessary transformations. As an example, the varf documentation is:
PURPOSE: estimates a vector autoregression of order n
and produces f-step-ahead forecasts
------------------------------------------------------------USAGE:yfor = varf(y,nlag,nfor,begf,x,transf)
where:
y
= an (nobs * neqs) matrix of y-vectors in levels
nfor = the forecast horizon
begf = the beginning date of the forecast
(defaults to length(x) + 1)
x
= an optional vector or matrix of deterministic
variables (not affected by data transformation)
transf = 0, no data transformation
= 1, 1st differences used to estimate the model
= freq, seasonal differences used to estimate
= cal-structure, growth rates used to estimate
e.g., cal(1982,1,12) [see cal() function]
------------------------------------------------------------NOTE: constant term included automatically
------------------------------------------------------------RETURNS:
yfor = an nfor x neqs matrix of level forecasts for each equation
-------------------------------------------------------------
Note that you input the variables y in levels form, indicate any of four
data transformations that will be used when estimating the VAR model,
and the function varf will carry out this transformation, produce estimates
and forecasts that are converted back to levels form. This greatly simplifies
the task of producing and comparing forecasts based on alternative data
transformations.
For the case of a growth rates transformation a calendar structure variable is required. These are produced with the cal function that is part of the
Econometrics Toolbox discussed in Chapter 3 of the manual. In brief, using
cstruct = cal(1982,1,12) would set up the necessary calendar structure if
the data being used were monthly time series that began in 1982.
Of course, if you desire a transformation other than the four provided,
such as logs, you can transform the variables y prior to calling the function
and specify transf=0. In this case, the function does not provide levels
forecasts, but rather forecasts of the logged-levels will be returned. Setting
245
transf=0, produces estimates based on the data input and returns forecasts
based on this data.
As an example of comparing alternative VAR model forecasts based on
two of the four alternative transformations, consider the program in example
6.11.
% ----- Example 6.11 Forecasting VAR models
y = load(test.dat); % a test data set containing
% monthly mining employment for
% il,in,ky,mi,oh,pa,tn,wv
dates = cal(1982,1,12); % data covers 1982,1 to 1996,5
nfor = 12; % number of forecast periods
begf = ical(1995,1,dates); % beginning forecast period
endf = ical(1995,12,dates); % ending forecast period
% no data transformation example
fcast1 = varf(y,nlag,nfor,begf);
% seasonal differences data transformation example
freq = 12; % set frequency of the data to monthly
fcast2 = varf(y,nlag,nfor,begf,[],freq);
% compute percentage forecast errors
actual = y(begf:endf,:);
error1 = (actual-fcast1)./actual;
error2 = (actual-fcast2)./actual;
fdates = cal(1995,1,12);
fprintf(1,VAR model in levels percentage errors \n);
tsprint(error1*100,fdates,vnames,%7.2f);
fprintf(1,VAR - seasonally differenced data percentage errors \n);
tsprint(error2*100,fdates,vnames,%7.2f);
In example 6.11 we rely on the calendar structure variables which are

inputs to the tsprint utility function that prints time series with date labels
as shown in the program output below. This utility function is part of the
Econometrics Toolbox discussed in Chapter 3 of the manual.
VAR model in levels percentage errors
Date
il
in
ky
mi
Jan95
-3.95
-2.86
-1.15
-6.37
Feb95
-5.63
-2.63
-3.57
-7.77
Mar95
-3.62
-1.75
-4.66
-5.49
Apr95
-3.81
-4.23
-7.11
-4.27
May95
-4.05
-5.60
-8.14
-0.92
Jun95
-4.10
-3.64
-8.87
0.10
Jul95
-4.76
-3.76 -10.06
1.99
Aug95
-8.69
-3.89
-9.86
4.85
Sep95
-8.73
-3.63 -12.24
0.70
oh
-5.33
-7.56
-5.67
-5.18
-5.88
-4.65
-1.23
-2.49
-4.33
pa
-7.83
-8.28
-6.69
-5.41
-3.93
-4.15
-5.06
-5.41
-6.28
tn
-0.19
-0.99
2.26
2.14
2.77
2.90
3.44
3.63
3.38
wv
-0.65
0.38
2.30
0.17
-1.11
-2.44
-3.67
-3.59
-4.04

Oct95
Nov95
Dec95
-11.11
-11.79
-12.10
VAR - seasonally
Date
il
Jan95
-6.53
Feb95
-4.35
Mar95
-1.12
Apr95
-0.38
May95
0.98
Jun95
-0.73
Jul95
-1.41
Aug95
-3.36
Sep95
-3.19
Oct95
-2.74
Nov95
-2.47
Dec95
-1.35
-3.23
-4.30
-5.56
-12.10
-11.53
-11.12
-7.38
-8.93
-13.11
-4.74
-4.90
-5.57
246
-8.34
-7.27
-8.78
3.21
3.60
2.13
-5.57
-5.69
-9.38
differenced data percentage errors

in
ky
mi
oh
pa
-0.52
-3.75
3.41
-1.49
-0.06
1.75
-6.29
0.35
-3.53
-2.76
2.61
-6.83
1.53
-2.72
2.24
-2.36
-7.03
-4.30
-1.28
0.70
-5.05
-3.90
-4.65
-1.18
2.02
-2.55
-2.04
-0.30
2.30
0.81
-0.36
-1.69
0.79
4.83
-0.06
2.36
-1.78
7.99
4.86
-1.07
3.47
-3.26
6.91
2.31
-1.44
3.27
-2.88
-2.14
2.92
-0.73
1.54
-2.63
-5.23
4.33
0.36
0.48
-3.53
-7.89
4.38
1.33
tn
3.86
4.46
2.96
5.55
6.49
3.96
7.68
8.75
8.30
9.00
9.02
7.03
wv
0.05
2.56
3.97
2.73
-0.43
-1.44
-4.24
-3.38
-3.02
0.08
0.64
-3.92
It is also possible to build models that produce forecasts that feed-in to

another model as deterministic variables. For example, suppose we wished
to use national employment in the primary metal industry (SIC 33) as a
deterministic variable in our model for primary metal employment in the
eight states. The following program shows how to accomplish this.
% ----- Example 6.12 Forecasting multiple related models
dates = cal(1982,1,12); % data starts in 1982,1
y=load(sic33.states); % industry sic33 employment for 8 states
[nobs neqs] = size(y);
load sic33.national;
% industry sic33 national employment
ndates = cal(1947,1,12);% national data starts in 1947,1
begs = ical(1982,1,ndates); % find 1982,1 for national data
ends = ical(1996,5,ndates); % find 1996,5 for national data
x = sic33(begs:ends,1); % pull out national employment in sic33
% for the time-period corresponding to
% our 8-state sample
begf = ical(1990,1,dates); % begin forecasting date
endf = ical(1994,12,dates); % end forecasting date
nfor = 12;
% forecast 12-months-ahead
nlag = 6;
xerror = zeros(nfor,1);
yerror = zeros(nfor,neqs);
cnt = 0; % counter for the # of forecasts we produce
for i=begf:endf % loop over dates producing forecasts
xactual = x(i:i+nfor-1,1); % actual national employment
yactual = y(i:i+nfor-1,:); % actual state employment
247
% first forecast national employment in sic33

xfor = varf(x,nlag,nfor,i); % an ar(6) model
xdet = [x(1:i-1,1)
% actual national data up to forecast period
xfor
];
% forecasted national data
% do state forecast using national data and forecast as input
yfor = varf(y,nlag,nfor,i,xdet);
% compute forecast percentage errors
xerror = xerror + abs((xactual-xfor)./xactual);
yerror = yerror + abs((yactual-yfor)./yactual);
cnt = cnt+1;
end; % end loop over forecasting experiment dates
% compute mean absolute percentage errors
xmape = xerror*100/cnt; ymape = yerror*100/cnt;
% printout results
in.cnames = strvcat(national,il,in,ky,mi,oh,pa,tn,wv);
rnames = Horizon;
for i=1:12; rnames = strvcat(rnames,[num2str(i),-step]); end;
in.rnames = rnames;
in.fmt = %6.2f;
fprintf(1,national and state MAPE percentage forecast errors \n);
fprintf(1,based on %d 12-step-ahead forecasts \n,cnt);
mprint([xmape ymape],in);
Our model for national employment in SIC33 is simply an autoregressive

model with 6 lags, but the same approach would work for a matrix X of
deterministic variables used in place of the vector in the example. We can
also provide for a number of deterministic variables coming from a variety
of models that are input into other models, not unlike traditional structural
econometric models. The program produced the following output.
national and state MAPE percentage forecast errors
based on 60 12-step-ahead forecasts
Horizon national
il
in
ky
mi
oh
1-step
0.27
0.70
0.78
1.00
1.73
0.78
2-step
0.46
1.02
1.10
1.15
1.95
1.01
3-step
0.68
1.22
1.26
1.39
2.34
1.17
4-step
0.93
1.53
1.45
1.46
2.81
1.39
5-step
1.24
1.84
1.63
1.74
3.27
1.55
6-step
1.55
2.22
1.70
2.05
3.41
1.53
7-step
1.84
2.62
1.59
2.24
3.93
1.68
8-step
2.21
3.00
1.56
2.34
4.45
1.82
9-step
2.55
3.30
1.59
2.58
4.69
1.93
10-step
2.89
3.64
1.74
2.65
5.15
2.08
11-step
3.25
3.98
1.86
2.75
5.75
2.29
12-step
3.60
4.36
1.94
2.86
6.01
2.40
pa
0.56
0.78
1.00
1.25
1.57
1.81
1.99
2.10
2.33
2.51
2.70
2.94
tn
0.88
1.06
1.16
1.35
1.53
1.64
1.76
1.89
1.99
2.12
2.27
2.23
wv
1.08
1.58
1.91
2.02
2.10
2.15
2.49
2.87
3.15
3.39
3.70
3.96
Consider that it would be quite easy to assess the contribution of using

national employment as a deterministic variable in the model by running
248
another model that excludes this deterministic variable.

As a final example, consider an experiment where we wish to examine
the impact of using different numbers of error correction variables on the
forecast accuracy of the EC model. Shoesmith (1995) suggests that one
should employ the number of error correction variables associated with the
Johansen likelihood ratio statistics, but he provides only limited evidence
regarding this contention.
The experiment uses time-series on national monthly employment from
12 manufacturing industries covering the period 1947,1 to 1996,12. Forecasts
are carried out over the period from 1970,1 to 1995,12 using the number
of error correction terms suggested by the Johansen likelihood ratio trace
statistics, as well as models based on +/-1 and +/-2 error correction terms
relative to the value suggested by the trace statistic.
We then compare the relative forecast accuracy of these models by examining the ratio of the MAPE forecast error from the models with +/-1 and
+/-2 terms to the errors from the model based on r relationships suggested
by the trace statistic.
Here is the program code:
% ----- Example 6.13 comparison of forecast accuracy as a function of
%
the # of co-integrating vectors used
load level.mat;
% 20 industries national employment
y = level(:,1:12);
% use only 12 industries
[nobs neqs] = size(y);
dates = cal(1947,1,12);
begf = ical(1970,1,dates); % beginning forecast date
endf = ical(1995,12,dates); % ending forecast date
nfor = 12;
% forecast horizon
nlag = 10; cnt = 1;
% nlag based on lrratio() results
for i=begf:endf;
jres = johansen(y,0,nlag); trstat = jres.lr1; tsignf = jres.cvt;
r = 0;
for j=1:neqs; % find r indicated by trace statistic
if trstat(j,1) > tsignf(j,2), r = j; end;
end;
% set up r-1,r-2 and r+1,r+2 forecasts in addition to forecasts based on r
if (r >= 3 & r <=10)
frm2 = ecmf(y,nlag,nfor,i,r-2); frm1 = ecmf(y,nlag,nfor,i,r-1);
fr
= ecmf(y,nlag,nfor,i,r); frp1 = ecmf(y,nlag,nfor,i,r+1);
frp2 = ecmf(y,nlag,nfor,i,r+2); act = y(i:i+nfor-1,1:12);
% compute forecast MAPE
err(cnt).rm2 = abs((act-frm2)./act); err(cnt).rm1 = abs((act-frm1)./act);
err(cnt).r
= abs((act-fr)./act); err(cnt).rp1 = abs((act-frp1)./act);
err(cnt).rp2 = abs((act-frp2)./act); cnt = cnt+1;
else
fprintf(1,time %d had %d co-integrating relations \n,i,r);
249
end; % end if-else; end; % end of loop over time

rm2 = zeros(12,12); rm1 = rm2; rm0 = rm2; rp1 = rm2; rp2 = rm2;
for i=1:cnt-1;
rm2 = rm2 + err(i).rm2; rm1 = rm1 + err(i).rm1;
rm0 = rm0 + err(i).r;
rp1 = rp1 + err(i).rp1;
rp2 = rp2 + err(i).rp2;
end;
rm2 = rm2/(cnt-1);
rm1 = rm1/(cnt-1);
rm0 = rm0/(cnt-1);
rp1 = rp1/(cnt-1);
rp2 = rp2/(cnt-1);
rnames = Horizon; cnames = [];
for i=1:12;
rnames = strvcat(rnames,[num2str(i),-step]);
cnames = strvcat(cnames,[IND,num2str(i)]);
end;
in.rnames = rnames; in.cnames = cnames; in.fmt = %6.2f;
fprintf(1,forecast errors relative to error by ecm(r) model \n);
fprintf(1,r-2 relative to r \n);
mprint(rm2./rm0,in);
fprintf(1,r-1 relative to r \n);
mprint(rm2./rm0,in);
fprintf(1,r+1 relative to r \n);
mprint(rp1./rm0,in);
fprintf(1,r+2 relative to r \n);
mprint(rp2./rm0,in);
The program code stores the individual MAPE forecast errors in a structure variable using: err(cnt).rm2 = abs((actual-frm2)./actual);, which
will have fields for the errors from all five models. These fields are matrices
of dimension 12 x 12, containing MAPE errors for each of the 12-step-ahead
forecasts for time cnt and for each of the 12 industries. We are not really interested in these individual results, but present this as an illustration.
As part of the illustration, we show how to access the individual results to
compute the average MAPE errors for each horizon and industry. If you
wished to access industry number 2s forecast errors based on the model using r co-integrating relations, for the first experimental forecast period you
would use: err(1).rm(:,2). The results from our experiment are shown
below. These results represent an average over a total of 312 twelve-stepahead forecasts. Our simple MATLAB program produced a total of 224,640
forecasts, based on 312 twelve-step-ahead forecasts, for 12 industries, times
5 models!
Our experiment indicates that using more than the r co-integrating relationships determined by the Johansen likelihood trace statistic degrades
the forecast accuracy. This is clear from the large number of forecast error
ratios greater than unity for the two models based on r + 1 and r + 2 ver-
250
sus those from the model based on r. On the other hand, using a smaller
number of co-integrating relationships than indicated by the Johansen trace
statistic seems to improve forecast accuracy. In a large number of industries
at many of the twelve forecast horizons, we see comparison ratios less than
unity. Further, the forecast errors associated with r 2 are superior to those
from r 1, producing smaller comparison ratios in 9 of the 12 industries.

forecast errors
r-2 relative to
Horizon
I1
1-step
1.01
2-step
0.92
3-step
0.89
4-step
0.85
5-step
0.82
6-step
0.81
7-step
0.79
8-step
0.78
9-step
0.76
10-step
0.76
11-step
0.75
12-step
0.74
r-1 relative to
Horizon
I1
1-step
1.01
2-step
0.92
3-step
0.89
4-step
0.85
5-step
0.82
6-step
0.81
7-step
0.79
8-step
0.78
9-step
0.76
10-step
0.76
11-step
0.75
12-step
0.74
r+1 relative to
Horizon
I1
1-step
1.01
2-step
0.99
3-step
0.99
4-step
0.99
5-step
0.98
6-step
0.98
7-step
0.98
8-step
0.98
9-step
0.97
10-step
0.97
11-step
0.97
12-step
0.97
r+2 relative to
Horizon
I1
1-step
1.00
2-step
1.00
3-step
1.00
4-step
1.00
5-step
1.01
6-step
1.01
7-step
1.01
8-step
1.00
9-step
1.01
10-step
1.01
11-step
1.01
12-step
1.00
relative to
r
I2
I3
0.99 1.00
1.01 0.99
1.04 1.00
1.03 0.99
1.03 0.98
1.05 0.97
1.07 0.96
1.04 0.95
1.03 0.93
1.01 0.92
1.00 0.91
0.99 0.90
r
I2
I3
0.99 1.00
1.01 0.99
1.04 1.00
1.03 0.99
1.03 0.98
1.05 0.97
1.07 0.96
1.04 0.95
1.03 0.93
1.01 0.92
1.00 0.91
0.99 0.90
r
I2
I3
1.00 1.02
1.02 1.01
1.01 1.01
0.99 1.01
0.98 1.03
0.98 1.03
0.98 1.04
0.96 1.05
0.95 1.05
0.96 1.05
0.97 1.05
0.97 1.05
r
I2
I3
1.01 1.02
1.05 1.00
1.02 1.01
0.99 1.01
0.97 1.03
0.95 1.04
0.96 1.06
0.93 1.08
0.92 1.09
0.92 1.09
0.93 1.09
0.93 1.09
251
error by ecm(r) model

I4
1.01
0.96
0.94
0.94
0.94
0.94
0.93
0.93
0.92
0.91
0.91
0.91
I5
1.00
1.03
1.03
1.05
1.03
1.01
0.99
0.98
0.97
0.96
0.95
0.94
I6
1.00
1.00
1.02
1.03
1.03
1.04
1.03
1.02
1.01
0.99
0.98
0.98
I7
1.01
1.02
1.01
1.02
1.04
1.04
1.05
1.04
1.02
1.01
1.01
0.99
I8
0.99
0.99
0.98
1.00
1.00
0.99
0.98
0.96
0.95
0.94
0.95
0.94
I9
1.00
1.01
0.99
0.97
0.97
0.97
0.97
0.96
0.95
0.94
0.94
0.93
I10
0.98
1.03
1.03
1.01
0.98
0.96
0.94
0.93
0.91
0.90
0.90
0.89
I11
0.97
0.99
1.00
1.00
1.02
1.03
1.03
1.03
1.01
0.99
0.99
0.98
I12
0.99
0.94
0.93
0.91
0.92
0.92
0.92
0.93
0.94
0.94
0.95
0.95
I4
1.01
0.96
0.94
0.94
0.94
0.94
0.93
0.93
0.92
0.91
0.91
0.91
I5
1.00
1.03
1.03
1.05
1.03
1.01
0.99
0.98
0.97
0.96
0.95
0.94
I6
1.00
1.00
1.02
1.03
1.03
1.04
1.03
1.02
1.01
0.99
0.98
0.98
I7
1.01
1.02
1.01
1.02
1.04
1.04
1.05
1.04
1.02
1.01
1.01
0.99
I8
0.99
0.99
0.98
1.00
1.00
0.99
0.98
0.96
0.95
0.94
0.95
0.94
I9
1.00
1.01
0.99
0.97
0.97
0.97
0.97
0.96
0.95
0.94
0.94
0.93
I10
0.98
1.03
1.03
1.01
0.98
0.96
0.94
0.93
0.91
0.90
0.90
0.89
I11
0.97
0.99
1.00
1.00
1.02
1.03
1.03
1.03
1.01
0.99
0.99
0.98
I12
0.99
0.94
0.93
0.91
0.92
0.92
0.92
0.93
0.94
0.94
0.95
0.95
I4
1.00
0.99
0.99
0.98
0.99
0.99
1.00
1.00
1.01
1.01
1.01
1.01
I5
1.00
0.99
1.00
1.01
1.01
1.01
1.01
1.02
1.02
1.02
1.02
1.02
I6
1.01
1.03
1.04
1.05
1.05
1.06
1.06
1.06
1.07
1.07
1.07
1.07
I7
1.01
1.00
1.00
1.01
1.01
1.00
1.00
0.99
0.99
0.98
0.98
0.98
I8
0.99
0.99
0.99
1.01
1.03
1.03
1.04
1.05
1.05
1.05
1.06
1.07
I9
1.00
0.99
0.98
0.97
0.97
0.97
0.97
0.97
0.96
0.96
0.95
0.95
I10
1.01
1.05
1.07
1.08
1.08
1.07
1.08
1.06
1.05
1.04
1.05
1.05
I11
1.02
1.03
1.03
1.04
1.04
1.04
1.04
1.04
1.04
1.04
1.04
1.04
I12
1.01
1.04
1.04
1.03
1.04
1.04
1.04
1.04
1.04
1.03
1.03
1.03
I4
1.01
0.97
0.96
0.97
0.98
0.99
1.00
0.99
0.99
0.99
0.99
0.99
I5
0.99
1.00
1.01
1.02
1.04
1.04
1.05
1.05
1.06
1.05
1.05
1.05
I6
1.01
1.03
1.06
1.07
1.08
1.10
1.10
1.10
1.11
1.11
1.12
1.12
I7
1.01
1.01
1.02
1.02
1.02
1.02
1.01
1.00
0.99
0.98
0.98
0.97
I8
0.99
1.00
1.02
1.04
1.07
1.08
1.09
1.10
1.11
1.11
1.13
1.13
I9
0.99
0.99
0.98
0.97
0.97
0.97
0.96
0.95
0.95
0.94
0.94
0.94
I10
1.05
1.11
1.13
1.14
1.15
1.15
1.15
1.15
1.14
1.13
1.14
1.15
I11
1.03
1.03
1.04
1.05
1.05
1.06
1.06
1.07
1.08
1.08
1.09
1.09
I12
1.01
1.06
1.06
1.05
1.04
1.04
1.03
1.02
1.02
1.01
1.00
0.99
6.5
252
An applied exercise
To illustrate using vector autoregressive modeling in a regional data example, we create a model that links national and state employment models.
Monthly employment time series for ten national 2-digit industry categories
is in a Bayesian vector autoregressive model to estimate and forecast employment one-year ahead for the period 1994,1 to 1994,12. We then use
the historical national data as well as the 12 months of forecasted values
as deterministic variables in a regional employment Bayesian vector autoregressive model that models monthly employment for eight contiguous states
in the midwest.
Forecast errors are computed for the 12 month forecast and compared to
the errors made by a Bayesian vector autoregressive model that is identical
except for the inclusion of the national employment information as deterministic variables. This should provide a test of the forecasting value of
national information in the regional model.
Example 6.14 shows the program to carry out the forecasting experiment. The program estimates and forecasts the ten industry national model
using no data transformations. That is, the data are used in levels form.
For the regional model a first-difference transformation is used. This transformation will be applied to the vector autoregressive variables, but not the
deterministic national variables in the model.
% ----- Example 6.14 Linking national and regional models
% load 10 sic national manufacturing employment
% time-series covering the period 1947,1 to 1996,12
% from the MATLAB data file level.mat
dates = cal(1947,1,12);
load level.mat;
% the data is in a matrix variable named level
% which we used to create the file (see make_level.m)
% 20 Food and kindred products
% 21 Tobacco manufactures
% 22 Textile, fabrics, yarn and thread mills
% 23 Miscellaneous fabricated textile products
% 24 Lumber and wood products
% 25 Furniture and fixtures
% 26 Paper and allied products
% 27 Printing and publishing
% 28 Chemicals, plastics, and synthetic materials
% 29 Petroleum refining and related industries
% produce a 12-month-ahead national forecast using a bvar model
begf = ical(1994,1,dates);
begs = ical(1982,1,dates);
253
nlag = 12; tight = 0.2; weight = 0.5; decay = 0.1;

nfor = bvarf(level,nlag,12,begf,tight,weight,decay);
national = [level(begs:begf-1,:)
nfor];
% use the national forecast as a deterministic variable
% in a regional forecasting model
load states.mat; % see make_states.m used to create states.mat
% il,in,ky,mi,oh,pa,tn,wv total state employment
% for 1982,1 to 1996,5
sdates = cal(1982,1,12);
snames = strvcat(il,in,ky,mi,oh,pa,tn,wv);
begf = ical(1994,1,sdates); nlag = 6;
% compute 12-month-ahead state forecast using a bvar model
% in 1st differences with national variables as deterministic
% and national forecasts used for forecasting
sfor = bvarf(states,nlag,12,begf,tight,weight,decay,national,1);
fdates = cal(1994,1,12);
% print forecasts of statewide employment
tsprint(sfor,fdates,1,12,snames,%8.1f);
% print actual statewide employment
tsprint(states(begf:begf+11,:),fdates,1,12,snames,%8.1f);
% compute and print percentage forecast errors
ferrors = (states(begf:begf+11,:) - sfor)./states(begf:begf+11,:);
tsprint(ferrors*100,fdates,1,12,snames,%8.2f);
% compare the above results to a model without national employment
sfor2 = bvarf(states,nlag,12,begf,tight,weight,decay,[],1);
% compute and print percentage forecast errors
ferrors2 = (states(begf:begf+11,:) - sfor2)./states(begf:begf+11,:);
tsprint(ferrors2*100,fdates,1,12,snames,%8.2f);
We can easily compute forecast errors because the vector autoregressive

forecasting functions always return forecasted series in levels form. This allows us to simply subtract the forecasted values from the actual employment
levels and divide by the actual levels to find a percentage forecast error.
Another point to note about example 6.11 is the use of the function ical
that returns the observation number associated with January, 1994 the time
period we wish to begin forecasting.
The results from the program in example 6.11 are shown below, where
we find that use of the national information led to a dramatic improvement
in forecast accuracy.
forecast values
Date
il
Jan94 53615.2
Feb94 53860.3
Mar94 54153.0
Apr94 54486.8
in
26502.0
26654.0
26905.9
27225.6
ky
15457.8
15525.7
15685.2
15862.7
mi
40274.7
40440.7
40752.1
41068.7
oh
49284.1
49548.7
50019.1
50528.2
pa
51025.3
51270.1
51634.6
52052.0
tn
23508.2
23676.3
24005.3
24206.0
wv
6568.9
6623.8
6698.2
6755.5

May94 54934.8 27473.4 15971.8 41517.4 51060.5 52455.3
Jun94 55118.3 27443.4 15943.5 41464.7 51085.9 52501.8
Jul94 54915.6 27315.6 15791.5 41033.6 50676.1 52073.2
Aug94 54727.4 27325.2 15802.0 40901.1 50493.6 51855.2
Sep94 54935.3 27591.3 15929.6 41188.7 50818.6 52045.4
Oct94 55128.2 27673.3 15993.3 41439.3 51082.4 52335.2
Nov94 55331.6 27769.2 16067.1 41480.2 51262.3 52453.1
Dec94 55508.7 27849.3 16097.4 41508.0 51411.4 52501.9
acutal values
Date
il
in
ky
mi
oh
pa
Jan94 52732.0 26254.0 15257.0 40075.0 49024.0 50296.0
Feb94 52999.0 26408.0 15421.0 40247.0 49338.0 50541.0
Mar94 53669.0 26724.0 15676.0 40607.0 49876.0 51058.0
Apr94 54264.0 26883.0 15883.0 40918.0 50221.0 51703.0
May94 54811.0 27198.0 16066.0 41497.0 50930.0 52140.0
Jun94 55247.0 27141.0 16100.0 41695.0 51229.0 52369.0
Jul94 54893.0 26938.0 15945.0 41236.0 50578.0 51914.0
Aug94 55002.0 27120.0 16047.0 41605.0 50742.0 51973.0
Sep94 55440.0 27717.0 16293.0 42248.0 51366.0 52447.0
Oct94 55290.0 27563.0 16221.0 42302.0 51700.0 52738.0
Nov94 55556.0 27737.0 16323.0 42547.0 51937.0 52938.0
Dec94 55646.0 27844.0 16426.0 42646.0 52180.0 52975.0
percentage foreast errors with national variables
Date
il
in
ky
mi
oh
pa
Jan94
-1.67
-0.94
-1.32
-0.50
-0.53
-1.45
Feb94
-1.63
-0.93
-0.68
-0.48
-0.43
-1.44
Mar94
-0.90
-0.68
-0.06
-0.36
-0.29
-1.13
Apr94
-0.41
-1.27
0.13
-0.37
-0.61
-0.67
May94
-0.23
-1.01
0.59
-0.05
-0.26
-0.60
Jun94
0.23
-1.11
0.97
0.55
0.28
-0.25
Jul94
-0.04
-1.40
0.96
0.49
-0.19
-0.31
Aug94
0.50
-0.76
1.53
1.69
0.49
0.23
Sep94
0.91
0.45
2.23
2.51
1.07
0.77
Oct94
0.29
-0.40
1.40
2.04
1.19
0.76
Nov94
0.40
-0.12
1.57
2.51
1.30
0.92
Dec94
0.25
-0.02
2.00
2.67
1.47
0.89
percentage foreast errors without national variables
Date
il
in
ky
mi
oh
pa
Jan94
-1.47
-0.88
-1.68
-0.53
-0.42
-1.52
Feb94
-1.15
-0.69
-1.14
-0.38
-0.09
-1.47
Mar94
-0.11
0.04
-0.26
-0.09
0.45
-0.91
Apr94
0.79
0.05
0.25
0.40
0.73
-0.03
May94
1.37
0.72
0.92
0.99
1.51
0.43
Jun94
2.14
0.94
1.49
1.74
2.34
1.11
Jul94
1.83
0.69
1.34
1.52
1.81
0.94
Aug94
2.17
1.18
1.58
2.46
2.25
1.14
Sep94
2.46
2.28
2.13
3.06
2.68
1.50
Oct94
1.86
1.46
1.32
2.49
2.77
1.44
Nov94
2.00
1.87
1.58
2.88
2.92
1.56
254
24407.6
24439.0
24293.8
24358.8
24568.3
24683.0
24794.5
24873.4
6889.0
6833.5
6845.9
6795.8
6800.8
6866.5
6884.3
6908.3
tn
23250.0
23469.0
23798.0
24009.0
24270.0
24317.0
24096.0
24315.0
24675.0
24622.0
25001.0
24940.0
wv
6414.0
6458.0
6554.0
6674.0
6894.0
6783.0
6807.0
6797.0
6842.0
6856.0
6983.0
6888.0
tn
-1.11
-0.88
-0.87
-0.82
-0.57
-0.50
-0.82
-0.18
0.43
-0.25
0.83
0.27
wv
-2.41
-2.57
-2.20
-1.22
0.07
-0.74
-0.57
0.02
0.60
-0.15
1.41
-0.29
tn
-1.12
-0.69
-0.25
0.32
0.97
1.33
1.01
1.54
2.14
1.59
2.77
wv
-2.06
-1.58
-0.74
0.92
2.82
2.41
2.69
2.90
3.45
2.66
4.23

Dec94
1.87
1.97
1.95
2.88
3.07
1.47
255
2.26
2.63
To continue with this example, we might wish to subject our experiment

to a more rigorous test by carrying out a sequence of forecasts that reflect
the experience we would gain from running the forecasting model over a
period of years. Example 6.15 shows a program that produces forecasts in
a loop extending over the period 1990,1 to 1994,12 for a total of five years
or 60 12-step-ahead foreasts. The mean absolute percentage errors for the
12-step-ahead foreast horizon are calculated for two regional models, one
with the national variables and another without.
% ----- example 6.15 Sequential forecasting of regional models
% load 10 sic national manufacturing employment
% time-series covering the period 1947,1 to 1996,12
% from the MATLAB data file national.mat
dates = cal(1947,1,12);
load national.mat;
% the data is in a matrix variable named data
% which we used to create the file (see make_level.m)
% produce a 12-month-ahead national forecast using a bvar model
load states.mat; % see make_states.m used to create states.mat
% il,in,ky,mi,oh,pa,tn,wv total state employment
% for 1982,1 to 1996,5
sdates = cal(1982,1,12);
begin_date = ical(1982,1,dates);
snames = strvcat(il,in,ky,mi,oh,pa,tn,wv);
begf1 = ical(1990,1,dates);
begf2 = ical(1990,1,sdates);
endf = ical(1994,12,sdates);
nlag = 12; nlag2 = 6;
tight = 0.2; weight = 0.5; decay = 0.1;
ferrors1 = zeros(12,8); % storage for errors
ferrors2 = zeros(12,8);
cnt = 0;
for i=begf2:endf; % begin forecasting loop
% national forecast
nfor = bvarf(data,nlag,12,begf1,tight,weight,decay);
national = [data(begin_date:begf1-1,:)
nfor];
begf1 = begf1+1;
% state forecast using national variables
sfor = bvarf(states,nlag,12,i,tight,weight,decay,national,1);
% compute errors
ferrors1 = ferrors1 + abs((states(i:i+11,:) - sfor)./states(i:i+11,:));
% state forecast without national variables
sfor = bvarf(states,nlag,12,i,tight,weight,decay,national,1);
% compute errors
256
ferrors2 = ferrors2 + abs((states(i:i+11,:) - sfor)./states(i:i+11,:));

cnt = cnt+1;
end; % end forecasting loop
in.cnames = snames;
in.rnames = strvcat(Horizon,step1,step2,step3,step4,step5,...
step6,step7,step8,step9,step10,step11,step12);
in.fmt = %6.2f;
mprint((ferrors1/cnt)*100,in);
mprint((ferrors2/cnt)*100,in);
A complication in the program results from the fact that the national
data sample begins in January, 1947 whereas the state data used in the regional model begins in January, 1982. This required the definitions begf1
and begf2 and begin date to pull out appropriate vectors of national variables for use in the regional model.
The mean absolute percentage forecast errors produced by the program
are shown below.
forecast MAPE
Horizon
il
step1
0.58
step2
0.75
step3
0.83
step4
0.95
step5
1.05
step6
1.15
step7
1.37
step8
1.46
step9
1.59
step10
1.74
step11
1.89
step12
2.06
forecast MAPE
Horizon
il
step1
0.54
step2
0.68
step3
0.74
step4
0.88
step5
0.86
step6
0.91
step7
0.95
step8
1.04
step9
1.14
step10
1.15
step11
1.18
step12
1.27
errors
in
0.47
0.64
0.83
0.99
1.13
1.35
1.49
1.63
1.77
1.90
2.02
2.07
errors
in
0.48
0.66
0.86
1.02
1.14
1.20
1.25
1.27
1.35
1.43
1.53
1.59
with national variables

ky
mi
oh
pa
0.57
0.51
0.39
0.40
0.82
0.70
0.56
0.54
1.04
0.84
0.71
0.66
1.20
1.05
0.94
0.81
1.42
1.21
1.10
1.03
1.70
1.35
1.26
1.21
1.82
1.53
1.41
1.34
1.89
1.66
1.48
1.43
1.95
1.72
1.55
1.52
2.11
1.94
1.70
1.55
2.23
2.05
1.85
1.59
2.36
2.11
1.97
1.69
with national variables
ky
mi
oh
pa
0.55
0.53
0.39
0.40
0.73
0.76
0.56
0.55
0.89
0.98
0.72
0.67
1.04
1.19
0.84
0.70
1.17
1.49
1.00
0.79
1.31
1.62
1.09
0.83
1.35
1.85
1.18
0.89
1.29
2.03
1.25
0.95
1.21
2.30
1.30
1.08
1.29
2.44
1.36
1.21
1.29
2.63
1.41
1.28
1.33
2.82
1.48
1.36
tn
0.47
0.64
0.84
0.99
1.24
1.44
1.69
1.82
1.96
2.13
2.20
2.35
wv
1.02
1.34
1.49
1.88
2.12
2.34
2.62
2.75
2.97
3.02
3.19
3.48
tn
0.47
0.59
0.76
0.93
1.10
1.19
1.30
1.44
1.50
1.62
1.74
1.95
wv
1.00
1.22
1.36
1.61
1.64
1.63
1.67
1.76
1.71
1.59
1.56
1.65
257
From these results we see the expected pattern where longer horizon forecasts exhibit larger mean absolute percentage errors. The results from our
single forecast experiment are not consistent with those from this experiment
involving sequential forecast over a period of five years. The inclusion of national variables (and forecast) lead to less accurate forecasting performance
for our regional model. The decrease in forecast accuracy is particularly
noticable at the longer forecast horizons, which is probably indicative of
poor national forecasts.
As an exercise, you might examine the accuracy of the national forecasts
and try to improve on that model.
6.6
Chapter Summary
We found that a library of functions can be constructed to produce estimates

and forecasts for a host of alternative vector autoregressive and error correction models. An advantage of MATLAB over a specialized program like
RATS is that we have more control and flexibility to implement spatial priors. The spatial prior for the rvar model cannot be implemented in RATS
software as the vector autoregressive function in that program does not allow you to specify prior means for variables other than the first own-lagged
variables in the model.
Another advantage is the ability to write auxiliary functions that process
the structures returned by our estimation functions and present output in a
format that we find helpful. As an example of this, the function pgranger
produced a formatted table of Granger-causality probabilities making it easy
to draw inferences.
Finally, many of the problems encountered in carrying out forecast experiments involve transformation of the data for estimation purposes and
reverse transformations needed to compute forecast errors on the basis of the
levels of the time-series. Our functions can perform these transformations
for the user, making the code necessary to carry out forecast experiments
quite simple. In fact, one could write auxiliary functions that compute alternative forecast accuracy measures given matrices of forecast and actual
values.
We also demonstrated how the use of structure array variables can facilitate storage of individual forecasts or forecast errors for a large number of
time periods, horizons and variables. This would allow a detailed examination of the accuracy and characteristics associated with individual forecast
errors for particular variables and time periods. As noted above, auxiliary
258
functions could be constructed to carry out this type of analysis.
6.7
References
Dickey, David A., Dennis W. Jansen and Daniel L. Thornton. 1991.
A primer on cointegration with an application to money and income, Federal Reserve Bulletin, Federal Reserve Bank of St. Louis,
March/April, pp. 58-78.
Doan, Thomas, Robert. B. Litterman, and Christopher A. Sims. 1984.
Forecasting and conditional projections using realistic prior distributions, Econometric Reviews, Vol. 3, pp. 1-100.
Engle, Robert F. and Clive W.J. Granger. 1987. Co-integration
and error Correction: Representation, Estimation and Testing, Econoemtrica, Vol. 55, pp. 251-76.
Johansen, Soren. 1988. Statistical Analysis of Co-integration vectors, Journal of Economic Dynamics and Control, Vol. 12, pp. 231254.
Johansen, Soren. 1995. Likelihood-based Inference in Cointegrated
Vector autoregressive Models, Oxford: Oxford University Press.
Johansen, Soren and Katarina Juselius. 1990. Maximum likelihood
estimation and inference on cointegration - with applications to the
demand for money, Oxford Bulletin of Economics and Statistics, Vol.
52, pp. 169-210.
LeSage, James P. 1990. A Comparison of the Forecasting Ability of
ECM and VAR Models, Review of Economics and Statistics, Vol. 72,
pp. 664-671.
LeSage, James P. and Anna Krivelyova. 1997. A Spatial Prior for
Bayesian Vector Autoregressive Models, forthcoming in Journal of
Regional Science, also available at: www.econ.utoledo.edu,
LeSage, James P. and Michael Magura. 1991. Using interindustry
input-output relations as a Bayesian prior in employment forecasting
models, International Journal of Forecasting, Vol. 7, pp. 231-238.
259
LeSage, James P. and J. David Reed. 1989a. Interregional Wage

Transmission in an Urban Hierarchy: Tests Using Vector Autoregressive Models, with J. David Reed, International Regional Science Review , Volume 12, No. 3, pp. 305-318.
LeSage, James P. and J. David Reed. 1989b The Dynamic Relationship Between Export, Local, and Total Area Employment, with J.
David Reed, Regional Science and Urban Economics, 1989, Volume
19, pp. 615-636.
LeSage, James P. and J. David Reed. 1990. Testing Criteria for
Determining Leading Regions in Wage Transmission Models, with J.
David Reed, Journal of Regional Science, 1990, Volume 30, no. 1, pp.
37-50.
LeSage, James P. and Zheng Pan. 1995. Using Spatial Contiguity as
Bayesian Prior Information in Regional Forecasting Models, International Regional Science Review, Vol. 18, no. 1, pp. 33-53.
Litterman, Robert B. 1986. Forecasting with Bayesian Vector Autoregressions Five Years of Experience, Journal of Business &
Economic Statistics, Vol. 4, pp. 25-38.
MacKinnon, J.G. 1994 Approximate Asymptotic Distribution Functions for unit-root and cointegration tests, Journal of Business &
Economic Statistics, Vol. 12, pp. 167-176.
MacKinnon, J.G. 1996 Numerical distribution functions for unit-root
and cointegration tests, Journal of Applied Econometrics, Vol. 11,
pp. 601-618.
Shoesmith, Gary L. 1992. Cointegration, error Correction and Improved Regional VAR Forecasting, Journal of Forecasting, Vol. 11,
pp. 91-109.
Shoesmith, Gary L. 1995. Multiple Cointegrating Vectors, error Correction, and Littermans Model International Journal of Forecasting,
Vol. 11, pp. 557-567.
Sims, Christopher A. 1980. Macroeconomics and Reality, Econometrica Vol. 48, pp. 1-48.
Appendix: Toolbox functions

The Econometric Toolbox is organized in a set of directories, each containing
a different library of functions. When your Internet browser unpacks the
compressed file containing the Econometric Toolbox the files will be placed
in the appropriate directories.
To install the toolbox:
1. create a single subdirectory in the MATLAB toolbox directory:
C:\matlab\toolbox\econ
Where we have used the name econ for the directory.
2. Copy the system of directories to this subdirectory.
3. Use the graphical path tool in MATLAB to add these directories to
your path. (On a unix or linux system, you may need to edit your
environment variables that set the MATLAB path.) the graphical
path tool in MATLAB to add these directories to your path. (On a
unix or linux system, you may need to edit your environment variables
that set the MATLAB path.)
A listing of the contents file from each subdirectory is presented on the
following pages.
260

The regression function library is in a subdirectory regress.
regression function library
------- regression program functions ----------ar_g
bma_g
boxcox
boxcox2
hmarkov_em
hwhite
lad
lm_test
logit
mlogit
nwest
ols
ols_g
olsar1
olsc
olst
probit
probit_g
ridge
rtrace
robust
sur
switch_em
theil
thsls
tobit
tobit_g
tsls
waldf
Gibbs sampling Bayesian autoregressive model

Gibbs sampling Bayesian model averaging
Box-Cox regression with 1 parameter
Box-Cox regression with 2 parameters
Hamiltons Markov switching regression
Halbert Whites heteroscedastic consistent estimates
least-absolute deviations regression
LM-test for two regression models
logit regression
multinomial logit regression
Newey-West hetero/serial consistent estimates
ordinary least-squares
Gibbs sampling Bayesian linear model
Maximum Likelihood for AR(1) errors ols model
Cochrane-Orcutt AR(1) errors ols model
regression with t-distributed errors
probit regression
Gibbs sampling Bayesian probit model
ridge regression
ridge estimates vs parameters (plot)
iteratively reweighted least-squares
seemingly unrelated regressions
switching regime regression using EM-algorithm
Theil-Goldberger mixed estimation
three-stage least-squares
tobit regression
Gibbs sampling Bayesian tobit model
two-stage least-squares
Wald F-test
-------- demonstration programs ----------------ar_gd

bma_gd
box_cox_d
boxcox2_d
demo_all
hmarkov_emd
hwhite_d
lad_d
lm_test_d
logit_d
mlogit_d
demonstration of Gibbs sampling ar_g

demonstrates Bayesian model averaging
demonstrates Box-Cox 1-parameter model
demonstrates Box-Cox 2-parmaeter model
demos most regression functions
demos Hamiltons Markov switching regression
H. Whites hetero consistent estimates demo
demos lad regression
demos lm_test
demonstrates logit regression
demonstrates multinomial logit
261

nwest_d
ols_d
ols_d2
ols_gd
olsar1_d
olsc_d
olst_d
probit_d
probit_gd
ridge_d
robust_d
sur_d
switch_emd
theil_d
thsls_d
tobit_d
tobit_gd
tsls_d
waldf_d
demonstrates Newey-West estimates

demonstrates ols regression
Monte Carlo demo using ols regression
demo of Gibbs sampling ols_g
Max Like AR(1) errors model demo
Cochrane-Orcutt demo
olst demo
probit regression demo
demo of Gibbs sampling Bayesian probit model
ridge regression demo
demonstrates robust regression
demonstrates sur using Grunfelds data
demonstrates switching regression
demonstrates theil-goldberger estimation
three-stage least-squares demo
tobit regression demo
demo of Gibbs sampling Bayesian tobit model
two-stage least-squares demo
demo of using wald F-test function
-------- Support functions -----------------------ar1_like

bmapost
box_lik
box_lik2
boxc_trans
chis_prb
dmult
fdis_prb
find_new
grun.dat
grun.doc
lo_like
maxlik
mcov
mderivs
mlogit_lik
nmlt_rnd
nmrt_rnd
norm_cdf
norm_pdf
olse
plt
plt_eqs
plt_reg
pr_like
prt
prt_eqs
used by olsar1
(likelihood)
used by bma_g
used by box_cox (likelihood)
used by box_cox2 (likelihood)
used by box_cox, box_cox2
computes chi-squared probabilities
used by mlogit
computes F-statistic probabilities
used by bma_g
Grunfelds data used by sur_d
documents Grunfelds data set
used by logit
(likelihood)
used by tobit
used by hwhite
used by mlogit
used by mlogit
used by probit_g
used by probit_g, tobit_g
used by probit, pr_like
used by prt_reg, probit
ols returning only residuals (used by sur)
plots everything
plots equation systems
plots regressions
used by probit
(likelihood)
prints everything
prints equation systems
262

prt_gibbs
prt_reg
prt_swm
sample
stdn_cdf
stdn_pdf
stepsize
tdis_prb
to_like
prints Gibbs sampling models

prints regressions
prints switching regression results
used by bma_g
used by norm_cdf
used by norm_pdf
used by logit,probit to determine stepsize
computes t-statistic probabilities
used by tobit
(likelihood)
The utility functions are in a subdirectory util.

utility function library
-------- utility functions ----------------------------accumulate
cal
ccorr1
ccorr2
fturns
growthr
ical
indicator
invccorr
lag
levels
lprint
lprintf
mlag
mode
mprint
mprint3
mth2qtr
nclag
plt
prt
sacf
sdiff
sdummy
shist
spacf
tally
tdiff
tsdates
tsprint
unsort
vec
vech
accumulates column elements of a matrix

associates obs # with time-series calendar
correlation scaling to normal column length
correlation scaling to unit column length
finds turning-points in a time-series
converts time-series matrix to growth rates
associates time-series dates with obs #
converts a matrix to indicator variables
inverse for ccorr1, ccorr2
generates a lagged variable vector or matrix
generates factor levels variable
prints a matrix in LaTeX table-formatted form
enhanced lprint function
generates a var-type matrix of lags
calculates the mode of a distribution
prints a matrix
prints coefficient, t-statistics matrices
converts monthly to quarterly data
generates a matrix of non-contiguous lags
wrapper function, plots all result structures
wrapper function, prints all result strucutres
sample autocorrelation function estimates
seasonal differencing
generates seasonal dummy variables
plots spline smoothed histogram
sample partial autocorrelation estimates
computes frequencies of distinct levels
time-series differencing
time-series dates function
print time-series matrix
unsorts a sorted vector or matrix
turns a matrix into a stacked vector
matrix from lower triangular columns of a matrix
263
264
-------- demonstration programs ------------cal_d.m

fturns_d
ical_d.m
lprint_d.m
lprintf_d.m
mprint_d.m
mprint3_d.m
sacf_d
spacf_d
tsdate_d.m
tsprint_d.m
util_d.m
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrated
cal function
fturns and plt
ical function
lprint function
lprintf function
mprint function
mprint3 function
sacf
spacf
tsdate function
tsprint function
some of the utility functions
-------- functions to mimic analogous Gauss functions ------------cols

cumprodc
cumsumc
delif
indexcat
invpd
matadd
matdiv
matmul
matsub
prodc
rows
selif
seqa
stdc
sumc
trimc
trimr
returns the # of columns in a matrix or vector

returns cumulative product of each column of a matrix
returns cumulative sum of each column of a matrix
select matrix values for which a condition is false
extract indices equal to a scalar or an interval
makes a matrix positive-definite, then inverts
adds non-conforming matrices, row or col compatible.
divides non-conforming matrices, row or col compatible.
multiplies non-conforming matrices, row or col compatible.
divides non-conforming matrices, row or col compatible.
returns product of each column of a matrix
returns the # of rows in a matrix or vector
select matrix values for which a condition is true
a sequence of numbers with a beginning and increment
std deviations of columns returned as a column vector
returns sum of each column
trims columns of a matrix (or vector) like Gauss
trims rows of a matrix (or vector) like Gauss
A set of graphing functions are in a subdirectory graphs.

graphing function library
-------- graphing programs --------------------------pairs
pltdens
tsplot
- scatter plot (uses histo)

- density plots
- time-series graphs
-------- demonstration programs -----------------------

pairs_d
pltdens_d
tsplot_d
265
- demonstrates pairwise scatter

- demonstrates pltdens
- demonstrates tsplot
------- support functions ----------------------------histo

plt_turns
- used by pairs
- plots turning points from fturns function
A library of routines in the subdirectory diagn contain the regression

diagnostics functions.
regression diagnostics library
-------- diagnostic programs --------------bkw
bpagan
cusums
dfbeta
diagnose
rdiag
recresid
studentize
BKW collinearity diagnostics

Breusch-Pagan heteroscedasticity test
Brown,Durbin,Evans cusum squares test
BKW influential observation diagnostics
compute diagnostic statistics
graphical residuals diagnostics
compute recursive residuals
standarization transformation
------- demonstration programs ------------bkw_d

bpagan_d
cusums_d
dfbeta_d
diagnose_d
rdiag_d
recresid_d
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
demonstrates
bkw
bpagan
cusums
dfbeta, plt_dfb, plt_dff
diagnose
rdiag
recresid
------- support functions -----------------ols.m

plt
plt_cus
plt_dfb
plt_dff
least-squares regression
plots everything
plots cusums test results
plots dfbetas
plots dffits
The vector autoregressive library is in a subdirectory var bvar.

vector autoregressive function library

------- VAR/BVAR program functions ----------becm_g
becmf
becmf_g
bvar
bvar_g
bvarf
bvarf_g
ecm
ecmf
irf
lratio
recm
recm_g
recmf
recmf_g
rvar
rvar_g
rvarf
rvarf_g
var
varf
Gibbs sampling BECM estimates

Bayesian ECM model forecasts
Gibbs sampling BECM forecasts
BVAR model
Gibbs sampling BVAR estimates
BVAR model forecasts
Gibbs sampling BVAR forecasts
ECM (error correction) model estimates
ECM model forecasts
impulse response functions
likelihood ratio tests for lag length
ecm version of rvar
Gibbs sampling random-walk averaging estimates
random-walk averaging ECM forecasts
Gibbs sampling random-walk averaging forecasts
Bayesian random-walk averaging prior model
Gibbs sampling RVAR estimates
Bayesian RVAR model forecasts
Gibbs sampling RVAR forecasts
VAR model
VAR model forecasts
------- demonstration programs ----------becm_d

becm_g
becmf_d
becmf_gd
bvar_d
bvar_gd
bvarf_d
bvarf_gd
ecm_d
ecmf_d
irf_d
irf_d2
lrratio_d
pftest_d
recm_d
recm_gd
recmf_d
recmf_gd
rvar_d
rvar_gd
rvarf_d
rvarf_gd
var_d
BECM model demonstration

Gibbs sampling BECM estimates demo
becmf demonstration
Gibbs sampling BECM forecast demo
BVAR model demonstration
Gibbs sampling BVAR demonstration
bvarf demonstration
Gibbs sampling BVAR forecasts demo
ECM model demonstration
ecmf demonstration
impulse response function demo
another irf demo
demonstrates lrratio
demo of pftest function
RECM model demonstration
Gibbs sampling RECM model demo
recmf demonstration
Gibbs sampling RECM forecast demo
RVAR model demonstration
Gibbs sampling rvar model demo
rvarf demonstration
Gibbs sampling rvar forecast demo
VAR model demonstration
266

varf_d
- varf demonstration
------- support functions -----------------johansen

lag
lrratio
mlag
nclag
ols
pftest
pgranger
prt
prt_coint
prt_var
prt_varg
rvarb
scstd
theil_g
theilbf
theilbv
trimr
vare
used by ecm,ecmf,becm,becmf,recm,recmf
does ordinary lags
likelihood ratio lag length tests
does var-type lags
does contiguous lags (used by rvar,rvarf,recm,recmf)
used for VAR estimation
prints Granger F-tests
prints Granger causality probabilities
prints results from all functions
used by prt_var for ecm,becm,recm
prints results of all var/bvar models
prints results of all Gibbs var/bvar models
used for RVARF forecasts
does univariate AR for BVAR
used for Gibbs sampling estimates and forecasts
used for BVAR forecasts
used for BVAR estimation
used by VARF,BVARF, johansen (in /util/trimr.m)
used by lrratio (vare uses /regress/olse.m)
The co-integration library functions are in a subdirectory coint.

co-integration library
------ co-integration testing routines -------adf
cadf
johansen
- carries out Augmented Dickey-Fuller unit root tests

- carries out ADF tests for co-integration
- carries out Johansens co-integration tests
------ demonstration programs ----------------adf_d

- demonstrates adf
cadf_d
- demonstrates cadf
johansen_d - demonstrates johansen
------ support functions ---------------------c_sja
c_sjt
cols
detrend
prt_coint
ptrend
rows
returns critical values for SJ maximal eigenvalue test

returns critical values for SJ trace test
(like Gauss cols)
used by johansen to detrend data series
prints results from adf,cadf,johansen
used by adf to create time polynomials
(like Gauss rows)
267

rztcrit
tdiff
trimr
ztcrit
268
returns critical values for cadf test

time-series differences
(like Gauss trimr)
returns critical values for adf test
The Gibbs convergence diagnostic functions are in a subdirectory gibbs.

Gibbs sampling convergence diagnostics functions
--------- convergence testing functions --------apm
coda
momentg
raftery
Gewekes chi-squared test

convergence diagnostics
Gewekes NSE, RNE
Raftery and Lewis program Gibbsit for convergence
--------- demonstration programs ---------------apm_d

coda_d
momentg_d
raftery_d
demonstrates
demonstrates
demonstrates
demonstrates
apm
coda
momentg
raftery
--------- support functions --------------------prt_coda

empquant
indtest
mcest
mctest
ppnd
thin
prints coda, raftery, momentg, apm output (use prt)

These were converted from:
Rafferty and Lewis FORTRAN program.
These function names follow the FORTRAN subroutines
Distribution functions are in the subdirectory distrib.

Distribution functions library
------- pdf, cdf, inverse functions ----------beta_cdf
beta_inv
beta_pdf
bino_cdf
bino_inv
bino_pdf
chis_cdf
chis_inv
chis_pdf
beta(a,b) cdf
beta inverse (quantile)
beta(a,b) pdf
binomial(n,p) cdf
binomial inverse (quantile)
binomial pdf
chisquared(a,b) cdf
chi-inverse (quantile)
chisquared(a,b) pdf

chis_prb
fdis_cdf
fdis_inv
fdis_pdf
fdis_prb
gamm_cdf
gamm_inv
gamm_pdf
hypg_cdf
hypg_inv
hypg_pdf
logn_cdf
logn_inv
logn_pdf
logt_cdf
logt_inv
logt_pdf
norm_cdf
norm_inv
norm_pdf
pois_cdf
pois_inv
pois_pdf
stdn_cdf
stdn_inv
stdn_pdf
tdis_cdf
tdis_inv
tdis_pdf
tdis_prb
probability for chi-squared statistics

F(a,b) cdf
F inverse (quantile)
F(a,b) pdf
probabililty for F-statistics
gamma(a,b) cdf
gamma inverse (quantile)
gamma(a,b) pdf
hypergeometric cdf
hypergeometric inverse
hypergeometric pdf
lognormal(m,v) cdf
lognormal inverse (quantile)
lognormal(m,v) pdf
logistic cdf
logistic inverse (quantile)
logistic pdf
normal(mean,var) cdf
normal inverse (quantile)
normal(mean,var) pdf
poisson cdf
poisson inverse
poisson pdf
std normal cdf
std normal inverse
std normal pdf
student t-distribution cdf
student t inverse (quantile)
student t-distribution pdf
probabililty for t-statistics
------- random samples ----------------------beta_rnd

bino_rnd
chis_rnd
fdis_rnd
gamm_rnd
hypg_rnd
logn_rnd
logt_rnd
nmlt_rnd
nmrt_rnd
norm_crnd
norm_rnd
pois_rnd
tdis_rnd
unif_rnd
wish_rnd
random beta(a,b) draws

random binomial draws
random chi-squared(n) draws
random F(a,b) draws
random gamma(a,b) draws
random hypergeometric draws
random log-normal draws
random logistic draws
left-truncated normal draw
right-truncated normal draw
contaminated normal random draws
multivariate normal draws
poisson random draws
random student t-distribution draws
random uniform draws (lr,rt) interval
random Wishart draws
269
-------- demonstration and test programs -----beta_d

bino_d
chis_d
fdis_d
gamm_d
hypg_d
logn_d
logt_d
pois_d
stdn_d
tdis_d
trunc_d
unif_d
demo
demo
demo
demo
demo
demo
demo
demo
demo
demo
demo
demo
demo
of
of
of
of
of
of
of
of
of
of
of
of
of
beta distribution functions

binomial distribution functions
chi-squared distribution functions
F-distribution functions
gamma distribution functions
hypergeometric distribution functions
lognormal distribution functions
logistic distribution functions
poisson distribution functions
std normal distribution functions
student-t distribution functions
truncated normal distribution function
uniform random distribution function
-------- support functions --------------------betacfj

betai
bincoef
com_size
gammalnj
is_scalar
used by fdis_prb
used by fdis_prb
binomial coefficients
test and converts to common size
used by fdis_prb
test for scalar argument
Optimization functions are in the subdirectory optimize.

Optimization functions library
--------------- optimization functions ----------------dfp_min
frpr_min
maxlik
pow_min
solvopt
Davidson-Fletcher-Powell
Fletcher-Reeves-Polak-Ribiere
general all-purpose optimization routine
Powell conjugate gradient
yet another general purpose optimization routine
--------------- demonstration programs ----------------optim1_d

optim2_d
optim3_d
- dfp, frpr, pow, maxlik demo

- solvopt demo
- fmins demo
--------------- support functions ----------------------apprgrdn

box_like1
gradt
- computes gradient for solvopt

- used by optim3_d
- computes gradient
270

hessian
linmin
stepsize
tol_like1
updateh
271
evaluates hessian
line minimization routine (used by dfp, frpr, pow)
stepsize determination
used by optim1_d, optim2_d
updates hessian
A library of spatial econometrics functions are in the subdirectory spatial.

------- spatial econometrics functions ----------casetti
darp
far
far_g
gwr
bgwr
lmerror
lmsar
lratios
moran
sac
sac_g
sacp_g
sact_g
sar
sar_g
sarp_g
sart_g
sem
sem_g
semp_g
semt_g
semo
sdm
sdm_g
sdmp_g
sdmt_g
slag
walds
xy2cont
Casettis spatial expansion model

Casettis darp model
1st order spatial AR model
- y = pWy + e
Gibbs sampling Bayesian far model
geographically weighted regression
Bayesian geographically weighted regression
LM error statistic for regression model
LM error statistic for sar model
Likelihood ratio statistic for regression models
Morans I-statistic
spatial model - y = p*W1*y + X*b + u, u = c*W2*u + e
Gibbs sampling Bayesian sac model
Gibbs sampling Bayesian sac probit model
Gibbs sampling Bayesian sac tobit model
spatial autoregressive model - y = p*W*y + X*b + e
Gibbs sampling Bayesian sar model
Gibbs sampling Bayesian sar probit model
Gibbs sampling Bayesian sar tobit model
spatial error model - y = X*b +u, u=c*W + e
Gibbs sampling Bayesian spatial error model
Gibbs sampling Bayesian spatial error probit model
Gibbs sampling Bayesian spatial error tobit model
spatial error model (optimization solution)
spatial Durbin model y = a + X*b1 + W*X*b2 + e
Gibbs sampling Bayesian sdm model
Gibbs sampling Bayesian sdm probit model
Gibbs sampling Bayesian sdm tobit model
creates spatial lags
Wald test for regression models
constructs a contiguity matrix from x-y coordinates
------- demonstration programs ----------casetti_d - Casetti model demo

darp_d
- Casetti darp demo
darp_d2
- darp for all data observations
far_d
- demonstrates far using a small data set
far_d2
- demonstrates far using a large data set

far_gd
far_gd2
gwr_d
gwr_d2
bgwr_d
bgwr_d2
lmerror_d
lmsar_d
lratios_d
moran_d
sac_d
sac_d2
sac_gd
sac_gd2
sacp_gd
sact_gd
sact_gd2
sar_d
sar_d2
sar_gd
sar_gd2
sarp_gd
sart_gd
sart_gd2
sdm_d
sdm_d2
sdm_gd
sdm_gd2
sdmp_g
sdmt_g
sem_d
sem_d2
sem_gd
sem_gd2
semo_d
semo_d2
semp_gd
semt_gd
semt_gd2
slag_d
walds_d
xy2cont_d
far Gibbs sampling with small data set

far Gibbs sampling with large data set
geographically weighted regression demo
GWR demo with Harrison-Rubinfeld Boston data
demo of Bayesian GWR
BGWR demo with Harrison-Rubinfeld Boston data
lmerror demonstration
lmsar demonstration
likelihood ratio demonstration
moran demonstration
sac model demo
sac model demonstration large data set
sac Gibbs sampling demo
sac Gibbs demo with large data set
sac Gibbs probit demo
sac Gibbs tobit demo
sac tobit right-censoring demo
sar model demonstration
sar model demonstration large data set
sar Gibbs sampling demo
sar Gibbs demo with large data set
sar probit Gibbs sampling demo
sar tobit model Gibbs sampling demo
sar tobit right-censoring demo
sdm model demonstration
sdm model demonstration large data set
sdm Gibbs sampling demo
sdm Gibbs demo with large data set
sdm Gibbs probit demo
sdm Gibbs tobit demo
sem model demonstration
sem model demonstration large data set
sem Gibbs sampling demo
sem Gibbs demo with large data set
semo function demonstration
semo demo with large data set
sem Gibbs probit demo
sem Gibbs tobit demo
sem tobit right-censoring demo
demo of slag function
Wald test demonstration
xy2cont demo
------- support functions ----------anselin.dat- Anselin (1988) Columbus crime data

boston.dat - Harrison-Rubinfeld Boston data set
latit.dat - latittude for HR data
longi.dat - longitude for HR data
c_far
- used by far_g
272

c_sem
c_sar
c_sdm
c_sac
darp_lik1
darp_lik2
elect.dat
ford.dat
f_far
f_sac
f_sar
f_sem
f_sdm
f2_far
f2_sac
f2_sar
f2_sem
f3_sem
f2_sdm
normxy
prt_gwr
prt_spat
scoref
wmat.dat
used by sem_g
used by sar_g
used by sdm_g
used by sac_g
used by darp
used by darp
Pace and Barry 3,107 obs data set
Pace and Barry 1st order contiguity matrix
far model likelihood (concentrated)
sac model likelihood (concentrated)
sar model likelihood (concentrated)
sem model likelihood (concentrated)
sdm model likelihood (concentrated)
far model likelihood
sac model likelihood
sar model likelihood
sem model likelihood
semo model likelihood
sdm model likelihood
isotropic normalization of x-y coordinates
prints gwr_reg results structure
prints results from spatial models
used by gwr
Anselin (1988) 1st order contiguity matrix
273

Spatial Econometrics

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Spatial Econometrics

Uploaded by

Document Information

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

Spatial Econometrics

Uploaded by

Copyright:

Available Formats

Spatial Econometrics

2 Spatial autoregressive models

3 Bayesian Spatial autoregressive models

4 Locally linear spatial models

5 Limited dependent variable models

6 VAR and Error Correction Models

An applied exercise . . . . . . . . . . . . . . . . . . . . . . . . 252

Demonstrate regression using the ols() function

Using the sarp g function . . . . . . . . .

Measurement error and spatial dependence

Spatial autoregressive fit and residuals

Vi estimates from the Gibbs sampler . . . . . .

Spatial x-y expansion estimates . . . . . . . . . . .

Average Vi estimates over all draws and observations

Results of plt() function . . . . . . . . . . . . . . . . . . . . . 199

Prior means and precision for important variables . . . . . . . 241

on approaches to implementing the estimation methods described. Another

Spatial dependence in a collection of sample data observations refers to the

Note that we allow the dependence to be among several observations,

Suppose that labor force is measured by place of residence and employment

Figure 1.1: Measurement error and spatial dependence

tribution across the spatial dimension of the data sample.

mean distance = 0.021

Figure 1.2: Distributions of house age versus distance

The term spatial heterogeneity refers to variation in relationships over space.

Where i indexes observations collected at i = 1, . . . , n points in space, Xi

the relationship, we must provide a specification for variation over space.

Quantifying location in our models

units are neighbors (have borders that touch) or represent observational

Quantifying spatial contiguity

Figure 1.3: An illustration of contiguity

The motivation for the standardization can be seen by considering what

Where represents a regression parameter to be estimated and denotes the

As an illustration, consider the following example which is intended to

We see that around 10 percent of the variation in the crime incidents

Quantifying spatial position

approaches to using location that allow locally linear regressions to be fit

This model can be estimated using least-squares to produce estimates of

Here it is clear that X, Z and J represent available information or data

Figure 1.4: Distance expansion estimates

One confusing aspect of this notation is that Wi y denotes an n-vector

Figure 1.5: GWR estimates

The MATLAB spatial econometrics library

As indicated in the preface, all of the spatial econometric methods discussed

W = standardized contiguity matrix

Using the results structure

The first use of prt produces a printout of results to the MATLAB

to developing a set of spatial econometric estimation functions can provide

Kelejian, H. H. and D. P. Robinson. 1995. Spatial Correlation:

Spatial autoregressive models

CHAPTER 2. SPATIAL AUTOREGRESSIVE MODELS

model is termed a first order spatial autoregression because its represents a

Letting W1 = 0 results in a regression model with spatial autocorrelation

The first-order spatial AR model

CHAPTER 2. SPATIAL AUTOREGRESSIVE MODELS

Can we show that this estimate is unbiased? If not, is it consistent?

CHAPTER 2. SPATIAL AUTOREGRESSIVE MODELS

2 = (1/n)(y W y)0 (y W y) in the likelihood (2.9) and taking logs

2 = (1/n)(y W y)0 (y W y). In the next section, we discuss a sparse