BIAS ASSESSMENT AND REDUCTION IN LINEAR ERROR-CORRECTION MODELS

The small-sample bias of the ordinary least-squares coefficient estimator for dynamic regression models with innovation errors and lagged-dependent and strongly-exogenous explanatory variables is approximated through both small disturbance and large-sample asymptotics. Results for the standard ARMAX(p, 0, k) model are obtained and also for such models under linear parameter constraints and variable transformations. These approximations are then used to construct corrected estimators for the parameters of interest in higher-order dynamic models, including the empirically highly relevant linear error-correction model. By simulating two empirical cases the corrected estimators obtained via large-sample asymptotics are shown to have more attractive location and efficiency properties than ordinary least-squares.