In Search of WTO Trade Effects: Preferential Trade Agreements Promote Trade Strongly, But Unevenly

Contributor Notes

Authors’ E-Mail Addresses:,

The literature measuring the impact of Preferential Trade Agreements (PTA) and WTO membership on trade flows has produced remarkably diverse results. Rose's (2004) seminal paper reports a range of specifications that show no WTO effects, but Subramanian and Wei (2007) contend that he does not fully control for multilateral resistance (which could bias WTO estimates). Subramanian and Wei (2007) address multilateral resistance comprehensively to report strong WTO trade effects for industrialized countries but do not account for unobserved bilateral heterogeneity (which could inflate WTO estimates). We unify these two approaches by accounting for both multilateral resistance and unobserved bilateral heterogeneity, while also allowing for individual trade effects of PTAs. WTO effects vanish and remain insignificant throughout once multilateral resistance, unobserved bilateral heterogeneity, and individual PTA effects are introduced. The result is robust to the use of alternative definitions and coding conventions for WTO membership that have been employed by Rose (2004), Tomz et al. (2007), or by Subramanian and Wei's (2007).


The literature measuring the impact of Preferential Trade Agreements (PTA) and WTO membership on trade flows has produced remarkably diverse results. Rose's (2004) seminal paper reports a range of specifications that show no WTO effects, but Subramanian and Wei (2007) contend that he does not fully control for multilateral resistance (which could bias WTO estimates). Subramanian and Wei (2007) address multilateral resistance comprehensively to report strong WTO trade effects for industrialized countries but do not account for unobserved bilateral heterogeneity (which could inflate WTO estimates). We unify these two approaches by accounting for both multilateral resistance and unobserved bilateral heterogeneity, while also allowing for individual trade effects of PTAs. WTO effects vanish and remain insignificant throughout once multilateral resistance, unobserved bilateral heterogeneity, and individual PTA effects are introduced. The result is robust to the use of alternative definitions and coding conventions for WTO membership that have been employed by Rose (2004), Tomz et al. (2007), or by Subramanian and Wei's (2007).

I. Introduction

Over the past half century, the hallmark of the World Trade Organization (WTO) and its predecessor (GATT) has been the reduction of trade barriers.2 Trade theory holds that tariff reductions should increase trade flows, but the empirical literature on the trade effects of WTO membership produces surprisingly ambiguous results. Rose (2004, 2005) jumpstarted the literature when he documented the absence of WTO effects on bilateral trade volumes. After updating Rose’s dataset to include both de jure and de facto WTO membership, Tomz, Goldstein and Rivers (2007, henceforth TGR) did find positive WTO trade effects. Alternatively, Subramanian and Wei (2007, henceforth SW) examined different groups of WTO members and reported positive WTO trade effects for industrialized countries only. This diverse literature suggests that identification of WTO trade effects depends on empirical specifications or coding conventions. Instead, we show in this paper that the approaches of Rose, SW and TGR all produce one consistent result: the WTO does not boost trade among its members. Rather, regional and bilateral preferential trade agreements (PTAs) strongly influence trade flows. However, in order to achieve this one consistent result, the empirical methodology needs to account comprehensively for three crucial trade determinants: multilateral resistance, unobserved bilateral heterogeneity, and distinct effects across PTAs.

First, multilateral resistance controls are needed because average trade costs faced by a country will influence its bilateral trading decisions.3 Failure to control for countries’ average trade costs thus induces omitted variable bias (Anderson and van Wincoop, 2003). SW highlight that multilateral resistance is crucial to their estimating a strong WTO trade effect.4

Second, unobserved bilateral heterogeneity needs to be accounted for in order to further reduce omitted variable bias. That is because trading partners’ bilateral relationship is typically defined by characteristics extending beyond distance, geography or others that can be explicitly controlled for. For instance, unobserved bilateral heterogeneity may include political, personal, and/or civic factors.5 Two countries may exhibit this natural (partly unobserved) affinity for mutual trade before they join a PTA. In this case, the omitted affinity will bias WTO and/or PTA estimates upwards (e.g., Egger, 2000 and Cheng and Wall, 2005). Baldwin (2006) judges unobserved bilateral heterogeneity to be of such crucial importance for policy purposes that he recommends ignoring results obtained without the appropriate controls, such as the country-pair fixed effects we include in this paper. In this regard, one of the drawbacks of SW is that, while controlling for multilateral resistance, they do not simultaneously control for unobserved heterogeneity.

Third, coefficient estimates are also subject to omitted variable bias, if (1) impacts of PTAs on trade vary across different PTAs and (2) PTAs are aggregated into just one dummy. Neither Rose (2004), TGR, and SW all include such an aggregate PTA dummy, and thereby do not account for distinct PTAs effects. In another paper, Rose (2005, Table A5) subsequently does account for individual PTAs and produces a positive WTO impact as a result. However, we will show that this positive WTO effect appeared only because he did not fully control for multilateral resistance.

To date, our paper is the first that simultaneously controls for the three described sources of omitted variable bias in a large bilateral trade dataset.6 Our approach highlights that the omission of any of the controls renders the WTO coefficients biased upward. On the contrary, if all controls are included, e.g. by adding both unobserved bilateral heterogeneity and differential PTA trade effects to SW’s analysis, we find that their original dataset and econometric specification delivers no WTO effects for industrialized countries. The result holds even after we include de facto WTO members, as in TGR. Our empirical results thus confirm Rose’s (2004 and 2004b) finding of no WTO accession effect on bilateral trade.

Since SW and Rose/TGR differ in their coding convention of WTO and PTA variables, we confirm our results using both conventions.7 Our findings show that these conventions do not affect results once estimates are purged of omitted variable bias. We do highlight, however, that SW-style coding is susceptible to producing biased WTO estimates under two specific conditions. The bias typically occurs when (a) industrialized and developing PTAs differ considerably in their trade effects and (b) PTA effects are constrained to one aggregate PTA coefficient. More importantly, this bias is especially detrimental when WTO membership effects are separated into industrialized and developing countries.

In our empirical methodology, we intentionally follow the original Rose/TGR and SW setups closely in order to reconcile their approaches. Our conclusion is that differences in results between these approaches can indeed be reconciled when appropriate controls render unbiased estimates. However, Rose’s original work subsequently paved the way for a voluminous literature using other approaches to examine WTO effects. This literature covers firm heterogeneity, sample selection bias, nonparametric methods, and missing trade (see, e.g., Carrere, 2006; Felbermayr and Kohler, 2006 and 2007; Helpman et al., 2007; Liu, 2007; Chang and Lee, 2007).

After presenting the data in Section 2, we extend SW to account for each individual PTA’s trade creation (Section 3) as well as unobserved bilateral heterogeneity (Section 4). In Section 5, we reexamine Rose (2004, 2005) and extend his analysis to fully account for multilateral resistance in the presence of individual PTA effects. Section 6 provides a detailed discussion of the impacts of individual PTAs on trade and highlights the diversity of their effects by region. In particular, we also observe that the most comprehensive specification returns the most reasonable PTA trade effects. Section 7 concludes.

II. Data

Our data are based on an updated version of SW’s unbalanced panel.8 Their bilateral trade values are derived from the IMF’s Direction of Trade Statistics, deflated by the U.S. consumer price index. The dataset features not only a WTO dummy, but also a dummy that represents industrialized countries’ unilateral trade concessions to developing trading partners under the GATT/WTO’s Generalized System of Preferences (GSP) from 1979 onwards. We update the SW dataset to attribute a value of zero to GSP country-pairs that represent an industrialized country exporting to a developing country. These pairs should be identified as “pure” WTO effects, since GSP is granted as a unilateral preference (for industrialized countries’ imports from developing countries only).9 We also identify Luxembourg as a member of the European Union (EU) in 2000, and correct other minor coding errors identified by TGR. These changes do not affect our or SW’s results qualitatively.

SW employ Rose’s definition of de jure WTO membership. However, TGR indicated that de facto WTO members should also be considered, and that their exclusion may bias WTO effects downwards. To illustrate that WTO effects vanish even when accounting for de facto membership, we use TGR’s WTO membership definition throughout and refer the interested reader to the tables in appendix B that features all results in the original SW and Rose coding. These tables show that our conclusions are not affected by the coding convention.

A single aggregate PTA indicator dummy has been prominent in a number of empirical trade flow studies (see e.g., Rose 2000, 2004, 2005, Glick and Rose 2002, SW, TGR), to capture the average effect of PTAs on trade flows. We extend the SW dataset and introduce a more extensive set of PTAs used by Rose (2005) and by Ghosh and Yamarik (2004) to properly account also for individual trade effects of less known PTAs. In section 4, we also add country- pair fixed effects to the dataset to control for unobserved bilateral heterogeneity in bilateral trading relationships. Further modifications to the dataset are introduced in Section 5. Specifically, we show that results are robust to Rose’s (2004, 2005) mutually inclusive coding of multilateral trade, while continuing to comprehensively account for multilateral resistance, TGR’s de facto WTO membership, and individual PTA effects.10 Note that following SW, this extension splits Rose’s WTO dummy into two groups, representing industrialized and developed WTO members. Rose (2004b) indicated that such a split could lead to selection bias, it “conditions on success” (since industrialized countries liberalized trade more than their developing counterparts). We show that the WTO dummy split does not drive results. The dimensionality of the dataset remains constant throughout, with 55,831 observations for 177 countries and 11,797 bilateral trade pairs in five year intervals from 1950 to 2000.

III. Extending the Empirical Framework to Account for Individual PTA Effects

We first extend the general framework of SW to fully account for the impacts of all trade agreements (WTO, GSP, and individual PTAs). Section 4 then extends the econometric model to account for unobserved heterogeneity in bilateral trading relationships. The SW setup has two important characteristics: First, time-varying fixed effects are introduced to capture multilateral resistance for importers, δmt, and exporters, λxt (where the “m”, “x”, and “t” subscripts identify importer, exporter and time, respectively).11 Multilateral resistance can be conveniently accounted for with these fixed effects, since any nation faces only one import/export price index at any point in time. The inclusion of these time-varying importer and exporter effects requires, however, the dependent variable to be bilateral imports, Importsmxt, instead of the commonly used average trade flow variable.12To maintain consistency our dependent variable is identical to SW’s.13 The second important characteristic of the SW approach is their coding convention.

SW code the trade agreement indicator dummies mutually exclusively to quantify “pure” GSP and WTO effects. The key assumption in SW is that PTA membership “represents the culmination of trade integration.” Thus SW code trade agreement indicators such that all trade creation is exclusively attributed to PTAs, even if both trading partners are currently (or were previously) WTO/GSP members. Specifically, when two trading partners are members of the WTO, GSP, and the same PTA, only the PTA dummy takes the value “1”. Coding is hierarchical throughout, so that when both WTO and GSP dummies could display a “1,” only the GSP variable takes that value. Since the dependent variable is bilateral imports, the dummies’ specifications are defined exclusively on the basis of importer participation in the WTO or GSP.

SW disaggregate the WTO dummy further to identify membership effects for industrialized and developed nations, WTOIndustrialmt and WTODevelopingmt respectively. Our baseline regression (SW’s preferred specification in their Table 4) is then


Note that the “*” superscript indicates that the coding of the variables is mutually exclusive. With the exception of the natural log of bilateral distance, Distancemx, all explanatory variables are 0-1 dummies. PTAmxt indicates joint membership in a trade agreement, while GSPmt indicates whether an industrial country importer grants GSP tariff preferences to the developing country exporter in question. CurrencyUnionmx is the dummy for a common currency union.

The remaining dummies are proxies introduced to control for some observed factors that affect a country pair’s inclination to engage in trade. These proxies include contemporaneous or historical colonial relationships, CurColonymxt and EverColonymx, respectively; common colonizer relationships post-1945, CommonColonizermx; shared official languages, CommonLanguagemx; and territorial dependency and contingency, CommonNationmx and Bordermx. Equation (1) lacks country-year specific regressors (for example, importer/exporter GDP) that are common in canonical gravity equations. These regressors have been absorbed into the time-varying importer/exporter fixed effects that control for multilateral resistance.

Our first extension is to modify (1) by replacing the aggregate PTA vector, PTAmxt, with dummies that allow each PTA to account for its individual effect on bilateral imports. This converts PTAmxt into a matrix and β4into a vector of regression coefficients that captures the trade impacts of individual PTAs. We disaggregate the PTA effects in two stages. First, we introduce only PTAs that are contained in SW’s and Rose’s (2005) aggregate PTA dummy.14 Then we enlarge the set of PTAs to one suggested by the recent PTA literature (see Ghosh and Yamarik, 2004).15

Note that SW’s mutually exclusive coding of trade agreements implies that the introduction of additional PTAs in this second stage diminishes the number of “1” entries in the WTO variables. Therefore, the introduction of additional PTAs may influence the WTO/GSP estimates for two reasons: a) by allowing individual PTAs to correct for omitted variable bias, and b) by reducing the WTO entries of PTA observations to “zero.” If omitted PTAs are strongly trade creating and PTA members have also joined the WTO, the WTO coefficient in SW would be expected to carry upward bias.

A. Results: WTO Trade Impact After Controlling for Multilateral Resistance and Individual PTA Effects (Hierarchical, Mutually Exclusive WTO Coding)

Regressions 1-3 in Table 1 report results for SW’s general WTO framework with the addition of individual PTA effects. Regression 1 is an analogue of SW’s preferred regression.16 The corrections to their dataset discussed in Section 2 increase the WTO and PTA coefficients slightly, but their original results are robust. In this specification, only industrialized countries are shown to be profiting from WTO membership through higher import activity. The implied trade benefit is large, 187% (=e1.053-1), while there is no significant WTO impact for developing countries. With the exception of a common border and same nation status, all regressors that control for observable bilateral heterogeneity are highly significant; and they remain so throughout.

Table 1:

WTO and PTA Effects: (Hierarchical, Mutually Exclusive Coding)

article image
Notes: *, **, *** are 10%, 5%, 1% significance levels; standard errors in parenthesis, fixed effect coefficients are suppressed. GSP and PTA coefficients are composite coefficients that report GSP and PTA effects net of WTO effects and are calculated by the Delta method (the original regression output is in Table 1a). All regressions include time-varying importer and exporter fixed effects.

Regression 1 also indicates a highly significant coefficient associated with the aggregate PTA dummy in the original SW specification. Given the hierarchical, mutually exclusive coding of trade agreements in SW, the identification of the actual average PTA effects is not straightforward, however. In the SW dataset, trading partners that are PTA members are also likely to be WTO members.17 Consequently, the aggregate PTA coefficient estimated in equation 1 (and reported in Table 1a) represents the change in bilateral trade generated by both PTA and WTO membership. Under mutually exclusive coding, the pure PTA effects in Table 1 can be obtained by calculating a composite coefficient that subtracts the WTO coefficient from its PTA counterpart. The direct regression output for PTA coefficients needed in this calculation is presented in Table 1a (standard errors are calculated using the Delta Method, see Greene 2003).

Table 1a:

Raw Regression Output GSP and PTA Coefficient Estimates including WTO Effects

(Mutually exclusive coding, corresponding to results in Table 1)

article image
Notes: *,**, *** are 10%, 5%, 1% significance levels. Standard errors clustered by country-pairs in parentheses. Coefficients of Fixed Effect controls are suppressed. Remaining coefficients (WTO regressors and historical/geographical controls) are as in corresponding regressions in Table 1. F-Statistics for the inclusion of additional PTAs cannot be computed due to the mutually exclusive coding of the WTO dummies. All regressions include time-varying importer and exporter fixed effects.

The SW regressions (regression 1, Table 1a) imply that bilateral imports increase 234% (=e1.205-1) for the average PTA and the average PTA member. The “pure” PTA effect, net of any WTO effects, is provided in Table 1, regression 1 for industrialized and developing countries. Here we find that bilateral trade increases at a dramatically different rate for industrialized and developing country PTA members. Industrialized countries see their trade increase by a meager 16% (=e0.152-1) while developing country trade increases 214% (=e1.143-1).

Regression 2 (Table 1), allows for individual PTA effects. We find that all multilateral trade dummies are heavily impacted, confirming the suspicion of potential upward bias in regression 1. Most notably, we find that the economic and statistical significance of WTO membership for industrialized countries was reduced by an order of magnitude. WTO induced trade creation for industrialized countries falls from 187% in regression 1 to 80% (=e0.588-1) in regression 2. This suggests that the sizable trade creation that was attributed to WTO membership in regression 1 is more accurately associated with individual PTAs. The precision of the individual PTA estimates together with the F-Statistic (which rejects regression 1 in favor of regression 2) supports the suspicion that regression 1 suffers from substantial omitted variable bias.

The effects of individual PTAs are all trade creating and highly statistically significant (with the exception of the EU, a case we will discuss at length below). However, PTAs promote trade strongly, but unevenly: Magnitudes of PTA coefficients differ dramatically, ranging from 770% (=e2.162-1) for NAFTA to 76% (=exp0 568-1) for the ASEAN free trade agreement (AFTA). Regression 3 controls for additional PTAs that were not included in SW (or Rose 2005), but feature prominently in the empirical PTA literature.18 EFTA, EEA, LAIA, and APEC are all shown to be highly trade creating. Once we account for the individual PTA effects with the most comprehensive set of PTAs, we find that the SW result of positive WTO effects in industrial countries vanishes completely.

Allowing for individual PTA effects also generates seemingly bleak insights for developing countries. The effects of WTO membership are estimated to be either non-existent (regressions 1, 2) or negative and statistically significant (-21% = e-0.235-1 in regression 3). On the upside, however, PTA coefficients indicate strong trade creation for all PTAs involving developing countries. It may well be the case that developing countries reoriented their import activity considerably towards PTA partners after joining PTAs. This reorientation might produce trade creation, but it might also include some trade diversion that redirected trade from WTO trade partners to fellow PTA members. If this is a common pattern among developing countries, such a reorientation would have a negative impact on the WTO estimate for developing countries, given the hierarchical, mutually exclusive coding of SW.19 The largest levels of trade creation in regression 3 are observed for PTAs that consist of developing nations (CACM, CARICOM, MERCOSUR, NAFTA-Developing and SPARTECA-Developing) where PTA-internal trade is estimated to be roughly 350% (=e1.5-1) greater than PTA-external trade.

B. Hierarchical, Mutually Exclusive Coding and SW’s “Implicit Industrialized PTA Dummy”

The results from our specifications that allow for individual PTA effects suggest that PTA trade creation is not homogeneous across countries. To the contrary, bilateral PTA trade is estimated to be significantly larger for PTAs that consist of developing countries than it is for PTAs among industrialized nations. Ignoring these trade creation differentials, introduces a specific type of omitted variable bias into SW style regressions. This section discusses the specific bias and shows its particular effect on the estimate for industrialized countries’ WTO trade creation.

We will show that the industrialized WTO coefficient in SW’s preferred regression (regression 1) functions simply as an error-correction term. The term allows for an implicit split of SW’s aggregate PTA coefficient into one for industrialized countries’ PTAs (with low trade creation) and one for developing countries’ PTAs (with high creation). The mechanics of this implicit split requires additional explanation. In effect, an implicit industrialized PTA dummy is generated by the interaction between a) SW’s coding convention, b) the nature of industrialized countries’ WTO/PTA accession, and c) multilateral resistance controls (time-varying importer dummies). To clarify this interaction, it is important to understand that all industrialized importers in the dataset joined WTO before joining a PTA. Given SW’s coding convention, this implies that after WTO accession, all industrialized importer observations feature the value “1” for either the WTO/GSP dummy, or the PTA dummy. This implies that, for industrialized countries, the linear combination of these two dummies is perfectly collinear with the time- varying importer dummy.

To confirm, we can perform a quick experiment by explicitly partitioning the aggregate PTA dummy into two: one for industrialized and another for developing importer observations featuring PTA relationships. The experiment, reported in regression 1a, results in perfect multicollinearity forcing the industrial PTA dummy to be dropped from the regression. The results of regression 1a are exactly identical to those in regression 1. The only difference is that it is possible to directly identify industrialized PTA trade creation in regression 1a (from the WTO importer dummy, since it is perfectly collinear with the industrialized PTA dummy).20

For regression 1, the collinearity between the time-varying importer dummies and the linear combination of the industrialized countries’ WTO and PTA dummies has important implications. Since only two (WTO and time-varying importer) of the three dummies are included, the third is implied. Given the nature of industrialized countries’ WTO accession, the omitted dummy happens to be identical to all industrialized countries that entered a PTA. This is the source of the implicit industrialized PTA dummy in regression 1. Regression 1’s industrialized WTO coefficient does not reveal WTO trade creation, but trade creation that is due to industrialized PTAs.

Note also, that there is a flipside to the implied industrialized PTA dummy. Regression 1a proves that developing countries’ PTA effects exclusively determined the “aggregate PTA” effect in regression 1 (aggregate in quotations, since we now know that this effect has been purged of all industrialized countries’ effects, which are accounted for by the “WTO coefficient” in regression 1). This of course is nothing more than the flipside of our insight that industrialized PTA effects are entirely expressed through the industrialized WTO dummy.

IV. Individual PTA Effects, Multilateral Resistance, and Unobserved Bilateral Heterogeneity

There exits substantial evidence that trade agreements tend to form between trading partners whose bilateral trade has been “naturally” elevated all along, due to unobserved characteristics (see e.g. Baier and Bergstrand, 2007). This implies that some trade with the agreement may be caused by “natural” trade-promoting characteristics and is independent of the agreement. SW attempted to control for such country-pair specific characteristics with the inclusion of a number of control variables. When the included controls do not account for all unobserved bilateral heterogeneity, and when omitted variables favor trade, the WTO and PTA coefficients are biased upwards.

The PTA literature has long considered the AP, LAIA and APEC estimates suspiciously high relative to the small tariff reductions associated with these agreements (e.g. Frankel, 1992; Frankel and Wei, 1993; Frankel, Stein and Wei, 1995; Frankel, 1997). Given that our estimates in Table 1 (regressions 1-3) are of similar magnitude as those found in the previous literature, our results confirm that such magnitudes are robust to the inclusion of both the individual PTA effects and multilateral resistance. Egger and Pfaffermayr (2003) and Cheng and Wall (2005) have previously confirmed that implausibly large estimates can be lowered in the canonical gravity equation by taking unobserved bilateral heterogeneity into account (neither study includes GSP or WTO effects).

To fully account for unobserved bilateral heterogeneity relationships in equation (1), we control for all time-invariant bilateral heterogeneity by introducing country-pair fixed effects. This can be achieved by replacing the intercept in equation (1) by a country-pair specific one, amx:


Equation (2) features fewer explanatory variables than equation 1 because all time-invariant regressors are now absorbed into the pair specific fixed effects.

A. WTO Trade Impact: Controlling for Multilateral Resistance, Unobserved Bilateral Heterogeneity and Individual PTA Effects (Hierarchical/Mutually Exclusive WTO Coding)

Regressions 4-6 in Table 1 present analogues of regressions 1-3 above. The only differences are the added country-pair specific fixed effects that constitute comprehensive controls for unobserved bilateral heterogeneity in regressions 4-6. Two areas are the focus of our interest: a) whether WTO trade effects are influenced by unobserved bilateral heterogeneity, and b) whether PTA effects are reduced to plausible ranges.

Regression 4 is a replication of the original SW specification (2007, their Table 4, regression 4) with the addition of country-pair fixed effects.21 The results show a substantial reduction of the WTO’s economic and statistical significance for industrialized countries. WTO trade creation falls from 178% in regression 1 to 48% (=e0.393 -1) for industrialized countries in regression 4, and its significance fails to reach the 1% level. The inclusion of specific PTA trade effects in regressions 5 and 6 also negates all WTO trade effects for industrialized countries. Regression 5 even shows that the simple disaggregation of SW’s own aggregate PTA dummy into individual agreements is sufficient to neutralize any industrialized WTO trade effects, once we account for heterogeneity in bilateral relationships. For developing countries, in sharp contrast, regressions 4-6 illustrate that the GSP and WTO effects are hardly impacted by unobservable heterogeneity. Their estimates are closely aligned to those in regressions 1-3.

The F statistics confirm the importance of the inclusion of comprehensive country-pair fixed effects as well as individual PTA effects at significance levels that exceed 0.001%. The added controls also generate reduced (and more plausible) trade impacts for individual PTAs. This provides evidence that unobserved bilateral heterogeneity is generally trade-enhancing. Average trade creation across PTAs drops from 234% in regression 1 to 123% in regression 4 (Table 1a). Most individual trade agreements see their trade effects at least halved in regressions 5 and 6. In contrast, trade creation in the EU increases, and the EU coefficient is now estimated with considerable precision.

Before we present a detailed discussion of individual PTA trade effects in Section 6, we conduct a robustness analysis. The expanded SW results that have been presented in regressions 1-6 give rise to the question as to whether the absence of WTO effects may be an artifact of the hierarchical, mutually exclusive coding in SW. As discussed above, when increases in trade flows are attributed to PTAs rather than to both, WTO and PTA membership, one may suspect that the SW coding convention underestimates WTO trade effects. In the next section, we apply alternative, mutually inclusive WTO coding (as in Rose, 2004) to allow for separate identification of WTO and PTA effects. This robustness analysis has two purposes. Not only will this settle whether the SW coding convention is driving the results, but it is also a substantive extension of Rose (2004, 2005), because we extend his specification to introduce both disaggregated PTAs as well as comprehensive multilateral resistance controls.

V. Individual PTA Trade Effects, Multilateral Resistance, and Unobserved Bilateral Heterogeneity (Mutually Inclusive WTO Coding)

Rose (2004 and 2005) controlled for unobserved bilateral heterogeneity through country-pair fixed effects. Nevertheless, SW lamented that he did not control fully for multilateral resistance. By introducing comprehensive controls for multilateral resistance to Rose’s dataset, we unify the SW and Rose approaches. Our approach controls for all three key determinants of trade under both coding conventions: unobserved bilateral heterogeneity, multilateral resistance, and individual PTA effects. The unified approach allows us to highlight whether any results are due to mutually inclusive (Rose) or mutually exclusive (SW) coding of WTO dummies. To allow for a comparison between Rose and SW coding results, we split Rose’s inclusive WTO-dummy into SW-style indicators for industrialized and developing importers’ WTO membership.

Our robustness analysis reexamines regressions 1-6 using inclusive trade agreement coding. Our first set of regressions, which includes multilateral resistance but excludes country-pair fixed effects, is then given by


The only difference between (1) and (1’) is the coding convention, where inclusive coding is denoted by a “**” superscript.22 WTOmxt** indicates that both trading partners are WTO members and GSPmxt**identifies industrial countries’ GSP imports. Crucial is that under Rose’s coding convention both WTOmxt**and GSPmxt**take the value “1” when the two conditions are fulfilled. In addition, when the same trading partners are members in a common PTA, inclusive coding assigns the value “1” to all three dummies. For comparison purposes, it is important to point out that inclusive coding delivers coefficient estimates that represent pure PTA effects, while pure PTA effects had to be established via composite coefficients in SW’s coding convention (as discussed in Section 3, Table 1 and 1a.)

We proceed again in stages. First we provide results based on equation (1’), and then we incorporate country-pair fixed effects to control for all time-invariant bilateral heterogeneity. Inclusion of country-pair effects again converts the constant a to a pair-specific one, amx :


A. WTO Trade Impact: Controlling for Multilateral Resistance, Unobserved Bilateral Heterogeneity and Individual PTA Effects (Mutually Inclusive WTO Coding)

Regression 7 in Table 2 establishes a baseline regression that represents our closest analogue to Rose’s default regression (2004, Table 1). Regression 7 is a robustness test of Rose’s findings that examines whether a comprehensive account of multilateral resistance affects his original results. At the same time, regression 7 also provides a robustness check of SW’s preferred regression (Table 1, regression 1) to examine whether results are affected by the coding convention.

Table 2:

WTO and PTA Effects:

(Inclusive Coding)

article image
Notes: *, **, *** are 10%, 5%, 1% significance levels. Standard errors clustered by country-pairs in parenthess. Cofficients of Fixed Effect controls are suppressed. All regressions include time-varying importer and exporter fixed effects.

Rose’s preferred regressions (2004, Table 1) report insignificant WTO effects throughout. The insertion of multilateral resistance controls does not change Rose’s conclusions regarding WTO effects. Regression 7 shows that trade creation due to WTO membership for both industrialized and developing countries is insignificant. Even when the WTO dummies in regression 7 are aggregated to one WTO dummy (for industrial and developing countries), the combined effect is insignificant.23 The only result that does change is that Rose’s GSP effect is now eliminated.

The comparison of regressions 1 and 7 highlights that the coding convention matters to the insights derived from the individual regressions, but it does not affect the conclusions. Mutually inclusive coding renders the WTO effect statistically and economically insignificant even when individual PTA effects are not introduced. On the other hand, mutually exclusive coding delivered insignificant WTO effects only after individual PTA effects have been fully taken into account. The reason lies in our discussion in Section 3.2. PTA coefficients under mutually inclusive coding provide net effects, which implies that the industrialized WTO dummy cannot function as a error-correction term. It is not possible to implicitly split the aggregate PTA variable into one for North-North PTAs (with net lower trade creation) and one for South-South PTAs (with higher net trade creation) as we had seen in Sections 3 and 4. As a result, mutually inclusive coding in Table 2 never delivers significant WTO coefficients.

The crucial insight is that when variables are coded to represent only their net effects on trade (as in Rose, 2004), the risk of omitted variable bias is greatly reduced. Under mutually exclusive coding, however, estimated PTA coefficients include both PTA and WTO effects. This is the reason why the individual PTA effects in Tables 1 and 2 are strikingly similar, while the “gross” PTA effects in Table 1a exhibit little resemblance to those in Table 2. Mutually exclusive coding thus holds the danger that WTO dummies are biased when the following two conditions hold: (1) industrialized and developing PTAs differ considerably in their trade effects and (2) individual PTA effects are constrained to an average coefficient associated with one aggregate PTA dummy.

Regressions 8 and 9 (the inclusive-coding analogues of regressions 2 and 3) introduce individual PTA effects and represent two further robustness tests. The first test is whether Rose’s (2005) results of a small, positive WTO effect are robust to controlling for multilateral resistance. At the same time, regressions 8 and 9 represent a second robustness test that examines whether SW’s WTO effect vanishes only because of their coding convention. Recall that under mutually exclusive coding the introduction of additional PTAs reduces the number WTO observations, which is not the case under Rose’s mutually inclusive coding. Regressions 8 and 9 overturn Rose’s (2005) result of a statistically significant WTO effect when individual PTA effects are considered.24 Hence Rose’s (2005) finding of small (but significant) WTO effects came about only because he did not control comprehensively for multilateral resistance. The regressions that allow for distinct effects across preferential trade agreements also highlight that the vanishing WTO effect in SW - after we controlled for individual PTA effects - was indeed only due to SW’s hierarchical and mutually exclusive coding. It is thus essential to validate our WTO results by using the most comprehensive set of controls in both datasets.

Regressions 10-12 are the analogs to regressions 4-6, where the only difference in the datasets is the coding convention. With inclusive coding we find no WTO trade effects in either of these regressions. Regressions 12 and 6 represent our preferred specifications overall, since they contain all three key controls: multilateral resistance, natural trading partner effects and individual PTA effects Both regressions confirm the absence of trade gains from WTO membership independent of the coding convention and these regressions are also clearly favored by the F-statistics. More generally, we find that, without exception, the inclusion of individual PTA trade effects always improves the estimation no matter which set of controls is selected.

VI. Individual PTA Trade Effects: Sensitivity to Unobserved Bilateral Heterogeneity and Multilateral Resistance Controls

The byproduct of the thorough investigation of WTO trade effects above is a set of estimates for individual PTA effects produced with the use of the most comprehensive set of controls to date. Given the voluminous literature on the effects of PTAs on trade flows, these estimates are important in their own right. In our discussion of the individual PTA effects below, our main focus will be on our preferred specification, regression 12. Regression 12 features the most comprehensive set of controls and its econometric specification is identical to regression 6. The two regressions differ only in terms of their coding convention and produce virtually identical results. PTA estimates in regressions 12 and 6 generally do not differ by more than one standard deviation. We will use regression 9 as a reference to regression 12 in order to highlight the importance of bilateral heterogeneity controls.

Before we start our discussion of specific PTA effects, it is important to recall the exact interpretation of the PTA coefficients in regression 12. They indicate how much PTA-internal trade has increased relative to trade with non-members after the agreement was instituted. The percentage increases below are thus mixtures of trade creation and trade diversion.25 While we find no statistically significant negative net trade effects from PTA membership, even insignificant PTA estimates do not imply that the PTA was necessarily ineffective in creating trade flows. It could simply indicate that PTA-internal trade grew at the rate of external trade (as, for example, in the case of the export oriented ASEAN countries).

The individual trade effects of PTAs in regression 12 are the most “reasonable” among all our regressions, in the sense that the net trade creation for most PTAs is estimated to range between 30% and 80%. Curiously, the Central and Latin American Trade Agreements (CACM, CARICOM, MERCOSUR and LAIA) show by far the largest increases in relative trade. They are also the only PTAs that report net increases in trade creation of over 100% (with the exception of SPARTECA, which reports an increase of 124% = e0.807 -1).

Most notable is the reduction in the estimated net trade creation for most PTAs, after we control for omitted variable bias. Comparing regression 9 to 12 reveals that, with the exception of the EU, CACM and LAIA, all PTA estimates are substantially reduced when we include country- pair fixed effects. In other words, our results suggest strongly that PTAs are formed between countries that have all along been sharing characteristics favorable to mutual trade. In this case, tariff reduction may simply be an afterthought. Controlling for unobserved bilateral heterogeneity also improves the precision of the estimates in all cases but NAFTA. The suspiciously large net trade creation of NAFTA (230%) in regression 9 is reduced to insignificance after we control for unobserved bilateral heterogeneity.

Controlling for unobserved bilateral heterogeneity also increased trade creation for three PTAs (EU, EFTA and CACM). Given these PTA member countries’ characteristics, their actual trade flows are not large enough relative to the prediction of the gravity model. For example, in the case of CACM, all countries share a common language, colonizer, and very low distance. The introduction of country-pair fixed effects resolves the systematic overprediction of the gravity model and allows a better assessment of the impact of PTA accession.

It may initially be surprising that predicted trade volumes also exceed observed levels in the EU. The fact that the EU trade impact is underestimated in both the traditional and multilateral resistance-augmented versions of the gravity equation is, however, well known (e.g., Aitken, 1973; Soloaga and Winters, 2001; and Rose, 2004). Our estimates show a statistically significant 37% (= e 0.312-1) increase in trade due to EU accession, once we control for unobserved bilateral heterogeneity. In the case of the EU, it is likely that the large market and the strong harmonization efforts allowed firms to overcome trade fixed costs that subsequently lead to strong trade creation between both member and nonmember countries (e.g. Freund, 2000; Melitz, 2003). The indisputable increase in absolute trade volume among EU members is then relatively small compared to the trade increase with non-members. It is also important to note that EU members automatically reaped another 34% trade benefit when they became members of the EEA in 1994 (since the overwhelming majority of EEA members are EU countries).

Another trade agreement that has been the subject of great interest in the PTA literature is APEC. The highly significant and truly exorbitant APEC trade creation estimate (around 300%) is common in gravity models, although the agreement is only a consultative forum without implications for tariffs (see e.g. Frankel and Wei, 1993; Frankel, Stein and Wei, 1995; Frankel, 1997). Regression 9 indicates that the inclusion of multilateral resistance lowers values for the APEC coefficient substantially to 123% (=e0.802. -1). The additional inclusion of unobserved bilateral heterogeneity controls shows that much of the trade creation originally attributed to APEC has actually been due to bilateral unobservables. While this has been the suspicion of Frankel and coauthors all along, their quest to identify these unobservable drivers has largely been unsuccessful. In our preferred regression 12, which controls for multilateral resistance, unobserved bilateral heterogeneity and individual PTA effects, APEC’s trade creation estimate drops to a meager 28% (=e0.244-1).

VII. Conclusion

This paper establishes robust estimates of the impact of WTO and PTA accession on bilateral trade flows. We control for omitted variable bias in three dimensions: individual PTA effects, multilateral resistance, and unobserved bilateral heterogeneity. Our results show a clear absence of WTO trade effects, and our estimates are nearly identical across the different coding conventions of the WTO dummy (Rose, 2004; Subramanian and Wei, 2007; or Tomz, Goldstein and Rivers, 2007). Our comprehensive account for omitted variable bias allows us to unify the previous WTO and PTA literatures. Our analysis highlights, that previously diverging results regarding WTO effects on bilateral trade flows (most prominently in Rose, 2004, 2005; Subramanian and Wei, 2007; and Tomz, Goldstein and Rivers, 2007) may have been generated by omitted variable bias. F-Tests indicate that the most extensive specification, which controls for all three dimensions of omitted variable bias, is preferred.

The magnitude of the resulting individual PTA estimates resolves a number of empirical puzzles. Most notably, the non-tariff reducing APEC is shown to exert comparatively little trade impact, and the strongly tariff reducing EU is shown to be trade creating. Trade theory motivates the inclusion of comprehensive multilateral resistance controls to pick up variations in relative trade costs. These controls are shown to be insufficient to generate unbiased estimates of trade agreements’ impacts on trade flows. Of crucial importance are also country- pair fixed effects that control for unobserved bilateral characteristics. While a significant share of trade is shown to be country-pair specific, the importance of such characteristics is not fully addressed in current trade theories.

In Search of WTO Trade Effects: Preferential Trade Agreements Promote Trade Strongly, But Unevenly
Author: Theo S. Eicher and Christian Henn