A comparison of dynamic panel data estimators: Monte Carlo evidence and an application to the investment function

In our analysis we discuss several dynamic panel data estimators proposed in the literature and assess their performance in Monte Carlo simulations. It is a well known fact that the natural choice, the least squares dummy variable estimator is biased in the context of dynamic estimation. The estimators taking into account the resulting bias can be grouped broadly into the class of instrumental estimators and the class of direct bias corrected estimators. The simulation results clearly favour the direct bias corrected estimators, especially the estimator proposed by Hansen (2001). The superiority of these estimators decreases with growing numbers of individuals in the simulation. This is the well known fact of large sample properties of the GMM-methods. In the case of endogenous predetermined regressors, the system-estimator proposed by Blundell and Bond is unbiased and most efficient, while direct bias corrected estimators perform similar to the GMM-estimator proposed by Arellano and Bond (1991). Turning to the empirical comparison, we find that the different estimators lead to the same conclusions concerning the investment behaviour of German manufacturing firms based on the Deutsche Bundesbank's Corporate Balance Sheet Statistics. Investment is strongly positive dependent on lagged investment and Q. Nevertheless, in detail the differences of the estimated parameters are not negligible.