EViews.com

Hello!

I've noticed that in PP, ERS and NP unit root tests HAC corrected variance (f0) is calculated by the same formulae. But if i have missing observations in series, PP and ERS give me the same result, but NP differs from them. Why does it happen?

Open attached workfile, execute the following code:
freeze(t1) y.uroot(pp, const, hac=bt)
freeze(t2) y.uroot(ers, const, hac=bt)
freeze(t3) y.uroot(ng, const, hac=bt)

Same issue as in the other cases. The NP code is using a smaller sample because it is doing sample exclusions early so it doesn't use the same pre-sample values for the transformation.

But it should, isn't it?
I don't understand why it shows only when I have missing observations.

As I said in the other thread (http://forums.eviews.com/viewtopic.php?f=4&t=7090) the various unit root tests handle the samples slightly differently. The presence of the missing values isn't the issue, it's the fact that the NP test uses the sample after we've done differences on the detrended series while the others use a sample which doesn't lose the differencing observation. It's not ideal, but both estimators are consistent.

Also as noted in the other thread, I'm going to line these up so that they use the same rule, but I don't want to do it mid-release.

Sorry, I thought you didn't understand me. If in new release everything will be consistent, I should stop to ask such questions.

Thank you.

No problem. It's useful having this kind of discussion as it highlights the fact that there are a lot of choices that go into computing something like the NP test and that there may be small differences across various approaches.