Revisiting the effects of regional trade agreements on trade flows with proper specification of the gravity model

(1)

Article

Reference

Revisiting the effects of regional trade agreements on trade flows with proper specification of the gravity model

CARRERE, Céline

Abstract

This paper uses a gravity model to assess ex-post regional trade agreements. The model includes 130 countries and is estimated with panel data over the period 1962–1996. The introduction of the correct number of dummy variables allows for identification of Vinerian trade creation and trade diversion effects, while the estimation method takes into account the unobservable characteristics of each pairs of trade partner countries, the endogeneity of some of the explanatory variables as well as a potential selection bias. In contrast to previous estimates, results show that regional agreements have generated a significant increase in trade between members, often at the expense of the rest of the world.

CARRERE, Céline. Revisiting the effects of regional trade agreements on trade flows with

proper specification of the gravity model. European Economic Review , 2006, vol. 50, no. 2, p.

223-247

DOI : 10.1016/j.euroecorev.2004.06.001

Available at:

http://archive-ouverte.unige.ch/unige:46660

Disclaimer: layout of this document may differ from the published version.

1 / 1

(2)

European Economic Review 50 (2006) 223–247

Revisiting the effects of regional trade agreements on trade ﬂows with proper

speciﬁcation of the gravity model

Ce`line Carre`re

CERDI-CNRS, Universite´ d’Auvergne, 65 Bd Franc-ois Mitterrand, 63000 Clermont-Ferrand Cedex 163009, France

Received 28 June 2002; accepted 7 May 2004 Available online 28 October 2004

Abstract

This paper uses a gravity model to assess ex-post regional trade agreements. The model includes 130 countries and is estimated with panel data over the period 1962–1996. The introduction of the correct number of dummy variables allows for identiﬁcation of Vinerian trade creation and trade diversion effects, while the estimation method takes into account the unobservable characteristics of each pairs of trade partner countries, the endogeneity of some of the explanatory variables as well as a potential selection bias. In contrast to previous estimates, results show that regional agreements have generated a signiﬁcant increase in trade between members, often at the expense of the rest of the world.

JEL classification:F11; F15; C23

Keywords:Regional trade agreements; Gravity equation; Trade creation; Trade diversion; Panel data

1. Introduction

After a long period of neglect from the late 1960s to the late 1980s, the gravity trade model has acquired a second youth. First, new theoretical foundations have

www.elsevier.com/locate/econbase

doi:10.1016/j.euroecorev.2004.06.001

Tel.: +33-473-17-74-00; fax: +33-473-17-74-28.

E-mail address:c.carrere@cerdi.u-clermont1.fr (C. Carre`re).

(3)

been proposed both with the advent of trade theories based on increasing returns to scale, in imperfectly competitive markets and ﬁrm-level product differentiation (Helpman and Krugman, 1985;Bergstrand, 1985, 1989;Baier and Bergstrand, 2001;

Evenett and Keller, 2002) and, within a perfect competition setting, with product differentiation at the national level (Deardorff, 1998;Anderson and Van Wincoop, 2003). Second, the gravity model has been used extensively to study trade patterns, as for example in the case of the drastic changes following the demise of central planning. Most recently, in the estimation of models of geography and trade, the gravity model is, once again, holding center stage (Hummels, 2001; Redding and Venables, 2004; Limao and Venables, 2001; Brun et al., 2002). In fact, the gravity model has also become a favored tool to assess the ex-post trade effects of a currency union (Glick and Rose, 2002;Rose and Van Wincoop, 2001), or the trade creating (TC) and trade diverting (TD) effects associated with regional trade agreements (RTAs) (Frankel, 1997; Soloaga and Winters, 2001). However, for reasons elaborated in this paper, previous estimates of TD and TC are likely to be unreliable.

Along with this renewal in interest, questions have been raised about the proper formulation of the model (choice of variables) as well as about proper econometric techniques, especially when the usual cross-country formulation is amended to include a temporal dimension. Indeed, the discussion about the proper econometric specification of the gravity model has shown that the conventional cross-section formulation without the inclusion of country specific effects is misspecified and so introduces a bias in the assessment of the effects of RTAs on bilateral trade (e.g., Matyas, 1997; Soloaga and Winters, 2001; Anderson and Van Wincoop, 2003).

However, it turns out that this panel speciﬁcation, with three speciﬁc effects (exporter, importer and time effects) is only a restricted version of a more general model which allows for country pair heterogeneity (e.g., Cheng and Wall, 1999;

Egger and Pfaffermayr, 2003).

In contrast to the traditional cross-section gravity model which includes time invariant trade impediment measures (e.g. distance, common language dummies, border, historical and cultural links as in most studies, see Frankel (1997)), this general proposed speciﬁcation is more adequate since it accounts for any time invariant (unobserved) bilateral effect. Hence, all factors that inﬂuence bilateral trade which were partially captured by regional dummies are now controlled for.

In this paper, I apply this more general panel specification on a recent gravity model specification derived byBaier and Bergstrand (2002)with the addition of: (i) a barrier-to-trade function similar to Limao and Venables (2001) instead of the traditional distance variable and common border dummy, and: (ii) three dummy variables for each RTA considered (intra-trade, imports and exports dummies) to allow for a correct identification of Vinerian trade effects. I show that the predictions of the effects of RTAs in terms of trade creation (TC) and trade diversion (TD) are very different whether one uses a cross-section or a panel specification that controls for the unobservable characteristics of each pair of countries (modeled as random effects). In this setting, one has to check for the potential correlation of some explanatory variables with the country-pair unobservable effects. I show that the use of the instrumental variable method proposed by Hausman and Taylor (1981) is

(4)

necessary to avoid estimation bias. Moreover, the selection bias that can appear in an unbalanced sample is tested and corrected for by the inclusion of a selection rule in the model’s estimation (e.g. Nijman and Verbeek, 1992; Guillotin and Sevestre, 1994).

Section 2 presents the canonical gravity model (with the barrier-to-trade function) and the modified cross-section version used for ex-post evaluations of RTAs (with the three dummies mentioned above that have to be included for each RTA). Section 3 specifies the alternative panel model and the econometric method. Finally, Section 4 presents the average effects of each RTA over 1962–1996 and the evolution of these effects over the same period, comparing cross-section and panel estimates. To anticipate the main conclusions, it turns out that: (i) the panel estimates yield more convincing estimations of the average effects of RTAs (as RTA dummies no longer capture unobservable bilateral trade patterns) as well as the evolution of these effects; (ii) globally, after the date of implementation, RTAs have generated a significant increase in trade between members, often at the expense of the rest of the world; (iii) a plausible pattern of estimates across the seven RTAs evaluated in the paper is evidenced. Insofar as the gravity model is the accepted model for estimating the efficiency effects of RTAs, attention should be paid to the proper specification of the model. Section 5 concludes.

2. The gravity model as an ex-post method to assess regional agreements

2.1. From the theoretical gravity model to estimable gravity equation

Following the recent exposition in Baier and Bergstrand (2002), I use a generalized version of the ‘‘standard’’ gravity model derived from a framework where firms in monopolistic competition maximize profits and consumers maximize utility according to Dixit–Stiglitz preferences. As shown by Baier and Bergstrand (2002), if the representative profit-maximizing firms in country j set product prices delivered to market i according to Eq. (2), one obtains the following equilibrium trade flow for each goods-producing firm in countryjselling to market i:

Mij¼ g jð1sÞ

Y_j

p_j Yi

p_jy_ij Pi

1s

½sjð1þtiÞ1þtijs

; (1)

Mij: c.i.f value of the aggregate merchandise trade ﬂow imported by countryifrom exporterj;

s: the elasticity of substitution in consumption in goods (Dixit–Stiglitz preferences);

g: the Cobb–Douglas preference parameter for goods;

j: ﬁxed cost facing each ﬁrm (including both capital and labor);

Y_iðjÞ: gross domestic product of countryiðjÞ;

p_j: exporter (countryj) price level of it representative good.

(5)

The price level of this good in countryi(c.i.f. price) is given by

p_ij¼p_jy_ij (2)

withy_ija barrier-to-trade function betweeniandjto be developed below.

P_i: referred to as the ‘‘multilateral resistance term’’ in the literature (e.g.Anderson and Van Wincoop, 2003), interpreted as an output-weighted measure of the remoteness (in terms of trade costs) of countryi:

P_i¼ X^N

k¼1

n_k½p_ky_ikð1þt_ikÞ^1s

" #1=1s

(3)

withnj being the number of varieties of goods produced inj;

t_ij: ad-valorem tariff rate by countryion the good produced inj(t_ii¼0 assumed);

s_j: real share of goods output in national product in countryj;

ti: share of tariff revenue relative to income.

Eq. (1) is the currently accepted theoretical foundation for the gravity equation in the presence of transportation costs and tariffs.¹As shown by Anderson and Van Wincoop (2003), assuming thatt_ij¼t_jiandy_ij ¼y_ji;the implicit solution to Eqs. (1) and (3) is²:

p_j ¼ s_jY_j njYw

1=1s

P¹_j : (4)

Substitutingp_j into Eq. (1),³and assumingt_i¼0 (in most countries, tariff revenue is a trivial share of GDP), yields:

M_ij¼ g jð1sÞ

1

Y_w s_jY_iY_jy^1s_ij ð1þt_ijÞ^s½P_iP_j^s1; (5) whereY_w is world output of goods.

Eq. (5) is remarkably close to the gravity model in the empirical literature. It suggests that the proper speciﬁcation should include:

(1) the logarithm of the product of the GDPs of countriesi ðYiÞandjðYjÞ;

(2) per capita GDP or population of the exporting country, Nj;as a proxy for the capital-endowment ratio (which determines the endogenous share of goods in national output i.e.s_j);

1Corresponds to Baier and Bergstrand’s equation (28) (2002); to Feenstra’s equation (5.26) (2003) or to Anderson and Van Wincoop’s equation (9) (2003).

2Resolution of the system inAnderson and Van Wincoop (2003)or inBaier and Bergstrand (2002).

3Eq. (1) can be rewrittenM_ij¼ ½g=jð1sÞn_jY_ip^1s_j y^1s_ij P^s1_i ð1þt_iÞð1þt_ijÞ^sasn_j¼s_jY_j=p_j:

(6)

(3) a proxy for the termyij;

(4) the product of the multilateral resistance term for each country-pair.⁴

Regardingyij;it is obvious that it is crucial to get the best handle possible on what constitutes the ‘‘barrier-to-trade’’ function, which is usually proxied either by distance,Dij;between trading partners (and the presence of a common border and a common language as for instance inBaier and Bergstrand (2002)orAnderson and Van Wincoop (2003)), or sometimes by the c.i.f./f.o.b. price ratio.⁵Because recent studies have shown that these variables are not the only determinants of trade costs, I model the barrier-to-trade function, between countriesiandj, as follows:⁶

yij¼ ðDijÞ^d¹ðINiÞ^d²ðINjÞ^d³½e^d⁴^Lîj^þd⁵Êⁱ^þd⁶Ê^j (6) with (expected signs on coefficients in parentheses):

Dij: distance between countries iandjðd140Þ;

Lij¼1 ifiandjshare a common border, otherwise 0ðd4o0Þ;

E_iðjÞ¼1 if the countryiðjÞis landlocked; otherwise 0 ðd₅40;d₆40Þ;

INiðjÞ: level of infrastructure of the countryi(j), computed as an average of the density of road, railway and the number of telephone lines per capitaðd2o0;d3o0Þ:

Turning to the modeling of ½P_iP_j; recent papers had suggested either a complex nonlinear estimation technique ofP_iandP_j(seeAnderson and Van Wincoop, 2003), or the introduction of country ﬁxed effects (see Rose and Van Wincoop, 2001;

Feenstra, 2003). Both approaches yield consistent estimates in a gravity equation (see Anderson and Van Wincoop, 2003;Feenstra, 2003). However, both suggestions are relevant only in a cross-section estimation, i.e. without time dimension in the data.⁷ Hence, as the equation needs to be estimated both with and without the time dimension (see later, Section 3), the following proxies for the multilateral resistance terms (called ‘‘remoteness’’ variables byBaier and Bergstrand (2002)), are introduced:

R_i¼ X^N

k¼1;kai

Y_kðD_ikÞ^1s

" #1=1s

; (7a)

R_j¼ X^N

k¼1;kaj

Y_kðD_kjÞ^1s

" #1=1s

; (7b)

4Polak (1996)shows that if one does not use a measure of the average distance between a country and its main partners as well as the absolute distance in assessing the effects of RTAs, one will underestimate trade between faraway countries and thus bias the estimated RTA coefﬁcient. As shown in the recent derivations of the gravity model, one should include the average distance of the importing countryifrom its main partners.

5Baier and Bergstrand (2001)use the c.i.f./f.o.b. ratio to model transport costs, but their study only deals with OECD countries which have better data. For a discussion about the problems associated with the use of c.i.f./f.o.b. data seeHummels (2001)andLimao and Venables (2001).

6E.g.Limao and Venables (2001).

7Actually, as described in Section 3, the model is estimated over 1962–1996. So, the assumption that the multilateral resistance is a constant country characteristic over the period is not relevant.

(7)

where Eqs. (7a) and (7b) are estimated using a central elasticity value ofs¼4 which corresponds to estimates proposed in empirical literature (e.g. Obstfeld and Rogoff, 2001). Results in the paper are robust to 2psp6:

Taking into account the modiﬁcations discussed above, after substitution of Eq. (6) into Eq. (5), the estimated reduced-form boils down to

lnM_ij¼b₀þb₁lnY_iþb₂lnY_jþb₃lnN_jþb₄lnR_iþb₅lnR_jþb₆lnD_ij þb₇lnL_ijþb₈lnE_iþb₉lnE_jþb₁₀ ln IN_iþb₁₁ ln IN_jþo_ij; ð8Þ where ½g=jð1sÞ1=Yw is absorbed in the constant term, oij is the error term (assumed to have a standard normal distribution), and with the expected signs:

b₁40;b₂40;b₃o0;b₄40;b₅40;b₆¼ ð1sÞd₁o0;b₇¼ ð1sÞd₄40;

b₈¼ ð1sÞd₅o0;b₉¼ ð1sÞd₆o0;b₁₀¼ ð1sÞd₂40 and b₁₁¼ ð1sÞd340:

2.2. The gravity model for ex-post assessment of regional trade agreements

Initially used by Aitken (1973) as an ex-post assessment for the European Economic Community, the gravity model seems well deﬁned for this issue for two reasons. Firstly, arguably, the model represents a relevant counterfactual to isolate the effects of an RTA. If the sample of countries is appropriately selected, the gravity equation suggests a ‘‘normal’’ level of bilateral trade for the sample. Then, dummy variables can be used to capture the ‘‘atypical’’ levels resulting from an RTA.

Secondly, thanks to the correct introduction of dummy variables in the model, one can isolate TC and TD effects of an RTA.

In a Vinerian world following an RTA, TC and TD will be reflected in trade flows as follows: (i) under pure TC, intra-regional trade increases and imports from the rest of the world (ROW) remain unchanged; (ii) under pure TD, the increase in intra- regional trade is entirely offset by a corresponding decrease in imports from the ROW; (iii) if there is both TC and TD, intra-regional trade increases more than imports from the ROW decrease. Because of second-best considerations, identification of TD and TC does not allow inference about the welfare consequences of an RTA for its members. Finally, for non-members, one should include a measure of the change in volume of exports from members to non-members (an increase signifying an improvement in welfare for non-members).⁸

Therefore, the correct ex-post assessment (e.g.Egger, 2002;Soloaga and Winters, 2001) of a RTA on the volume of trade should include the following dummy variables (associated coefﬁcients in parentheses):⁹

8SeeWinters (1997). Note that, if looking at the exports to the ROW is more important than imports from the ROW in terms of non-member welfare, the critical welfare variable is the terms of trade.

9Most authors (e.g.Bayoumi and Eichengreen, 1997; Frankel, 1997; Krueger, 1999) have not included enough dummy variables to distinguish between exports and imports, so they fail to isolate TD and TC effects.

(8)

(i) DIðaIÞ ¼1 if both partners belong to the same RTA [zero otherwise] (capturing intra-bloc trade);

(ii) DMðaMÞ ¼1 if importing countryibelongs to the RTA and exporting countryj, to the ROW [zero otherwise] (capturing bloc imports from the ROW);

(iii) DXðaXÞ ¼1 if exporting countryjbelongs to the RTA and importing countryi to the ROW [zero otherwise] (capturing bloc exports to the ROW).

Suppose thataI40;which corresponds to more intra-bloc trade than predicted by the reference and which can be in substitution to domestic production or to exports from the ROW. Hence, to conclude on whether this corresponds to TC or TD, one needs to examine the signs of the coefﬁcientsa_M anda_X:Then,a_I40 along with a lower propensity to import from the ROW (a_Mo0Þindicates TD, and if the increase in intra-regional trade is entirely offset by a decrease in regional imports from the ROW, this is a pure TD. If intra-regional trade increases more than imports from the ROW decrease, there is both TC and TD. And withaI40 andaMX0;there is pure TC. Finally, comparing aI and aX can lead to inferences about welfare for non- members. For example, ðaI40;aXo0Þ would indicate a dominant ‘‘export diversion’’ and hence a decrease in welfare for non-members. In sum, following a RTA, [aI40 and aMX0ðaXX0Þ] indicates pure TC in terms of imports (exports) and [aI40 and aMo0ðaXo0Þ], indicates TD in terms of imports (exports).

3. Data and estimation

The model is estimated with data for 130 countries over the period 1962–1996.

Trade data are from UN COMTRADE (total bilateral imports in current dollars).

Data sources for the explanatory variables along with data transformations are presented in Section A.1. Once the missing values are taken out,¹⁰the sample covers 130 countries (a list of the countries in the sample is presented in Section A.2). There are thus 240 691 observations for 14 387 pairs of countries.

3.1. Panel specification

It has been observed repeatedly (see Polak, 1996; Matyas, 1997; Bayoumi and Eichengreen, 1997) that regional dummy variables in cross-country estimates capture everything speciﬁc to the importing or exporting countries not captured by the variables included in the equation that inﬂuence the level of trade (e.g. historical, cultural, ethnic, political or geographical factors)¹¹ which is troublesome since the

10Countries which do not declare their imports from a partner or which do not import from this partner are identiﬁed in the same way, i.e. with a missing value. Hence, our data are not censored at zero. The actual number of observations (240 691) represents around 50% of potential number. This implies a potential selection bias which is tested (and corrected for) in Section 3.2.

11If these factors are also correlated with gravity variables (GDP, populations, distance), estimations which do not include them will have an endogeneity bias, because the omitted variables are correlated with the level of bilateral trade and with the explanatory variables (see below).

(9)

dummy variables should really isolate TD and TC effects. Omitting countries’

heterogeneity, or the effects specific to country-pairs in bilateral trade relations may introduce a bias. By contrast, a panel data method enables one to identify the country-pair specific effects and to isolate them. The usual correction introduces three specific effects: exporter, importer and time effects (e.g.Matyas, 1997;Soloaga and Winters, 2001). But the model with three specific effects is only a restricted version of the more general model which allows for country-pair heterogeneity adopted here (e.g. Cheng and Wall, 1999;Endoh, 1999). Then a bilateral term,m_ij; specific to each pair of countries and common to each year (and different according to the direction of trade:m_ijam_ji), is included in the previous model (8) specified for panel data:

lnM_ijt¼a₀þa_tþb₁lnY_itþb₂lnY_jtþb₃lnN_jtþb₄lnR_itþb₅lnR_jt þb₆lnDijþb₇lnLijþb₈lnEiþb₉lnEjþb₁₀ln INit

þb₁₁ ln INjtþb₁₂ ln RERijtþm_ijþnijt: ð9Þ

a₀: effect common to all years and pairs of countries (constant);

a_t: effect speciﬁc to yeartbut common to all the pairs of countries to capture common shocks (e.g. changes in oil prices);

m_ij: effect speciﬁc to each pair of countries and common to all the years;

nijt: error term assumed to be log-normally distributed.

Note the introduction of the bilateral real exchange rateðRER_ijtÞin Eq. (9). In a model with panel data that span a long time period (here 35 years), it is essential to capture the evolution of competitiveness (e.g.Soloaga and Winters, 2001;Bayoumi and Eichengreen, 1997). An increase in the bilateral real exchange rate reﬂecting a depreciation of the importing country’s currency against that of the exporting country, one would expectb₁₂o0:

Following the specification check, the three dummy variables discussed above are introduced in the model to detect TD and TC for a selection of RTAs. The RTAs considered in the paper are: EU, ANDEAN, NAFTA, CACM, MERCOSUR, ASEAN and LAIA (see Appendix A.2 for definition and members). To assess the total effect of these RTAs, I consider the trade between RTAs members (and with the ROW) even before the implementation of the agreements, in order to look for break points around the important dates of the agreements (and notably before and after the implementation date). Actually, it is likely that an increase in trade between members a few years just before the official implementation of an agreement would signal an ‘‘anticipation effect’’. Moreover, even with specific country-pair effects in the specification, one cannot be sure to capture all specific links between members (other than the agreement) that can influence intra-RTA trade (in particular if these links evolve with time). Hence, it is necessary to look at intra-RTA trade before the implementation of the agreement to avoid an ‘‘artificial break point’’ and hence attribute a wrong interpretation to the regional dummy coefficient. Therefore, the

(10)

regional dummies capture the trade between RTAs members (and with the ROW) over the entire period 1962–1996.

3.2. Econometric methods

This section summarizes the main steps followed in the estimation to justify the results reported in Section 4. Intermediate results are reported in Table 2 in Appendix A.3 (Appendix A.3 more formally details the econometric method and the different steps required by the results of appropriated statistical tests).

The Within equation treats the bilateral specific effects as fixed, thereby giving unbiased parameter estimates for time-varying variables.¹² However, since the regional dummies are defined over the whole period of the RTA, these variables would only vary when there are changes in membership during the period. So the fixed-effects model does not allow the estimation of the effects of RTAs with fixed membership (i.e. NAFTA, CACM, MERCOSUR, ASEAN and LAIA). Another problem with the fixed-effects model is that, since the within-method ignores the cross-sectional nature of the data, the interpretation of the regional dummies coefficients does not exactly answer the question of this paper, namely what are the total effects of RTAs. This last point is developed in Section 4.1 about the interpretation of the effects of the EU and the ANDEAN pact RTAs which have been ‘‘enlarged’’ during the period 1962–1996.

Hence, modeling the bilateral effects as random variables is more appropriate. In the absence of correlation between the explanatory variables and the specific bilateral effects, the Generalized Least Squares (GLS) estimation provides consistent estimates of the coefficients. However, variables like GDPs or infrastructure may be correlated with bilateral specific effects. Even, ‘‘the FTA dummy variables may be endogenous by being correlated with unobservable (omitted) variables that are correlated also with the decision to trade’’ (Baier and Bergstrand, 2002, p. 5).

Actually, if cultural, political or historical ties between countries increase their propensity to form a trade agreement as well as their bilateral volume of trade, then there would still be a bias in the coefﬁcient for intra-RTA trade. The Hausman test (1978), based on differences between Within and GLS estimators, conﬁrms that GLS estimator is biased and then that some explanatory variables are endogenous.

The usual way to deal with this issue is to consider instrumental variables estimation such as that proposed byHausman and Taylor (henceforth HT) (1981), though here it is adapted to the case of an unbalanced sample according to the method proposed byGuillotin and Sevestre (1994). The HT estimator is based upon an instrumental variable estimator which uses both the between and within variation of the strictly exogenous variables as instruments. More specifically, the individual mean of the strictly exogenous variables are used as instruments for the time-variant variables that are correlated with the bilateral specific effects. The definition of the explanatory variables as exogenous or endogenous is a testable hypothesis. Notably, a Hausman test of over-identification, based on the comparison of the HT estimator

12All these coefﬁcients are signiﬁcant at the 1% level and have the expected sign (column 1,Table 2).

(11)

and the Within estimator, is carried out. According to different tests (detailed in Section A.3), the ‘‘ﬁnal’’ regression considers as endogenous the variables of GDP, population, infrastructure and the seven intra-RTA trade dummies.

A last potential estimation bias deserves scrutiny: the unbalanced sample could be subject to a non-ignorable selection rule.¹³ In this case, the selection bias can be tested and corrected by the inclusion of the selection rule in the model estimation. I use a method proposed by Nijman and Verbeek (1992), which approximates the Heckman correction term by adding variables that reﬂect each country-pair presence in the sample (see Guillotin and Sevestre, 1994). So the following additional variables are included: (i) PRES: number of years of presence of the coupleij’s in the sample; (ii) DD: dummy that takes the value 1 if ij is observed during the entire period, 0 otherwise; (iii) PA_t: dummy that takes the value 1 if ij was present in t1 ðPA₀¼0Þ:

Results are reported in the first column ofTable 1. Concerning traditional gravity variables, coefficients have the expected sign and are significant at a 1% level (except forEi). Import volume ofifrom jincreases with GDP and coefficients are close to unity as suggested by the theory. The exporting country population variable has the expected negative sign. The elasticity of bilateral trade to distance is close to unity (0:98)¹⁴and the coefficients for the remoteness variables are significantly positive.

The volume of trade increases with the level of infrastructure of each country, as in Limao and Venables (2001). Sharing a land border allows countries to trade 3 times more than expected from the gravity equation ð¼e^1:09Þ: Likewise, imports from a country without direct access to the sea are 28% lower. Finally, a real depreciation ofiwith respect tojlowersi’s imports fromj. Concerning the variables introduced to avoid a selection bias in the coefficients of regional dummies, the coefficients associated to the variables PRES and PA_t are significantly different from zero.

Hence, ceteris paribus, country-pairs which have at least two years of consecutive available data (and a fortiori if they are present over several years) have more bilateral trade than country-pairs with interruption in their data.

4. Application to the assessment of the effects of regional trade agreements

I proceed in two steps: ﬁrst I comment the average effects of each RTA over the sample period (1962–1996) reported in Table 1, then I decompose these average effects to look for the evolutions over the period (notably around the important dates).

13I.e. that the probability of a pair of countries being included in the sample is not independent of model error, and in particular to the unobserved bilateral effects.

14According to Eq. (8),b₆¼ ð1sÞ:d₁:The elasticity of transport costs with respect to distance is usually estimated in the range 0:2od₁o0:4 (e.g.Limao and Venables, 2001; Hummels, 2001). Combined with an elasticity of substitution between goods of about 4 (Rose and Van Wincoop (2001), use an estimate ofs¼5;Obstfeld and Rogoff (2001), suggest a consensus estimate ofsbetween 4 and 6), the implied distance coefﬁcient would be in the range 1:2ob₆o0:8 which is almost identical to the estimates ofb₆inTable 1.

(12)

Table 1

Comparison of the results with regional dummies (1962–1996)

Variables Mijt

Panel HT^a Cross-section OLS

Coeff. t Average coeff. Max Min

lnYit 1.08 78.1 0.80 0.95 0.68

lnYjt 1.14 84.2 1.00 1.26 0.76

lnNjt 0.67 23.7 0.08 0.03 0.22

lnR_it 0.13 14.0 0.15 0.27 0.07

lnRjt 0.19 10.2 0.10 0.57 0.00

lnDij 0.98 22.9 1.04 0.54 1.29

Lij 1.09 5.2 0.85 1.48 0.45

Ei 0.13 1.5 0.12 0.15 0.52

E_j 0.33 3.4 0.35 0.02 1.10

ln INit 0.04 5.4 0.20 0.47 0.09

ln INjt 0.03 5.8 0.41 0.63 0.19

ln RERijt 0.005 4.1 — — —

PRES 0.05 14.2 — — —

DD 0.09 0.5 — — —

PAt 0.48 49.1 — — —

EUintra 0.71 14.6 0.24 0.44 1.02

EUimports 0.20 10.5 0.90 1.09 0.55

EUexports 0.29 12.6 1.02 1.62 0.26

ANDEANintra 0.60 4.8 1.35 2.31 0.06

ANDEANimports 0.97 25.0 0.43 1.18 0.86

ANDEANexports 0.92 21.4 0.49 0.23 1.01

CACMintra 0.70 2.9 2.27 2.79 1.60

CACMimports 0.80 4.7 0.47 0.14 0.87

CACMexports 0.21 1.2 0.15 0.37 0.91

LAIAintrab 0.55 2.7 1.23 1.60 0.07

LAIA_imports 1.58 6.6 1.05 0.46 1.87

LAIAexports 0.42 1.6 0.58 0.66 2.22

MERCOSURintra 0.90 0.9 0.52 1.18 1.68

MERCOSURimports 1.09 9.1 0.01 0.62 0.54

MERCOSURexports 0.18 1.0 0.14 1.16 0.92

NAFTAintra 0.65 0.48 0.77 2.15 0.50

NAFTA_imports 0.50 2.7 0.32 0.68 0.12

NAFTAexports 0.06 1.1 0.07 0.63 0.65

ASEANintra 0.88 5.7 1.50 2.40 1.11

ASEAN_imports 0.48 3.5 0.34 0.89 0.01

ASEANexports 0.76 9.0 0.47 1.20 0.25

Number of obs (NT) 240 691 6877 8472 6012

Number of bilateral (N) 14 387 — — —

R² 0.69 0.65 0.73 0.60

Oij(mean) 0.91 — — —

andsigniﬁcant at the 1% and 5% levels, respectively (t-student next to correspondent coefﬁcient). The time dummy variables and the constant are not reported in order to save space.

aHausman–Taylor estimator. Endogenous variables¼lnYit;lnYjt; lnNjt;ln INit; ln INjt;and the 7 RTAs intra.

bAs all the members of ANDEAN and MERCOSUR belong also to LAIA, the evolution of trade of the two former RTAs is isolated in computing the dummies for LAIA as follows (i.e.Soloaga and Winters, 2001): LAIAintra¼LAIAintraANDEANintraMERCOSURintra;LAIAimports¼LAIAimports ANDEANimportsMERCOSURimports;LAIAexports¼LAIAexportsANDEANexportsMERCOSURexports:

(13)

4.1. Average effects over the period 1962–1996

Table 1 reports the coefficients for dummy variables for two sets of regressions, one in cross-section (corresponding to most uses of the gravity model for ex-post assessments of RTAs), yielding 35 separate regressions (one for each year), the other with the panel specification as defined in the previous section.

Results are quite different depending on the estimation method. For instance, on average over 1962–1996, intra-EU trade is 104%ð¼100 ðe^0:711ÞÞabove what is predicted by the panel gravity model, whereas it is 21% below the expected level according to the cross-section gravity model. The latter negative result is obtained in other cross-section studies (e.g. Frankel, 1997; Soloaga and Winters, 2001).

Likewise, in the panel estimation, the ANDEAN and the ASEAN blocs present a trade between members about, respectively, 1.2 and 2.4 times above the reference prediction, associated with a propensity to import from the ROW inferior by 62%

and 38%. By contrast, in the cross-section estimations, the positive intra-trade for these two RTAs is associated, on average, with a positive propensity to import from the ROW. Contrasting results are also obtained according to the estimator used for the NAFTA and the MERCOSUR. Concerning the LAIA and the CACM the predictions are the same with the two estimators: increase in intra-trade with an import diversion. However, the order of magnitude of the level of intra-trade due to the agreement is not credible in the case of OLS estimations: the intra-LAIA (CACM) is 240% (868%) above the norm vs. 72% (99%) predicted in the panel estimates. In sum, one must conclude that the estimates generated by the panel estimates are far more plausible than those generated by OLS.

Turn next to the two RTAs with varying membership (EU and ANDEAN pact).

Recall that the objective is to measure the ‘‘total’’ trade effects of RTAs. These include the effects of RTAs on: (i) countries joining (leaving) the RTA, and; (ii) countries which have belonged to the RTA over the entire sample period [and hence are time invariant—e.g. the ‘‘6’’ of the EU]. Hence, the choice is to (i) deﬁne regional dummies that vary over time depending on new members and (ii) assess these dummies with the Hausman–Taylor method to capture both effects (on original and new members). However, one can try to distinguish the two effects. In the case of the EU, the Within estimate (which exploits variation over time) reveals that the average effect over the period for a new member of the EU is a trade volume with the other members around 1.75 times (+75%) more than before the adhesion to the agreement (column 1 inTable 2).¹⁵On the other hand, if the EU dummy is redeﬁned to capture only intra-trade for the six older members (column 8 in Table 2), the average effect over the period of the EU is a trade between the 6 original members around 25% above the norm. The sum of these two effects is equal to 100%, which is a good approximation of the total effect obtained inTable 1with the HT estimator (+104%).¹⁶ Arguably, there estimates are, once more plausible.

15SeeGlick and Rose (2002, pp. 1130–1131)for a similar interpretation of the Within coefﬁcients in the case of monetary unions.

16Of course this exercise only provides a very rough proxy of the real decomposition.

(14)

4.2. Evolution of the effects during the RTA’s existence

Since relevant inferences about TD and TC require inspection of the evolution of these coefficients over time and around the period when RTAs are implemented, I break down regional dummy variables into two-year periods with these variables replacing the global regional dummies.¹⁷To ease exposition and interpretation, the estimated regional dummies are plotted over time (annual dummies from cross- section estimations and two-year dummies from estimation on panel data). Note that all the non-significant estimated coefficients are graphed as zero and that the variation between two consecutive coefficients is always tested as being statistically significant in the panel estimations. I comment on the results in Figs. 1–3for the three RTAs considered to be most significant in terms of depth of integration: EU, NAFTA and MERCOSUR. For other RTAs, I provide a brief summary (and corresponding figures are reported in Appendix A.4).

In the case of the EU, the cross-section analysis (Fig. 1a) displays a counter- intuitive negative trend in intra-EU trade until 1980 before it turns positive with the propensity to export to the ROW declining over the period. Overall, one would conclude exports TD but no evidence of import-TD. This result is similar to what Soloaga and Winters (2001)obtained using the same estimation method.

By contrast, panel estimates (Fig. 1b) suggest three rather distinct periods in terms of TC and TD. From 1967 to 1973, intra-trade presents no clear tendency, as for the trade with the ROW. However, following the first (and second) enlargement, the model predicts a significant positive trend in intra-trade (a_I increases and turns significantly positive in 1978, the pattern continuing with the deep integration following the EC-92 program). In parallel, there is first a stagnation of imports of members from the ROW until 1985 and then a negative trend (a_M became negative in 1990). Hence, the model suggests that, if the first enlargement of the EU (from six to nine members in 1974) resulted in a pure TC, the second enlargement (with Spain and Portugal in 1986) presents signs of significant TD, in terms of imports and exports. These results are quite different fromBayoumi and Eichengreen (1997), and arguably more sensible, since many studies have raised concerns about overall TD (e.g. the discussion byWinters (1993)on the TD effects of the common agricultural policy which would become stronger with time). This pattern of results revealed by the panel estimates seem very plausible indeed, especially when compared with those obtained from previous studies in cross-section.

Comparing the results from both estimation methods is even more striking in the cases of MERCOSUR and NAFTA. Here, the cross-section estimates show largely unexplainable volatility throughout the time-period whereas the panel estimates

17Each RTA dummy is multiplied by 18 time dummies of two years (except the last one that capture only the year 1996). The HT method is used to be consistent with the first step as this section decomposes an average effect already estimates with the HT method. With this procedure, coefficient estimates for traditional explanatory variables of the gravity model are identical to the ones inTable 1as, for each agreement, the addition of the new dummy variables introduced is, by definition, equal to the former aggregate dummy variable. Note that one can also use the within estimator in this section. This does not change the qualitative conclusion.

(15)

capture much more clearly the expected effects of an RTA around the time of announcement or of implementation: an increase in intra-trade and a decrease in imports from the ROW. The difference in patterns is particularly evident for NAFTA, which reveals largely insigniﬁcant dummies until the ﬁrst trade policy reforms in Mexico, and the announcement of NAFTA negotiations. As to

-3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

EU-intra EU-imports EU-export s -3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

EU-int ra EU-import s EU-exports Second enlargement First enlargement

Second enlargement First enlargement

(a) (b)

Fig. 1. Evolution of EU dummies over 1962–1996 (a_I;a_Manda_X): (a) in cross-section and (b) in panel.

-3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

M ER C OS UR -intra M ER C OS UR -im po rts M ER C OS UR -e xpo rts

-3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

M ER C OS UR -intra M ER C OS UR -im po rts M ER C OS UR -e xpo rts

Asuncion Treaty Protocols Arg-Brazil

(a) (b)

Fig. 2. Evolution of MERCOSUR dummies over 1962–1996 (a_I;a_Manda_X): (a) in cross-section and (b) in panel.

-3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

NAFT A-int ra NAFT A-imports NAFT A-export s

-3 -2 -1 0 1 2 3

1962 1966 1970 1974 1978 1982 1986 1990 1994

NAFT A-int ra NAFT A-import s NAFT A-export s

NAFTA signed NAFTA signed

trade liberalisation

Mexico’s unilateral trade liberalisation

(a) (b)

l ico’s unilatera Mex

Fig. 3. Evolution of NAFTA dummies over 1962–1996 (aI;aM andaX): (a) in cross-section and (b) in panel.

(16)

MERCOSUR, panel estimates capture both the increase in intra-trade and the diversion of import from the ROW observed in the more disaggregated analysis in Yeats (1998). At the same time, there is some evidence of an increase of exports for NAFTA and MERCOSUR to the ROW. These results probably capture the non- discriminatory reduction in trade barriers to non-partners that were taking place at the same time that they were entering into RTAs.

This pattern of import (and sometimes export) TD was also found for other RTAs reported in Section A.4. In the case of the ANDEAN agreement, the model ﬁnds import-TD over the period 1969–1981, for the CACM over the period 1962–1974, and for the LAIA over the period 1968–1980. Concurrently, over the same period, an export-TD is observed for the ANDEAN, whereas there is some evidence of an increase of the propensity to export towards the ROW for CACM. ASEAN is the only examples of pure TC over the period (some TC was detected for LAIA after 1980). Overall, these patterns also appear reasonable with TC observed for ASEAN, an RTA that was outward-looking in contrast with the typical inward-looking ﬁrst- wave South–South RTAs.¹⁸

5. Conclusions

This paper has paid particular attention to the specification and the estimation of the gravity model to correct for biases present in previous studies. The panel estimation with bilateral specific random effects was revealed to be statistically justified after correcting for the endogeneity of the income, size, infrastructure and intra-RTA trade variables. Moreover, dummies were introduced to take into account the selection rule of the sample. Arguably, these modifications lead to a better formulation of the counterfactual against which one assesses the trade performance of RTAs.

Comparison of panel estimates with the more usual cross-section estimates revealed a far more plausible pattern of trade effects associated with RTAs. In general, the ﬁndings of this study, covering seven RTAs, show that most of these RTAs resulted in an increase in intra-regional trade beyond levels predicted by the gravity model, often coupled with a reduction in imports from the rest of the world, and at times coupled with a reduction in exports to the rest of the world, suggesting evidence of trade diversion.

Acknowledgements

I thank Jaime de Melo for guidance, and J.F. Brun, Patrick Guillaumont, Alan Winters, and three anonymous referees for their valuable comments.

18deMelo and Panagariya (1993)review the record of the ﬁrst-wave RTAs of the 1960s and 1970s. They note that preferences were granted behind high tariff walls suggesting that these must have been mostly of the trade-diverting variety.

(17)

Appendix A

A.1. Sources and definitions of data

Mijt: COMTRADE, aggregate total bilateral imports of countryifrom countryj (current US$).

YiðjÞt: CD-ROM WDI, World Bank 1999, GDP of countryiat timet(constant US

$ 1995).

N_it: CD-ROM WDI, World Bank 1999, total population of countryiat timet.

D_ij: Data for distance are extracted from the software developed by the company CVN. The distance is orthodromic, i.e., taking into account the sphericity of Earth, and measured in kilometers between the main economic city of the countryiand that of countryj.

L_ij: Dummy equal to one if the countriesiandjshare a common land border, 0 otherwise.

EiðjÞ: Dummy equal to one if the countryi(j) does not have a direct access to the sea, 0 otherwise.

INiðjÞt: This index is built using four variables from the database constructed by

Canning, 1998: the density of roads, of paved roads, of railways, and the number of telephone lines per capita of countryi(j) at timet, each variable being normalized to have a mean equal to one. An arithmetic average is then calculated over the four variables, for each country and each year (the computation is similar toLimao and Venables (2001)).

R_iðjÞt: Multilateral trade resistance or ‘‘remoteness’’ of countryi(j), computed

according to Eqs. (7a) and (7b) withs¼4:

RER_ijt: The nominal exchange rate for each country against US$(NER_i=$;country i’s currency value of 1 US$) is extracted from the IFS data set, as the consumption price index for countryiðCPIiÞfor each year from 1962 to 1996. If the CPI is not available for a country, the GDP deﬂator of the country is used. The bilateral real exchange rate (RER) is computed as following:

RER_i=j¼ ðCPI_jÞ=ðCPI_iÞ ðNER_i=$=NER_j=$Þ:

For each pair of countries, the RER is speciﬁed such as its mean over the period is 100.

A.2. Countries in the sample and definition of the RTAs studied

OECD Sub-Saharan Latin America and Asia and the Others

Africa the Caribbean Paciﬁc

Australia Angola Argentina^e;g Bangladesh Albania

Austria^a South Africa Bahamas Brunei Armenia

Belgium+ Burundi Barbados Bhutan Azerbaijan

(18)

Luxembourg^aBenin Belize China Bulgaria

Canada^f Burkina Faso Bolivia^b;e Fiji Belarus

Germany^a Central African Rep. Brazil^e;g Hong Kong Czech Rep.

Denmark^a Ivory Coast Chile^b;e Indonesia^c Algeria

Spain^a Cameroon Colombia^b;e India Saudi Arabia

Finland^a Congo Costa Rica^d Cambodia Egypt

France^a Comoros Dominican Rep. Lao PDR Estonia

UK^a Cape Verde Dominica Macao Georgia

Ireland^a Djibouti Ecuador^b;e Mongolia Greece^a

Iceland Ethiopia + Eritrea Grenada Malaysia^c Bosnia and

Italy^a Gabon Guatemala^d Nepal Herzegovina

Japan Ghana Guyana Pakistan Hungary

Korea, Rep. Guinea Honduras^d Philippines^c Iran United States^fGuinea–Bissau Haiti Papua New Guinea Israel

Netherlands^a Gambia Jamaica Singapore^c Jordan

Norway Equatorial Guinea Mexico^e;f Salomon Islands Kazakstan New Zealand Kenya Nicaragua^d Thailand^c Kyrgyz Rep.

Portugal^a Madagascar Panama Vietnam Kuwait

Sweden^a Mali Peru^b;e Western Samoa Lithuania

Switzerland+ Mozambique Paraguay^e;g Sri Lanka Latvia Liechtenstein Mauritania El Salvador^d Tonga Macedonia

Mauritius Suriname Kiribati Morocco

Malawi Trinidad and TobagoVanuatu Malta

Niger Uruguay^e;g Oman

Nigeria St. Vincent and Poland

Rwanda The Grenadines Romania

Sudan Venezuela^b;e Russian Federation

Senegal St. Lucia Slovenia

Sierra Leone Antigua and Slovak Rep.

Sao Tome´ and Principe Barbuda Syrian Rep.

Seychelles St. Kitts and Tajikistan

Somalia Nevis Turkmenistan

Chad Tunisia

Togo Turkey

Tanzania Ukraine

Uganda Uzbekistan

Zaire Zambia Zimbabwe

Countries written in italic are not available as reporter countries in COMTRADE(only as partners).

South Africa includes bilateral trade of the group of countries: South Africa+Lesotho+Botswa- na+Namibia+Swaziland.

aEU member (European Union, 1957) except UK, Denmark, Ireland (1973), Greece (1981), Spain, Portugal (1986), Austria, Finland and Sweden (1995);

bANDEAN member (1969), except Chile (1969–1976), Venezuela (1973), Peru (1969–1992);

cASEAN member (Association of Southeast Asian Nations, 1967);

dCACM member (Central American Common Market, 1960);

eLAIA member (Latin American Integration Association, 1980+former LAFTA, 1960);

fNAFTA member (North American Free Trade Agreement, 1992);

gMERCOSUR member (Mercado comu`n del Sur, 1991).

(19)

A.3. Econometric appendix

A.3.1. Hausman and Taylor (1981) estimator

Mijt¼XijtcþZijfþm_ijþnijt; (A.1)

where i¼1;2;. . .;N; j¼1;2;. . .;M; t¼1;. . .;T; m_ij is IIDð0;s²_mÞ and n_ijt is IIDð0;s²_nÞboth independent of each other and among themselves.

Let XðZÞ the matrix of explanatory variables variant (invariant) over time and suppose that among the variablesXandZ, there exist:

(i) X_ijt: k₁ ðk₂Þexogenous (endogenous) variables, denotedX₁(X₂);

(ii) Zij: g₁ ðg₂Þexogenous (endogenous) variables denotedZ1 (Z2).

If the conditionk1Xg₂is satisfied, then the equation is identified and Eq. (A.1) can be estimated using½QX₁;QX₂;PX1;Z1as instruments (seeBreusch et al., 1989) with Qthe matrix that computes the deviations from individual means andPthe matrix that computes the observation across time for each pair of countries. As the resulting estimator is consistent but not efficient (it does not correct for heteroskedasticity and serial correlation due to the presence of random bilateral specific effects), Hausman and Taylor suggest using this first round of estimates to compute the variance of the specific effectðs²_mÞ;and the variance of the error termðs²_nÞ:Note that in this papers²_m and s²_n have to be corrected for the bias of heteroskedasticity specific to the unbalanced sample (seeGuillotin and Sevestre, 1994).

The instrumental variable estimator is then applied to the following transformed equation:

M_ijt ð1O_ijÞM_ij:¼ ½X_ijt ð1O_ijÞX_ijcþO_ijZ_ijfþO_ijm_ij þ ½nijt ð1OijÞnij:

withO_ij¼ ðs²_n=T_ijs²_nþs²_nÞ¹⁼²:

The average value ofOij is systematically presented in theTable 2.

A Hausman test of over-identification, based on the comparison of the HT estimator and the Within estimator, must be carried out. If the null hypothesis cannot be rejected, the instruments are legitimate (in the sense of no bias due to a correlation between specific bilateral effects and the explanatory variables), and the HT estimator is the most efficient estimator (e.g.Baltagi, 2001). Note that canonical correlations are also a useful device for comparing different sets of instruments. In fact, one should use instruments for which the geometric average of the canonical correlations with the regressors are maximized (e.g.Baltagi, 2001; Mairesse et al., 1999;Hall et al., 1996).

Finally, Guillotin and Sevestre (1994) also recommend comparing the HT estimator to the GLS estimator with a Hausman test. If the nullH₀is rejected, one can conclude that the instrumented model gives better estimations then the GLS model (without any instrumentation). Thus, instrumented variables are actually endogenous. All these tests are reported in Table 2 and commented in the next section.