Testing the stochastic convergence of Italian regions using panel data

Abstract

This paper examines stochastic convergence in real per capita GDP for Italian regions using recent non-stationary panel data methodologies over the period 1951 to 2002. Economies stochastically converge when regional differences across economies are not persistent, and long-run movements in a region's real per capita GDP are driven by technological shocks. Four panel unit root tests are used to evaluate stochastic convergence and Monte Carlo simulation is performed to obtain critical values. The results indicate that there is no stochastic convergence among Italian regions. They also provide some evidence in favour of stochastic convergence of regions in the Northern part of Italy.

Notes

¹ The Italian regions are: Piedmonte, Valle D’Aosta, Lombardia, Trentino Alto Adige, Veneto, Friuli Venezia-Giulia, Liguria, Emilia-Romagna, Toscana, Umbria, Lazio, Marche, Abruzzo, Molise, Campania, Puglia, Basilicata, Calabria, Sicilia, Sardegna.

² Investigation is carried out at two different levels in order to evaluate the stochastic convergence in the poorer part of the country.

³ O’Connel (1998) shows that the procedure of subtracting cross-sectional mean from the series proposed by LL and IPS does not reduce cross-sectional dependence and size distortion of those tests.

⁴ These two tests are used in addition to LL and IPS since the contemporaneous correlation matrices (Appendix), indicates group wise cross-sectional dependence for the log real GDP per capita data used in the paper. Österholm (2004) used MADF tests in order to test the presence of unit roots in four US macroeconomic time series. Chen and Lu (2003) used MADF tests to test the validity of Gibrat's Law.

⁵ Neumayer and Cole (2003) provided an important contribution to the economic growth literature by studying the convergence hypothesis focusing on people rather than countries. They weight their regressions by populations.

⁶ Toscana is the richest regions in terms of real per capita GDP.

⁷ We use a two-step procedure in order to reduce considerably computational costs. Estimation can also be obtained by iterative procedure. For further details see Greene (2000, pp. 616–36) and Hayashi (2000, pp. 535–7).

⁸ The Monte Carlo simulation used to obtain critical values for the LL, IPS, SUR and MADF tests is described in more detail in Section IV. The critical values for t-bar test are obtained using μ _t and (the mean and variance of each t _{ρ i} statistic) reported in Table 2 of IPS (1997).

⁹ TS apply SUR estimator that is essentially a multivariate GLS (generalized least square) using an estimate of the contemporaneous covariance matrix of the disturbances obtained from individual ordinary least square estimation.

¹⁰ Indeed, one value is found below critical values at 10% significant level for the IPS test, but this does not really alter the results.