Economic geography in the presence of preferential trade agreements: an empirical analysis of OECD data

Abstract

In the presence of preferential trade agreements (PTAs), there are two distinct levels of economic geography–global economic geography (GEG) between a PTA bloc and the world and local economic geography (LEG) between a PTA member and its PTA bloc. Using OECD data, we empirically examine whether both types of economic geography can help to explain the international production structure. Our findings provide some support for the importance of distinguishing between LEG and GEG when testing for the presence of economic geography.

Notes

¹ For example, political commitment and leadership by governments, especially those of Germany and France, to foster economic integration as a means of preventing conflicts and wars played a central role in the birth and evolution of the European Union (EU), the world's most successful PTA.

² At a theoretical level, Baldwin and Venables (1995) and Puga and Venables (1997) show that lower inter-bloc trade costs can shift production toward a PTA bloc and sufficiently small intra-bloc trade costs can lead to geographically concentrated production even within a PTA bloc.

³ Although Davis and Weinstein (1996) use the world demand for a good, we follow Davis and Weinstein (2003) in using the demand for a good in the rest of the world. For example, if we are looking at Japan, the world demand includes Japan's demand but demand in the rest of the world excludes Japan's demand. Davis and Weinstein (2003) point out that using world demand may lead to serious identification problems and thus replace it with demand in the rest of the world.

⁴ Davis and Weinstein (1996) explain the relationship between the estimated coefficient and the three hypotheses. If β ₂ = 0, then we are in a world of frictionless comparative advantage in which transport costs do not matter. If 0 < β ₂ ≤ 1, then we are in a world in which transport costs and demand patters affect the location of production, but there are no economies of scale. Finally, if β ₂ > 1, then we are in a world in which home market effects associated with economies of scale help determine the location of production.

⁵ Please refer to Footnote 4. The only difference is that β ₃ > 1 supports the presence of GEG whereas β ₂ > 1 supports the presence of local geography.

⁶ However, the degree of correlation between the dependent variable and the explanatory variables may differ considerably. Indeed the Spearman correlation coefficients indicate that this is the case.

⁷ Results from the more conventional Durbin-Hu–Hausman test for endogeneity confirm the results of the C-test. Please refer to Davidson and Mackinnon (1993) for an extended discussion of the Durbin-Hu–Hausman test. The results are available from the authors upon request.

⁸ This condition is specified in p. 93 of Wooldridge (2002) as well as p. 4 of Baum et al . (2003).

⁹ However, GMM assumes that the error terms are orthogonal, i.e. independent of the instruments.

¹⁰ This means we have to consider the relationships among the error terms in addition to modelling heteroskedasticity.

¹¹ But we do not follow Davis and Weinstein's implementation of instrumental variables because doing so would lead to inconsistent and biased estimates. Please refer to pp. 90–1 in Wooldridge (2002).

¹² We can solve the model as the following system of equations and substitute Equations b to d into Equation a.

¹³ In fact, Davis and Weinstein (2003) note that the coefficient estimates in Davis and Weinstein (1996) show larger SEs and deviate from unity whenever endowments are included in the regression. They identify the cause as the multicollinearity between SHARE and endowments. To take care of multicollinearity, Davis and Weinstein (2003) exclude SHARE when endowments are included as instruments and regressors.

¹⁴ The countries and industries in our sample are identical to those of Davis and Weinstein (1996), which provides a more detailed discussion of the COMTAP data set.

¹⁵ Indeed our preliminary empirical analysis confirmed that using per-capita production rather than total production sharply reduces heteroskedasticity. Furthermore, our analysis also indicated multicollinearity among explanatory variables when we used total values rather than per-capita values. We obtained the population data required to compute per-capita values from World Development Indicator (CD-ROM, 2004).

¹⁶ The earlier version of the Penn World Tables used by Davis and Weinstein did not contain capital stock data so they had to compute the stocks from investment and price data. Source OECD provides more comprehensive energy data on OECD countries than the UN data source used by Davis and Weinstein.

¹⁷ As reported in the UNESCO Statistical Yearbook, the levels of education for each country are the same for different years and industries. We already incorporate this effect in the panel regressions when we address the country-specific effects so we cannot use it as an instrumental variable. However, we still need a third instrument for our model to be identified. Hence, we use another measure of education. The data are from the World Development Indicator (CD-ROM, 2004).