Re: [Gretl-users] counting

Sven Schreiber schrieb: > Artur T. schrieb: > >> Allin Cottrell schrieb: >>

>>>> >>> <script> >>> nulldata 100 # or whatever >>> scalar alpha = 0.99 >>> scalar level = 0.05 # or whatever >>> scalar reject = 0 >>> scalar replics = 1000 >>> loop replics --progressive --quiet >>> genr a = 10*uniform() >>> genr y = 1*a + normal() >>> genr y = alpha*y(-1) + normal() >>> adf 1 y --nc --quiet >>> genr pv = $pvalue >>> reject += (pv < level) >>> print pv >>> store pvals.gdt pv >>> endloop >>> >>> printf "rejections = %d (%.2f%%)\n", reject, 100*reject/replics >>> </script> >>> >>> >From running this a few times, it looks as if this test is >>> over-rejecting quite seriously (if I'm interpreting the results >>> correctly). I'll look into this further. >>>

>> >> @Allin >> I was also surprised how often the null is rejected. But actually, as is >> known, the result depends heavily on the number of observations and how >> close %alpha is to the unit. >> >> > What actually are you trying to analyze? Are you aware that the adf test > with --nc is mis-specified, or is this the whole point? >

> Yes, actually I wanted to test the validity of the test under misspecification, sample size and structural breaks. The code was only one part of the total story.

>>> genr a = 10*uniform() >>> genr y = 1*a + normal()