An Econometric View on the

Estimation of Gravity Models and

the Calculation of Trade Potentials
Peter Egger

1. INTRODUCTION

I N the last decade, the application of gravity models enjoyed a big revival.
This was not so much driven by its more rigorous theoretical foundation
(Anderson, 1979; Bergstrand, 1985, 1989 and 1990; Helpman and Krugman,
1985; and Helpman, 1987, etc.) but by the opportunity to project bilateral trade
relations (see Wang and Winters, 1991; Hamilton and Winters, 1992; Baldwin,
1994; and successors). The first applications were undertaken within the context
of the Fall of the Iron Curtain and the new potential integration effects between
the EU (OECD) and the former COMECON member states. According to the
traditional concept of the gravity equation, bilateral trade can be explained by
GDP and GDP per capita figures and both trade impediment (distance) and
preference factors (common border, common language, etc.). The economic
framework in most cases was cross-section analysis (Wang and Winters, 1991;
Hamilton and Winters, 1992; Brulhart and Kelly, 1999; and Nilsson, 2000, etc.).
Only a few authors made use of (random effects) panel econometric methods
(Baldwin, 1994; Gros and Gonciarz, 1996; Matyas, 1997; and Egger, 2000).1
Two conceptually different approaches were followed. First, some authors
argued that in the initial stage of the systemic and economic transformation, e.g.
of the Central and Eastern European countries (CEEC), these countries behaved
differently from developed countries such as the EU or OECD members. A
gravity model was estimated for EU or OECD countries and the parameters were
used to project natural trade relations between these countries and the CEEC.

PETER EGGER is from the Austrian Institute of Economic Research. Helpful comments from
M. Pfaffermayr, an anonymous referee and the editor are gratefully acknowledged.
1
Matyas (1997 and 1998) provides insights in the question of proper econometric specification
without dealing with the issue of trading potentials.

Blackwell Publishers Ltd 2002, 108 Cowley Road, Oxford OX4 1JF, UK
and 350 Main Street, Malden, MA 02148, USA. 297
298 PETER EGGER

The difference between the observed and the predicted trade flows was then
interpreted as the un-exhausted trade potential. I think that this was a valuable
experiment, especially at the early stage of transformation of the CEEC. This
technique was used by Wang and Winters (1991), Hamilton and Winters (1992)
and Brulhart and Kelly (1999), and I refer to it as the out-of-sample projection
approach. Second, and more recently, authors have included the transformation
countries under consideration already in the regression analysis. Then, the
residual of the estimated equation was interpreted as the difference between
potential and actual bilateral trade relations (Baldwin, 1994; and Nilsson, 2000). I
call this the in-sample projection approach.
This paper intends to provide insights into the choice of the adequate
estimation technique and into the problems related to in-sample projections of
trade potentials. In general, the estimator choice is an important issue for the
interpretation of the gravity coefficients, which depends on the underlying
interests. For future research, three important econometric problems should be
considered:
1. The traditional cross-section approach is probably affected by a severe
problem of misspecification. Matyas (1997) notes that the most natural
representation of bilateral trade flows is a three-way specification. Then,
eliminating one of the three dimensions (time) implies that the natural
representation of a time-averaged gravity model is a two-way panel with
(fixed or random) exporter and importer effects. Since these are the most
important dimensions of variation, convenient OLS estimates are very
likely to result in inconsistent estimates. This renders conclusions on OLS-
based trade potentials as problematic and affects both the in-sample and
the out-of-sample prediction concept.
2. We should care about the association of different estimators with short-
term and long-term time-horizons when comparing results (see Pirotte,
1999). Whereas fixed effects (and consistent random effects) model
estimates reflect short-run parameters, between model estimates are closer
to long-run parameters.
3. Finally, from a consistent and efficient estimator we should expect white-
noise residuals, which do not have any more systematic variation. If an
estimator reveals large systematic differences between observed and in-
sample predicted values (such as large un-exhausted East-West trade
potentials), this should be interpreted as an indication for misspecification
and parameter inconsistency.2

2
The application of the random effects approach is problematic because of the likelihood of its
inconsistency due to correlation between some of the explanatory variables and the unobserved
individual effects.

Blackwell Publishers Ltd 2002

 GRAVITY MODELS AND TRADE POTENTIALS 299

I focus on panel estimators. In my application, the previously used estimators

result in large unused trade potentials at least for intra-CEEC trade. I demonstrate
that the large in-sample actual-to-potential trade ratios stem from two sources.
First, the correlation of the explanatory variables with the unobserved effects
leading to inconsistent parameter estimates for the random effects model (REM)
as used by Baldwin (1994) and Gros and Gonciarz (1996). Second, from serial
correlation of the residuals. The consistent and efficient estimator in the
application is an AR(1) model in the spirit of Hausman and Taylor (1981), which
has not been used previously either in trade or in other fields of economic
research. This estimator eliminates the systematic difference between observed
and in-sample predicted trade flows. However, the in-sample approach to the
calculation of trade potentials is misleading since properly specified econometric
models cannot obtain systematic variations in residuals at all.
The paper is organised as follows. Section 2 briefly introduces the
specification. Section 3 presents the estimation results and in-sample projections
of export potentials of the EU member states (EU15), Hungary, Poland, and the
Czech Republic (henceforth CEEC3) in 10 CEEC (CEEC10).3 Section 4
concludes.

2. THE ECONOMETRIC SPECIFICATION

According to the endowment-based new trade model with Dixit and Stiglitz
(1977) preferences, bilateral trade is an increasing function of bilateral sum of
factor income G, relative country size S, and the difference in relative factor
endowments (R; compare Helpman and Krugman, 1985; Helpman 1987; and
others). Accordingly, bilateral exports can be estimated by
Yijt 0 1 Gijt 2 Sijt 3 R ijt 4 Vit 5 Vjt 6 RL it 7 RL jt
8 Eijt 9 Dij 10 Bij 11 Lij t uijt ; 1
where all variables are real figures and expressed in logs, and the error term can
be written as:
uijt ij vijt ; 2
with ij as the (fixed or random) unobserved bilateral effect and vijt as the
remaining error. In line with Helpman (1987) the Heckscher-Ohlin determinants
can be formulated in the following way:

3
The 10 CEEC are: Bulgaria, Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland,
Romania, Slovenia, and Slovakia.

Blackwell Publishers Ltd 2002

300 PETER EGGER

Gijt logGDPit GDPjt

 2  2 !
GDPit GDPjt
Sijt log 1 3
GDPit GDPjt GDPit GDPjt
    
 GDPjt 
 GDPit
R ijt log log 
 Nit Njt 

where N denotes a countrys population and GDP per capita is as commonly used
as a proxy for a countrys capital-labour ratio. Moreover, I use four variables,
which reflect a countrys freedom with respect to international exchange and
therefore transport costs in a broad sense.4 These are exporter (importer) viability

of contracts Vit ; Vjt and exporter (importer) rule of law RL it ; RL jt .
They influence an exporters (importers) costs of international exchange, since a
higher level of contract viability reduces a firms risk (i.e. of bankruptcy) and a
higher level of rule of law also reduces the probability of losing money, since
contract breakers have to take their responsibility and are more likely to lose their
case. Additionally, bilateral trade is affected by more traditional measures of
transport costs, which are commonly in use. These are the real bilateral exchange

rate Eijt , distance Dij , common borders Bij and common language

Lij , where the latter two are dummy variables. 5

For the panel econometric projection of potential bilateral trade, researchers

have concentrated on the random effects model (REM), which requires that ij 
0; 2 , vijt  0; 2v , and the ij are independent of the vijt . Moreover, the Xijt
(i.e. the explanatory variables) have to be independent of the ij and vijt for all
cross-sections ij and time periods t. Whereas the fixed effects model (FEM) is
always consistent in the absence of endogeneity or errors in variables, the REM is
only consistent if the above-mentioned orthogonality conditions are fulfilled.
Then, the REM has the advantage of more efficiency as compared to the FEM. If
these conditions do not hold, only the FEM is consistent since it wipes out all the
time-invariant effects ij . The decision between FEM and REM can be based
on a Hausman (1978) test. However, in the FEM time-variant variables cannot be
estimated any longer and it wastes a lot of degrees of freedom, since the ij may
be correlated only with a few explanatory variables. Therefore, Hausman and
Taylor (1981) provide an alternative which makes use of the several dimensions
of panel data in order to overcome this correlation without any variables from
outside the model. The appropriateness of the latter can be based on a Hausman
and Taylor test for over-identifying restrictions. Finally, the mentioned models
assume that there is no serial correlation of the error term vijt and the only
4
See the next section for more details on data sources.
5
Note that the econometric arguments below are fully independent of the underlying theoretical
context and also hold for the traditional set-up a` la Linnemann (1966).

Blackwell Publishers Ltd 2002

 GRAVITY MODELS AND TRADE POTENTIALS 301

correlation over time is due to equicorrelation (i.e. the presence of the same
individuals over time).6 If vijt follows an autoregressive process and this is
ignored, it results in consistent but inefficient parameter estimates and standard
errors also rendering the Hausman (1978) and Hausman and Taylor (1981) tests
inappropriate, since they require to use the efficient estimator under the null.7
When trade potentials are projected, researchers usually focus on residuals
rather than parameters. This paper demonstrates, that the choice of the
econometric set-up is of great relevance for the calculation of bilateral trade
potentials (especially, in the in-sample prediction approach). I estimate the
following models: an FEM, an REM, the corresponding Hausman and Taylor
model (HTM), and a Between model all assuming no autocorrelation of the error
term. According to Pirotte (1999) the consistent FEM (and consequently also the
REM or the HTM) can be associated with short-term parameter estimates,
whereas the Between estimator gives parameter estimates, which reflect the long
run. Since I find autocorrelation of the residuals, an REM and HTM for the case
of first-order autocorrelation (AR(1)) are estimates in addition (see the Appendix
for a derivation of the latter).

3. DATA, ESTIMATION RESULTS AND PROJECTIONS

I estimate a panel of exports of OECD countries to other OECD members and

the 10 Central and Eastern European countries over the period 19861997. I use
real exports, GDP, and exchange rates with 1995 as the base year. Nominal
exports in current USD are from OECD (Monthly Statistics of International
Trade), IMF (Direction of Foreign Trade), and the Vienna Institute of
Comparative Economic Studies (hereafter WIIW). Nominal GDP in USD and
GDP deflators are from OECD (Economic Outlook and National Accounts
Volume 1), IMF (International Financial Statistics), and WIIW. Exchange rate
indices are collected from IMF (International Financial Statistics) and WIIW.
Population numbers are available from OECD (Economic Outlook and National
Accounts Volume 1), IMF (International Financial Statistics) and WIIW. The
economic freedom variables are provided by Economic Freedom Network
(Economic Freedom of the World) and account for legal structure and property
rights (Area V of the database) and international exchange (part of Area VI of the
respective database). Transport costs in a narrow sense are proxied by distance
between two countries capitals in miles.
6
Baltagi (1995) and Matyas (1996) provide an overview of the literature on autocorrelation in
panel data.
7
Noteworthy, if one finds systematic country or country-pair specific differences between
observed and predicted trade flows this automatically implies the presence of autocorrelation in the
residuals.

Blackwell Publishers Ltd 2002

 TABLE 1
Panel Regression Results for Bilateral OECD Exports in OECD and 10 Central and Eastern European Countries

Independent Variables1 Fixed Effects Random Effects Hausman-Taylor Between First-Order Autocorrelation AR(1)

Random Effects Hausman-Taylor

Bilateral Sum of GDP 2.470*** 1.817*** 2.454*** 1.757*** 1.815*** 2.624***

(0.114) (0.027) (0.109) (0.030) (0.027) (0.154)
Similarity in Country Size 0.504*** 0.853*** 0.498*** 0.918*** 0.879*** 0.556***
(0.087) (0.031) (0.084) (0.034) (0.031) (0.109)
Difference in Relative 0.377*** 0.055* 0.383*** 0.606*** 0.138*** 0.338***
Factor Endowments (0.062) (0.031) (0.059) (0.129) (0.033) (0.077)
Exporter Viability of Contracts 0.039 0.290*** 0.028 2.499*** 0.331*** 0.059
(0.045) (0.043) (0.050) (0.297) (0.053) (0.057)
Importer Viability of Contracts 0.684*** 0.653*** 0.688*** 0.186 0.548*** 0.615***
(0.038) (0.038) (0.037) (0.255) (0.043) (0.044)
Exporter Rule of Law 0.091** 0.081* 0.001 0.242 0.104** 0.134**
(0.045) (0.043) (0.025) (0.230) (0.051) (0.053)
Blackwell Publishers Ltd 2002

Importer Rule of Law 0.175*** 0.152*** 0.176*** 0.236* 0.098*** 0.107***

(0.028) (0.028) (0.027) (0.140) (0.030) (0.032)
Real Exchange Rate 0.023*** 0.005 0.025*** 0.263*** 0.003 0.027***
(0.006) (0.006) (0.006) (0.091) (0.007) (0.008)
Distance 0.915*** 0.783 0.178*** 0.914*** 0.002
(0.027) (17.079) (0.046) (0.027) (0.008)
 Blackwell Publishers Ltd 2002 Common Border 0.308** 17.375 0.931*** 0.290** 0.911**
(0.123) (119.945) (0.029) (0.122) (0.383)
Common Language 0.731*** 0.367 0.291** 0.698*** 0.474
(0.130) (66.611) (0.121) (0.129) (0.298)
Constant 49.091*** 23.939*** 55.999 28.217*** 23.488*** 55.994***
(3.063) (0.723) (134.058) (1.964) (0.735) (5.353)

Number of Observations 9375 9375 9375 9375 9375 9375

Number of Bilateral Relations 837 837 837 837 837 837
R2 0.9994 0.9966 0.9992 0.9988 0.9978 0.9996
Average i 0.857 0.997 0.772 0.984
Time Effects: F(11, 8520) 5.39*** 92.91*** 6.01*** 5.92*** 82.71*** 2.86***
Bilateral Effects: F(836, 8520) 49.20***
Hausman Test: 2 (18) 328.17*** 5213.36***
Honda Lagrange Multiplier Test2 170.53***
Hausman Over-identification Test:
2 (3) 0.59 0.88
Canonical Correlations3 0.30 0.37
Durbin-Watson4 0.82 0.85
Baltagi-Wu LBI5 1.11 1.13
Notes:
1
Standard errors in parantheses.
2
One-sided test statistic based on square root figures of the traditional Breusch-Pagan test statistics; asymptotically following a standard normal distribution.
3
Geometric mean of canonical correlation coefficients.
4
According to Bhargava et al. (1982).
5
Locally best invariant test statistic according to Baltagi and Wu (1999).
*** Significant at 1 per cent.
** Significant at 5 per cent.
* Significant at 10 per cent.
304 PETER EGGER

Table 1 presents the estimation results for six different panel estimators. Note
that the Between model should reflect long-term influences. All other estimators
reflect short-run impacts if the parameters can be consistently estimated (see
Pirotte, 1999). According to the test statistics we should not ignore cycle effects
(F-tests for time effects) and the presence of heteroscedasticity in the cross-
section (F-test and Honda test).8 However, the Hausman test statistic reveals that
the REM suffers from correlation and gives inconsistent parameter estimates
(which are closer to the Between estimates). The HTM seems most appropriate as
is indicated by the over-identification test. It comes close to the FEM in terms of
parameter estimates and it seems efficient. The time-invariant variables cannot be
estimated significantly, which is due to the explanatory variables at hand. Testing
for autocorrelation reveals that the HTM is not efficient and both the Hausman
test statistic and the Hausman and Taylor test for over-identification are
inappropriate. Due to large differences in the parameters between the FEM and
the REM, the Hausman test also rejects the appropriateness of the REM AR(1). In
our application, the HTM AR(1) is consistent and efficient among the short-term
estimators. It additionally allows us to estimate the impact of common borders
with more success.
In order to underpin the relevance of the model choice for the in-sample
projection of bilateral trade potentials, I calculate trade potentials of EU15
countries and three Central and Eastern European economies (CEEC3) in the
CEEC10. Usually, the exponent of minus one times the bilateral residual is
interpreted as the bilateral trade potential and, for the moment, it is interpreted in
the traditional manner.
Obviously, in terms of actual-to-potential trade ratios the REM lies in between
the HTM and the Between model. As compared to the HTM, the REM
overestimates the trade potential of the CEEC3 in the CEEC10 by more than 210
per cent. In contrast, it underestimates the potential of the EU15 by about six per
cent. This is even more pronounced if autocorrelation is accounted for. As
compared to the HTM AR(1) approach, the REM AR(1) overestimates the EU15
trade potential in the CEEC10 by about 49 per cent and that of the CEEC3 by
about 367 per cent. According to the HTM AR(1) model there is no export

8
The latter reveals the inappropriateness of pure OLS analysis. Hence, one should specify the
problem as a two-way panel also when estimating a time-averaged specification. If one finds
which is very likely the case correlation among country-specific effects and exogenous variables,
there are two possibilities: When estimating a new trade theory model a` la Helpman (1987), all
important explanatory variables vary in the bilateral rather than the country-specific dimension.
Then, the model could be estimated with fixed exporter and importer effects. If a specification a` la
Linnemann (1966) is used, the coefficients for country-specific factors like GDP per capita can no
longer be estimated. In contrast, distance, common borders, and common language can be
estimated. Then, the proper model would be a two-way HTM in order to obtain properly estimated
coefficients for GDP and GDP per capita. This model could provide consistent estimates of the
GDP and GDP per capita coefficients in contrast to usually applied OLS.

Blackwell Publishers Ltd 2002

 Blackwell Publishers Ltd 2002

TABLE 2
Actual-to-Potential Export Ratio in 10 Central and Eastern European Countries

Exporting Country Random Effects Hausman-Taylor Between First-Order Autocorrelation AR(1)

Random Effects Hausman-Taylor

Belgium-Luxembourg 0.61 0.88 0.74 0.59 0.88

Denmark 0.94 0.88 1.15 0.89 0.88
Germany 0.63 0.95 0.63 0.56 1.22
Finland 0.59 0.83 0.87 0.58 0.85
France 1.44 0.90 1.43 1.33 0.90
Greece 0.75 0.79 0.55 0.77 0.98
Great Britain 1.14 0.91 1.34 1.09 0.92
Ireland 0.73 0.83 1.19 0.74 0.85
Italy 0.85 0.83 1.03 0.82 0.84
Netherlands 0.57 0.93 0.64 0.54 0.91
Austria 0.73 0.96 0.90 0.68 1.27
Portugal 2.59 0.75 4.14 2.71 0.78
Sweden 0.72 0.89 0.94 0.69 0.88
Spain 1.40 0.75 1.82 1.39 0.79

EU15 0.98 0.92 1.24 0.72 1.07

Hungary 0.33 0.98 0.34 0.36 1.14

Poland 0.45 0.95 0.34 0.48 1.15
Czech Republic 0.12 1.04 0.13 0.13 1.37

CEEC3 0.32 1.00 0.29 0.27 1.26

306 PETER EGGER

potential to CEEC10 left for both the EU15 and the three CEEC. The EU15
export potential to CEEC10 is about 34 per cent (16 per cent) higher in the HTM
(the HTM AR(1)) approach than in the Between set-up, where the former can be
associated with the short-run and the latter with the long-run. The opposite holds
true for export potentials of the CEEC3, which are 71 per cent (76 per cent for
AR(1)) smaller in the short-run than in the long-run.
Unfortunately, from a pure econometric point of view such in-sample
prediction experiments of thought are inappropriate, since proper specification
should always result in white-noise residuals, which is the case for the HTM
AR(1). This reveals the difficulty of the in-sample prediction approach to trade
potentials as it is commonly in use. I propose to interpret in-sample evidence for a
large overlap of predicted over observed trade relations as an indication of
misspecification rather than of unused trade potentials. In the above example the
REM, the REM AR(1), and the Between model exhibit non-white-noise residuals
indicating specification and consistency/efficiency problems.
The above arguments raise the question about a possible guide for practitioners
on how to do it right? and how to proceed in the future. I think several important
conclusions can be drawn. First and independently of the researchers interest,
consistent estimation is a must, and I think that we should not base our
conclusions on simple OLS estimates. In the gravity cross-section case, this
requires a two-way fixed or random exporter and importer effects panel data
model (compare Egger, 2001). If the data vary over time, a two-way model with
(fixed) time and (fixed or random) bilateral effects is appropriate. If
autocorrelation is present, one should control for it. Whether exporter, importer
or bilateral effects should be treated as random or fixed depends on the
consistency of the REM and on whether one is interested in the estimation of
time-invariant effects (see also Egger, 2000). In general, I would recommend the
HTM or the HTM AR(1). If this is not consistent (which is testable by the
mentioned over-identification test), the FEM is the only valid alternative. Second,
the in-sample approach to the prediction of trade potentials is inappropriate. If the
underlying model is consistent, there is no systematic difference between
observed and in-sample predicted trade flows. Third, whether out-of-sample
predictions make sense depends on the researchers suggestion about the stage of
the countries transformation process. For the CEEC, out-of-sample predictions
of EU-CEEC trade potentials probably become less and less appropriate
(regarding the development e.g. in Hungary or Slovenia). This approach is most
appropriate in the early stage of transformation. Fourth, the calculation of
counterfactuals remains a useful tool. Interesting questions are influence of
changes in the explanatory variables on bilateral trade flows (e.g. catching up in
GDP per capita in the CEEC, the reduction of trade costs via infrastructure
investments or tariffs, etc.). For this line of research, the difference between
short-term and long-term influences could provide interesting insights.

Blackwell Publishers Ltd 2002

 GRAVITY MODELS AND TRADE POTENTIALS 307

4. CONCLUSIONS

This paper provides insights into the relevance of the appropriate estimator
choice for the analysis of bilateral trade flows. I compare different estimators part
of which have been used in previous studies and provide insights that none of the
previously used estimators is appropriate in terms of consistency and/or
efficiency in my application. I argue that three problems should be kept in mind
when estimating gravity models and/or calculating trade potentials. First,
traditionally estimated time-averaged cross-section gravity models are very
likely to be misspecified since they ignore the presence of exporter and importer
effects without testing for their relevance. Second, one should be careful with
comparing estimation results between different econometric concepts, which
refer to different time horizons with respect to responses of trade flows on
changes in the explanatory variables. Third and in contrast to previous research, I
do not see any way to derive information about so-called trade potentials of the
in-sample prediction approach. Rather, I suggest that any large systematic
difference between observed and in-sample predicted trade flows indicates
misspecification of the econometric model instead of unused (or overused) trade
potentials.
In the present application, the consistent and efficient model is a Hausman and
Taylor AR(1) estimator, which has never been used before. According to
econometric theory, which demands for white-noise residuals in the case of
proper (i.e. consistent and efficient) specification, this estimator fails to identify
large systematic differences between residuals among country groups. Large
unused in-sample export potentials, which have been identified previously in the
context of European integration reveal nothing other than inherent problems of
misspecification in terms of consistency and efficiency of the estimators and the
econometric models in use. Nevertheless, the gravity model remains a useful tool
for counterfactual simulation analysis. Future applications could focus on
simulations of changes in the explanatory variables such as convergence in terms
of GDP per capita or trade cost reductions.

APPENDIX

A Hausman and Taylor AR(1) Model

Following Baltagi and Wu (1999) in the notation, we have to Prais-Winsten
transform the data by:9

9
Note that in contrast to Baltagi and Li (1991), Baltagi and Wu (1999) allow for unequally spaced
panel data and missing observations, which is relevant in my application.

Blackwell Publishers Ltd 2002

 2 3
1 0 ... 0 0
6  2ti;2 ti;1 1=2  1=2 7
6  1 7
61 ... 0 0 7
6 1 2ti;2 ti;1 1 2ti;2 ti;1 7
6 7
6 7
6 .. .. .. .. .. 7
6
2 1=2 6 . . . . . 7

Ci  1 
6 7 4
 1=2 7
6 1 7
6 ... 7
6 0 0 0 7
6 1 2ti;ni 1 ti;ni 2 7
6 7
6  2ti;n ti;n 1 1=2  1=2 7
4  i i 1 5
0 0 ...
1 2ti;ni ti;ni 1 1 2ti;ni ti;ni 1
Blackwell Publishers Ltd 2002
 GRAVITY MODELS AND TRADE POTENTIALS 309

Accordingly, one can transform the Amemiya Within-type residuals from the
2" .
initial FEM to obtain the variance component of the remainder disturbance ^
Therefore, we have to define:
u diagCi u diagCi diagni  diagCi v; 5

!0
1 ti;2 ti;1 1 ti;ni ti;ni 1
gi Ci ni 1  2 1=2
1; ;...;
1 2ti;2 ti;1 1=2 1 2ti;ni ti;ni 1 1=2
6
0 1 0
and Pgi gi gi gi gi ; Qgi Ini Pgi in order to obtain:
0
u diagQgi u
^2" N ; 7
X
ni 1
i1

where N refers to the number of cross-sections and ni is the number of

observations in cross-section i. This corresponds to "0 ", where " is the residual
vector from the OLS regression on the Within transformed model where each
variable corresponds to:
X
ni
gi;s yi;ti;s

ywi;ti; j yi;ti; j gi; j s1
: 8
Xni
2
gi;s
s1

In the presence of correlation between (some of) the explanatory variables X

and the unobserved effects ij we have to average the Within residuals over
time (i.e. to construct pseudo-averages) and to run 2SLS of these residuals on the
time-invariant, Prais-Winsten transformed variables with the exogenous time-
variant variables as instruments.10 This regression not only obtains a parameter
estimate for the time-invariant variables, but it also produces residuals, which are
used to derive the second required variance component. I call the residuals from
this second regression  . We can obtain an estimate of the second required
variance component by:
0
^2!  diagPgi  : 9

10
In contrast to Hausman and Taylor (1981) we consider all time-invariant variables as correlated
with the ij . According to Cornwell et al. (1992), we call the correlated variables as singly
exogenous and the uncorrelated ones as doubly exogenous.

Blackwell Publishers Ltd 2002

310 PETER EGGER

Accordingly, one can derive an estimate for the cross-sectional variance

component via:
0
 diagPgi  N ^2"
^2 ; 10
XN
0
gi gi
i1

which gives:
0
^2i gi gi ^2 ^2"
! 11

and
 2 1=2
^"
^i 1 : 12
!^2i
Finally, we can transform our data according to Fuller and Battese (1973 and
1974) by premultiplying the Prais-Winsten transformed data by ^"
1=2 to get
y ^"
1=2 y with the typical elements
X
ni
gi;s yi;ti;s
y  ^ s1
i;ti; j yi;ti; j i gi; j Xni : 13
2
gi;s
s1

Running 2SLS on the transformed model with the proper set of instruments A
yields the consistent and efficient AR(1) estimator in the spirit of Hausman and
Taylor (1981). A consists of the Within transformed time-variant variables
(according to (8)) and of pseudo-averages over time of the doubly exogenous,
time-variant variables (in our case Sijt , R ijt , Vit , Vj;t , RL it , RL jt , and Eijt ). The latter
are derived from the transformation:
X
ni
gi;s yi;ti;s
yi;ti; j gi; j s1
: 14
Xni
2
gi;s
s1

In our application, the set of time-invariant, singly exogenous variables comprises

Dij , Bij , and Lij . Since we have more time-variant, doubly exogenous variables
than time-invariant, singly exogenous variables at hand, the AR(1) Hausman and
Taylor-type estimator is efficient and the corresponding AR(1) Within (FEM)
estimator is not.

Blackwell Publishers Ltd 2002

 GRAVITY MODELS AND TRADE POTENTIALS 311

REFERENCES

Anderson, J.E. (1979), A Theoretical Foundation of the Gravity Equation, American Economic
Review, 69, 1, 10616.
Baldwin, R. (1994), Towards an Integrated Europe (London: Centre for Economic Policy
Research).
Baltagi, B.H. (1995), Econometric Analysis of Panel Data (Wiley, Chichester).
Baltagi, B.H. and Q. Li (1991), A Transformation That Will Circumvent the Problem of
Autocorrelation in an Error Component Model, Journal of Econometrics, 48, 3, 38593.
Baltagi, B.H. and P.X. Wu (1999), Unequally Spaced Panel Data Regressions with AR(1)
Disturbances, Econometric Theory, 15, 6, 81423.
Bergstrand, J.H. (1985), The Gravity Equation in International Trade: Some Microeconomic
Foundations and Empirical Evidence, Review of Economics and Statistics, 67, 3, 47481.
Bergstrand, J.H. (1989), The Generalized Gravity Equation, Monopolistic Competition, and
Factor Proportions Theory in International Trade, Review of Economics and Statistics,
71, 1, 14353.
Bergstrand, J.H. (1990), The Heckscher-Ohlin-Samuelson Model, the Linder Hypothesis and the
Determinants of Bilateral Intra-industry Trade, Economic Journal, 100, 403, 121629.
Bhargava, A., L. Franzini and W. Narendranathan (1982), Serial Correlation and the Fixed Effects
Model, Review of Economic Studies, 49, 4, 53349.
Brulhart, M. and M.J. Kelly (1999), Irelands Trading Potential with Central and Eastern European
Countries: A Gravity Study, Economic and Social Review, 30, 2, 15974.
Cornwell, C., P. Schmidt and D. Wyhowski (1992), Simultaneous Equations and Panel Data,
Journal of Econometrics, 51, 12, 15181.
Dixit, A. and J.E. Stiglitz (1977), Monopolistic Competition and Optimum Product Diversity,
American Economic Review, 67, 3, 297308.
Egger, P. (2000), A Note on the Proper Econometric Specification of the Gravity Equation,
Economics Letters, 66, 1, 2531.
Egger, P. (2001), Proper Estimation of Cross-Section Gravity Models: A Note, mimeo (Vienna:
Austrian Institute of Economic Research), [Link]
Fuller, W.A. and G.E. Battese (1973), Transformations for Estimation of Linear Models with
Nested Error Structure, Journal of the American Statistical Association, 68, 343, 62632.
Fuller, W.A. and G.E. Battese (1974), Estimation of Linear Models with Crossed-Error Structure,
Journal of Econometrics, 2, 1, 6778.
Gros, D. and A. Gonciarz (1996), A Note on the Trade Potential of Central and Eastern Europe,
European Journal of Political Economy, 12, 4, 70921.
Hamilton, C.B., and A.L. Winters (1992), Opening up International Trade with Eastern Europe,
Economic Policy, 14, 77116.
Hausman, J.A. (1978), Specification Tests in Econometrics, Econometrica, 46, 6, 125171.
Hausman, J.A. and W.E. Taylor (1981), Panel Data and Unobservable Individual Effects,
Econometrica, 49, 6, 137798.
Helpman, E. (1987), Imperfect Competition and International Trade: Evidence from Fourteen
Industrial Countries, Journal of the Japanese and International Economies, 2, 1, 6281.
Helpman, E. and P.R. Krugman (1985), Market Structure and Foreign Trade (Cambridge, Mass.:
MIT Press).
Honda, Y. (1985), Testing the Error Components Model with Non-normal Disturbances, Review
of Economic Studies, 52, 4, 68190.
Matyas, L. (1996), Error Components Models, in L. Matyas and P. Sevestre (eds.), The
Econometrics of Panel Data. A Handbook of the Theory with Applications (Dordrecht: Kluwer
Academic Publishers), 5076.
Matyas, L. (1997), Proper Econometric Specification of the Gravity Model, The World Economy,
20, 3, 36368.
Matyas, L. (1998), The Gravity Model: Some Econometric Considerations, The World Economy,
21, 3, 397401.

Blackwell Publishers Ltd 2002

312 PETER EGGER

Nilsson, L. (2000), Trade Integration and the EU Economic Membership Criteria, European
Journal of Political Economy, 16, 4, 80727.
Pirotte, A. (1999), Convergence of the Static Estimation Toward the Long Run Effects of Dynamic
Panel Data Models, Economics Letters, 63, 2, 15158.
Wang, Z.K. and A.L. Winters (1991), The Trading Potential of Eastern Europe, Discussion Paper
No. 610 (London: CEPR).

Blackwell Publishers Ltd 2002