Size distortions of tests for heteroskedasticity, cross-selectional correlation and autocorrelation in dinamic panel data models

Abstract

This paper investigates the extent to which different error covariance structures can be identified in finite samples in a dynamic panel data context. Specifically, the size of several known tests for heteroskedasticity, cross-sectional correlation and autocorrelation is evaluated using a dynamic fixed-effects model. Except in a few cases, the size of the tests is found to be seriously distorted. The problem of groupwise heteroskedasticity can be reasonably identified for AR parameter values of 0.5 and 0.7 as the size of the tests is quite close to nominal values. For the same AR values, the size of the tests for cross-sectional and autocorrelation is appreciably distorted towards the over rejection of the null hypothesis, the exception being Baltagi's (1995) test for autocorrelation in the case where β = 05, N = 15, T = 100. For β = 0.9 the size of all tests is greatly distorted.