The Accuracy of Inference in Small Samples of Dynamic Panel Data Models

Through Monte Carlo experiments the small sample behavior is examined of various inference techniques for dynamic panel data models when both the time-series and cross-section dimensions of the data set are small. The LSDV technique and corrected versions of it are compared with IV and GMM regarding: coefficient bias, accuracy of variance estimators - both of the disturbances and of the coefficient estimators - and the actual size of coefficient tests. A reasonably simple and consistent bias adjusted LSDV estimator, for which we find an analytical and a bootstrap consistent estimator of its variance, performs relatively well. Further higher-order refinements of the bias correction do not improve the accuracy considerably. Most techniques show substantial size distortions for asymptotic t tests. Finally, it is illustrated how these findings help to interpret empirical results on the relationship between so-called dynamic externalities and local economic activity in Moroccan urban areas.