Abstract
Second-order approximations are obtained to the biases of several common estimators of the error variance σ2 and autocorrelation coefficient ρ in the linear regression model with AR(1) errors. The estimation methods considered are maximum likelihood and three versions of two-stage and iterated feasible GLS (Prais-Winsten). A simple relation between the approximate biases of ρ2 and ρ is noted. The accuracy of these approximations is assessed by simulation.