The Euro Effect on Trade

is not as Large as Commonly Thought

Maurice J.G. Bun

University of Amsterdam and Tinbergen Institute
and
Franc J.G.M. Klaassen∗
University of Amsterdam and Tinbergen Institute

Forthcoming in the Oxford Bulletin of Economics and Statistics

February 21, 2006

Abstract
Existing studies on the impact of the euro on goods trade report increments
between 5% and 40%. These estimates are based on standard panel gravity
models for the level of trade. We show that the residuals from these models
exhibit upwards trends over time for the euro countries, and that this leads
to an upward bias in the estimated euro effect. To correct for that, we extend
the standard model by including a time trend that may have different effects
across country-pairs. This results in an estimated euro impact of only 3%.

Key words: currency union, dynamic OLS, EMU, gravity model, panel data,
robust standard errors, time trend.
JEL classification: C23; F15; F33.

∗
Corresponding author. Address: Department of Economics, Roetersstraat 11, 1018 WB Amster-
dam, The Netherlands; tel. +31-20-5254191; fax +31-20-5254254; e-mail: [email protected].
We would like to thank Richard Baldwin, Peter Boswijk, Harry Flam, Henk Jager, Jan Magnus, Michel
van de Velden and an anonymous referee for very useful and constructive comments and Andrew Rose
for his valuable comments and for making the Glick and Rose (2002) dataset available via the internet.
An earlier version has appeared as Tinbergen Institute Discussion Paper 2003-086/2 under the title of
“The Importance of Accounting for Time Trends when Estimating the Euro Effect on Trade”.
1 Introduction
Since the introduction of the euro in 1999, several studies have estimated the impact
of the euro on bilateral goods trade within the euro zone. Micco, Stein and Ordoñez
(2003) estimate an increase between 5 and 20 percent, Flam and Nordström (2003)
report estimates between 8% and 15%, Barr, Breedon and Miles (2003) calculate 29%
and Bun and Klaassen (2002) report an increase of 38% for the trade effect of the euro.1
This suggests that the euro effect is positive and lies in the range of 5% to 40%.
These estimates are useful for evaluating the benefits of the euro for existing euro
zone countries (European Commission, 2003). They have also affected the debate in
non-euro countries on whether to join the euro (HM Treasury, 2003, for the U.K.). Also
for the new European Union (EU) members the potential trade benefits of adopting
the euro are relevant.
Because of this policy relevance, it is important to verify the robustness of the
current euro estimates: do they really represent the impact of the euro, or are they
driven by something else? The models that are typically used are all quite similar in
the sense that they essentially explain trade by income and a euro dummy (which is
one if the countries involved have the euro), while correcting for some other factors. It
might be that some variables omitted from this common modelling approach have led
to bias in all estimated trade benefits given earlier.
An indication of such omitted variables bias follows from the variation in the esti-
mates given earlier in relation to the number of time periods in the sample. Micco et al.
(2003) and Flam and Nordström (2003) find the lowest estimates using data over about
1992-2002, Barr et al. (2003) derive the middle estimate from data over 1978-2002, and
Bun and Klaassen (2002) report the largest one from data over 1965-2001. Therefore,
the longer the data period, the higher the euro estimate. This is difficult to explain
from an economic point of view. Because in longer samples the time series character-
istics of the variables involved usually have more impact, we suspect the euro estimate
to be biased by some misspecification of the time series characteristics of trade.
In this study we investigate a particularly important time series characteristic,
namely the trends in trade flows over time. It is well known from the time series litera-
ture that trend misspecification can lead to substantial bias. We examine whether euro
estimates are biased upwards, because the euro dummy (which is one only at the end
of the sample) picks up increasing trends in trade that are actually caused by omitted
variables.
1
More precisely, by the euro effect on trade we mean the trade effect of entering stage three of the
Economic and Monetary Union (EMU).

2
 To some extent omitted trending variables bias is already avoided in the studies
mentioned earlier. After all, they include an income regressor and other trending
variables to explain the trend in trade. Moreover, they use panel data so that they
can include time effects to correct for any residual trend common to all bilateral trade
flows.
However, trending behavior of trade flows may also be affected by variables not
included in the specification and trends may vary across country-pairs, both due to
country-specific and country-pair specific factors. Examples of such factors can be de-
rived from the generalized gravity equation of Bergstrand (1989), which is particularly
interesting because the models in the existing euro studies are closely related to this
model. Although Bergstrand’s model is a one-period model, let us imagine that it holds
several periods after each other. Bergstrand derives that trade depends on productiv-
ity parameters, capital/labor ratios, transport costs, tariffs, among other things. The
first two factors are nation specific, and they tend to increase over time, leading to
country-specific trends in trade. Transport costs depend on distance and the goods
composition of trade, which are both different across country-pairs. Because transport
costs have decreased over time, the transport cost term in the gravity model is one
source of country-pair specific trend growth in trade. The tariff term in the model
is another source, because trade liberalization usually occurs gradually and at differ-
ent speeds across country-pairs. Because such variables, and potentially many others,
such as telecommunications costs and the trade costs mentioned by Anderson and van
Wincoop (2004), are not included in existing models used to estimate the euro impact,
it is unlikely that the standard trend corrections mentioned earlier are sufficient to
completely avoid omitted trending variables bias.
To correct for this bias, one could include proxies for the variables just mentioned.
However, these may be difficult to construct, and it is unlikely that one can find proxies
to capture all omitted trending variables. The way we correct for this is based on the
fact that a major part of their signal is the deterministic time trend, or drift term.
Hence, in our panel model we add the time variable t and, to account for both country
and country-pair trending variables, we allow it to have heterogeneous coefficients across
country pairs. This extension is novel in the gravity literature, but has already been
used elsewhere (Cornwell, Schmidt and Sickles, 1990, and Mark and Sul, 2003, among
others).
The set up of the paper is as follows. In the next section we discuss an empirical
panel gravity model including country-pair specific time trends. In Section 3 the esti-
mation results are presented, both for a dataset involving euro data (to study whether

3
existing euro estimates are biased upwards) and for the Glick and Rose (2002) sample,
which has data on non-euro currency unions (to study the robustness of the typical find-
ing that such currency unions increase bilateral trade a lot, namely by about 90%). In
Section 4 we examine the robustness of the main findings. The final section concludes.

2 A gravity model with country-pair specific time trends

2.1 Model and estimation

The panel gravity model occurs in several variants. Rose (2000) explains bilateral
trade (exports plus imports) by national income of both countries, their incomes per
capita, free trade area and currency union dummies, and time-invariant variables such
as distance. Glick and Rose (2002) extend the model by using fixed country-pair specific
intercepts to correct for all time-invariant trade determinants, and in a robustness check
they also include fixed time effects to account for all country-pair invariant variables.
Micco et al. (2003) and Barr et al. (2003) use a similar model to examine the currency
union in Europe.
The model of Bun and Klaassen (2002), also used by Flam and Nordström (2003),
is somewhat different from that of Glick and Rose (2002), as it takes exports instead
of trade as dependent variable. Moreover, in contrast to the other studies, Bun and
Klaassen (2002) account for the dynamics in trade data by including lagged dependent
and explanatory variables besides contemporaneous values.
These latter two differences are not important for the point we want to make in
the current study, because our conclusions turn out to be essentially the same for
models with trade or exports as dependent variable and for static or dynamic panel
data models (see Section 4.1). For the sake of comparison, the benchmark model in
the present study is therefore based on the static panel gravity model for trade used
by Glick and Rose (2002). We now describe that model in more detail and generalize
it by introducing the country-pair specific time trends.
The dependent variable is T RADEijt , the logarithm of real bilateral trade between
countries i and j in year t, where real bilateral trade is the sum of nominal bilateral
exports and imports, both in U.S. dollars, divided by the U.S. producer price index.2
The first explanatory variable is the log of the product of the countries’ real GDP, both
expressed in U.S. output; it is denoted by GDPijt . We also include the log of GDP per

2
This is the common approach in the currency union literature. This focus on the U.S. is not
important for our main conclusion, because Section 4.1 shows that using real exports measured in
exporter’s output as dependent variable leads to a similar conclusion.

4
capita (GDP CAPijt ), which is GDPijt minus the log of the product of the countries’
population sizes. To measure the euro effect, we follow existing studies by including
EU ROijt , a dummy that is one if i and j have the euro in year t (hence it can only
be one from 1999 onwards). Thus, we model the euro impact as a permanent break in
the level of trade for the euro country pairs (in Section 4.1 we show that our results
are robust to somewhat more sophisticated approaches). To correct for trade increases
from free trade area arrangements, we include a dummy F T Aijt that is one in case
the countries have free trade with each other. Finally, we account for the effects of all
possible time-invariant determinants of trade (such as distance) by a fixed “individual”
effect η ij for country-pair ij, and we use a fixed time effect λt to correct for the impact
of all possible country-pair invariant trade determinants (such as the state of the world
economy). These are all quite standard elements and definitions in panel gravity models
nowadays, and we follow these choices to ensure that our results can be easily compared
to the existing literature.
Apart from the aforementioned set of regressors, there may be many other trade
determinants. A particularly important group of variables may be the group of trending
trade determinants other than GDPijt and GDP CAPijt , because trends are strong
signals so that leaving them out of the model may have substantial effects on the results.
A subset of all trending determinants, the country-pair invariant ones, are accounted
for by the time effects λt . To extend the standard model, we thus concentrate on
the country and country-pair specific ones, such as factor productivity, capital/labor
ratios, transportation costs and tariffs. To simplify terminology, we will refer to all
such variables as “country-pair” variables.
To approximate the impact of all country-pair specific omitted trending variables,
we focus on one of their main characteristics, the time trend t, as motivated in the
introduction. We therefore ignore other potentially relevant characteristics, such as
stochastic trends. Incorporating such refinements would go beyond the purpose of this
study. The country-pair dependence of the trend effects is represented by τ ij . These
effects are considered to be fixed (instead of random), just as η ij and λt .
This results in

T RADEijt = β 1 GDPijt + β 2 GDP CAPijt + δ 1 EU ROijt + δ 2 F T Aijt

+η ij + τ ij · t + λt + εijt , (1)

where εijt is allowed to be heteroskedastic (across country-pairs and time), serially

correlated and cross-sectionally correlated (both contemporaneous and lagged). We
assume that εijt is stationary and treat all regressors as strictly exogenous with respect

5
to εijt (Sections 4.3 and 4.4 show that these assumptions are not important for the
main conclusion of our study).
The parameter of interest is δ 1 , which represents the impact of the euro on trade
between euro member states. The difference in their trade before and after the intro-
duction of the euro is used to identify δ 1 .
We estimate model (1) by least-squares after transforming away the nuisance effects
η ij , τ ij · t and λt . This is an LSDV (least-squares dummy variables) type approach.
Note that the standard within transformation to wipe out η ij , which subtracts country-
pair specific means over time from each variable, does not work here, because that will
not remove τ ij · t. To nevertheless wipe out τ ij · t, we use the fact that the within
transformation is actually a projection of all variables on the null-space of the matrix
of dummy variables corresponding to all η ij ; see Wansbeek and Kapteyn (1989). We
apply this projection argument to our model, so that we project all variables in model
(1) on the null-space of the matrix of dummy/time variables corresponding to all η ij ,
τ ij · t and λt . This transforms away all fixed effects.
To compute standard errors that are robust to heteroskedasticity as well as serial
and cross-sectional correlation, we follow Driscoll and Kraay (1998) in combination
with Newey and West (1987, 1994).3

2.2 Comparison with existing approaches

The model commonly used to estimate the euro effect is the special case of (1) where
τ ij = 0 for all country-pairs. If there happen to be no omitted trending regressors in
reality (the true value of τ ij is 0), the estimated δ 1 in the general model will be equal
to that of the standard model on average (although the standard error will be larger).
Hence, the fact that we leave τ ij unrestricted does not cause a bias of the estimated
euro effect.
Another difference with existing models concerns the moments of the innovations
εijt . One usually allows for arbitrary heteroskedasticity of εijt across country-pairs
and time. However, because of, for instance, entrance and exit barriers to trade due
to sunk costs and habit formation among consumers, past trade presumably has an
impact on current trade that is not captured by the regressors and effects in model (1);

3
In essence, the method takes the sample moment conditions on which the least-squares procedure
is based, averages them across country-pairs so that a single time series results, and computes a het-
eroskedasticity and autocorrelation consistent variance matrix for that series (we take the Newey and
West (1987) algorithm with the Newey and West (1994) optimal lag selection rule, which results in one
lag). This gives a robust estimate of the long run variance of the moment conditions. As usual, pre-
and postmultiplication by the sample second moment of the regressor vector (Hessian) then gives the
total estimated limiting variance, which delivers the standard errors.

6
see Bun and Klaassen (2002) for empirical evidence. Hence, εijt is probably serially
correlated. Moreover, εijt may be cross-sectionally correlated, because regional trade
shocks affect several trade flows jointly and nation-specific shocks potentially affect
trade flows with all trading partners, for example. We therefore allow for both serial
correlation and cross-sectional correlation of εijt in addition to the usual correction for
heteroskedasticity.
Finally, model (1) is related to Baltagi, Egger and Pfaffermayr (2003). In a model for
exports from country i to j, they add fixed effects indexed by it and jt, say, ξ it + μjt , so
that each country has a separate parameter for each time period when it is an exporter
and another set of parameters when it is an importer. Because we have summed exports
and imports into trade, μjt cannot be distinguished from ξ jt , so their approach in our
context means having ξ it + ξ jt . This is very flexible in the it and jt dimensions of
the panel, because the effects correct for all possible nation-specific variables (such
as institutional characteristics, factor endowments, government policy, and cultural
aspects) and these are allowed to move unrestrictedly over time. In the cross-sectional
(ij) dimension, however, our approach is more flexible, because it allows the trade
development over time to be driven by other than purely national factors, such as the
transportation cost and tariff variables mentioned in the introduction. Because we
want to study the effect of omitted trends, we want to account for such trends in the
ij dimension. As linear trends usually capture the major part of the time development
of trending variables, we thus prefer our full flexibility in the ij dimension at the cost
of imposing linearity for the trend instead of allowing for unrestricted time variation at
the cost of restricting the ij dimension. Our linearity assumption is supported by the
fact that the euro estimates remain essentially the same when allowing for quadratic
trends (see Section 4.1).

3 The importance of accounting for time trends for the

euro estimate
This section describes the data and then estimates model (1), both under τ ij = 0 (in
Section 3.2) and with τ ij unrestricted (in Section 3.3). By comparing both estimates
for the euro dummy EU ROijt we get an idea of the trend robustness of the estimated
euro effect, which is the main purpose of the study.

7
 Table 1: Numbers of observations across country pairs

Country pair Period Total

Pre-euro Euro

Two euro countries 1,780 200 1,980

Euro and non-euro countries 2,816 352 3,168

Two non-euro countries 896 112 1,008

Total 5,492 664 6,156

A country is named “euro country” if it has the euro in 2002, so

including Greece, which entered the euro zone in 2001.

3.1 Data

We have data on all bilateral combinations of 19 countries, namely all EU countries prior
to the May 2004 expansion, Norway, Switzerland, Canada, Japan and the U.S., where
Belgium and Luxembourg are taken together because trade data are only available at
the Belgium-Luxembourg Economic Union (BLEU) level. This gives N = 171 country-
pairs. We have annual data from 1967 through 2002 (including four years of the euro),
so that there are T = 36 time periods. The panel is balanced, so that we have 6,156
observations. Table 1 provides more details. It implies that identification of the euro
effect δ 1 is driven by the behavior of 1,780 versus 200 observations.
Data for T RADEijt come from the IMF Direction of Trade Statistics (DOTS)
in combination with the U.S. producer price index from the OECD Main Economic
Indicators. Data on GDPijt are from the OECD Economic Outlook. Population data
used to construct GDP CAPijt are from the U.S. Bureau of the Census website. The
F T Aijt dummy is based on the trade agreement chronology given in Bun and Klaassen
(2002).
For comparison, we also use two other samples. The first one is the 1992-2002
subset of the data just described. This approximates the Micco et al. (2003) dataset.
The main difference is that they also use data on Australia, Iceland and New Zealand,
but that is not expected to affect the results much. The second sample is the Glick and
Rose (2002) dataset. It is an unbalanced panel of N = 11, 178 country-pairs from 1948
through 1997, resulting in 219,558 observations. This sample includes many different
currency unions, mostly involving small and poor countries, but not the euro area.
Hence, the EU ROijt dummy in (1) is substituted by CUijt , which is one if the trading
partners have a currency union in year t. Although the focus of the study is on the

8
euro, the results from the Glick and Rose sample give some insight into the robustness
of our conclusions.

3.2 Estimation without country-pair specific time trends

The estimates for model (1) using the three samples are shown in Table 2. The columns
headed by “No trends” contain the results under the restriction τ ij = 0, so that they
are the estimates one would obtain using the standard panel gravity model. Note that
this restricted model still allows for some trend, as it includes time effects λt , but this
trend is restricted to be common to all country-pairs. The estimated euro and general
currency union effects are 0.41, 0.16 and 0.62, which are similar to the ones reported in
Bun and Klaassen (2002), Micco et al. (2003) and Glick and Rose (2002), respectively.4
Because T RADEijt is the logarithm of trade, these estimates correspond to a relative
change of trade itself of (exp(δ 1 ) − 1 =) 51%, 18% and 86%, respectively.
Table 2 reports two types of standard errors. The first one, in braces, represents
the common approach in the gravity literature of allowing for conditional and cross-
sectional heteroskedasticity. However, for reasons discussed in Section 2.2, the residuals
will presumably also exhibit serial correlation and are correlated across country pairs,
implying that the common standard errors are invalid. The second type of standard
errors, in brackets, is robust to heteroskedasticity as well as serial and cross-sectional
correlation.
The usefulness of the additional robustness is demonstrated by Table 2. The com-
mon standard errors turn out to be roughly three times smaller than the robust ones.
A more detailed analysis reveals that this is caused by both neglected serial and ne-
glected cross-sectional correlation. Nevertheless, even with robust standard errors, the
euro and currency union estimates in the model without heterogeneous trends are all
significant at the 5% level (the level we use throughout the study).
Instead of moving directly to the estimates of the unrestricted model, we first an-
alyze the standard model in more detail in the remaining part of this section. The
purpose is to obtain some preliminary insights into the relevance of our suggestion that
omitted upward trending trade determinants in combination with a euro dummy that
is only one at the end of the sample may lead to an upward bias in the estimated euro

4
The difference between 0.41 and the long-run estimate of 0.33 in our (2002) paper is caused by the
fact that that paper uses a model for exports instead of trade, takes account of dynamics, and has a
slightly smaller dataset.
Even though we use the data underlying the Glick and Rose (2002) paper, our 0.62 differs slightly from
their 0.59 (which they obtain when using year controls, see their Table 5). The reason is that we have
left out their current colony variable. This simplification does not alter the main pattern of results in
the present paper.

9
 Table 2: Estimation results for trade model (1)

OWN DATA GLICK-ROSE DATA

Whole period Micco et al. period Whole period

1967-2002 1992-2002 1948-1997

No trends Trends No trends Trends No trends Trends

EU ROijt /CUijt δ1 0.410 0.032 0.164 0.018 0.622 0.223

(currency union) {0.028} {0.014} {0.013} {0.013} {0.043} {0.055}
(0.075) (0.016) (0.032) (0.016) (0.079) (0.052)

F T Aijt δ2 0.41 0.06 — — 0.85 0.32

(free trade area) {0.02} {0.01} {0.03} {0.03}
(0.09) (0.03) (0.13) (0.08)

GDPijt β1 1.41 0.70 1.99 0.12 0.46 0.86

(product GDP) {0.10} {0.15} {0.31} {0.95} {0.02} {0.05}
(0.39) (0.36) (0.63) (0.77) (0.08) (0.10)

GDP CAPijt β2 −0.68 −0.23 −1.51 0.25 0.53 −0.13

(product GDP capita) {0.09} {0.15} {0.33} {0.96} {0.02} {0.05}
(0.37) (0.35) (0.69) (0.76) (0.10) (0.10)

#observations 6, 156 6, 156 1, 881 1, 881 219, 558 219, 558

#fixed effects 206 376 181 351 11, 227 21, 304

White standard errors in braces (robust for arbitrary heteroskedasticity over time and country-pairs),
and Driscoll-Kraay-Newey-West standard errors in parentheses (robust for arbitrary heteroskedasticity
over time and country-pairs, serial correlation and lagged and contemporaneous cross-sectional corre-
lation; see footnote 3).
“No trends” denotes model (1) under τ ij = 0 and “Trends” is the model with τ ij unrestricted. CUijt
indicates that the Glick and Rose (2002) data are about currency unions other than the euro area. The
F T Aijt effect cannot be estimated with the 1992-2002 subsample, because it is constant for that period
for each country-pair. The number of fixed effects is computed after removing unidentified parameters.

effect.
We first consider the sample length T . If the euro effect is biased upwards by
omitted trends in trade, then one would expect a larger estimate from a long sample
than from a short sample, because the signal coming from trends is stronger for longer
samples. Indeed, the estimate of 0.41 based on the complete sample exceeds the 0.16
from the much smaller subsample starting in 1992.
When we use a more gradual reduction of the sample period by taking as starting
years 1970, 1980, and 1990, then the estimated euro effects become 0.38, 0.25, and
0.18, respectively. Similarly, we can reduce the Glick and Rose dataset. Because most
currency unions are in the first part of the sample, however, we move the ending instead
of starting years. If the ending years are 1990, 1980, 1970 and 1960, then the estimated

10
currency union effects are 0.61, 0.49, 0.39, and -0.13. Hence, the shorter the sample,
the smaller the estimate. This is difficult to justify from an economic point of view.
However, it corroborates that trends matter for the magnitude of euro (or general
currency union) estimates.
The standard model does correct for trends in some respect. After all, it includes
time effects λt to account for omitted trending variables that are common to all country-
pairs. If the omitted trends in trade were all driven by this general trend, then there
would be no reason for bias in the euro estimate. The sample length dependence just
discussed thus suggests that cross-sectional variation in the omitted trending variables
is relevant.
If there exist omitted country-pair specific trending trade determinants, then one
expects to see varying trends in the country-pair residual series from a model that
does not account for that. Hence, we study the residuals from model (1) estimated
under τ ij = 0. Plotting the residuals by country-pair over time reveals that there are
indeed time trends left in the residuals and that these vary across country-pairs. This
is confirmed by country-pair specific regressions of the residuals on the time variable
t. They yield t-values for the time variable that are smaller than -2 in 42% and larger
than 2 in 33% of the cases for the whole sample (34% and 29%, respectively, for the
post-1992 subsample, and 29% and 30% for the Glick and Rose data).
The mere existence of omitted country-pair trends does not necessarily result in an
upward bias of the euro effect. Only if the trends are upwards for the euro countries,
our argumentation could explain an upward bias in the euro effect. To check whether
the euro dummy is misused to help capture upward trends, we reestimate model (1)
under τ ij = 0 but without the euro dummy. The residuals are plotted in the left graph
of panel A of Figure 1, where the solid line plots the residual series averaged across
the 55 country pairs involving two euro countries, and the dotted line refers to the
other 116 country pairs. There is a clear upward residual trend for the euro country
pairs, and the diverging trends across the two groups of country pairs lead to a large
difference in the residuals at the end of the sample period. We then take the residuals
of the model including the euro dummy. The right graph of panel A shows that the
euro dummy is used to explain the difference in the residuals at the end of the sample,
so that the estimated euro effect is indeed driven by the diverging trends.
A more detailed analysis reveals that 40 of the 55 euro country-pairs have an upward
estimated trend in the residuals (43 for the subsample). Removing these 40 pairs and
estimating model (1) under τ ij = 0 with the euro dummy reduces the euro effects from
0.41 to 0.06 (0.16 to 0.01 for the subsample). This again suggests that the many upward

11
 Panel A: no country−pair trends in model (1)
0.3 0.3
EURO not in model (1) EURO in model (1)
ijt ijt
0.2 0.2

0.1 0.1

0 0

−0.1 −0.1

−0.2 −0.2
1970 1980 1990 2000 1970 1980 1990 2000

Panel B: country−pair trends in model (1)

0.3 0.3
EURO not in model (1) EURO in model (1)
ijt ijt
0.2 0.2

0.1 0.1

0 0

−0.1 −0.1

−0.2 −0.2
1970 1980 1990 2000 1970 1980 1990 2000

Figure 1: Residuals from model (1) averaged across 55 euro country pairs (solid lines)
and 116 other pairs (dotted lines)

sloping omitted trending variables for the euro countries cause an upward bias in the
estimated euro effect.
We redo this analysis for the Glick and Rose sample. In contrast to our sample,
currency unions in that sample sometimes break down during the sample period and
sometimes (though less frequently) are formed. Under the aforementioned claim that
upward residual trends combined with currency union formation lead to an upward
bias, downward residual trends combined with currency union dissolution also lead to an
upward bias. Hence, we treat both combinations as one group. Of the 131 country-pairs
that have changes in the currency union dummy, 85 belong to that group.5 Removing
them from the sample and reestimating the standard model with the currency union
dummy reduces the estimate from 0.62 to -0.55, so almost the opposite. This confirms

5
In the underlying regressions we have only used country-pairs with 10 observations or more to have
at least some degrees of freedom.

12
the suggestion about the trend relevance from our own data.

3.3 Estimation with country-pair specific time trends

The results from the previous section indicate that it pays to correct for country-pair
specific omitted trending variables. As motivated in Section 2, we do this by including
country-pair specific trends τ ij · t. The columns headed by “Trends” in Table 2 present
the results.
The euro and currency union effects on trade for the three datasets under considera-
tion become 0.03, 0.02 and 0.22. We see that the existing euro estimates obtained from
the standard gravity specification (so without τ ij · t) are substantially different: 51%
versus 3% and 18% versus 2% for the two samples. Given the small standard errors,
the estimates are rather precise and are around the insignificant/significant bound.
In addition, the currency union estimate based on the Glick and Rose (2002) dataset
changes from 86% to 25%. However, the currency union effect is still considerable and,
in our opinion, the magnitude of the new estimate is more realistic than the existing
one from an economic point of view.
Another sign of the relevance of including the trends follows from varying the time
dimension T of the sample. Recall from the previous section that without trends a
gradual reduction of T changes the euro estimate. With trends and starting the sample
in 1970, 1980, and 1990, gives euro estimates of 0.01, 0.01, and 0.03, respectively. For
the Glick and Rose dataset we again move the ending years from 1990 to 1980, 1970,
and 1960. The estimates are 0.16, 0.25, 0.16, and 0.26. Hence, the estimates with
unrestricted τ ij no longer depend on T in a systematic way.
Not surprisingly, also the country-pair residual series no longer have a trend as they
had before; compare panel B of Figure 1 to panel A. Hence, the euro dummy can no
longer be misused to explain a difference in residuals resulting from trend divergence.
Based on the results so far, we argue that upward trends in omitted trade determi-
nants have caused a substantial upward bias in the existing euro estimates, and that
the magnitudes of those estimates are to a large extent driven by the lengths of the
sample periods considered. When we add country-pair trends to the standard model,
the estimate changes from 51% to 3%, so the euro effect is not as large as one would
conclude from the literature so far. Hence, it is important to account for time trends
when estimating the effect of the euro on trade. This is the main claim of the study,
which is confirmed by the more recent study of Berger and Nitsch (2005).
Finally, we briefly discuss the effects of allowing for τ ij on the other estimates
in Table 2. The estimated effect of F T Aijt has become substantially lower. This is

13
presumably explained by the fact that trade integration between two countries often
takes up a major part of the time series available for the country-pair and gradually
increases over time, so that projecting out τ ij · t also removes these trade integration
effects to a great extent. For instance, European integration has existed over the whole
sample period and gradually deepened, so that its trade enhancing effects may be
captured more by τ ij · t than by the dummy variable F T Aijt . Nevertheless, F T Aijt
still seems to have some positive effect.
The estimates for GDPijt and GDP CAPijt have become more homogeneous across
the three samples. This is presumably due to the fact that adding time trends relieves
the included trending regressors from the burden of capturing the trend of omitted
variables as well, so that the true income and income per capita elasticities are more
cleanly detectable. Moreover, Table 2 shows that using 36 years of data yields more
precise estimates of the GDPijt and GDP CAPijt impacts than taking 11 years, as
expected.

4 Sensitivity analysis
We now examine the robustness of the euro and currency union estimates presented in
Section 3.3, namely 0.03, 0.02, and 0.22. Section 4.1 examines whether the estimates
change if we generalize model (1), so as to verify that this specification is sufficient for
appropriate estimation of the euro impact. Section 4.2 examines alternatives for our
country-pair trend specification. Section 4.3 discusses the effect of explicitly accounting
for the nonstationarity and cointegration features of the data. Finally, Section 4.4
discusses potential endogeneity issues.

4.1 Generalizing the model

Although the model allows for heterogeneous fixed effects and time trends, the para-
meters for the economic variables are assumed to be homogeneous. Allowing them to
be heterogeneous as well, and estimating the model for each of the 55 euro country-
pairs separately, gives an average euro effect of 0.02 for the whole sample, 0.02 for the
post-1992 sample, and 0.26 for the Glick and Rose data (Table 3, row 1). The averages
are similar to the estimates from model (1), which are replicated in row 0 of the table.
We have also assumed homogeneity over time, in particular for the euro effect.
One might argue that the euro benefits gradually increase over time. Perhaps there
were already advantages in the process towards the euro because in the middle of the
nineties it was already known that a number of countries would presumably qualify for

14
the common currency. The simple term δ 1 · EU ROijt in model (1) does not allow for
such time variation. Therefore, we now let the euro effect depend on time by interacting
three sets of dummy variables, similar to the approach in Micco et al. (2003). Let 1[.]
denote the indicator function that is one if the condition inside brackets is valid, define
∗ be one if i and j have the euro in
Y EARt as the year number at time t, and let EU ROij
2002 (so it is already positive in all years before the euro era) and zero otherwise. Then
P
we substitute δ 1 · EU ROijt in (1) by 2002 ∗
τ =1993 δ τ · 1[Y EARt = τ ] · EU ROij · 1[Y EARt ≥
1993]. It yields different euro estimates δ 1993 , . . . , δ 2002 from 1993 onwards. For the
model with τ ij = 0, they are 0.25, 0.28, 0.32, 0.31, 0.34, 0.39, 0.44, 0.46, 0.47, 0.51,
with standard errors of 0.06. The euro effect thus seems to increase over time. This
corresponds to the claims of Micco et al. (2003) and Flam and Nordström (2003).
However, these estimates may also be biased due to omitted trending variables, just
like the existing estimates for δ 1 are biased. Indeed, leaving τ ij unrestricted yields
0.00, 0.01, 0.04, -0.00, -0.01, 0.03, 0.05, 0.03, 0.02, 0.05, with standard errors of about
0.03. Improving the model specification thus removes the gradual increase in the euro
benefits. Interestingly, the last five estimates tend to be higher than the first four, and
the simple δ 1 · EU ROijt term with the estimate of 0.03 is apparently quite appropriate.
Using the 1992-2002 sample, we also see a gradual increase in the euro estimates
if τ ij = 0 is imposed. Allowing for unrestricted τ ij leads to identification problems.
The reason is as follows. If one wants to test whether the euro has led to a gradual
increase in trade from 1993 through 2002, then one needs a reference path of trade for
that whole period that indicates how large trade would have been if there had been no
euro. This reference path depends on τ ij . However, there is not enough data before
1993 to estimate it, so that from the 1992-2002 sample one cannot identify which part
of the realized trade increase is caused by the euro and which part is simply normal
growth in trade. One needs a sufficiently long period before the period of analysis for
proper identification. This is another motivation for our choice for using data starting
in the sixties.
The next robustness check of the euro and currency union estimates deals with
the idea that the euro period may have coincided with a slowdown in intra-euro-area
trade resulting from other factors. We do control for many of such potential factors,
but our correction is necessarily imperfect. Suppose that factors for which we do not
control would have caused a trade reduction of 10% if there had been no euro. Then,
our estimate of 3% would actually represent a trade-enhancing effect of the euro of
13%. Though it is impossible to completely correct for such factors, one can obtain an
approximate correction by assuming that the trade impact of such factors is the same

15
 Table 3: Sensitivity of euro estimate to generalizations of model (1)

Model generalization OWN DATA GLICK-ROSE

1967-2002 1992-2002 1948-1997

0: Baseline 0.032 0.018 0.223

(0.016) (0.016) (0.052)

1: Country-pair specific regression parameters 0.024 0.023 0.257

(0.078) (0.058) (0.736)

2: Year-specific euro effects [-0.006, 0.053] — —

([0.024, 0.042])

3: Year-specific EU effects added 0.047 0.040 —

(0.018) (0.030)

4: Quadratic country-pair trends added −0.013 −0.002 0.239

(0.014) (0.015) (0.056)
¡ ¢
5: Unrestricted country trends ξ it + ξ jt added 0.050 0.058 0.101
(0.012) (0.022) (0.053)

6: Dynamic model (long-run effects) 0.015 −0.001 0.167

(0.039) (0.014) (0.086)

7: Export as dependent variable 0.012 −0.004 —

(0.019) (0.016)

Driscoll-Kraay-Newey-West standard errors in parentheses; see note to Table 2.

1: The numbers are averages of the estimates and standard errors resulting from country-pair specific
estimations of (1) without time effects. For the Glick-Rose data we only use currency union country-
pairs with at least 10 observations. 2: Intervals give range of estimates and standard errors; for the
post-1992 subsample there are identification problems: see main text of Section 4.1. 6: The dynamic
model adds two lags of T RADEijt , GDPijt , GDP CAPijt as regressors to (1); as there is no residual
serial correlation now, standard errors are based on zero lags. 7: This variant explains EXP ORTijt as
a function of GDPjt , GDP CAPjt , EUROijt , F T Aijt , the bilateral real exchange rate RERijt , and the
fixed effects of (1). There is no estimate for the Glick-Rose data set, because it does not include data
on uni-directional trade flows.

for country-pairs that have the euro and country-pairs that do not have the euro but
are in the European Union. We want to allow for unrestricted intra-EU specific trade
developments from 1993 onwards (taking 1999 instead of 1993 yields similar results).
P
Therefore, we add 2002 ∗ ∗
τ =1993 ατ · 1[Y EARt = τ ] · EUij · 1[Y EARt ≥ 1993], where EUij
∗ before. As
is one if i and j are in the EU in 2002, similar to the definition of EU ROij
row 3 in Table 3 shows, the euro estimates become 0.05 and 0.04 for the long and short
samples, respectively, so similar to the baseline.
Next, we examine the linearity of the country-pair time trends. One may argue that
the results underestimate the true euro effect, because of a slowdown in the growth of
international trade (for instance, after the first half of the seventies), so that the euro

16
parameter may now capture the fact that trade is below the linear trend at the end of
the sample. However, a general slowdown in the growth is captured by the time effects
λt , and adding quadratic country-pair specific trends to the model (as in Cornwell et
al., 1990) hardly affects the estimates (they become -0.01, -0.00 and 0.24 for the three
samples).
One can also generalize the trends in the country instead of country-pair direction,
along the lines of Baltagi et al. (2003), as discussed in Section 2.2. They suggest in-
cluding fixed effects to control for all possible individual country time-varying variables,
where the triple (ijt) dimensionality of the trade panel allows them to let these fixed
effects move unrestrictedly over time. This in fact generalizes the year effect λt into
ξ it + ξ jt . We add this to model (1), which results in an encompassing combination
of the Baltagi et al. and our approaches. As row 5 of Table 3 shows, the estimated
euro effects become 0.05 for the whole sample and 0.06 for the post-1992 sample. De-
spite the substantial number of additional parameters (612 for the complete sample),
the estimates are in line with the ones from model (1) reported in Table 2. For the
Glick and Rose sample we obtain a currency union estimate of 0.10. This is somewhat
lower than the baseline, so that for the Glick-Rose data nation-specific factors matter.
Interestingly, the currency union estimate is not statistically different from the euro
estimates.
Next, we extend (1) by adding lags of T RADEijt , GDPijt , and GDP CAPijt as
regressors to capture the dynamic nature of trade more directly instead of indirectly
through the adjustment of standard errors. Two lags turn out to be enough, and we
have no evidence that EU ROijt and F T Aijt need lags. As before, we estimate the
model by LSDV. It is well known that in dynamic panel models this estimator is biased
and only consistent when the number of time periods is large. Because we use 36 years
of data and inspecting the simulation results by Judson and Owen (1999), we expect
finite sample bias to be acceptably small in our case. The estimates are 0.02 and -0.00,
and for the Glick-Rose sample we get 0.17 (see row 6 of Table 3). These are in line
with the baseline estimates.
Finally, we use real exports instead of trade as dependent variable (following Bun
and Klaassen, 2002), where real exports are measured in exporter’s output. The esti-
mated euro effects become 0.01 and -0.00. They are similar to our baseline estimates,
although they tend to be somewhat lower.
In summary, we conclude that several model generalizations leave the estimated euro
effect essentially unchanged, so that model (1) seems general enough for appropriate
estimation of the euro impact on trade.

17
4.2 Alternative trend specifications

We have suggested using unrestricted country-pair specific linear trends to make the
estimated euro effect more robust. However, there may exist alternative trend specifica-
tions with which one can attain the same goal. This section considers such alternatives.
More specifically, we use two alternative approaches. First, we examine simplifi-
cations of our approach to examine whether model (1) is too general, as we currently
use many (171) parameters τ ij . Even though the standard error for the euro estimate
is quite small, one may try to obtain more precise estimates by economizing on the
number of trend parameters. Such simplifying restrictions may also provide insights
into the factors that are really crucial for avoiding bias in the euro estimate. Second,
we compare our specification with the trend approach of Baltagi et al. (2003), which
uses ξ it + ξ jt instead of τ ij · t.
To organize the results, we consider all alternatives as restrictions upon a single
general model:

T RADEijt = β 1 GDPijt + β 2 GDP CAPijt + δ 1 EU ROijt + δ 2 F T Aijt

+η ij + τ ij · t + ξ it + ξ jt + εijt , (2)

which encompasses the Baltagi et al. (2003) model and our specification (1). For
instance, the latter imposes ξ it = λt (for all i). We check the validity of each restriction
by comparing the restricted estimates to those from a more general variant of (2). We
confine ourselves to the complete sample; the results for the post-1992 period and the
Glick-Rose sample are similar.
The first column of Table 4 gives the estimates for the encompassing model (2),
which we take as the benchmark. (These are exactly equal to those underlying row 5
of Table 3.)
For ease of comparison, the next two columns replicate the estimates of model (1)
in Table 2 with and without trends. Note that the latter model only has η ij + λt as
effects. As discussed earlier, the euro estimate is biased upwards in that case. Because
all restricted versions of η ij + λt , such as e
ηi + e
η j , give similarly high estimated euro
impacts, the third column represents all those incomplete specifications.
The next three columns concern specifications with various restrictions on τ ij .
Apart from that, all variants restrict ξ it = λt in model (2). First, we analyze whether
we can replace the country-pair specific trends by time-varying effects of distance, i.e.
imposing τ ij = αDISTij where DISTij is the logarithm of distance obtained from
the Glick-Rose data set. We find the counterintuitive result of slightly significantly in-
creasing (in absolute value) distance coefficients over time. However, this is consistent

18
 Table 4: Sensitivity of estimates to trend specifications in (2)

∗
τ ij t + ξ it + ξ jt τ ij t — DISTij t EUROij τi + h
t (h τ j ) t ξ it + ξ jt

EU ROijt δ1 0.050 0.032 0.410 0.399 0.119 0.063 0.055

(currency union) (0.012) (0.016) (0.075) (0.055) (0.032) (0.022) (0.020)

F T Aijt δ2 0.10 0.06 0.41 0.39 0.31 0.23 0.33

(free trade area) (0.02) (0.03) (0.09) (0.08) (0.02) (0.06) (0.06)

GDPijt β1 — 0.70 1.41 1.61 1.60 0.60 —

(product GDP) (0.36) (0.39) (0.35) (0.10) (0.41)

GDP CAPijt β2 — −0.23 −0.68 −0.87 −0.89 −0.10 —

(product GDP capita) (0.35) (0.37) (0.33) (0.09) (0.39)

#fixed effects 988 376 206 206 206 224 836

All trend specifications include year effects λt . Driscoll-Kraay-Newey-West standard errors in paren-
theses; see note to Table 2.

with earlier empirical evidence reporting stable distance coefficients with no tendency
to decline over time (Leamer and Levinsohn, 1995) or increasing distance coefficients
(Brun et al., 2005). It may indicate that distance does not completely capture trans-
portation costs. Anyway, the pattern of the remaining estimates corresponds closely
with that of the model restricting τ ij = 0. In particular, the euro effect is 0.40. The
residuals again show trends and we conclude that this parsimonious specification of the
trends is not sufficient.
The second restriction on τ ij is motivated by Figure 1. It suggests that the euro
estimate is biased because of faster unexplained trade growth over the whole period
for country-pairs that now have the euro compared to other country-pairs, for instance
due to increased economic cooperation irrespective of the euro. We model this by
allowing for a difference in trend between the group of euro country-pairs and the other
∗ , where EU RO ∗ has been defined in the previous section. The
pairs: τ ij = αEU ROij ij
estimated α is indeed positive and the euro effect moves from 0.41 to 0.12. However, it
is still significantly higher than the estimated euro impact from the benchmark model.
We see that allowing for more flexibility regarding trends affects the euro estimate
substantially. However, we conclude that the single group-specific trend variable is not
sufficiently general.6 Nevertheless, it could be that after the introduction of additional

6
Micco et al. (2003) use a related variable called EUT rendijt , which is t if both i and j belong to
the EU at time t and 0 otherwise. Adding this variable to our specification yields a euro estimate of
0.26. Again, the estimate is lower than in case of no country-pair trend correction, but it is significantly
higher than the estimate from the more general benchmark model. Hence, also EUT rendijt does not

19
group-specific trends the trend variation becomes sufficiently captured. In any case,
the general approach of using a trend for each country-pair separately provides a useful
reference point.7
Third, we let the trends be country instead of country-pair specific. That is, we
restrict τ ij = e
τi + e
τ j , where e
τ i is the trend coefficient for country i irrespective of the
partner country. This reduces the number of trend effects from 171 to 19. Column
four presents an estimate of 0.06. This does not differ significantly from the benchmark
case. Hence, the trending behavior of country-specific omitted regressors is the most
important reason for the bias of the euro effect in models with no trends.
This result motivates another direction for testing the robustness of our euro esti-
mate. The Baltagi et al. (2003) approach controls for all country variables irrespective
of their development over time. That is, their approach generalizes the linear country
τi + e
trends, (e τ j ) · t, into unrestricted country trends ξ it + ξ jt . It restricts τ ij = 0. Table
4 demonstrates that the euro estimate is similar to the one resulting under linearity, de-
spite the 612 additional parameters. Therefore, linearity of the country-specific trends
is fine. An additional advantage of the linearity restriction is that the effects of vari-
ables consisting of country-specific components only, such as GDPijt (the sum of the
log of real GDP of both countries), can be estimated, whereas the presence of ξ it + ξ jt
makes them unidentified.
The obvious question is then why to use country-pair specific instead of country-
specific trends. First, trade data are country-pair oriented by nature, so it seems natural
to start with country-pair instead of country trends, particularly in a model that has
already country-pair intercepts η ij . At least, allowing for country-pair trends provides
a way to check whether the restriction τ ij = e
τi + e
τ j is valid.
A second reason for using country-pair trends comes from a comparison of esti-
mates across trend specifications for the free trade area dummy F T Aijt . Table 4 shows
that the most general specification (2) yields an estimated FTA effect of 0.10. Imposing
homogeneity on the country-specific effects (ξ it = λt ), as our model does, makes no sub-
stantial difference: 0.06. However, excluding the country-pair trends τ ij · t changes the

completely avoid the upward bias.

7
Micco et al. (2003) and Flam and Nordström (2002) have added another dummy to the model to
capture trade diversion effects of the euro, that is, a shift in a euro member’s trade with a non-member
to a member. That dummy is one if exactly one of the trading partners has the euro. If there is trade
diversion, this dummy has a negative effect. However, contrary to the authors’ expectations, they get
a positive estimate. Since the dummy is only one at the end of the sample, it may be biased upwards
for the same reason as the euro dummy estimate. Indeed, if we add the trade diversion dummy to our
model, then it has a positive impact if no country-pair specific trends are allowed (0.20 for complete
sample, 0.08 for post-1992 sample), but a zero impact once trends are in the model (0.02 and -0.01,
respectively). Hence, the inclusion of trends solves the surprisingly positive trade diversion estimate.

20
FTA estimate into 0.23 and 0.33 for the linear and unrestricted country trends, respec-
tively. Apparently, correction for country-pair variation of omitted trending variables
is important.
We conclude that country-pair trends are sufficient to obtain a robust euro estimate,
and that they are necessary for a robust estimate of the FTA dummy. Moreover, they
make it possible to estimate the impacts of nation-specific factors such as GDP, in
contrast to models with nation-year-specific fixed effects.

4.3 Panel cointegration estimation

The analysis up to now has ignored potential unit-root nonstationarity features of the
variables in model (1). This is the standard approach in the gravity literature. Because
T RADEijt , GDPijt and GDP CAPijt are presumably nonstationary, we have thus es-
sentially implicitly approximated the distribution of the estimator by the asymptotic
distribution for an infinite cross-section dimension N but a finite time dimension T .
This may be appropriate for the 1992-2002 sample, because there T is rather small.
However, we have also seen in Sections 3.3 and 4.1 that it is advisable to use the com-
plete sample, and for that sample (more than 30 years) the approximation for infinite
T may be better. In that case, nonstationarity issues are relevant for inference. For
instance, if the three variables are nonstationary and cointegrated, which seems quite
plausible from an economic point of view, the limiting variance of the least squares esti-
mator of the cointegrating vector depends on the long run covariance between changes in
the regressors (∆GDPijt and ∆GDP CAPijt ) and the error term εijt , which invalidates
standard inference (see Mark and Sul, 2003). Even though we are ultimately interested
in the euro estimate and not in cointegrating vector estimation, problems regarding
the latter may carry over to the euro estimate. Therefore, this section investigates the
nonstationarity and whether it affects the estimated euro effect.
We first test for unit roots and cointegration in T RADEijt , GDPijt and GDP CAPijt .8
The panel unit root tests of Harris and Tzavalis (1999), testing the null hypothesis of
a unit root, and of Hadri (2000), testing the null of stationarity, both indicate nonsta-
tionarity (only in the case of T RADEijt there are conflicting outcomes). The panel
cointegration tests of Pedroni (1999) reveal cointegration.
To solve the resulting least squares inference problems mentioned earlier, one can use
fully modified OLS (FMOLS) or dynamic OLS (DOLS) techniques for panel data (see
Kao and Chiang, 2000). Because various authors report satisfactory results from DOLS
(Stock and Watson, 1993, Maddala and Kim (1998, p.184), we follow that approach. In

8
The tests are computed using the package NPT 1.3 of Chiang and Kao (2002).

21
particular, we follow Mark and Sul (2003) who allow for a similar specification as model
(1). They show that the panel DOLS estimator is asymptotically normally distributed,
so that standard inference can be made.
The way we use DOLS to estimate the euro effect consists of two steps. First, we
use DOLS to estimate the cointegrating vector. That is, we correct for the covariance
between changes in the nonstationary regressors and the error term εijt by including
two leads and two lags of these changes directly in the regression equation using het-
erogeneous coefficients (see the note to Table 5 for a motivation for the number of
P
leads and lags). Thus we add 2s=−2 γ ijs1 ∆GDPij,t−s + γ ijs2 ∆GDP CAPij,t−s to the
right-hand-side of model (1) (while removing the stationary euro and FTA dummies).
Panel DOLS estimation then gives estimates for the cointegrating vector parameters
β 1 and β 2 with standard errors, which we base on the Newey and West (1987, 1994)
long-run variance estimator.
In the second estimation step we substitute the estimates for β 1 and β 2 into model
(1) and estimate the impacts δ 1 and δ 2 of EU ROijt and F T Aijt . Because the equi-
librium error is stationary, this step is a stationary panel regression, so one can use
standard inference. As before, we use the Driscoll-Kraay-Newey-West approach to
obtain robust standard errors.9
The estimation results are in Table 5. The conclusions drawn in the previous
section remain valid. In particular, the euro effect is again substantially different when
heterogeneous trends are included, and in the model with trends it is again about 3%.

4.4 Endogeneity

So far, we have assumed that the general disturbance term εijt in model (1) is uncor-
related with the regressors. This may be problematic. First, countries that trade a lot
with each other may experience high economic growth, may be more likely to adopt
a common currency, enter into a free trade agreement, and so on. Hence there may
be a causality from εijt via T RADEijt to the regressors, that is, simultaneity. Second,
omitted variables may be correlated with included regressors. Both sources of endo-
geneity can lead to biased estimates if the estimation method does not account for it.
This problem has yet to be effectively addressed in the gravity literature, though some
authors have contributed to solving the issue (Rose, 2000, among others). This sec-

9
We follow a two-step instead of single-step procedure, because adding the leads and lags removes
observations at the beginning and end of the sample and (because the euro dummy is only one at the
end of the sample) that would lead to a severe loss of euro observations in a single-step approach. In
the second step of our regression, there are no leads and lags, so that no euro observations are lost.
Because the estimator of β 1 and β 2 is superconsistent, the asymptotic distribution of the estimator for
the euro and FTA effects is not affected by the two-step nature of the approach.

22
 Table 5: Cointegration based estimation results for trade model (1)

OWN DATA GLICK-ROSE DATA

Whole period Micco et al. period Whole period

1967-2002 1992-2002 1948-1997

No trends Trends No trends Trends No trends Trends

EU ROijt /CUijt δ1 0.374 0.034 0.153 0.015 0.586 0.171

(currency union) (0.064) (0.018) (0.031) (0.016) (0.079) (0.055)

F T Aijt δ2 0.38 0.05 — — 0.86 0.31

(free trade area) (0.08) (0.04) (0.16) (0.06)

GDPijt β1 0.59 0.94 2.03 0.92 0.56 0.94

(product GDP) (0.18) (0.26) (0.44) (2.17) (0.03) (0.08)

GDP CAPijt β2 0.20 −0.49 −1.78 −0.68 0.61 0.31

(product GDP capita) (0.16) (0.26) (0.47) (2.17) (0.03) (0.08)

See the notes to Table 2. However, the estimates and standard errors for GDPijt and GDP CAPijt
are now based on the Mark and Sul (2003) panel dynamic OLS estimator using the Newey and West
(1987, 1994) method for long run variance estimation.
DOLS requires choosing the number of leads and lags. We have used all combinations between (0,0),
(3,0), (0,3), and (3,3), but the eventual euro estimate is very robust. Therefore, we only present results
for two leads and two lags, following the choice by Mark and Sul (2003). For the 1992-2002 data the
DOLS endogeneity correction is restricted to be homogeneous across country-pairs and uses only one
lag and one lead to maintain enough degrees of freedom.

tion provides an alternative approach, based on cointegration and instrumental variable

estimation.
Regarding the regressors GDPijt and GDP CAPijt we already corrected for en-
dogeneity in the panel cointegration model given above. More in particular, using
the DOLS procedure we specified that εijt is correlated with some leads and lags of
the changes in the nonstationary regressors GDPijt and GDP CAPijt . Because we
established that T RADEijt , GDPijt , and GDP CAPijt are cointegrated, the super con-
sistency and limiting distribution of the DOLS estimators of β 1 and β 2 are robust to
potential endogeneity of GDPijt and GDP CAPijt .
This argument does not hold for the stationary regressors EU ROijt and F T Aijt .
To correct for endogeneity we estimate by Instrumental Variables (IV) the second es-
timation step of the procedure of the previous section using one-period lagged values
EU ROij,t−1 and F T Aij,t−1 as instruments. A complication is that the disturbance term
of the second step exhibits autocorrelation, so that such instruments are not valid. To
remove the residual autocorrelation we include two lags of the dependent variable, i.e.
the equilibrium error, as regressors. This makes the instruments valid. Relying on large

23
T asymptotic theory, as in Section 4.3, the resulting dynamic panel data model can be
consistently estimated by the proposed IV procedure.
The IV euro estimates are in Table 6. We include also the corresponding OLS
(LSDV) estimates, which are inconsistent in case of endogeneity. As can be seen, IV
and OLS estimates do not differ much, particularly for the complete sample. Hence
we conclude that endogeneity is of limited importance. Furthermore, the estimates
corroborate those in Table 2, so that the conclusions drawn in Section 3 remain valid.
In particular, the euro effect is again substantially different after the inclusion of the
heterogeneous trends, and in the model with trends it is now about 1.5%.
There are several explanations for the minor relevance of endogeneity here. The
economic reasons for exchange rate stabilization and integration policies, such as euro
and FTA membership, mainly depend on the importance of trade between countries i
and j for their economies. This is only partly determined by the absolute trade level,
i.e. the dependent variable in the model, so that the effect of εijt on the regressors is
not so direct.
In addition, much of the potential endogeneity of the EU ROijt and F T Aijt regres-
sors is absorbed by the fixed unobserved heterogeneity effects, i.e. the country-pair
specific constants and time trends and the year effects. To explain this, assume that
euro and FTA membership are determined by the importance of trade between coun-
tries i and j for their economies. This depends on (1) the importance of bilateral
trade relative to multilateral trade, and (2) the importance of multilateral trade for the
economies involved. Firstly, bilateral trade relative to multilateral trade (as measured
by trade shares) is largely determined by gravity type of variables (such as distance),
and cultural and institutional characteristics. These variables are more or less constant
over time, so that the regressors EU ROijt and F T Aijt are presumably more correlated
with the country-pair specific effect η ij than with εijt . The fact that bilateral trade for
some country pairs has nevertheless grown faster than for others is captured by τ ij t.
Secondly, the importance of multilateral trade has increased over time for all countries,
reflecting globalization of the international economy, so that the regressors are more
affected by the time effect λt than by εijt .
Finally, monetary and trade policies are also affected by cultural and political cir-
cumstances, which are most likely exogenous of nature. This is particularly true for the
introduction of the euro, because a major reason was to get closer to a political union
in Europe.

24
 Table 6: IV based euro estimates for trade model (1)

OWN DATA GLICK-ROSE DATA

Whole period Micco et al. period Whole period

1967-2002 1992-2002 1948-1997

No trends Trends No trends Trends No trends Trends

OLS 0.369 0.014 0.115 0.034 0.583 0.133

(0.094) (0.039) (0.038) (0.016) (0.095) (0.084)

IV 0.379 0.019 0.138 0.081 0.547 0.016

(0.070) (0.052) (0.053) (0.064) (0.106) (0.095)

See the notes to Table 2. The estimates concern long-run effects. As there is no residual serial correlation
now, standard errors are based on zero lags.

5 Conclusion
This study has revisited the question whether the euro has increased trade. Existing
estimates show trade benefits between 5% and 40% and the magnitude of the euro
effect positively depends on the length of the sample used. Using data on 171 industrial
country-pairs over 1967-2002 we have first replicated these findings with a commonly
used panel gravity model.
The residuals from that model exhibit trends that vary across country-pairs. Most of
the euro country-pairs have upward trends, and we have shown that the euro dummy,
which is one only at the end of the sample, captures part of these upward residual
trends. This leads to an upward bias in the estimated euro effect.
To avoid such omitted trending variables bias, we have proposed extending the
standard model by country-pair specific time trends τ ij · t. The estimated euro effect
then becomes 3%, which is on the significant/insignificant bound. The estimate is
robust to various model generalizations and is no longer driven by the length of the
sample period.
Hence our main conclusions are that omitted trending variables have biased existing
euro estimates and that the magnitude of that bias depends on the length of the sample.
Including country-pair time trends avoids both and shows that the euro effect is not as
large as one would conclude from the literature so far.
In addition, we have improved on existing standard error computations by mak-
ing the standard errors not only robust to heteroskedasticity, but also to serial and
cross-sectional correlation. We have accounted for the nonstationarity and cointegra-
tion features in the data by using panel dynamic OLS estimation. Finally, we have

25
contributed to the solution of endogeneity bias in the gravity literature by combining
cointegration with instrumental variables estimation.
Our finding that it is important to account for country-pair specific time trends
may be relevant for other applications of the panel gravity model as well. For instance,
the current study has shown that generalizing the model in this direction changes
the estimated benefit of non-euro currency unions from 86% to 25% using the Glick
and Rose (2002) sample. Moreover, including trends may be relevant for research on
the effect of trade integration or the benefits of accession to the EU for the Eastern
European countries, as well as for studies using general non-trade panel models for
trending data. These issues are left for future research.

26
References
Anderson, J.E. and van Wincoop, E. (2004). ‘Trade costs’, Journal of Economic Liter-
ature, Vol. 42, pp. 691-751.
Baltagi, B.H., Egger, P. and Pfaffermayr, M. (2003). ‘A generalized design for bilateral
trade flow models’, Economics Letters, Vol. 80, pp. 391-397.
Barr, D., Breedon, F. and Miles, D. (2003). ‘Life on the outside: economic conditions
and prospects outside Euroland’, Economic Policy, Vol. 37, pp. 573-613.
Berger, H. and Nitsch, V. (2005). ‘Zooming out: the trade effect of the euro in historical
perspective’, CESifo Working Paper No. 1435.
Bergstrand, J.H. (1989). ‘The generalized gravity equation, monopolistic competition,
and the factor-proportions theory in international trade’, Review of Economics
and Statistics, Vol. 71, pp. 143-153.
Brun, J.-F., Carrère, C., Guillaumont, P. and de Melo, J. (2005). ‘Has distance died?
Evidence from a panel gravity model’, The World Bank Economic Review, Vol.
19, pp. 99-120.
Bun, M.J.G. and Klaassen, F.J.G.M. (2002). ‘Has the euro increased trade?’ Tinbergen
Institute Discussion Paper No. 02-108/2, University of Amsterdam.
Chiang, M.-H. and Kao, C. (2002). ‘Nonstationary Panel Time Series Using NPT 1.3
- A User Guide’, Center for Policy Research, Syracuse University.
Cornwell, C., Schmidt, P. and Sickles, R.C. (1990). ‘Production frontiers with cross-
sectional and time-series variation in efficiency levels’, Journal of Econometrics,
Vol. 46, pp. 185-200.
Driscoll, J.C. and Kraay, A.C. (1998). ‘Consistent Covariance Matrix Estimation with
Spatially Dependent Panel Data’, Review of Economics and Statistics, Vol. 80,
pp. 549-560.
European Commission (2003). Quarterly report on the euro area, No. III/2003.
Flam, H. and Nordström, H. (2003). ‘Trade volume effects of the euro: aggregate and
sector estimates’, mimeo, Institute for International Economic Studies, Stock-
holm University.
Glick, R. and Rose, A.K. (2002). ‘Does a currency union affect trade? The time series
evidence’, European Economic Review, Vol. 46, pp. 1125-1151.
Hadri, K. (2000). Testing for stationarity in heterogeneous panel data, Econometrics
Journal, 3, 148-161.
Harris, R.D.F. and Tzavalis, E. (1999). ‘Inference for unit roots in dynamic panels
where the time dimension is fixed’, Journal of Econometrics, Vol. 91, pp. 201-
226.

27
HM Treasury (2003), EMU and trade, http://www.hm-treasury.gov.uk/media/C4AA1
/adsuf03_456.pdf.
Judson, R.A. and Owen, A.L. (1999). ‘Estimating dynamic panel data models: a guide
for macroeconomists’, Economics Letters, Vol. 65, pp. 9-15.
Kao C. and Chiang, M.-H. (2000). ‘On the estimation and inference of a cointegrated
regression in panel data’, in Baltagi, B.H. (ed), Nonstationary panels, panel coin-
tegration and dynamic panels, Advances in Econometrics, Vol. 15, pp. 179-222.
Leamer, E. and Levinsohn, J. (1995). ‘International trade theory: the evidence’, in:
Grossman, G.M. and Rogoff, K. (eds), Handbook of International Economics,
Vol. 3, Elsevier, North Holland, pp. 1339-1394.
Maddala, G.S. and Kim, I.M. (1998). Unit Roots, Cointegration and Structural Change,
Cambridge University Press, Cambridge, UK.
Mark, N.C. and Sul, D. (2003). ‘Cointegration vector estimation by panel DOLS and
long-run money demand’, Oxford Bulletin of Economics and Statistics, Vol. 65,
pp. 655-680.
Micco, A., Stein, E. and Ordoñez, G. (2003). ‘The currency union effect on trade: early
evidence from EMU’, Economic Policy, Vol. 37, pp. 315-356.
Newey, W.K. and West, K.D. (1987). ‘A simple, positive semi-definite, heteroskedas-
ticity and autocorrelation consistent covariance matrix’, Econometrica, Vol. 55,
pp. 703-708.
Newey, W.K. and West, K.D. (1994). ‘Automatic lag selection in covariance matrix
estimation, Review of Economic Studies, Vol. 61, pp. 631-653.
Pedroni P. (1999). ‘Critical values for cointegration tests in heterogeneous panels with
multiple regressors’, Oxford Bulletin of Economics and Statistics, Vol. 61, pp.
653-678.
Rose, A. (2000). ‘One money, one market: the effect of common currencies on trade’,
Economic Policy, Vol. 30, pp. 9-45.
Stock, J.H. and Watson, M.W. (1993). ‘A simple estimator of cointegrating vectors in
higher order integrated systems’, Econometrica, Vol. 61, pp. 783-820.
Wansbeek, T. and Kapteyn, A. (1989). ‘Estimation of the error-components model
with incomplete panels’, Journal of Econometrics, Vol. 41, pp. 341-361.

28