Bias Correction in Autoregressive Roots

This paper develops a novel method to correct small-sample bias in autoregressive roots of AR(p) models. We evaluate median-bias properties and variability of the bias-adjusted parameters by examining the accuracy of bias-adjusted impulse responses. Our simulation results show that bias correction in roots achieves a substantial reduction in the median-bias of impulse response functions relative to the least-squares estimates. Furthermore, correcting the bias in roots significantly improves the median-bias in half-life, quarter-life and up-life estimates. The procedure pays a negligible-to-small price in terms of increased standard deviation for its improved median-bias properties