The WTO puzzle, multilateral resistance terms and multicollinearity

Abstract

Since Rose’s (2004) striking finding of negligible WTO trade effects, numerous studies have attempted to solve the so-called WTO puzzle. These studies have progressively improved model specifications to control for potential sources of bias, but they often lead to conflicting results. Multilateral resistance terms (MRTs) are considered to be one of the most crucial factors to be accounted for to avoid the omitted variable bias in the gravity model. What has gone unnoticed, however, is that the control for MRTs leads to near-prefect multicollinearity because of the structural relationship between the variables that measure the General Agreement on Tariffs and Trade/World Trade Organization (GATT/WTO) membership status of any country pairs. This multicollinearity contributes to the fragility of the WTO effect estimates, as it dramatically increases SEs. In this article, we explain how this multicollinearity arises and provide evidence of it.

Keywords:

• World Trade Organization
• gravity model
• WTO puzzle
• multilateral resistance terms
• multicollinearity

JEL Classification:

• F13
• F15

Acknowledgement

We are very grateful to Xuepeng Liu for kindly providing his data set.

Notes

¹ Liu (2009) tries to account for MRTs using the remoteness variable; but as mentioned in Anderson and van Wincoop (2003), this approach is atheoretical as it only captures trade barriers from the geographical distance, which is not the focus of the analysis in this literature.

² In their latest version, Dutt et al. (2013) report only the results with MRTs.

³ Nonparametric methods do not suffer from this multicollinearity problem. Chang and Lee (2011) use a matching method and obtain positive WTO trade effects using Rose’s data. Unobserved country heterogeneity, which is another source of bias in the estimations, however, can not be easily incorporated into the matching method. Moreover, parametric methods are still more popular in this literature (e.g. see Dutt et al., 2013).

⁴ The perfect negative cross-sectional correlation also exists between and . Thus, using Nonein instead of Onein in the regression does not avoid the multicollinearity problem.

⁵ Roy (2011) also uses practically the same data set. Liu’s (2009) data cover more countries than those from Rose (2004); Subramanian and Wei (2007), among others.

⁶ The correlation is obtained using the residuals from regressing the WTO membership variables on CTFE dummy variables. We have also computed the correlation using the Bonus vetus OLS method suggested by Baier and Bergstrand (2009) to control for MRTs instead of using CTFEs. However, that merely reduces the conditional correlations between Onein and Bothin from –1 to –0.99.

⁷ The literature points out many other sources of bias including the definition of memberships, unobserved heterogeneity and zero trade (see e.g. Tomz et al. (2007); Helpman et al. (2008); Eicher and Henn (2011)). However, as our main purpose is to highlight the multicollinearity problem, which exists independent of other sources of bias, we do not consider those other sources of bias in our estimations.