Properties of Fixed Effects Dynamic Panel Data Estimators for a Typical Growth Dataset

This paper investigates the properties of dynamic panel data (DPD) estimators in the context of a typical growth dataset. Using Monte Carlo simulations, we compare the performance of various DPD estimators, namely the Anderson-Hsiao (AH) and Arellano-Bond's General Method of Moment (GMM) one-step and two- step estimators, using the least-square dummy variable (LSDV) as a benchmark. We arrive at three conclusions. First, LSDV produces biased estimates and the biases are significant even for a moderate-sized time dimension. Second, there is no immediately obvious choice to replace LSDV among the estimators considered here. For one, there is the bias-efficiency trade-off. In addition, differences in the characteristics of data influence the performances of the various estimators. Finally, serial correlations in the error terms, even at a low degree, can introduce significant biases to the estimations.