The effect of nonreciprocal preferential trade agreements on benefactors’ exports

For the list of NRPTAs and when they entered into force, see Gil-Pareja et al. (2014).

In a recent survey on panel econometric methods to gravity modeling, Baltagi et al. (2015, p. 610) note that “even an analysis of a cross-section of county pairs (based on a single year s or an average across several years) in essence involves panel data, ...”. As these authors point out, Pöyhönen (1963) is probably the first to control for country-specific fixed effects in cross-sectional data.

The exporter and importer effects capture all time-invariant observable and unobservable country characteristics. These effects could either be treated as fixed (Mátyás 1997) or as random (Mátyás 1998) and being part of the error term. Egger (2000) shows that the fixed-effects model is the right choice on conceptual and econometric grounds. Time effects control for common shocks and trend shared by all countries. They are treated as fixed.

The bilateral effects control for the impact of any time-invariant determinant of trade (observed or not).

Baltagi et al. (2003) propose for the first time the inclusion of fixed exporter-time and importer-time effects. As they point out, these effects are included to capture country-specific time-varying effects like the exporter (importer) country’s business cycle, its cultural, political or institutional characteristics and unobserved factor endowment variables.

Baldwin and Taglioni (2006) generalize Anderson and van Wincoop (2003) framework (which is limited to cross-section data), to allow for panel data and then show that multilateral resistance terms can be dealt with using country-and-time dummies (it, jt) with omitted determinants of bilateral trade being dealt with by time-invariant bilateral dummies.

In this paper we estimate partial (or direct) effects, not general equilibrium effects as in Anderson and van Wincoop (2003), Baier and Bergstrand (2009), Egger and Larch (2011), and Bergstrand et al. (2013).

We use data at four-year intervals as Bergstrand et al. (2013) and Gil-Pareja et al. (2014) do and akin to Chen and Wall (2005), Baier and Bergstrand (2007), Subramanian and Wei (2007), Eicher and Henn (2011a, b), Beahr and Cirera-i-Crivillé (2013) and Kohl (2014) use of data for every five years. The use of this kind of data addresses the concern raised by Chen and Wall (2005, p. 52): “Fixed-effects estimation is sometimes criticized when applied to data pooled over consecutive years on the grounds that dependent and independent variables cannot fully adjust in a single year’s time.”

The sample considered by Bergstrand et al. (2013) only includes 41 trading partners and four years (1990, 1994, 1998 and 2002), which allow them to use the first approach. The large datasets considered by Baier et al. (2014) and Gil-Pareja et al. (2014) preclude to estimate that specification since the huge number of fixed effects required is beyond the capability of commonly used statistical software. The same applies for this paper.

In this study we use the expression “preferential trade agreement” to also refer to other agreements involving a higher degree of economic integration. In fact, most economic integration agreements considered in the sample are free trade agreements.

See, http://​www.​agoa.​gov/​eligibility/​country_​eligibility.​html for membership in AGOA and http://​ec.​europa.​eu/​trade/​wider-agenda/​development/​generalised-system-of-preferences/​everything-but-arms for EBA.

http://​ec.​europa.​eu/​trade/​wider-agenda/​development/​economic-partnerships.

The index is defined as: (% Protestants in country i * % Protestants in country j) + (% Catholics in country i * % Catholics in country j) + (% Muslims in Country i * % Muslims in country j).

The Wald test rejects, at the 10 % level of significance, the null hypothesis of equality between the estimated coefficients for XNRPTA and MNRPTA.

These authors argue that countries likely select endogenously into free trade agreements (FTA) and that the most plausible estimates for the effect of an FTA on trade are obtained using panel data with bilateral fixed effects (or differenced panel data) and country-and-time fixed effects.

This result is consistent with that provided by Eicher and Henn (2011a, b) who find smaller estimated coefficients when they add unobserved bilateral heterogeneity controls. In contrast, Baier and Bergstrand (2007) find a larger estimated coefficient for free trade agreements accounting for both country-and-time fixed effects and country-pair fixed effects (0.48) than using OLS with time dummies (0.27).

In the random effects specification the country-pair effects are considered as part of the error term. The main drawback of this specification lies in its restrictive exogeneity assumption because, if there is correlation between the covariates and the country-pair effects, parameter estimates with the random effects model are biased and inconsistent, while the fixed country-pair effects estimator is immune to the potential problem of correlation.

In the specifications with lags we report the sum of the estimated coefficients from current and lagged values.

The first difference approach clearly reveals the importance of incorporating lagged effects. The results with first-differenced data without accounting for lagged effects (not reported) show that only the variable of interest presents a coefficient estimate that is positive (0.091) and statistically significant at least at the 5 per cent level of significance. Moreover, the coefficient of the variable PTA is 0.062 (statistically significant at the 10 per cent level), whereas neither CU, GATT/WTO nor XNRPTA dummies have positive significant effects at conventional levels. This result is in line with Kohl (2014, p. 453) who finds, using first differences with data at five-year intervals over the period 1950–2010, that the impact of economic integration agreements on trade is only statistically significant when allowing for phase-in periods with lags. In particular, he finds a positive (0.058) but statistically insignificant coefficient estimate without lags and a cumulative treatment effect equal to 0.39 with two lags.

Baier et al. (2014) do not consider currency unions or membership in GATT/WTO in their analysis of the effects of various types of economic integration agreements on trade flows and trade margins.

Monte Carlo simulations lead Head and Mayer (2013, p. 50) to conclude that “While Poisson PML has many virtues... it should not replace OLS as the new workhorse estimator of gravity equations. Rather, Poisson PML should be used as part of a robustness-exploring ensemble...”

Since the Poisson estimator did not achieve convergence including time-varying fixed effects in addition to country-pair fixed effects, in the regressions with pair effects we include instead GDPs.