Lag Selection and Multiple Tests

Abstract

The chapter is concerned with two related issues. The first concerns the need to select a lag or truncation parameter when there is serial dependence in the errors. Unit root tests can become severely mis-sized in the presence of serially correlated errors, especially if there is an MA component to the errors; however, increasing the lag length in an ADF regression, or increasing the bandwidth in a semi-parametric estimator of the long-run variance, may have costs in terms of a loss of power if the some of the included lags are superfluous. It is important, therefore, to have an assessment of different methods of selecting the ‘truncation’ parameter.