Trade and regional growth in Spain: panel cointegration in a small sample

Abstract

In this article, we test for the existence of a relationship between per capita Gross Domestic Product (GDP) and trade, for 15 Spanish Autonomous Communities between 1988 and 2004, using a panel cointegration methodology. In particular, we implement several panel unit root tests (Maddala and Wu, 1999; Levin et al., 2002; Im et al., 2003) and panel cointegration tests (Pedroni, 1999, 2004), with a special attention to their behaviour in a small sample. We also develop a Seemingly Unrelated Regression Estimation (SURE) residual based test, in order to explicitly take into account the cross regional correlation pattern. Appropriate confidence intervals are estimated with a sieve bootstrap designed for our small time sample, preserving the dependence structure among cross sectional units. Our cointegration tests reject the existence of a significant relationship between GDP per capita and exports. However, we do find some evidence of a significant relationship between GDP per capita and imports or with total trade.

Acknowledgements

The research presented in this article was done at the Center for Operation Research and Econometrics (CORE). I would like to thank Jacques Thisse, Dominique Peeters, Hylke Vandenbussche and Luc Bauwens for their helpful comments and suggestions. This text presents research results of the Belgian Program on Interuniversity Poles of Attraction initiated by the Belgian State, Prime Minister's Office, Science Policy Programming. The scientific responsibility is assumed by the author.

Notes

¹ See Lewer and Van den Berg (2003) for an empirical review of literature.

² For instance, the international specialization of production, the agglomeration forces shaping the location of industry, external scale economies (such as technological externalities) are all dynamic regional processes.

³ To the best of our knowledge, Paluzie and Lafourcade (2005) is the only other paper using this database.

⁴ Indeed, Shan and Tian (1998) have shown the importance of controlling for imports to avoid a spurious causality result between exports and growth.

⁵ As shown in Xu (1996), the failure to support causality between export and growth may be attributable to the choice of lags in the VAR system.

⁶ But accordingly, this technique identifies neither the particular channels governing this relationship nor the direction of causality.

⁷ In Laurin (2007b), we show some cointegration tests based on the Phillips–Perron unit-root tests. Results are rarely significant. The nonparametric correction for serial correlation proposed by Phillips–Perron might be more sensitive in a small sample; in particular, the estimation of the long-term variance could be more imprecise with a reduced number of time observations.

⁸ In fact, the MWF test can combine the p-value of any unit root test (Maddala and Wu, 1999). They show that their test has a higher power than the IPS and LLC tests.

⁹ Which corresponds, in Pedroni's words, to pooling along the between dimension of the panel.

¹⁰ T = 60 is chosen to replicate the number of observations generally available in annual cross-country macroeconomic databases used by empirical papers (the Penn World Tables for example, starting in 1950).

¹¹ This test must be interpreted with caution since the null hypothesis implies a stationary process for all units.

¹² Usually, the empirical literature imposes homogeneity of the long-term coefficients in the cointegration equation (Equation 1). Considering the theoretical literature on trade and regional growth, we rather take the opposite view. The relationship equilibrium between growth and trade can be heterogeneous across regions, because of divergent industrial structures, economic geography, productivity gaps, capacity to adapt to trade liberalization, differences in geography, the cumulative process of agglomeration, etc.

¹³ Bootstrap statistics for the other specifications are available upon request.

¹⁴ With the exceptions of the Pedroni variance ratio test and Panel rho-stat.: in a small sample, it might be difficult to estimate the long-term variance correctly.

¹⁵ Data on intra-Spanish trade between ACs themselves are unfortunately not available.

¹⁶ Andalusia, Aragon, Asturias, Basque country (P. Vasco), Cantabria, Catalonia, Galicia, Castilla y Leon (C. Leon), Castilla y Mancha (C. Mancha), Comunidad Floral de Navarra (Navarra), Comunidad Valenciana (Valenciana), Extremadura, La Rioja (Rioja), Madrid and Region of Murcia (Murcia).

¹⁷ The detailed unit root tests for each variable can be obtained in Laurin (2007b).

¹⁸ Usual spatial autocorrelation tests, such as Moran's statistic, are usually applied in cross-section, not in time series. Here, the SURE variance–covariance matrix allows for any type of correlation between regions. In Laurin (2007a), we investigate further the pattern of this cross-correlation between regions.

¹⁹ See for example Balassa (1985) and Moschos (1989).

²⁰ See for example Calderón et al. (2001) and Lee and Huang (2002).