Multiple break tests

[Doe]

Hello:

In the multiple break tests, the global l vs none alternative, in the output,

Is the “significant f statistics largest breaks” referring to the supF statistic of Bai and Perron?

And, for the number of breaks greater than 10, how are are the the critical values calculated? How is the approximation done? I.e., are the critical values reliable after 10?

Thank you.

[EViews Glenn]

Yes.
We use the value for 10 as that is the best available. So they are really just illustrative at that point. We considered using the Bai-Perron 2003 response surface regressions but there are issues with implementing those as well.

[Doe]

Thank you for your reply.

So, just to double check, “yes”, the statistics is supF.

And, we can trust the critical values up to 10, but after that, the critical values are mostly illustrative.

Correct?

Are the critical values up to 10 from Perron’s website, or sth?

Thank you.

[EViews Glenn]

Yes, it is the supF statistic.

Mostly yes to the second. I actually went and looked at my notes and the code more carefully and need to update my answer somewhat.

. First, the results are based on the unpublished tabular results from Bai-Perron 2003 (dated Feb. 2003) which, I am pretty certain were from Perron's website.

. Second, we originally were going to use the Bai-Perron response surface regression results for the computation of these critical values, but found that we were unhappy with the performance of these results so we switched to the table lookup. Unfortunately, some of the checks for the maximum number of breaks differ for the two cases, and we didn't update our checks to account for the switch back to the table. So yes, the higher order break results are not reliable, and in fact are sometimes very misleading for some large k.

. Lastly, note that the Bai-Perron results actually have different numbers of maximums depending on the trimming percentage. So while you noted 10 as the maximum number, the actually number is trimming dependent.

I am going do some cleanup to the code and most importantly will ensure that the approximations are as accurate as possible for the very high number of breaks and that there are comments in the output when we are using the maximum number of simulated breaks for higher order values. This will appear in the next EV13 update.

Thanks for your comments and patience.

[EViews Glenn]

Oh, I just realized that I should add that "supF" doesn't quite describe the test as Bai-Perron use SupF as part of two different test procedures:

. The supF we are discussing above is the supF for m vs. 0.
. There is a second supF for sequential l+1 vs. l.

Note that there are different critical values for the two statistics.