Alternative general method of moments estimators in dynamic panel data models

Abstract

The choice of initial weighting matrix and the problem of instruments proliferation are crucial issues in generalized method of moments (GMM) estimators of dynamic panel models. In this study, we propose alternative system GMM estimators that utilize suboptimal initial weighting matrices together with reduced instruments set (specifically lag-limited and partially collapsed instruments). Comparison of the performance of the proposed estimators against the conventional estimator was undertaken in terms of bias, root mean squared error (RMSE) and coverage probabilities through Monte Carlo simulations. Our simulation results revealed that sub-optimally weighted system GMM estimator adopting partially collapsed instruments outperforms the standard GMM estimator in terms of both bias and RMSE for large T and large variance ratios as the coefficient of the lagged dependent variable gets close to zero. Under these scenarios, the system GMM estimators based on reduced instruments were also found to perform well in terms of coverage probabilities. As the process approaches a random walk, however, there is no considerable gain from the use of suboptimal initial weights matrix and reduced instruments in general.