Performance of various estimators in dynamic panel data models

Abstract

This paper investigates the performance of various estimators in dynamic panel data models, in small samples and highly persistent AR processes. Tt is found that the different estimators of the AR parameter are either downward biased or imprecise. 'The IV estimator is nearly unbiased, but it becomes extremely imprecise as the true AR parameter approaches one. On the other hand, the GMM and LSDV estimators are relatively efficient, but they are downward biased. Even though they can compete in terms of a mean squared error criterion, the LSDV is by far more efficient. Kiviet's LSDVc estimator, as implemented in this study, performs quite poorly for AR parameter values close to one. The results of this investigation suggest that a natural way to asses the bias/efficiency problems in a highly persistent AR context, would be to use bias-corrected LSDV estimators. This approach is implemented in Cermeño (1997).