GMM Estimation of Short Dynamic Panel Data Models With Error Cross-Sectional Dependence

This paper considers the issue of GMM estimation of a short dynamic panel data model when the errors are correlated across individuals. We focus particularly on the conditions required in the cross-sectional dimension of the error process for the dynamic panel GMM estimator to remain consistent. To this end, we demonstrate that cross-sectional independence (or uncorrelatedness) is not necessary - rather, it suffices that, if there is such correlation in the errors, this is weak. We define a stochastic scalar sequence to be cross-sectionally weakly correlated at any given point in time if the sequence of the covariances of the observations across individuals i and j at time t, given the conditioning set of all time-invariant characteristics of individuals i and j, converges absolutely as N grows large. Spatial dependence satisfies this condition but factor structure dependence does not. Consequently, the dynamic panel GMM estimator is consistent only in the first case. Under cross-sectionally weakly correlated errors, an additional, non-redundant, set of moment conditions becomes relevant for each i - specifically, instruments with respect to the individual(s) which unit i is correlated with. We demonstrate that these moment conditions remain valid when the errors are subject to both weak and strong correlations, in which situation the standard moment conditions with respect to individual i itself are invalidated - meaning that the dynamic panel GMM estimator is inconsistent. Simulated experiments show that the resulting method of moments estimators largely outperform the conventional ones in terms of both median bias and root median square error.