Lag Length for Unit Root Test

[upani]

Dear All,

When we are checking the stationarity test for a variable at level, is it required to give lag 0? Why we choose various lag when we try to test the stationarity with a first difference, when most of the data shows first difference is stationary in nature.

With sincere regards,
Upananda

[TomDoan]

The asymptotics of the D-F test rely on the residuals being serially uncorrelated (and homoscedastic). Not simply stationary, but serially uncorrelated, hence the use of the augmenting lags. The Phillips-Perron test is an alternative that estimates the unit root regression without trying to eliminate the serial correlation by use of lags, but it requires adjustments based upon non-parametric estimates of the serial correlation to get the asymptotics correct.

[upani]

Thanks a lot clearing my doubt. Which lag selection method is best? Is it best to use @adfautoselect to choose the lag length.

[TomDoan]

That would be my recommendation since it shows four different criteria plus the ADF statistic. You can then see whether you get similar decisions about the unit root test based upon the lag lengths chosen by the different criteria.