Feasible bias-corrected OLS, within-groups, and first-differences estimators for typical micro and macro AR(1) panel data models

Abstract

Dynamic panel data (DPD) models are usually estimated by the generalized method of moments. However, it is well documented in the DPD literature that this estimator suffers from considerable finite sample bias, especially when the time series is highly persistent. Application of the asymptotically equivalent continuous updating method eliminates this problem but the resultant estimator exhibits too much variability in small samples. Thus, other estimation methods are considered in this paper. Focussing in the AR(1) case with no exogenous regressors, we analyze several alternative ways of correcting the bias of the traditional estimators utilized in non-dynamic settings, showing how to construct feasible bias-adjusted ordinary least squares, within-groups, and first-differences estimators. We obtain very promising results for some of these estimators in a Monte Carlo simulation study involving data with the qualities normally encountered by both microeconomists and macroeconomists.