Improved Coefficient and Variance Estimation in Stable First-Order Dynamic Regression Models

In dynamic regression models the least-squares coefficient estimators are biased in finite samples, and so are the usual estimators for the disturbance variance and for the variance of the coefficient estimators. By deriving the expectation of the initial terms in an expansion of the usual expression for the asymptotic coefficient variance estimator and by comparing these with an approximation to the true variance we find an approximation to the bias in variance estimation from which a bias corrected estimator for the variance readily follows. This is also achieved for a bias corrected coefficient estimator and allows to compare analytically the second-order approximation to the mean squared error of the least-squares estimator and its counterpart for the first-order bias corrected coefficient estimator. Two rather strong results on efficiency gains through bias correction for AR(1) models follow. Illustrative simulation results on the magnitude of bias in coefficient and variance estimation and on the scope for effective bias correction and efficiency improvement are presented for some relevant particular cases of the ARX(1) class of models.