Estimation for Unit Root Testing

Abstract

We revisit estimation and computation of the Dickey Fuller (DF) and DF-type tests. Firstly, we show that the usual one step approach, based on the "DF autoregression", is likely to be subject to misspecification. Secondly, we clarify a neglected two step approach for estimation of the DF test. (In fact, we introduce a new two step DF autoregression.) This method is always correctly specified and efficient under the circumstances. However, it is either neglected or misused in unit root testing literature. The commonly employed hybrid of the (correct) two step method is shown to be inefficient, even asymptotically. Finally, we further improve/robustify the proposed two step method by employing the missing initial observations. Our finally proposed method is to be used in unit root testing, since it is a new DF autoregression that retains the missing observations.