Generalized Least Squares Model Averaging

This paper proposes a method of averaging generalized least squares (GLS) estimators for linear regression models with heteroskedastic errors. We derive two kinds of Mallows' Cp criteria, calculated from the estimates of the mean of the squared errors of the tted value based on the averaged GLS estimators, for this class of models. The averaging weights are chosen by minimizing Mallows' Cp criterion. We show that this method achieves asymptotic optimality. It is also shown that the asymptotic optimality holds even when the variances of the error terms are estimated and the feasible generalized least squares (FGLS) estimators are averaged. Monte Carlo simulations demonstrate that averaging FGLS estimators yields an estimate that has a remarkably lower level of risk compared with averaging least squares estimators in the presence of heteroskedasticity, and it also works when heteroskedasticity is not present, in nite samples.