Do the most frequently used dynamic panel data estimators have the best performance in a small sample? A Monte Carlo comparison

Abstract

. Differenced GMM and system GMM estimators are the two most frequently used dynamic panel estimators. Regardless the fact that both estimators are proposed for samples with a large N and short T, both of them are frequently used for small samples. Therefore, this paper compares the small sample properties of these two estimators with standard dynamic LSDV and LSDV bias-corrected estimators to examine the justification of their frequent use. Data set dimensions are formed considering dimensions of previous empirical studies that use dynamic panel data on small samples. The results show that LSDV bias-corrected estimator has the smallest RMSE in almost every design while in terms of bias, the results are mixed. LSDV bias-corrected outperforms both GMM estimators in terms of bias in design when the number of individuals is 10 and the number of time periods is 30. GMM estimators show somewhat better properties in terms of bias in design when the number of individuals is 30 and the number of time periods is 10.