Choices between OLS with robust inference and feasible GLS in time series regressions

Abstract

We consider the practice of estimating static regressions by OLS from time series data and using robust standard errors for inference. Depending on the form of exogeneity being violated, the asymptotic bias of OLS can exceed that of GLS. Feasible GLS, where the error process is approximated by a sieve autoregression, can dominate the OLS approach with robust standard errors both in terms of bias and MSE for some regions of the parameter space.