See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4723496

Gravity Models of the Intra-EU Trade: Application of the Hausman-Taylor

Estimation in Heterogeneous Panels with Common Time-specific Factors

Article · April 2004

Source: RePEc

CITATIONS READS

30 800

2 authors:

Laura Serlenga Yongcheol Shin

Università degli Studi di Bari Aldo Moro The University of York
50 PUBLICATIONS   365 CITATIONS    118 PUBLICATIONS   43,717 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Technical Efficiency Clubs and Spatial Convergence in Development: The Case of China View project

All content following this page was uploaded by Yongcheol Shin on 12 December 2013.

The user has requested enhancement of the downloaded file.

 Gravity Models of the Intra-EU Trade: Application of
the Hausman-Taylor Estimation in Heterogeneous
Panels with Common Time-specific Factors∗
Laura Serlenga
School of Economics, University of Edinburgh
Yongcheol Shin
School of Economics, University of Edinburgh
February 2004

Abstract
In this paper we follow recent developments of panel data studies and explicitly
allow for the existence of unobserved common time-specific factors where their indi-
vidual responses are also allowed to be heterogeneous across cross section units. In
the context of this extended panel data framework we generalize the Hausman-Taylor
estimation methodology and develop the associated econometric theory. We apply our
proposed estimation technique along with the conventional panel data approaches to
a comprehensive analysis of the gravity equation of bilateral trade flows amongst the
15 European countries over 1960-2001. Empirical results clearly demonstrate that our
proposed approach fits the data reasonably well and provides much more sensible re-
sults than the conventional approach based on the fixed time dummies. These findings
may highlight the importance of allowing for a certain degree of cross section depen-
dence through unobserved heterogeneous time specific common effects, otherwise the
resulting estimates would be severely biased.

JEL Classification: C33, F14.

Key Words: Gravity Models of Trade, Heterogeneous Panel Data, Hausman-Taylor Es-
timation, Time-specific Common Factors, Intra-EU Trade.

∗
We are grateful to Colin Roberts, Ron Smith, Andy Snell and seminar participants at University of
Edinburgh for their helpful comments. Partial financial support from the ESRC (grant No. R000223399) is
gratefully acknowledged. The usual disclaimer applies.
1 Introduction
The gravity model of international trade flows has been widely used as a baseline model for
estimating the impact of a variety of policy issues related to regional trading groups, currency
unions and various trade distortions, e.g. Bougheas, Demetriades and Morgenroth (1999), De
Grauwe and Skudelny (2000), Glink and Rose (2002), Martinez-Zaroso and Nowak-Lehmann
(2003) and De Sousa and Disdier (2002). Since the seminal paper by Anderson (1979), some
attempts have also been made explicitly to derive the prediction of the gravity model from
different theoretical models such as Ricardian models, Heckscher-Olin models and Increasing
Returns to Scale models, e.g. Bergstrand (1990), Markusen and Wigle (1990) and Leamer
(1992). As argued by Davis (2000), it is nowadays remarkable to observe that in the space
of a little more than a decade the gravity model has gone from theoretical orphan to having
several competing claims to maternity.
Recently, it is criticised that the use of conventional cross-section estimation is misspec-
ified since it is not able to deal with bilateral (exporter and/or importer) heterogeneity,
which is extremely likely to be present in bilateral trade flows. In this regard a panel-
based approach will be desired because heterogeneity issues can be modelled by including
country-pair “individual” effects. In particular, Matyas (1997) argues that the correct econo-
metric specification should be the so-called “triple-way model”, where time, exporter and
importer effects are specified as fixed and unobservable. But, Egger and Pfaffermayr (2002)
demonstrate that when the Matays’ triple-way model is extended to include bilateral trade
interaction effects, then this three-way specification reduces to a conventional two-way model
with time and bilateral effects only. Although a number of panel estimation techniques such
as the pooled OLS, the Fixed Effects Model, the Random Effects Model have been applied
in various contexts, the assumption that unobserved individual effects are uncorrelated with
all the regressors is convincingly rejected in almost all studies. Therefore, the Fixed Effects
estimation has been the most preferred estimation method in order to avoid the potentially
biased estimation, e.g. Cheng and Wall (2002).
However, it is worth noting that the Fixed Effects approach does not allow for estimat-
ing coefficients on time invariant variables such as distance or common language dummies,
though the consistent estimation of such effects are equally important in many situations.
Cheng and Wall (2002) simply suggest to estimate the regression of the (estimated) individ-
ual effects on individual-specific variables by the OLS, though this approach clearly ignores
the potential correlation between individual specific variables and (unobserved) individual
effects such that the resulting estimates are likely to be severely biased. In order to properly
address this issue we need to employ the Hausman and Taylor (1981, hereafter HT) instru-
mental variable estimation technique, see for example Brun, Carrere, Guillaumont and de
Melo (2002) for the HT estimation of gravity models of international trade. Most recent
empirical studies also emphasise the importance of explicitly allowing for the presence of
time specific effects in order to capture business cycle effects or to deal with globalization
issues. The conventional approach extends the benchmark model simply by incorporating
the fixed T − 1 time dummies in the panel regression, e.g. Matyas (1997), De Sousa and
Disdier (2002) and Egger (2002).
In this paper we follow recent developments of panel data studies surrounding the use

[1]
of unobserved common time effects, e.g. Ahn, Lee and Schmidt (2001), Bai and Ng (2002),
Pesaran (2002) and Phillips and Sul (2002), and advance an alternative estimation frame-
work in which we explicitly allow for the existence of observed and/or unobserved common
time-specific factors and also allow the individual responses to those common factors to be
heterogeneous across country pairs. This approach has an additional advantage in accom-
modating certain degrees of cross section dependence through heterogeneous time-specific
factors, in which case we may avoid the potential bias of the uncorrected estimates. In par-
ticular, we aim to generalize the HT estimation in this extended panel data setup, develop
the underlying econometric theory but also propose an alternative source of instruments in
addition to the (internal) instruments suggested by HT; namely, some of (consistently es-
timated) heterogeneous time-specific factors under the assumption that they are correlated
with individual specific variables but not with unobserved individual effects.
We apply our proposed HT estimation technique along with the conventional panel data
approaches to a comprehensive analysis of the sources of bilateral trade amongst the 15
European countries over 1960-2001. We use both the triple and the double indexed versions
of the gravity equation, where we consider as the dependent variable the logarithm of real
export in the former and the logarithm of total trade in the latter. First, we use the
basic specification and consider the impacts of core explanatory variables such as GDP and
population, and the distance. We then augment the basic specification by adding various
variables such as common language, common border, free trade area and currency union
membership dummies. Finally, we follow recent theoretical developments [e.g. Helpman
(1987) and Egger (2002)] and include variables measuring both similarity in relative size of
trading countries and differences in relative factor endowments.
Our main empirical findings are summarised below. First, the impact of the GDP vari-
ables is always significantly positive, whereas the impact of population variables is found to
be mostly insignificant. Second, the impacts of free custom union membership are all posi-
tively significant, whilst the results are mixed for the impacts of EMU. Third, the impact
of similarity in relative size of trading countries are mostly significant and positive, while
the impact of differences in relative factor endowments (RLF ) are somewhat ambiguous.
Turning to the estimation results for individual specific variables, the impacts of distance,
common language dummy and common border dummy are mostly significantly negative, pos-
itive and positive, respectively, as expected. A notable finding is that once the correlation
between the common language dummy and unobserved individual effect is accommodated
by the HT estimation, there is evidence that the effects of the variables that may proxy for
geographical distance, i.e. distance and common border dummy, might be compensated each
other, whereas the role of cultural affinities proxied by common language dummy becomes
more significant. On the other hand, when using the conventional approach using the T − 1
fixed time dummies, the HT estimates of the impact of distance are surprisingly positive but
insignificant, the impacts of common language dummy are significant but seem to be too
large, and common border dummy loses its statistical significance. This observation may
reflect the practical importance of properly incorporating the time-specific effects.
Furthermore, we find that our proposed HT estimation results produce more sensible
predictions on the impacts of differences in factor endowments and of the common currency
dummy on intra-EU trade flows than the conventional approach; namely the impacts of both

[2]
EMU and RLF are found to be convincilngly insignificant only in the extended HT model.
First, considering that the total trade flows are the sum of inter- and intra-industry trades,
and RLF is positively correlated only with the intra-industry trade, we may argue that the
impact of RLF on total trade flows would not be unambiguous. Secondly, empirical evidence
on the impact of EMU on trade flows has been at most mixed in the literature, see Rose
(2002) and Glink and Rose (2002) for a rather large positive effect of currency union on
trade, and Persson (2001), Pakko and Wall (2002) and De Nardis and Vicarelli (2003) for its
negative or insignificant effects on trade. In particular, as argued by de Souza (2002), the
(evaluation) periods are too short after an introduction of the Euro to use the EMU dummy
as an adequate proxy for monetary union membership and therefore, we also expect that the
impacts of EMU are yet to be significantly materialised. This observation may indicate the
importance of properly accommodating a certain degree of cross section dependence through
unobserved heterogeneous time effects, otherwise the resulting estimates would be severely
biased.
The plan of the paper is as follows: Section 2 presents an overview on gravity models
of international trade flows. Section 3 develops the extended HT estimation methodology
for heterogeneous panels with both observed and unobserved common factors. Section 4
presents a comprehensive empirical application to the gravity model of an intra-EU trade.
Section 5 concludes with further discussions.

2 Overview on Gravity Models of International Trade

Since early 1940s, the gravity model has been applied to a wide variety of goods and factors
of production moving across regional and national boundaries under differing circumstances.
For example, the model has been successfully applied to explain the determinants of varying
types of flows, such as migration, flows of buyers to shopping centers, recreational traffic or
commuting flows and patient flows to hospitals. In the context of international trade flows,
the gravity model states that the size of trade flows between two countries is determined
by supply conditions at the origin, demand conditions at the destination and stimulating
or restraining forces related to the trade flows between the two countries. Empirically, the
gravity model has been well suited for trade policy analysis and thus it has been widely
used as a baseline for estimating the impact of a variety of policy issues regarding regional
trading groups, currency unions and various trade distortions, e.g. Bougheas et al. (1999),
De Grauwe and Skudelny (2000), Glink and Rose (2002) and De Sousa and Disdier (2002).
Core explanatory variables used to explain the volume of trade across a pair of countries are
measures of economic size of trading partners and of the distance between them. Moreover,
empirical works to date are often augmented by various variables such as common language,
common border, free trade area and currency union membership dummies.
Despite its widespread empirical use, the gravity model was earlier criticized because
it lacked theoretical foundations. Nowadays, it is certainly no longer true that the gravity
model is without a theoretical basis. Since the seminal paper by Anderson (1979) it has been
increasingly recognized that the prediction of the gravity model can be derived from different
structural models such as Ricardian models, Heckscher-Olin (H-O) models and increasing
returns to scale (IRS) models of the New Trade Theory. These three types of models differ

[3]
by the way product specialization is obtained in equilibrium: technology differences across
countries in Ricardian model, factor proportion differences in the H-O model, and increasing
returns at the firm level in the IRS model, see Helpman (1987), Bergstrand (1990), Markusen
and Wigle (1990), Leamer (1992) and Eaton and Kortum (2002).
Although the gravity model per se cannot be used to test the validity any of these trade
theories against each other, its empirical success is mainly due to its ability to incorporate
most of empirical phenomena observed in international trade. In order to reconcile theory
and empirical evidence, Evenett and Keller (2002) address the so called ‘model identification’
issue and try to determine which models generate gravity-like trade predictions. The H-O
model predicts that the trade will be exclusively inter-industry (defined as trade in goods with
different factor intensities), whereas the IRS model anticipates that trade is intra-industry.
Using a cross-sectional study on a sample of almost all industrialized countries Evenett and
Keller find a robust evidence that an IRS-based trade theory provides an important reason
why the gravity equation fits trade flows well. This implies that volume of international trade
among industrialized countries is likely to be determined mainly by the extent of product
specialization and factor proportions differences, though it is also acknowledged that these
findings do not rule out the possibility that Ricardian technology differences might be what
is really behind intra-industry trade. See also Deardorff (1998). As highlighted by Davis
(2000), it is remarkable to observe that in the space of a little more than a decade the gravity
model has gone from a theoretical orphan to having several competing claims to maternity.
We now turn to the issue of econometric specifications in details. Most of earlier empir-
ical studies relied upon the use of cross-section estimation techniques. We begin with the
following typical gravity equation of the international trade:

yhf t = α0 + θt + β01t xhf t + β 02t xht + β03t xf t + β04t zhf + uhf t , (2.1)

for h = 1, ..., N, f = 1, ..., N, h 6= f , t = 1, ..., T , where yhf t is the dependent variable

(say, the volume of trade from home country h to target country f at time t), xhf t are
explanatory variables with variation in all the three dimensions (say, exchange rates between
local currencies), xht , xf t are explanatory variables with variation in h or f and t (say, GDP
or population), zhf are explanatory variables that do not vary over time but vary in h and
f (say, distance), and the disturbance terms uhf t are assumed to be iid with zero mean and
constant variance across all h, f , t. Then, (2.1) is estimated by the cross-section OLS for
each year, where α0 and θt cannot be separately identified. However, it is well-known that
this cross-section OLS estimation will ignore any of heterogeneous characteristics related to
bilateral trade relationship. For instance, a country would export different amounts of the
same product to the two different countries, even if their GDPs are identical and they are
equidistant from the exporter. Since the cross-section OLS estimates clearly fail to account
for these heterogeneous factors, they are likely to suffer from substantial heterogeneity bias.
A panel-based approach will be more desirable in order to deal with heterogeneity issues
because the effects of such determinants can be modelled by including country-pair “indi-
vidual” effects. Imposing β jt = β for all t and j = 1, ...4, and θt = 0 in (2.1), we obtain the
following pooled panel data model:

yhf t = α0 + β01 xhf t + β02 xht + β03 xf t + β 04 zhf + uhf t . (2.2)

[4]
The pooled OLS estimator obtained from (2.2) does not still deal with the issue of hetero-
geneity bias.
Matyas (1997) claims that the gravity model based on the pooled specification (2.2) is
misspecified, and proposes that the proper econometric specification of the gravity model
should be a three-way model:

yhf t = α0 + αh + γf + θt + β 01 xhf t + β02 xht + β 03 xf t + β04 zhf + uhf t . (2.3)

where one dimension is time-specific effect (θt ), and the other two are time invariant export
and import country-specific effects (αh and γf ), and it is assumed that these effects are
unobservable and thus specified as fixed effects. Clearly, the unduly strict restrictions αh =
γf = θt = 0 for all h, f , and t are imposed in (2.2). Estimating both models (2.2) and (2.3),
he finds a statistically significant evidence against restrictions, αh = γf = θt = 0.
Egger and Pfaffermayr (2002) demonstrates that when the Matays’ model (2.3) is ex-
tended to include bilateral trade interaction effects such as

yhf t = α0 + αh + γf + θt + αhf + β01 xhf t + β02 xht + β03 xf t + β04 zhf + uhf t ,
(2.4)

then this generalized three way specification is in fact identical to a two way model with
time and bilateral effects only. This implies that the Matyas’ model (2.3) is also likely
to be mis-specified, since it does not span the whole vector space of possible treatments of
explaining variations in bilateral trade and ignoring such bilateral trade interactions may lead
to biased estimation. In general, the bilateral effect accounts for any time invariant historical,
geographical, political, cultural and other bilateral influences which will lead to deviations
from country pair’s ‘normal’ propensity to trade. Since most of these influence usually remain
unobserved, including bilateral interaction effects is the natural way of controlling them.
Cheng and Wall (2002) also focus on the issue of heterogeneity bias and propose the
following fixed effects model (FEM):

yhf t = α0 + αhf + θt + β 01 xhf t + β02 xht + β03 xf t + β04 zhf + uhf t . (2.5)

It is argued that the fixed effects are a result of ignorance because we do not know which
variables are responsible for heterogeneity bias in practice. Indeed, those cultural, historical
and political factors are difficult to observe and measure. Thus, they suggest to allow
each pair of countries to have its own dummy variable that may be correlated with both
the bilateral trade and explanatory variables. The main feature that distinguishes it from
Matyas’ model is the inclusion of country-pair effects which are allowed to differ accordingly
with the direction of trade, i.e. αhf 6= αf h . In this regard, (2.3) can be seen as a special case
of (2.5), where arbitrary cross-country restrictions on the country-pair effect are imposed,
i.e. αhf = αh + γf . Cheng and Wall also consider the two other models: the symmetric fixed
effect (SFE) and the difference fixed effect model (DFE). The former specification imposes
the restriction that country-pair effects are symmetric, i.e. αhf = αf h , whilst the latter
model applies first differencing to (2.5) so as to eliminate the fixed effects. Based on the
statistical finding that the restrictions imposed in (2.2), the symmetry restriction on the
country-pair effects and those needed to obtain the DFE specification are all significantly

[5]
rejected, they conclude that the FEM (2.5) will be the most robust econometric specification
of the gravity model of international trade.
However, it is worth noting that the fixed effects approach does not allow for estimating
coefficients on time invariant variables such as distance, common border or language dum-
mies. Although it is difficult to find an appropriate measure of economic distance and of
controlling for contiguity (e.g., considering Canada and the US, China and Russia, and Ar-
gentina and Chile are all equivalently contiguous pairs), it is still important to find relatively
precise effects on trade flows of those variables. Cheng and Wall (2002) simply suggest to
estimate the additional regression of the (estimated) individual effects on individual-specific
variables by the OLS. See also Martinez-Zaroso and Nowak-Lehmann (2003) for a similar
two-step approach to an analysis of determinants of bilateral trade flows between European
Union and Mercosur countries. However, this approach clearly ignores the potential correla-
tion between individual specific variables and (unobserved) individual effects and therefore,
the resulting estimated impacts are likely to be biased. In order to properly address the issue
of correlation between regressors (including both time-varying and time-invariant) and un-
observed individual effects we need to employ the Hausman and Taylor (1981, hereafter HT)
instrumental variable estimation technique, which allows us to obtain consistent estimation
of the coefficients on time invariant variables as well. In this context, Brun et al. (2002)
attempt to apply the HT estimation by using infrastructure and population as instruments
for standard trade-barrier function such as distance, common language and common border
dummies, assuming that they are not correlated with individual effects.
The triple index model as given in (2.5) is not the only way of representing the panel data-
based gravity model of international trade. A more conventional double index-based panel
data specification have also been applied in which case explanatory variables are expressed
as a combination of characteristics of trading partners, e.g. Egger (2001) and Glink and
Rose (2002). Thus we now consider the following double index panel data model:

yit = β0 xit + γ 0 zi + αi + θt + εit , i = 1, ..., N, t = 1, ..., N, (2.6)

where an index i represents each country-pair hf such that αi = αhf = αh + γf as in Cheng

and Wall (2002). Notice that variables in xit are defined as a combination of features of the
countries in each pair, but importantly embrace variables, xhf t that vary in all the three
dimensions, and variables, xht and xf t that vary only with one partner of trade and time,
respectively. Time invariant regressors such as distance, common language and common
borders dummies are now included in zi that coincide with zhf . For instance, De Sousa and
Disdier (2002) use (2.6) to investigate the role of consumer’s preferences as well as tariff
and non-tariff barriers in explaining border effects on trade flows among Hungary, Romania
and Slovenia, European Union (EU) and Central European Free Trade Agreement (CEFTA)
countries, and apply the HT estimation to consistently estimate the impacts of individual
country’s characteristics like distance, common border or language. In particular, they find
that once the correlations between regressors and unobserved individual effects are properly
accommodated, the significance of the distance is strongly reduced and the coefficient of on
common border becomes insignificant.
Motivated by the New Trade Theory initiated by Krugman (1979), who attempts to
explain trade patterns under monopolistic competition and increasing returns, Helpman

[6]
(1987) suggests that the share of intra-industry trade in bilateral trade flows should be larger
for countries with similar incomes per capita or similar characteristics in general. Helpman
estimates (2.1) by the cross-section OLS estimation for 14 countries for every year from 1970
to 1981, where the share of intra-industry trade is used as the dependent variable and some
combined measures of trading partners’ incomes and relative country size are suggested as the
regressors that are meant to proxy for size, similarity in size and difference in relative factor
endowments of trading partners, and finds that there is a positive correlation between the
share of intra-industry trade and similarity in income per capita. Hummels and Levinsohn
(1995) extend Helpman’s analysis into a panel data framework. In similar veins, Egger
(2002) attempts to explain the total volume of export (the sum of inter- and intra-industry
volumes) in terms of the geographical distance between two trading partners, the relative
factor endowments, the relative size of two countries (GDP) and their overall economic space.
His empirical findings generally confirm the importance of allowing for both heterogeneity
and correlation between explanatory variables and individual effects.
In summary, we may conclude that the FEM along with the HT is the most preferred
estimation technique in the analysis of gravity model of international trade, because we
need to deal with unobserved heterogeneous individual effects and its correlation with both
time-varying and time invariant regressors to avoid any potential biases. In next section
we will generalize the HT estimation in presence of both observed and unobserved common
time-specific factors.

3 The Hausman-Taylor Estimation in Heterogeneous

Panels with Time-specific Factors
Noticing that both triple and double index versions of the gravity model of trade, (2.5) and
(2.6), can be expressed as a conventional double index panel-data model, we begin with
yit = β 0 xit + γ 0 zi + εit , i = 1, ..., N, t = 1, ..., T, (3.1)

εit = αi + θt + uit , (3.2)

where the error term εit is composed of three parts; namely, αi is an individual effect that
accounts for the effect of all possible time invariant determinants and might be correlated
with some of the explanatory variables xit and zi , θt is the time-specific effects common to
all cross section units that is meant to correct for the impact of all the individual invariant
determinants such as potential trend or business cycle, and uit is a zero mean idiosyncratic
random disturbance uncorrelated across cross section units and over time periods. The
conventional assumptions are that these three components are independent of each other.
We now generalize (3.2) such that the individual responses to variations of the common
time-specific effects are heterogeneous. This suggests that we extend (3.2) to
εit = αi + λi ft + uit , (3.3)
where λi capture heterogeneous responses that trade flows between trading countries might
have with respect to the time-specific common factors, ft . It is clearly seen that the pooled

[7]
or fixed effects estimation of β and γ in (3.1) will be less efficient without properly ac-
commodating the error component structure given by (3.3). More importantly, in the case
where some or all of the regressors in xit are likely to be correlated with ft , the uncorrected
estimator will be severely biased. There is now a growing number of panel studies using
(3.3) explicitly, e.g., Ahn, Lee and Schmidt (2001), Bai and Ng (2002), Pesaran (2002) and
Phillips and Sul (2002). Additional advantage of this approach is to allow for certain degrees
of cross section dependence via heterogeneous time-specific effects.
To accommodate this potentially important issue, we now combine (3.1) and (3.3). Here
we consider the two cases. First, we simply assume that all of the time-specific common
effects are observable in which case we have

yit = β 0 xit + γ 0 zi + λ0i ft∗ + εit , i = 1, ..., N, t = 1, ..., T, (3.4)

εit = αi + uit , (3.5)

where ft∗ are observed multiple time-specific factors. The distinguishing features of the above
model are: (i) it considers explicitly the impacts of time-specific factors ft∗ instead of the
conventional fixed time effects to investigate the business cycle or the globalization issues,
and (ii) it does not impose the homogeneous restriction on the coefficients on ft∗ . Considering
that ft∗ usually measure the common macro shocks or policies, it is natural to expect that
individual’s responses will be different from each other. Secondly, in the case where we have
both observed and unobserved common time-specific effects, we follow the pooled correlated
common effect (PCCE) estimation approach advanced by Pesaran (2002), and extend the
model (3.4) to

yit = β0 xit + γ 0 zi + λ0i ft + αi + uit , i = 1, ..., N, t = 1, ..., T, (3.6)

where we assume that there is a single unobserved time-specific common effect in εit and then
ft is the augmented set including ft∗ and the cross sectional averages of yit and xit , namely
P −1 PN
y t = N −1 N i=1 yit and xt = N i=1 xit . Pesaran (2002) shows that the PCCE estimation
(also called the generalised within estimator) will provide the consistent estimator of β,
though it does not provide a consistent estimator of γ.
In what follows we will work on (3.6) without loss of generality. Here notations are: xit =
(x1,it , x2,it , ..., xk,it )0 is a k × 1 vector of variables that vary over individuals and time periods,
zi = (z1,i , z2,i , ..., zg,i )0 is a g × 1 vector of individual-specific variables, ft = (f1,t , f2,t , ..., fl,t )0
is an l × 1-vector of time-specific variables, and β = (β1 , β2 , ..., βk )0 , γ = (γ1 , γ2 , ..., γg )0 ,
λi = (λ1,i , λ2,i ..., λl,i )0 are conformably defined column vectors of parameters, respectively.
Finally, we follow Hausman and Taylor and rewrite (3.6) by

yit = β01 x1it + β02 x2it + γ 01 z1i + γ 02 z2i + λ0i ft + αi + uit , (3.7)

where xit = (x01it , x02it )0 , zi = (z01i , z02i )0 , x1it , x2it are k1 - and k2 -vectors, z1i , z2i are g1 - and
g2 -vectors, and β1 , β 2 , γ 1 , γ 2 are conformably defined column vectors.
We now make the following assumptions:
Assumption 1. (i) uit ∼ iid (0, σu2 ). (ii) αi ∼ iid (α, σα2 ). (iii) E (αi ujt ) = 0 for all i, j, t.
(iv) E (xit ujs ) = 0, E (ft uis ) = 0 and E (zi ujt ) = 0 for all i, j, s, t, so all the regressors are

[8]
exogenous with respect to the idiosyncratic errors, uit . (v) x1it , z1i and ft are uncorrelated
with αi for all i, t, whereas x2it and z2i are correlated with αi . (vi) Both N and T are
sufficiently large.
Assumption 1 is standard in the panel data literature. In particular, we need to use
prior information to distinguish columns of x and z which are correlated with the individual
unobservable effect, αi and those which are not. Assumption (vi) is necessary to consistently
estimate (nuisance) heterogenous parameters, λi .
We now develop the estimation theory for all the parameters in (3.7), which involves the
two steps. First, we rewrite (3.4) as

yit = αi + β0 xit + γ 0 zi + λ0i ft + uit , i = 1, ..., N, t = 1, ..., T, (3.8)

and obtain the consistent estimator of β by

à !−1 à !
P
N P
N
β̂F E = x0i MT xi x0i MT yi , (3.9)
i=1 i=1

where ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
yi1 1 f10 x0i1
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ yi2 ⎟ ⎜ 1 ⎟ ⎜ f20 ⎟ ⎜ x0i2 ⎟
yi = ⎜ .. ⎟ ; 1T = ⎜ .. ⎟, f = ⎜ .. ⎟ , xi = ⎜ .. ⎟,
⎜ ⎟ ⎜ ⎟ (T ×l) ⎜ ⎟ ⎜ ⎟
(T ×1) ⎝ . ⎠ (T ×1) ⎝ . ⎠ ⎝ . ⎠ (T ×k) ⎝ . ⎠
yiT 1 fT0 x0iT
HT = (1T , f) is a T × (l + 1) matrix and MT = IT − HT (H0T HT )−1 H0T . Next, the consistent
estimators of λi can be obtained from the following regression:

ỹit = bi + λ0i ft + ũit , i = 1, ..., N, t = 1, ..., T, (3.10)

0
where ỹit = yit − β̂ F E xit and bi = αi + γ 0 zi .
Assuming that all the underlying variables are stationary, in which case under fairly
standard conditions, the consistency and the asymptotic normality of the FE estimator of β
can be easily established. In the current context, as (N, T ) → ∞ jointly, we have
√ ³ ´
a
NT β̂ F E − β ∼ N (0, ΣβF E ) , (3.11)

where the consistent estimator of ΣβF E is given by

à N
!−1
1 X x0i MT xi
Σ̂βF E = σ̂u2 , (3.12)
N i=1 T

where σ̂u2 is a consistent estimator of σu2 provided by

PN
û0i ûi
i=1
σ̂u2 = , (3.13)
N (T − 1 − l) − k
0 0 0
ûi = (ûi1 , ..., ûiT )0 with ûit = ỹit −b̂i −λ̂i ft = yit −β̂F E xit −b̂i −λ̂i ft for i = 1, ..., N, t = 1, ..., T ,
and b̂i , λ̂i are the OLS estimators of bi , λi obtained from (3.10).

[9]
 However, the above FE estimation will wipe out any individual specific variables in Zi
from (3.7). In order to consistently estimate γ 1 and γ 2 on individual specific variables, we
notice that (3.8) can be written as

dit = αi + γ 01 z1i + γ 02 z2i + uit , i = 1, ..., N, t = 1, ..., T, (3.14)

where dit = yit − β 0 xit − λ0i ft for i = 1, ..., N and t = 1, ..., T . Using Assumption 1(ii), we
rewrite (3.14) as

dit = α + γ 01 z1i + γ 02 z2i + αi∗ + uit = α + γ 0 zi + ε∗it , i = 1, ..., N, t = 1, ..., T,

(3.15)

where αi∗ ∼ (0, σα2 ) and ε∗it = αi∗ + uit is a zero mean process by construction. Rewriting
(3.15) in matrix notation we have

di = α1T + z1i 1T γ 1 + z2i 1T γ 2 + ε∗i , i = 1, ..., N, (3.16)

d = α1NT + Z1 γ 1 + Z2 γ 2 + ε∗ , (3.17)

where
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
d1 1T zj1 1T ε∗1
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ d2 ⎟ ⎜ 1T ⎟ ⎜ zj2 1T ⎟ ⎜ ε∗2 ⎟
d =⎜
⎜ .. ⎟ ; 1N T
⎟ =⎜
⎜ .. ⎟ , Zj
⎟ =⎜
⎜ .. ⎟,
⎟ j = 1, 2, ε∗ =⎜
⎜ .. ⎟.
⎟
(NT ×1) ⎝ . ⎠ (N T ×1) ⎝ . ⎠ (T ×g) ⎝ . ⎠ (NT ×1) ⎝ . ⎠
dN 1T zjN 1T ε∗N
n o
Replacing d by its consistent estimate, d̂ = dˆit , i = 1, ..., N, t = 1, ..., T, , where dˆit =
0 0
yit − β̂ xit − λ̂i ft for i = 1, ..., N, t = 1, ..., T , we now have

d̂ = α1NT + Z1 γ 1 + Z2 γ 2 + ε = Cδ + ε∗ , (3.18)

where C = (1NT , Z1 , Z2 ) and δ = (α, γ 01 , γ 02 )0 . Here we notice that approximation errors

stemming from the use of d̂ in (3.18) are (asymptotically) negligible. To deal with the
nonzero correlation between Z2 and α or α∗ , we need to find the following NT ×(1 + g1 + h)
matrix of instrument variables:

W = (1NT , Z1 , W2 ) ,

where W2 is an NT × h matrix of instrument variables for Z2 with h ≥ g2 for identification.

The advantage of the HT estimation is that the instrument variables for Z2 can be obtained
internally, and they suggest to use QX1 as the instruments for Z2 . See also Amemiya and
MaCurdy (1986) and Breusch, Mizon and Schmidt (1989) for additional source of instru-
ments.
We now suggest to use an alternative source of instruments as follow: For this we rewrite
(3.8) as

yit = bi + β0 xit + λ1i f1t + λ2i f2t + · · · + λli flt + uit , (3.19)

[10]
where bi = αi +γ 0 zi . Define θ̂jit = λ̂ji fjt for j = 1, ..., l, i = 1, ..., N and t = 1, ..., T , where λ̂ji
are consistent estimates of heterogenous factor loadings λji , and similarly define the NT × 1
matrix, ⎛ ⎞ ⎛ ⎞
fj λ̂j1 fj,1
⎜ ⎟ ⎜ ⎟
⎜ fj λ̂j2 ⎟ ⎜ fj,2 ⎟
⎜ ⎟
Θ̂j = ⎜ .. ⎟ , fj = ⎜ . ⎟
⎜
⎟ , j = 1, ..., l.
⎜
⎝ . ⎟ (T ×1) ⎝ .. ⎠
⎠
fj λ̂jN fj,T
We now make the following assumption:
Assumption 2. λji , j = 1, ..., l1 , are correlated with z2i , but not correlated with αi ,
whilst λji , j = l1 + 1, ..., l, are correlated with both z2i and αi .
Assumption 2 implies that some of individuals’ heterogeneous responses with respect to
common factors ft are correlated with Z2 , but not with individual effects. In fact, the nature
and implication of this assumption is basically the same as those of Assumption 1(v). Under
Assumption 1(v) and Assumption 2, we now obtain the following instrument matrix for Z2 ,
³ ´
W2 = QX1 , Θ̂1 , Θ̂2 , ..., Θ̂l1 ,

where the dimension of W2 is N T × h with h = k1 + l1 . Then, the consistent estimator of

δ is obtained by the GLS-IV estimation. Premultiplying W0 by (3.18), we have
∗
W0 d̂ = W0 Cδ + W0 ε . (3.20)

and therefore we obtain the GLS estimator of δ by

h i−1 0
δ̂ GLS = C0 WV−1 W0 C C0 WV−1 W d̂, (3.21)

where V = V ar (W0 ε∗ ). The feasible GLS estimator is obtained by replacing V by its

consistent estimator. We first obtain an initial consistent estimation of δ̂ by the OLS esti-
mator from
³
(3.18) and construct
´0
a consistent estimate of ε∗ by ε̂∗OLS = d̂ − Cδ̂ OLS , where
ε̂∗OLS = ε̂∗0 ∗0
OLS,1 , ..., ε̂OLS,N . Then, we estimate the initial consistent estimate of V by

N
X
V̂(1) = wi0 ε̂∗OLS,i ε̂∗0
OLS,i wi , (3.22)
i=1

where wi is the T × (1 + g1 + h) instrument matrix for individual i, defined in W =

0 0
(w10 , ..., wN ) , and estimate the feasible GLS (FGLS) estimator of δ by
∙ ¸−1 0
(1) −1 −1
δ̂ F GLS = C0 WV̂(1) W0 C C0 WV̂(1) W d̂. (3.23)

(1)
Next, we construct the GLS residuals by ε̂GLS = d̂ − Cδ̂ F GLS , and estimate V and δ further
by
N
X
V̂(2) = wi0 ε̂∗GLS,i ε̂∗0
GLS,i wi .
i=1

[11]
 ∙ ¸−1
(2) 0 −1 0 −1 0
δ̂ F GLS = C WV̂(2) W C C0 WV̂(2) W d̂. (3.24)
¯ ¯
¯ (j) (j−1) ¯
This iteration will be repeated until the convergence occurs, e.g. ¯δ̂ − δ̂ F GLS ¯¯ < 0.0001,
¯ F GLS
j = 1, 2, ... Once we have obtained the final converged FGLS estimator, its covariance matrix
will be computed by
(∙ ¸−1 )
³ ´ −1
0
V ar δ̂ F GLS = C WV̂F GLS W0 C . (3.25)

Under fairly standard conditions the consistency and the asymptotic normality of the FGLS
estimator of δ can also be easily established. As (N, T ) → ∞ jointly, we have
√ ³ ´
a
NT δ̂ F GLS − δ ∼ N (0, ΣδF GLS ) , (3.26)

where the consistent estimator of ΣδF GLS is given by

⎡ Ã !−1 ⎤−1
C0 W V̂F GLS W0 C ⎦
Σ̂δF GLS =⎣ . (3.27)
NT NT NT

4 Empirical Application to the Intra-EU Trade

In this section we will provide a comprehensive analysis of the determinants of bilateral trade
flows amongst the fifteen European countries using both triple and double indexed versions
of the gravity equation, (2.5) and (2.6), where we consider as the dependent variable the
logarithm of real export in (2.5) and the logarithm of total trade in (2.6). (For detailed
definition of all the variables see the Data Appendix.)
In each case we consider the three different specifications. First, the basic model specifies
that bilateral export or trade only depends on the mass of the countries (measured by GDP
and population) and barrier to trade (measured by distance). A high level of income in the
exporting country indicates a high level of production, which increase availability of goods
for exports, whereas a high level of income in importing country suggest higher imports.
Therefore, we expect the positive impacts of those variables on trade flows. The effect of
population is not unambiguous as disputed in the literature. Here we follow Bergstrand
(1989) and interpret that a positive (negative) impact of exporter population indicates that
the exports tend to be labor (capital) intensive goods, whilst a positive (negative) impact
of importer population indicates that the exports tend to be necessity (luxury) goods. As
noted by Baldwin (1994), however, both impacts might be negative as larger countries are
sometimes regarded as self-efficient. On the other hand, the effect of transportation costs
proxied by geographical distance between capital cities is certainly expected to be negative
on trade flows. Notice that in the double indexed version both GDP and population are
expressed as a combined measure of trading partners.
Second, we consider the augmented specification, where trade flows are also allowed to
depend on variables that take into account free trade agreements and common currency

[12]
union as well as time invariant dummies for common language and common border. The
variable CEE is a dummy that is equal to one when both countries belong to the European
Community and is expected to exert a positive impact. See also De Grauwe and Skudelny
(2000), Martinez-Zaroso and Lehmann (2001), Cheng and Wall (2002) and De Sousa and
Disdier (2002) for an analysis of the effects of regional trading blocks. We also consider
the time-varying dummy variable EMU which is equal to one when both trading partners
adopt the same currency. The issue on the benefits of joining a common currency union has
recently been getting more attention since the introduction of the Euro in 1999. Since an
official motivation behind the EMU project (European Commission, 1990) is that a single
currency will reduce the transaction costs of trade within member countries, the impact of
EMU on trade flows is expected to be positive. But, the empirical evidence is mixed. Rose
(2002) and Glink and Rose (2002) have analysed the trade data for almost all countries in
the world and found evidence of a rather large positive effect of currency union on trade.
Interestingly, this finding is not consistent with the earlier studies that fail to find a significant
link between exchange rate stability and trade, e.g. Branda and Mendez (1988) and Frankel
and Wei (1993). See also a number of recent studies that find negative or insignificant
effects on trade of a monetary union, e.g. Persson (2001) and Pakko and Wall (2002). In
particular, de Souza (2002), and De Nardis and Vicarelli (2003) investigate the effect of
the EMU in the euro area over the last two decades and find no significant evidence of a
robust relationship between EMU and trade. The common language dummy (Lan) has a
value equal to one when both countries speak the same official language and is meant to
capture similarity in cultural and historical backgrounds between trading countries. The
shared border dummy (Bor) is equal to one when the trading partners share a border, which
is a proxy for geographical proximity. Obviously, both effects on bilateral trade flows are
expected to be positive.
Finally in the full specification version of the gravity equation, we also aim to follow
recent developments of the New Trade Theory advanced by Helpman (1987), Hummels and
Levinsohn (1995) and Egger (2001, 2002) and thus add variables such as RLF and SIM.
The variable RLF measures the difference in terms of relative factor endowments (proxied
by per capita GDPs) between two countries and takes a minimum value of zero when there is
equality in relative factor endowments. The larger is this difference, the higher is the volume
of inter-industry (and the total) trade will be, and the lower the share of the intra-industry
trade. The variable SIM captures the relative size of two countries in terms of GDP. This
index is bounded between zero (absolute divergence in size) and 0.5 (equal country size).
The larger this measure is (meaning that the more similar two countries are), the higher
the share of the intra-industry trade will be. We note in passing that these variables have
been considered to mainly explain trends of the intra-industry trade share. For example,
Helpman (1987) finds a negative correlation between the intra-industry trade share and
RLF , and a positive correlation between the intra-industry trade share and SIM, which
is interpreted as supporting evidence of the theory of IRS and imperfect competition in
international trades. Since our analysis aims to explain the patterns of both intra-industry
trades and the total trade flows (sum of inter- and intra-industry trades), the impact of RLF
might not be unambiguous on total trade flows. We also consider the impact of (logarithm
of) real exchange rates (RER) between two countries, which is defined as the price of the

[13]
foreign currency per the home currency unit and which is meant to capture the relative price
effects. A depreciation of the home currency relative to the foreign currency (an increase in
RER) should lead to more export and less import for the home country. The effect of real
exchange rates on total trade flow will be positive (negative) if the export component of the
total trade is significantly larger than the import component. For similar lines of studies
see De Grauwe and Skudelny (2000) and Egger and Pfaffermayr (2002). Here we drop the
population variables from the full specification in order to avoid collinearity as RLF is a
linear combination of GDP and population.

4.1 Explanatory Data Analysis

The data used cover a period of 42 years (1960-2001) whereas the country sample con-
tains all of the 15 EU member countries, namely Austria, Belgium, Denmark, Finland,
France, Germany, Greece, Ireland, Italy, Luxemburg, Netherlands, Portugal, Spain, Sweden,
United Kingdom where Belgium and Luxemburg are treated as a single country, counting
182 country-pairs in the triple index version of the gravity model (2.5) and 91 country-pairs
in the double index version (2.6).
Table 1 reports some of summary figures presented in the Statistical Yearbook (Eurostat,
1997) and shows that the intra-EU trade has always been a considerable part of EU’ s
total trade (currently it is almost two-thirds). Since 1960, there have been only three time
periods during which an intra-EU trade share declined as a percentage of the total EU trade.
During the periods 1973-1975 and 1979-1981, the relative importance of the intra-EU trade
fell sharply due to price increases in primary goods. As a result, the total value of the extra-
EU imports went up, raising total value of extra-EU trade. Even when the internal market
was introduced in 1993, the relative importance of intra-EU trade has declined. But this
may be a purely statistical phenomena due to the fact that the collection of the intra-EU
data has been reorganized since 1993.
In general, the intra-EU trade volumes were positively affected by the enlargement of
the European Community, e.g. with the accession of new member states (Greece, Portugal
and Spain) in the 1980s and with the German unification at the beginning of the 1990s,
see Single Market Review (European Commission, 1997). Also, the enlargement of the EU
in 1995 with Austria, Sweden and Finland has significantly increased the intra-EU trade
volume: for example, the intra-EU share of total EU trade before the three new member
states joined the EU was 58% in 1994, whereas it reached around 64% in 1995, see External
and intra-European Union trade: Statistical Yearbook (Eurostat, 1996). This clearly suggests
that one of main factors behind the increasing importance of intra-EU trade within the total
EU trade is clearly the stronger link among member states over the last few decades.

Table 1 about here

Table 1 also shows that an intra-EU trade trends along with the total EU GDP. But,
the fact that the trade volume between EU countries grows faster than GDP is further
evidence of the increasing integration of EU market. The Single Market Review (European
Commission, 1997) summarizes that the growth of the Intra-EU trade, initiated by the
programme to complete the single market implemented in the mid-1980s, leads to major

[14]
changes for the European economies. The measures taken consist mainly of a liberalization of
trade in products and services through the abolition of non-tariff barriers, border formalities,
a liberalization of public procurement practices and the mutual recognition of technical
standards. Also included are the liberalization of factors movements and deregulation of
sectors formerly subject to tight national regulation. The anticipation by economic agents
of the completion of the single market caused a drive towards strong industrial restructuring
at the microeconomic level, notably through merges and acquisitions by both European and
non-European companies. Liberalization would also tend to lower prices through increased
competition and foster a concentration of resources in more efficient use. These effect would
translate into sizable welfare gains, increases in GDP, and increase competitiveness vis-a-
vis non-member states. On the other hand, as Jacquemin and Sapir (1990) notice, the
concentration of European industries might also create or foster dominant position which
lead to higher domestic prices. This lowers trade barriers against imports from the rest of
the world, meaning more extra-EU and less inter-EU import. Table 1 actually shows that
in our sample the share of exports is generally higher than the share of imports within EU
trades. In light of these figures we therefore expect that positive effects of an increase in
real exchange rates on exports will dominate negative impacts on imports. As a result its
influence on total trades is expected to be positive.
The Single Market Review (European Commission, 1997) further reports that the removal
of barriers to the mobility of goods leads to an increase in trade flows within the Community,
and most increases are of the intra-industry type. Intra-industry, boosted by similarity
of the trading nations, may lead to cost-free adjustments, increased efficiency and welfare
gains associated with variety. In contrast, inter-industry trade, traditionally associated with
comparative advantages of nations, may lead to more costly adjustments, as trade and
specialization move factors from contested, export-oriented industries. Figure 1 shows the
evolution of trade in intra-EU trade between 1980 and 1994.1 At the beginning of the
1980s the most important trade was the inter-industry type (share of 45%), but it started
to decline from the mid-1980s onwards. The resulting increase in the share of intra-industry
is essentially due to a trade boost in vertically differentiated products that are predominant
in the largest European countries, e.g., Germany and France since 1986 and the UK since
1989. This is consistent with evidence that intra-industry trade accounts for a substantial
fraction of total trade among industrialized countries, see Deardorff (1984) and Evenett
and Keller (2002). Molle (1997) states that contrary to what some had expected, both EU
and EFTA has not produced specialization among countries along lines of traditional trade
theory predicting that one country will be specialized in one good and the other in the other
1
The share of intra-industry trade is measured by the traditional Grubel-Lloyd (1975) index, whereas
inter-indusrty trade is represented by the so called ‘one way trade’. The Grubel-Lloyd index is defined
X −M
as GL = 1 − Xjj +Mjj and measures the amount of intra-industry trade in a particular product group j.
The value ranges from zero to unity representing a situation of zero and 100 percent intra-industry trade,
respectively. When Xj or Mj equal to zero, there is no overlap of export or import so no intra-industry
trade will take place. On the other hand if Xj = Mj , matching will be completed and GL = 1. Total trade
is decomposed in three trade types according to their similarity in price (proxy for quality) and to overlap
in trade: two-way trade in similar products (significant overlap and low price differences); two-way trade
in vertically differenced products (significant overlap and high price differences); one-way trade (no or no
significant overlap).

[15]
good on the basis of comparative advantages. In fact, at the beginning of the 1960s, it
became clear that specialization occurred within sectors and consumers have benefited from
the resulting increased range of products available. The more similar the demand structures
of two countries are, the more intensive are the potential trade between them, see Linder
(1961).

Figure 1 about here

4.2 Estimation results

We now briefly discuss alternative estimation procedures used to estimate (2.5) and (2.6):
namely, the pooled OLS (POLS), the between estimation (BTW), the fixed effect model
(FEM), the random effect model (REM) and Hausman and Taylor (HT) instrumental vari-
able estimation. The POLS estimation is likely to gain in efficiency due to the increased
number of observations but estimation results would be biased due to neglected (individual)
heterogeneity. The between estimator runs an OLS regression on the time averages of cross
section pairs, but is also likely to subject to the potential heterogeneity bias. The FEM
explicitly takes into account the bilateral trade heterogeneity by specifying that all explana-
tory variables are assumed to be correlated with unobserved fixed individual effects, though
it also wipes out any of time invariant variables. On the other hand, under the stronger
assumption that unobserved individual effects are randomly distributed but uncorrelated
with all regressors, the REM allows us to estimate the parameters on both time-varying and
time-invariant variables, simultaneously. The validity of this assumption should be tested
by using the Hausman (1978) test , and when this assumption is rejected, we will use the
HT estimator to consistently estimate the impacts of time-invariant variables.
We consider the two different scenarios: First, we estimate both (2.5) and (2.6) without
including any time-specific effects, which we call the benchmark case. Secondly, we follow
most recent empirical studies that also emphasise the importance of explicitly allowing for the
time specific effects in (2.5) and (2.6), e.g., Matyas (1997) and Egger (2002). Since we analyse
the trade data over the longer time span, this issue should be addressed properly in order
to capture business cycle effects or deal with the generic globalization issues. We consider
the three extensions: we extend the benchmark model by incorporating the conventional
fixed time-specific dummies in the regressions. We will also use our proposed approach
described in Section 3, namely by incorporating observed and unobserved common time
factors, respectively.
Tables 2 present alternative estimation results for triple and double index gravity models
of bilateral trades amongst the 15 EU countries. Since the validity of the REM assumption
that there is no correlation between explanatory variables and unobserved individual effects
is convincingly rejected in all cases considered, we will discuss estimation results mainly with
the FEM results. For overwhelmingly similar empirical evidence see Egger (2001), Cheng
and Wall (2002) and De Sousa and Disdier (2002) and Glink and Rose (2002)
Starting from the full specification of the triple index version (see Table 2(a)), almost
all the FEM estimation results are statistically significant and consistent with our a priori
expatiations. Both GDPs of home and foreign country have a positive effect on real exports

[16]
and a depreciation of the home currency leads to an increase in export flows. Similarity
in size and relative difference in factor endowments between trading partners help to boost
real exports although the impact of RLF is much smaller than that of SIM. This finding
clearly reflects the fact that the intra-industry trade is the main part of the total EU trade as
described in subsection 4.1. A trade union membership also boosts real exports significantly,
though the effect of EMU appears negative but insignificant. Although both REM and the
POLS estimation results are likely to be biased because of correlation between regressors
and unobserved individual effects, both estimation results are relatively consistent with the
corresponding FEM results. Only the coefficient on EMU is positive but insignificant.
Next, the BTW estimates appear to be mostly insignificant (SIM, RLF , CEE, EMU).
This may be a clear indication of severe bias problem expected over the relatively long time
span considered in our estimation, though we might expect to obtain different results over
different time periods since the between estimator is based on a regression on time averages
of cross section pairs. Turning briefly to the basic and augmented specifications, we find
that only the impact of importer population is significant and negative, which leads us to
conclude that the exports within EU countries are most likely to be luxury goods.
Table 2(b) reports the estimation results for the double index version, (2.6). Though they
are mostly consistent with those of the triple index model, there are two notable differences.
First, the impact of EMU on the total trade is now positive and significant. Hence, EMU
seems to have a more positive impact on imports than on exports contrary to the evidence
observed after the completion of the single market. Secondly, the impact of SIM on the
total trade (mostly via the impact on the intra-industry trade) is much higher (1.17 versus
0.35). Once again the effect of income variable is highly significant, whereas the impact of
population is insignificant. This reinforces the previous finding in the triple index version,
but may also imply that the mass effect is likely to be captured mostly by income vari-
able rather than population. (Considering that both GDP and population are proxies for
the economic size of trading partners and they are highly correlated, this might indicate a
certain degree of collinearity.) We also note that the magnitude of the FEM coefficient on
the total GDP is somewhat larger than its OLS counterpart, a consistent finding with the
previous empirical study by Matyas, Konya and Harris (2000) who argue that allowing for
heterogeneous bilateral effects is likely to increase the magnitude of the impact of GDP .2
One of our main purposes of the current study is an investigation of consistent esti-
mation (and thus precise evaluation) of the impacts of individual specific variables. We
consider both (inconsistent) OLS and (consistent) HT estimations and summarise such es-
timation results in Table 2(c). Here we assume a priori that Lan is the only time invariant
variable correlated with unobserved individual effects (as common language is a proxy for
cultural, historical, linguistic proximity, it is highly likely to be correlated with unobserved
individual effects). We employ two different sets of instrument variables. The first instru-
ment set (HT1) contains only real exchange rates (RER), the second set (HT2) adds size
related variables such as GDP s, SIM and RLF . Following de Sousa and Disdier (2002) we
do not consider time-varying dummy variables as valid instruments. As expected a priori,
all estimation results show that distance has a negative effect on exports and trades, while
2
Most empirical studies find that estimates of the income coefficient are well over unity, e.g., Matyas
(1997), Cheng and Wall (2002) and Martinez-Zaroso and Nowak-Lehmann (2003).

[17]
common language and common border have positive effects on them. Here a notable finding
is that once the correlation between Lan and unobserved individual effect is accommodated
by the HT estimation, then the impacts of distance decrease (in absolute value) as com-
pared to the OLS counterpart, whilst the impacts of both common language and common
border dummies increase, especially the former. Furthermore, when we use the broad set
of instruments (HT2), the distance variable loses significance. This result might be plausi-
ble given the fact that both distance and common border proxy geographical distance, the
effects of which might compensate each other (the correlation coefficient between them is
about 0.6). Overall, this result suggests that the role of cultural affinities will become more
important in explaining the pattern of bilateral trade flows once the correlation between Lan
and unobserved individual effect is appropriately handled.

Table 2 about here

Next, we consider an extended model in order to capture business cycle effects or deal
with globalization issues. We first follow the conventional approach and include the T − 1
fixed dummy variables (not T dummies to avoid multicollinearity) in (2.5) and (2.6), that are
common to all country pairs. Notice that the impacts of fixed time dummies are assumed to
be homogeneous. Table 3 reports the related estimation results. Although most estimation
results for both triple and double index specifications follow similar patterns as obtained in
Table 2, there are a few notable discrepancies (mainly in the context of the FEM estimation
results). First, the impact of EMU is now mostly significantly positive. Second, the impact
of the GDPs seem to be somewhat too large. Third, the impact of SIM on exports is
no longer significant (see Table 3(a)), whereas the impact of SIM on total trades is still
significant and larger (see Table 3(b)). Finally, turning to the HT estimation of the impacts
of individual specific variables, we find that the estimates of the impact of distance are
surprisingly positive but insignificant, the impact of common language dummy are much
larger than in Table 2(c), but common border dummy loses its statistical significance.

Table 3 about here

We move to address an alternative approach of allowing for common time factors; namely
we consider our proposed extended HT approach as developed in Section 3. We find from
Table 1 that the share of EU trade with the US has always been a consistent part of the
extra-EU trade. For example, it is reported in Trade policy review of the European Union:
A Report by the Secretariat of the WTO (2002) that the percentage of export (import) from
Europe to the US increases from around the 10% (10%) of the total volume of EU export
in 1960 to around the 25% (20%) in 2000. Hence, we expect that certain characteristics of
the US will also help in further explaining the pattern of the intra-EU exports and/or total
trades. In this regard, we consider the EU and the US as two main trade blocks and then
augment the model with the US reference variables, which we regard as observed common
time factors. Here we simply choose the (logarithm of) real exchange rates (RERTt ) that
will capture any of the relative price effects between the European currencies and the US
dollar.3 We expect that a depreciation of the European currency with respect to the US
3
Here the home currency is the European currency, i.e. ECU till 1998 and Euro from 1999 to 2001, and

[18]
dollar (an increase in RERTt ) should result in more extra-EU exports to and less extra-
imports from the US, though its impact on the intra-EU trade will be ambiguous. We thus
consider the model (3.7) for both triple and double index versions, where ft = RERTt , and
focus only on the FEM combined with the HT estimation results. Under our maintained
assumption that common language dummy is only correlated with unobserved individual
effects, we consider the four different instrument sets, denoted HT1, HT2, HT3 and HT4,
respectively, where HT1 and HT2 are exactly the same as before, namely HT 1 = {RER}
and HT 2 = {RER, GDP s, SIM, RLF }, whilst HT3 and HT4 are the sets combining HT1
and HT2 respectively with λ̂i RERTt . Remind that we follow our theoretical discussion in
Section 3 and use λ̂i RERTt as an additional source of instrument in HT3 and HT4.
Table 4 summarizes there results. First, looking at the results for the triple index model
(Table 4(a)), we find that signs and significances of coefficients are preserved, though the
magnitudes of the coefficients are somewhat different from the previous estimates reported
in Table 2(a). But, the coefficient on EMU is surprisingly negative and significant. The
HT estimates of coefficients on individual specific variables all show the expected signs, but
the language dummy loses its statistical significance. Next turning to the double index
model (Table 4(b)), most FEM estimates are similar to those shown in Table 2(b) with the
following main difference: the coefficients on EMU and RLF are both insignificant. The HT
estimates of the impacts of individual specific variables show more or less the similar patterns
to Table 2(c), namely, the distance variable becomes insignificant whilst the language variable
becomes more important in explaining the pattern of trade flows.

Table 4 about here

We notice in passing that the choice of observed common factors might be somewhat
arbitrary in general and that there is always a possibility of missing factors. In this regard,
there is still a room for further improving previous estimation results, and we now take an
alternative approach based on the assumption that the common time factors are unobserved
and their impacts are heterogeneous. This approach has two advantages: First, we may
avoid inevitable arbitrariness and difficulty in selecting observed common factors. Secondly
and more importantly, this approach is also able to accommodate certain degrees of cross
section dependence via heterogeneous time-specific effects, and thus to avoid the potential
bias of uncorrected estimates as described earlier. Here we follow the PCCE estimation
methodology advancedn by Pesaran (2002) to deal with
o this issue and thus consider the model
(3.7), where ft = y t , T GDP t , SIM t , RLFt , RERt and the bar over the variable indicates
P
the cross sectional average of the variable of interest, namely y t = N −1 Ni=1 yit and so on.
4

As before, we focus only on the FEM combined with the HT estimation results and maintain
the assumption that common language dummy is only correlated with unobserved individual
effects. We now consider the following four different instrument sets: HT 1 = {RER} and
HT 2 = {RER, GDP s, SIM, RLF }, whilst HT3 and HT4 are the sets combining HT1 and
the foreign currency is the US Dollar. See also Data Appendix. We have also tried different US reference
variables such as the US GDP, and found the qualitatively similar results.
4
We do not include cross sectional average of the CEE and EM U dummies to avoid the potential
multicollinearity problem. We also notice that T GDP t = GDP ht = GDP f t .

[19]
 n o
HT2 respectively with λ̂1i y t , λ̂2i T GDP t , λ̂3i SIM t , λ̂4i RLFt , λ̂5i RERt .5
We provide these estimation results in Table 5. From Table 5(a) for the triple index
model, we find that the impacts of foreign GDP, RLF and EMU are all insignificant, while
the impact of CEE is smaller than reported in Table 2(a). The HT estimation results show
that the distance is significantly more negative while both common language and border
dummies become insignificant. Turning to Table 5(b) for the double index model, most
FEM estimates are quite similar to those reported in Table 2(b). Main differences are:
the coefficients on EMU and RLF are both insignificant while the impact of CEE is now
much smaller. The HT estimates of the impacts of individual specific variables confirms
similar findings to
n those reported in Table 2(c). Interestingly,o once the instrument set is
augmented with λ̂1i y t , λ̂2i T GDP t , λ̂3i SIM t , λ̂4i RLFt , λ̂5i RERt , we find that all individual
specific variables (distance, common language and border) become strongly significant with
expected signs. This may indicate the potential importance of using additional source of
instruments.

Table 5 about here

Comparing and evaluating the above three extended estimation results in light of our a
priori expectations, we may reach to the following conclusion: First, the results obtained
using the conventional T − 1 fixed dummies (with their homogeneous impacts) are least
satisfactory, which might indicate that the conventional approach may be too limited to
accommodate the time effects. Second, the estimation results with an observed time factor
are somewhat mixed in the sense that most estimation results are relatively sensible for the
double index model, but not quite for the triple index model. Finally, the estimation results
with unobserved time factor (in conjunction with the PCCE estimation) are similar to but
more sensible than those obtained using the observed common time factor. In particular,
the results of Table 5(b) for the double index model for explaining the patterns of bilateral
total trade flows are mostly sensible. Therefore, this overall observation may suggest the
potential advantage of our proposed approach over the conventional one based on the fixed
time dummies.
We now summarise our main findings in a broad context combining all of the above
estimation results together but mainly focussing on estimation results in Tables 2 and 5. We
begin with the triple index model in explaining the pattern of bilateral real exports. The
impact of the GDP variables is mostly significant and positive with the total impact being just
under 2. Only the impacts of foreign population are found to be significant but negative. The
impact of similarity in relative size of trading countries are mostly significant and positive,
ranging between 0.16 and 0.35. The impact of differences in relative factor endowments are
mostly significant and positive, ranging between 0.01 and 0.03. The impacts of CEE are all
positive and significant, mostly around 0.3. The results are mixed for the impacts of EMU,
but mostly insignificant in both Tables 2(a) and 5(a). The impacts of distance are mixed
in Table 2(c), but become significantly negative in Table 5(a). The impacts of common
n o
5
In practice, the subset of λ̂1i y t , λ̂2i T GDP t , λ̂3i SIM t , λ̂4i RLFt , λ̂5i RERt can be parsimoniously used
as instruments.

[20]
language are mixed in Table 2(c) but become insignificant in Table 5(a). The impacts of
common border are mostly significant and range between 0.49 and 0.76.
Next, we move on to the double index model in explaining the pattern of bilateral real
total trades. The impacts of GDP are all significant and positive, ranging between 1.63 and
2.02. The impacts of population are insignificant. The impacts of SIM are all significant
and positive, ranging between 1.11 and 1.4, which are significantly larger than its impacts on
exports only. The impacts of RLF are significantly positive in Table 2(b), but insignificant
in Table 5(b). The impacts of CEE are all significantly positive. The impact of EMU is
significantly positive in Table 2(b), but becomes insignificant in Table 5(b). The impacts of
distance, common language and common border are mostly significantly negative, positive
and positive, respectively.
Though the above estimation results and their interpretations are more or less consistent
with our a priori expectations, we notice that there are two conflicting findings between
the benchmark estimation results in Table 2 and the extended HT estimation results in
Table 5; namely, the role of the RF L and EMU variables. The impacts of RLF are found
to be significant and positive in Table 2, but become insignificant in Table 5, whilst the
impacts of EMU are found to be mostly insignificant, but only become significantly positive
in Table 2(b). As mentioned earlier, the impact of RLF on total trade flows might not
be unambiguous since the total trade flows are the sum of inter- and intra-industry trades.
Next, we earlier discussed that empirical evidence on the impact of EMU on trade flows is
mixed. In particular, de Souza (2002) argues that either the periods are too short after an
introduction of the Euro to use the EMU dummy as an adequate proxy for monetary union
membership, or forward looking agents anticipate and thus discount the increase of trade
associated with the EMU membership. In this regard we also expect that the impacts of
EMU are yet to be significant. Along this line of logics we may conclude that the estimation
results obtained using our proposed HT methodology seem to be much more sensible.

5 Conclusions
In this paper we follow recent developments of panel data studies surrounding the use of
common time effects, and advance an alternative estimation framework in which we explicitly
allow for the existence of observed and/or unobserved common time-specific factors and
individual responses to those common factors are heterogeneous across country pairs. We
then generalize the HT estimation methodology in the context of the extended panel data
model and develop the underlying econometric theory.
We apply our proposed HT estimation technique along with the conventional approaches
to a comprehensive analysis of the gravity equation of bilateral trade amongst the 15 Eu-
ropean countries over 1960-2001. Empirical results clearly demonstrate that our proposed
approach fits the data reasonably well and its estimations results are sensible in a number of
different dimensions. In particular, our proposed (extended) HT estimation provides much
more sensible results than the conventional approach based on the fixed time dummies, es-
pecially in terms of the impacts of individual sepcific variablse such as distance, common
border and langauage dummies. We further notice that our proposed HT estimation results
produce more sensible predictions on the impacts of differences in factor endowments and

[21]
of the common currency dummy on intra-EU trade flows than the conventional approach
with and without fixed time dummies. This observation may indicate the importance of
properly accommodating a certain degree of cross section dependence through unobserved
heterogeneous time effects, otherwise the resulting estimates would be severely biased.
A couple of extensions will be desirable. First, it would be worth investigating the effect of
globalization on transport costs more explicitly. For instance, transport and communication
revolutions should lead to a dispersion of economic activity. Although this dispersion did not
occur with the reduction in transportation costs during the first wave of the globalization in
the 20th century, the second wave of globalization associated with recent information and
communication technologies revolution should lead to an integrated equilibrium view of the
‘death of the distance’. Hence, it would be interesting to study the effect of an ‘augmented’
trade-barrier function which make transport costs both dependent on and independent of
distance in addition to the standard trade-barrier function that only comprehend variables
like distance, common language and common border dummies as employed in the current
paper, see for example Brun et al. (2002). Secondly, it would be interesting to analyse the
gravity models of international trade over different time periods. For instance, the impacts
of intra- and inter-industry trades will be different over different time periods, and thus we
might expect that the role of certain explanatory variables such as RLF and EMU changes
accordingly. Of the particular importance will be to reexamine the issue concerning the
impacts of the Euro on the bilateral intra-EU trade once the data over the longer time
periods will be available, as we argue that the insignificantly estimated impact of the EMU
dummy might be due to the shortage of observations.

[22]
 Data Appendix
We now describe how the variables are constructed. All variables are converted in constant
dollar prices with 1995 as³ the ´base year. Bilateral ³ exports
´ and imports are defined as
R R R R
logarithms of real export Xhf t and real imports M hf t , Xhf t and Mhf t are obtained by
R N 100 R N 100 N N
Xhf t = Xhf t × XP IU S , Mhf t = Mhf t × MP IU S , where Xhf t and Mhf t are bilateral export
and import measured in millions of current US dollars, and XP IUS and MP IUS are the
US³ export and ´import price indices. Then, the total volume of trade is given by T rade =
R R R
ln Xhf t + Mhf t . GDP of home and foreign country are defined as logarithms of GDPht and
GDPfRt , where GDPht R
and GDPfRt are gross domestic products at constant³
dollar of country
´
R
h and f , respectively, and the total GDP is defined as T GDPit = ln GDPht + GDPfRt .
GDP’s are originally expressed in million Euro for the twelve countries that joined the
European Monetary Union (Austria, Belgium, Finland, France, Germany, Greece, Ireland,
Italy, Luxemburg, Netherlands, Portugal, Spain) and in millions of current national currency
for Denmark, Sweden and UK (GDP N ). In the last three cases the original nominal values
of GDP have been deflated by the GDP deflator (GDP D, 1995 = 100) of the respective
countries whereas for the remaining countries the European GDP deflator has been used.
We also convert GDPs in US dollar at the exchange rate of 1995 (mean over period) in order
to exclude the effect of a dollar depreciation or appreciation as follow:
à !
R N 100 US$
GDPhf t = GP Dhf t × × ,
GDP Dht NCh 1995

where NCh stands for national currency of the home country. Population of home and
foreign countries are defined as logarithms of P OPht and P OPf t , where P OPht and P OPf t
are the population of country h and f measured in million of inhabitants and the total
population is defined as T P OPit = ln (P OPht + P OPf t ). Next, we construct SIMit and
RLFit respectively by
⎡ Ã !2 Ã !2 ⎤
GDPhtR GDPfRt
SIMit = ln ⎣1 − R
− ⎦,
GDPht + GDPfRt GDPfRt + GDPht
R

¯ ¯
RLFit = ln ¯¯P GDPfRt − P GDPht
R¯
¯,
where P GDP is per capita GDP. Real exchange rates in constant dollars at 1995 are defined
as RERit = NERit × XP IU S , where NERit is nominal exchange rate between currencies h
and f in year t in terms of dollars. Lastly, the distance between countries is measured as
the great circle distance between national capitals in kilometers.
The data sources are as follows: Export and import price indices are collected from
OECD Economic Outlook, GDP deflators from World Bank World Development Indicators,
and bilateral nominal export and import data (X N and M N ) from OECD, Statistical Com-
pendium, Main Economic Indicator, Yearly Statistic of Foreign Trade in current dollars,
GDP from IMF International Financial Statistics, Economic Concept View, National Ac-
counts, per capita GDP (already converted in constant dollars) from the World Bank World
Development Indicators, population from the World Bank World Development Indicators,
and NER from OECD, National Accounts, Volume I.

[23]
 Table Appendix

Table 1 Descriptive and Summary Statistics

Panel A 1960 1970 1980 1990 2000

Share of US on Extra-EU trade 16.5∗ 26.3∗ 33.8∗∗ 19∗∗ 21.9∗∗∗
Share of Intra-EU on EU trade 37.2∗ 49.8∗ 50.5∗∗ 59.7∗∗ 61.7∗∗∗
Share of Export on Intra-EU trade 52.4∗ 51.6∗ 51.1∗∗ 49.7∗∗ 51.2∗∗∗
Panel B 60/70 70/80 80/90 90/00
Average Growth of GDP 8.9 16.4 7.8 3.5
Average Growth of Intra-EU trade 11.5 17.3 9.3 5.8
Average Growth of Total EU trade 10.3 20.1 7.2 3.9
Average Growth of Bilateral Exchange Rate 0.12 7.9 -1.4 -3.7
Notes: Source: Trade Policy Review of the European Union: a Report by the Secretariat of the WTO, WTO
(2002) and Statistical Yearbook, Eurostat (1997). ‘∗ ’ denotes values for EU9 (Austria, Belgium, Finland,
France, Germany, Ireland, Italy, Luxemburg, Netherlands), ‘∗∗ ’ for EU12 ( EU9 plus Greece, Portugal and
Spain) and ‘∗∗∗ ’ for EU15 (EU12 plus Denmark, Sweden and United Kingdom) countries, respectively.

Table 2(a). Alternative Panel Data Estimation Results for Triple Index Models

Basic Model Augmented Model Full Model

OLS BTW FEM REM OLS BTW FEM REM OLS BTW FEM REM
Con −11.8∗ −6.1∗ −16.5∗ −12.7∗ −8.4∗ −16∗ −10∗ −10∗ −15.3∗
(.241)2 (1.89) (.59) (.266) (1.94) (.745) (.211) (1.24) (.73)
GDPh 0.73∗ .49∗ .64∗ .78∗ .68∗ .51∗ .55∗ .67∗ .76∗ .73∗ .49∗ .73∗
(.015) (.106) (.035) (.03) (.015) (.101) (.034) (.029) (.007) (.041) (.03) (.021)
GDPf 1.25∗ .99∗ 1.5∗ 1.55∗ 1.21∗ 1.03∗ 1.4∗ 1.4∗ .87∗ .85∗ 1.43∗ 1.18∗
(.015) (.106) (.035) (.03) (.015) (.101) (.034) (.029) (.007) (.041) (.03) (.021)
POPh .01 .27∗ .06 −.01 .04∗ .25∗ −.02 .07
(.019) (.124) (.127) (.058) (.019) (.121) (.124) (.057)
POPf −.52∗ −.25 ∗
.72∗ −.61∗ −.48 ∗
−.26 ∗
.64∗ −.54∗
(.019) (.124) (.127) (.058) (.019) (.121) (.124) (.057)
SIM .11∗ .04 .35∗ .30∗
(.013) (.071) (.051) (.041)
RLF .03∗ .02 .03∗ .03∗
(.007) (.05) (.007) (.007)
RER .1∗ .09∗ .08∗ .09∗
(.003) (.019) (.007) (.007)
CEE .28∗ −.12 .28∗ .29∗ .33∗ −.08 .31∗ .32∗
(.019) (.202) (.013) (.014) (.019) (.19) (.014) (.013)
EMU .07 −2.21 −.01 −.01 .17∗ −1.1 −.01 .01
(.044) (1.44) (.023) (.023) (.043) (1.49) (.023) (.024)
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗
Dist −1.05 −1.2 −.92 −.79 −.92 −.71 −.68 −.77 −.57
(.015) (.086) (.079) (.018) (.109) (.097) (.019) (.11) (.096)
Lan .51∗ .48∗ .59∗ .25∗ .27∗∗ .41∗
(.028) (.16) (.155) (.029) (.163) (.152)
Bor .45∗ .52∗ .41∗ .51∗ .60∗ .44∗
(.029) (.167) (.161) (.028) (.163) (.156)

Notes: Here the dependent variable is logarithm of real export. OLS stands for the pooled OLS estimator,
BTW the between estimator; FEM fixed effects estimator and REM random effects estimator, respectively.
Figures in (.) indicate the standard error. ‘∗ ’ denotes coefficient significant at the 5% level of significance.
‘∗∗ ’denotes coefficient significant at the 1% level of significance. Hausman statistic rejects the null hypothesis
of no correlation between explanatory variables and unobserved individual effects in all cases considered.

[24]
 Table 2(b). Alternative Panel Data Estimation Results for Double Index Models

Basic Model Augmented Model Full Model

OLS BTW FEM REM OLS BTW FEM REM OLS BTW FEM REM
Con −7.5∗ .02 −16.5∗ −7.9∗ −2.3 −15.4∗ −10.9∗ −9.7∗ −13.9∗
(.361) (2.64) (.94) (.39) (2.78) (1.18) (.247) (1.55) (.88)
GDP 1.7∗ .98∗ 2.2∗ 2.3∗ 1.5∗ 1.08∗ 1.9∗ 2.1∗ 1.57∗ 1.5∗ 1.81∗ 1.79∗
(.03) (.234) (.031) (.023) (.032) (.231) (.031) (.025) (.012) (.08) (.019) (.018)
∗
POP −.52 .24 .03 −.83∗ −.38∗ .12 .01 −.74 ∗
(.039) (.281) (.164) (.099) (.039) (.279) (.154) (.095)
SIM .88∗ .81∗ 1.17∗ 1.14∗
(.017) (.104) (.055) (.045)
RLF .03∗ .01 .03∗ .03∗
(.008) (.06) (.008) (.008)
RER .09∗ .09∗ .06∗ .07∗
(.004) (.023) (.009) (.008)
CEE .47∗ .14 .14∗ .36∗ .32∗ −.03 .31∗ .32∗
(.03) (.321) (.028) (.016) (.022) (.244) (.016) (.016)
EMU .22∗ .89 .31∗ .14∗ 0.2∗ −1.4 .08∗ .09∗
(.069) (.778) (.016) (.028) (.051) (1.84) (.027) (.027)
∗ ∗ ∗ ∗ ∗ ∗ ∗
Dist −1.2 −1.3 −.97 −.93 −1.03 −.77 −.64 −.71∗ −.6 ∗
(.022) (.129) (.124) (.028) (.169) (.156) (.022) (.13) (.116)
Lan .36∗ .24 .51∗ .23∗ .24 .41∗
(.043) (.255) (.247) (.034) (.201) (.185)
Bor .42∗ .58∗ .39 .52∗ .61∗ .44∗
(.045) (.269) (.258) (.034) (.201) (.19)

Notes: Here the dependent variable is logarithm of real export; Hausman statistic rejects the null hypothesis
of no correlation between explanatory variables and unobserved individual effects in all cases considered. See
also notes to Table 2(a).

Table 2(c). Hausman and Taylor Estimation Results

Triple index model Double index model

OLS HT1 HT2 OLS HT1 HT2
∗ ∗ ∗ ∗
Dist −.57 −.43 −.34 −.6 −.38 −.34∗∗
(.026)2 (.208) (.208) (.021) (.192) (.199)
Lan 0.45∗ 1.05 1.57∗ 0.45∗ 1.56∗ 1.8∗
(.041) (.755) (.72) (.034) (.707) (.695)
∗ ∗∗ ∗ ∗ ∗ ∗
Bor 0.43 0.53 0.61 0.43 0.6 0.64
(.042) (.282) (.289) (.035) (.258) (.275)

Notes: Here we consider only the full specifications, and those slope coefficients are already reported as FEM
estimates in Tables 2(a) and 2(b). The set of instrument variables used in the HT estimation are as follows:
{RER} for HT1 and {RER, GDP, SIM, RLF } for HT2. See also notes to Tables 2(a) and 2(b).

[25]
 Table 3(a). Alternative Panel Data Estimation Results for Triple Index Models with Time Dummies
Basic Model Augmented Model Full Model
OLS FEM REM OLS FEM REM OLS FEM REM
Con −7.15∗ −19.3∗ −9.64∗ −20.2∗ −10.6∗ −18.7∗
(.314) (.859) (.323) (.996) (.215) (.997)
GDPh 0.51∗ 0.97∗ 0.91∗ 0.54∗ 0.91∗ 0.86∗ 0.73∗ 1.04∗ 0.84∗
(.018) (.044) (.041) (.017) (.043) (.041) (.007) (.045) (.029)
GDPf 1.03∗ 1.82∗ 1.68∗ 1.06∗ 1.76∗ 1.62∗ 0.84∗ 1.97∗ 1.32∗
(.018) (.044) (.041) (.017) (.043) (.041) (.007) (.045) (.029)
POPh 0.24∗ 0.03 −.14∗ 0.21∗ −.007 −.11∗∗
(.021) (.123) (.063) (.021) (.12) (.062)
POPf −.28∗ 0.69∗ −.73∗ −.32∗ 0.65∗ −.7 ∗
(.021) (.123) (.063) (.021) (.12) (.062)
SIM 0.07∗ 0.02 0.25∗
(.012) (.052) (.041)
RLF 0.02∗ 0.02∗ 0.02∗
(.007) (.007) (.006)
RER 0.09∗ 0.08∗ 0.06∗
(.003) (.009) (.008)
CEE 0.11∗ 0.29∗ 0.29∗ 0.15∗ 0.33∗ 0.31∗
(.022) (.014) (.015) (.022) (.015) (.015)
EMU 0.13∗ 0.19∗ 0.23∗ 0.18∗ 0.17∗ 0.23∗
(.061) (.031) (.032) (.059) (.032) (.032)
Dist −1.15∗ −.86∗ −.89∗ −.61∗ −.75 ∗
−.52∗
(.015) (.081) (.018) (.099) (.019) (.1)
Lan 0.51∗ 0.64∗ 0.28∗ 0.62∗
(.027) (.156) (.028) (.154)
Bor 0.44∗ 0.41∗ 0.53∗ 0.31∗
(.028) (.162) (.028) (.157)

Notes: Here we augment the models in 2(a) by adding the time-specific fixed effects. Hausman statistic
rejects the null hypothesis of no correlation between explanatory variables and unobserved individual effects
in all cases considered. See also notes to Table 2(a).

Table 3(b). Alternative Panel Data Estimation Results for Double Index Models with Time Dummies
Basic Model Augmented Model Full Model
OLS FEM REM OLS FEM REM OLS FEM REM
∗
Con −1.75 −15.5∗ −3.37∗
−16.5∗ −10.2∗ −20.2∗
(.12) (.75) (.44) (1.45) (.257) (1.2)
GDP 1.03∗ 2.5∗ 2.3∗ 1.07∗ 2.48∗ 2.23∗ 1.53∗ 3.05∗ 2.22∗
(.038) (.084) (.077) (.037) (.081) (.074) (.013) (.078) (.053)
POP 0.18∗ −.49∗ −.96∗ 0.11∗ −.44∗ −.9 ∗
(.045) (.154) (.111) (.045) (.147) (.108)
SIM 0.84∗ 1.42∗ 1.27∗
(.017) (.055) (.049)
RLF 0.02∗ 0.02∗ 0.02∗
(.008) (.007) (.007)
RER 0.09∗ 0.09∗ 0.06∗
(.003) (.01) (.009)
CEE 0.17∗ 0.35∗ 0.31∗ 0.17∗ 0.32∗ 0.31∗
(.034) (.036) (.017) (.026) (.017) (.017)
EMU 0.11 0.32∗ 0.37∗ 0.21 ∗
0.22 ∗
0.28∗
(.091) (.017) (.036) (.07) (.034) (.035)
Dist −1.3∗ −1.01∗ −1.05∗ −.76∗ −.69∗ −.44∗
(.021) (.125) (.027) (.158) (.022) (.123)
Lan 0.35∗ 0.52∗ 0.26∗ 0.65∗
(.041) (.248) (.034) (.189)
Bor 0.43∗ 0.41 0.54∗ 0.28
(.043) (.258) (.033) (.195)

Notes: Here we augment the models in 2(b) by adding time-specific fixed effects; Hausman statistic rejects
the null hypothesis of no correlation between explanatory variables and unobserved individual effects in all
cases considered. See also notes to Tables 2(a) and 2(b).

[26]
 Table 3(c). Hausman and Taylor Estimation Results with Time Dummies
Triple index model Double index model
OLS HT1 HT2 OLS HT1 HT2
Dist −.14∗ 0.27 0.52 −.07 0.38 0.67
(.036) (.331) (.388) (.045) (.43) (.529)
Lan 0.92∗ 3.1∗ 4.71∗ 0.94∗ 3.27∗ 5.06∗
(.057) (1.17) (1.31) (.072) (1.52) (1.78)
Bor .06 0.4 0.7 0.03 0.39 0.71
(.059) (0.44) (.54) (.074) (.583) (.734)

Notes: See notes to Tables 2(c), 3(a) and 3(b).

Table 4. FEM and HT Estimation Results with an Observed Time Factor

Table 4(a). Triple index model
FEM1 OLS HT1 HT2 HT3 HT4
GDPh 1.09∗2
(.035)
GDPf 0.88∗
(.035)
∗
SIM 0.21
(.055)
RLF 0.01∗
(.005)
RER 0.16∗
(0.01)
CEE 0.33∗
(.011)
EMU −.06∗
(.018)
Dist −.43∗ −.23 −.43∗ −.13 −.43∗
(.02) (.344) (.161) (.367) (.161)
Lan 0.25∗ 1.26 0.25 1.85 0.24
(.032) (1.68) (.472) (1.81) (.473)
∗ ∗ ∗ ∗ ∗
Bor 0.49 0.65 0.49 0.74 0.49
(.033) (.312) (.222) (.319) (.222)

Notes: Here we augment the models in 2(a) by adding RERT . See also notes to Table 2(a). The set of
instrument variables used in the HT estimation are as follows: {RER} for HT1,
n {RER,
o GDP, SIM, RLF }
for HT2, and HT3 and HT4 are HT1 and HT2 respectively combined with λ̂i RERTt .

Table 4(b). Double index model

FEM1 OLS HT1 HT2 HT3 HT4
GDP 2.022∗
(.03)
SIM 1.4∗
(.062)
RLF 0.009
(.006)
RER 0.11∗
(.011)
CEE 0.31∗
(.012)
EMU −.008
(.019)
Dist −.4∗ −.38∗∗ −.19 −.35∗∗ −.17
(.024) (.208) (.227) (.211) (.227)
Lan 0.39∗ 0.44 1.45∗ 0.64 1.6∗
(.038) (.641) (.671) (.627) (.081)
Bor 0.43∗ 0.45 0.61∗ 0.48 0.63∗
(.041) (.317) (.309) (.312) (.311)

[27]
Notes: Here we augment the models in 2(b) by adding the single observed time-specific factor, RERT . See
also notes to Table 4(a).

Table 5 FEM and HT Estimation Results with an Unobserved Time Factor

Table 5(a). Triple index model

FEM1 OLS HT1 HT2 HT3 HT4
GDPh 1.26∗2
(.075)
GDPf −.03
(.075)
SIM 0.16∗
(.081)
RLF −.0001
(.005)
RER 0.03∗
(.011)
CEE 0.12∗
(.013)
EMU −.001
(.018)
Dist −.97∗ −.72∗ −.92∗ −.6∗ −.94∗
(.039) (.299) (.268) (.302) (.254)
Lan 0.31∗ 1.57 0.5 2.25∗∗ 0.46
(.062) (1.26) (1.12) (1.21) (1.01)
Bor 0.56∗ 0.76∗ 0.59 0.87∗ 0.59
(.064) (.385) (.371) (.396) (.364)

Notes:
© Here we follow Pesaran (2002) ª and augment the model in 2(a) by adding multiple factors,
y t , T GDP t , SIM t , RLFt , RERt , where the bar over the variable indicates its cross-section average. See
also notes to Table 2(a). The set of instrument variables used in the HT estimation are: {RER} for
HT1,
n {RER, GDP, SIM, RLF } for HT2, and HT3 o and HT4 are HT1 and HT2 respectively combined with
λ̂1i y t , λ̂2i T GDP t , λ̂3i SIM t , λ̂4i RLFt , λ̂5i RERt .

Table 5(b). Double index model

FEM OLS HT1 HT2 HT3 HT4

GDP 1.63∗
(.115)
SIM 1.11∗
(.093)
RLF −.001
(.005)
RER 0.03∗
(.014)
CEE 0.14∗
(.013)
EMU −.01
(.017)
Dist −.73∗ −.38∗∗ −.41∗∗ −.46∗ −.48∗
(.023) (.23) (.233) (.218) (.209)
Lan 0.51∗ 2.33∗ 2.17∗ 1.89∗ 1.82∗
(.037) (0.78) (.76) (.638) (.628)
Bor 0.44∗ 0.72∗ 0.7∗ 0.65∗ 0.65∗
(.038) (0.28) (.266) (.255) (.245)
© ª
Notes: Here we follow Pesaran (2002) and augment the model in 2(b) by adding y t , T GDP t , SIM t , RLFt , RERt .
See also notes to Table 5(a).

[28]
 60

50

40

One-way
30
GL

20

10

0
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
Figure 1 Evolution of trade in intra-EU trade 1980-1994
Notes: The Grubel-Lloyd indicator measures the share of intra-industry trade and one-way trade rep-
resents inter-industry trade. See also the footnote 1. Source: Single Market Review, European Commission
(1997): Trade Patterns inside the Single Market, page 50.

[29]
References
[1] Ahn, S.G., Lee, Y.H. and P. Schmidt (2001), “GMM Estimation of Linear Panel Data Models with
Time-varying Individual Effects,” Journal of Econometrics, 101, 219-255.
[2] Anderson, P.S. (1979), “A Theoretical Foundation for the Gravity Equation,” American Economic
Review, 69, 106-116.
[3] Amemiya T. and T. MaCurdy (1986), “Instrumental Variables Estimation of an Error Components
Model,” Econometrica, 54, 869-880.
[4] Bai, J. and S. Ng (2002), “Determine the Number of Factors in Approximate Factor Models,” Econo-
metrica, 70, 191-221.
[5] Baldwin, R.E. (1994), “Towards an Integrated Europe” London, CEPR.
[6] Bergstrand, J.H. (1989), “The Generalized Gravity Equation, Monopolistic Competition, and Factor
Proportions Theory of International Trade,” Review of Economics and Statistics, 71, 143-153.
[7] Bergstrand, J.H. (1990), “The Heckscher-Olin-Samuelson Model, the Linder hypothesis and the Deter-
minants of Bilateral Intra-industry Trade,” Economic Journal, 100, 1216-29.
[8] Bougheas, S., P.O. Demetriades and E.L.W. Morgenroth (1999), “Infrastructure, Transport Costs and
Trade,” Journal of International Economics, 47, 169-189.
[9] Brada, J.C. and J.A. Mendez (1988), “Exchange Rate Risk, Exchange Rate Regime and the Volume of
International Trade,” Kylos, 41, 263-80.
[10] Breusch T., G. Mizon and P. Schmidt (1989) “Efficient Estimation using Panel Data,” Econometrica,
57, 695-700.
[11] Brun, J., C. Carrere, P. Guillaumont and J. de Melo (2002), “Has Distance Died? Evidence from a
Panel Gravity Model,” unpublished manuscript, University of d’Auvergne.
[12] Cheng, I. and H.J. Wall (2002), “Controlling Heterogeneity in Gravity Models of Trade,” Working Paper
1999-010C, Federal Reserve Bank of St. Louis.
[13] Davis, D.R. (2000), “Understanding International Trade Patters: Advantages of the 1990s,” unpublished
Manuscript, Columbia University.
[14] De Grauwe, P. and F. Skudelny (2000), “The Impact of EMU on Trade Flows,” Weltwirtschaftliches
Archiv, 136, 381-397.
[15] De Nardis, S. and C. Vicarelli (2003), “The Impact of the Euro on Trade,” ENEPRI, Working Paper
No. 17.
[16] De Sousa, J and A. Disdier (2002), “Trade, Border Effects and Individual Characteristics: A Proper
Specification,” unpublished Manuscript, University of Paris 1.
[17] de Souza L.V. (2002), “Trade Effects of Monetary Integration in Large, Mature Economies: a Primer
on the European Monetary Union,” Keil Working Paper No.1137.
[18] Deardorff, A.V. (1984), “Testing Trade Theories and Predicting Trade Flows,” in Handbook of Interna-
tional Economics, 1.
[19] Deardorff, A.V. (1998), “Determinants of Bilateral Trade: Does Gravity Work in a Neoclassical World?”
NBER, Working Paper No.5377.
[20] Eaton, J. and S. Kortum (2002), “Technology, Geography, and Trade,” Econometrica, 70, 1741-1779.
[21] Egger, P. (2001) “On the Problem of Endogenous Unobserved Effects of the Gravity Models,” Working
Paper, Austrian Institute of Economic Research, Vienna.
[22] Egger, P. (2002), “A Note on the Proper Econometric Specification of the Gravity Equation,” Economics
Letters, 66, 25-31.

[30]
 [23] Egger, P. and M. Pfaffermayr (2002), “The Proper Econometric Specification of the Gravity Equation: A
Three-way Model with Bilateral Interaction,” Working Paper, Austrian Institute of Economic Research.
[24] Evenett, S.J. and W. Keller (2002), “On the Theories Explaining the Success of the Gravity Equation,”
Journal of Political Economy, 110, 281-316.
[25] Frenkel, J.A. and S. Wei (1992), “Trade Blocs and Currency Blocs,” NBER Working Paper No. 4335.
[26] Glink, R. and A.K. Rose (2002), “Does a Currency Union Affect Trade? The Time Series Evidence,”
European Economic Review, 46, 1125-51.
[27] Grubel, H.G. and P.J. Lloyd (1975), Intra-industry Trade, the Theory and Measurement of International
Trade in Differentiated Products. MacMillan: London.
[28] Hausman, J.A. (1978), “Specification Tests in Econometrics,” Econometrica, 46, 1251-1271.
[29] Hausman, J.A. and W.E. Taylor (1981), “Panel Data and Unobservable Individual Effect,” Economet-
rica, 49, 1377-1398.
[30] Helpman, E. (1987), “Imperfect Competition and International Trade: Evidence from Fourteen Indus-
trialized Countries,” Journal of the Japanese and International Economies, 1, 62-81.
[31] Hummels, D. and J. Levinsohn (1995), “Monopolistic Competition and International Trade: Reconsid-
ering the Evidence,” Quarterly Journal of Economics, 110., 798-836.
[32] Jacquemin, A. and A. Sapir (1990), “Competition and Imports in the European Market,” in European
Integration: Trade and Industry. Cambridge University Press
[33] Krugman, P.R. (1979), “Increasing Returns, Monopolistic Competition and International Trade,” Jour-
nal of International Economics, 9, 469-79.
[34] Leamer, E. (1992), “The Interplay of the Theory and Data in the Study of International Trade,” in
Issues in Contemporary Economics, 2, New York University Press.
[35] Linder, S.B. (1961), An Essay on Trade and Transformation. John Wiley and Son: New York.
[36] Markusen, J. and R. Wigle (1990), “Expanding the Volume of North-South Trade,” Economic Journal,
100, 1206-15.
[37] Martinez-Zaroso, I. and F. Nowak-Lehmann (2003), “Augmented Gravity Model: An Application to
Mercosur-European Union Trade Flows,” Journal of Applied Econometrics, 18, 291-316.
[38] Matyas, L. (1997), “Proper Econometric Specification of the Gravity Model,” The World Economy, 20,
363-68.
[39] Matyas, L., Konya L. and M.N. Harris (2002), “Modelling Export Activity of Eleven APEC Countries,”
Melbourne Institute Working Paper No. 5/00.
[40] Molle, W. (1997), The Economics of European Integration: Theory, Practice and Policy. Ashgate Pub-
lishing Limited: ??.
[41] Pakko, M.R. and H.J. Wall (2002), “Reconsidering the Effects of a Currency Union,” Federal Reserve
Bank of St. Louis
[42] Persson, T. (2001), “Currency Union and Trade: How Large is the Treatment Effect?” Economy Policy,
33, 435-448..
[43] Pesaran, M.H (2002), “Estimation and Inference in Large Heterogeneous Panels with Cross Section
Dependence,” University of Cambridge DAE Working Paper No.0305.
[44] Phillips, P.C.B. and D. Sul (2002), “Dynamic Panel Estimation and Homogeneity Testing Under Cross
Section Dependence,” Cowles Foundation Discussion Paper No.1362, Yale University.
[45] Rose, A.K (2002), “The Effect of Common Currency on International Trade: Where Do We Stand?”
unpublished manuscript, Haas School of Business, University of California, Berkeley.

[31]

View publication stats