Autocorrelation pre-testing in the linear model. Estimation, testing and prediction

This paper explores by means of a Monte Carlo experiment the consequences of autocorrelation pre-testing on estimation, hypothesis testing and prediction in the linear regression model with first-order autoregressive disturbances. We find that overall, pre-testing is preferable to pure OLS regression techniques and generally compares favourably with the strategy of always correcting for possible autocorrelation. More surprising findings include the degree to which the regression matrix affects the relative performance of the various strategies and the degree to which the familiar OLS based t-test can lose power in the presence of autocorrelation.