Small Sample Bias Propreties of the System GMM Estimator in Dynamic Panel Data Models

This paper examines analytically and experimentally why the system GMM estimator in dynamic panel data models is less biased than the first differencing or the level estimators even though the former uses more instruments. We find that the bias of the system GMM estimator is a weighted sum of the biases in opposite directions of the first differencing and the level estimator. We also find that an important condition for the system GMM estimator to have small bias is that the variances of the individual effects and the disturbances are almost of the same magnitude. If the variance of individual effects is much larger than that of disturbances, then all GMM estimators are heavily biased. To reduce such biases, we propose bias-corrected GMM estimators. On the other hand, if the variance of individual effects is smaller than that of disturbances, the system estimator has a more severe downward bias than the level estimator.