Bias-Correction in Time Series Quantile Regression Models

This paper examines the small sample properties of a linear programming estimator in time series quantile regression models. Under certain regularity conditions, the estimator produces consistent and asymptotically normally distributed estimates of model parameters. However, despite these desirable asymptotic properties, we find that the estimator performs rather poorly in small samples. We suggest the use of a subsampling method to correct for a bias and discuss a simple rule of thumb for setting a block size. Our simulation results show that the subsampling method can effectively reduce the bias at very low computational costs and without significantly increasing the root mean squared error of the estimated parameters. The importance of bias correction for economic policy is highlighted in a growth-at-risk application.