Autocorrelation and masked heterogeneity in panel data models estimated by maximum likelihood

Abstract

In a panel data model with random effects, when autocorrelation in the error is considered, (Gaussian) maximum likelihood estimation produces a dramatically large number of corner solutions: the variance of the random effect appears (incorrectly) to be zero, and a larger autocorrelation is (incorrectly) assigned to the idiosyncratic component. Thus heterogeneity could (incorrectly) be lost in applications to panel data with customarily available time dimension, even in a correctly specified model. The problem occurs in linear as well as nonlinear models. This article aims at pointing out how serious this problem can be (largely neglected by the panel data literature). A set of Monte Carlo experiments is conducted to highlight its relevance, and we explain this unpleasant effect showing that, along a direction, the expected log-likelihood is nearly flat.