Power Comparison of Autocorrelation Tests in Dynamic Models

Abstract

The four most readily available tests of autocorrelation in dynamic models namely Durbin’s M test, Durbin’s H test, Breusch Godfrey test ( BGF ) and Ljung & Box ( Q ) test are compared in terms of their power for varying sample sizes, levels of autocorrelation and significance using Monte Carlo simulations in STATA. Power comparison reveals that the Durbin M test is the best option for testing the hypothesis of no autocorrelation in dynamic models for all sample sizes. Breusch Godfrey’s test has comparable and at times minutely better performance than Durbin’s M test however in small sample sizes, Durbin’s M test outperforms the Breusch Godfrey test in terms of power. The Durbin H and the Ljung & Box Q tests consistently occupy the second last and last positions respectively in terms of power performance with maximum power gap of 63 & 60% respectively from the best test ( M test).