Re: st: xtlogit: panel data transformation's recast to double makes model incomputable

[Tom <tommedema@gmail.com>]

Hi Jay,

Per request these are the results of the "offending IVs" alone:

1) clogit depc_gpf30 close_g100, group(ticker_id) gradient hessian
trace showstep showtolerance
http://pastebin.com/cMx7sCm6

2) clogit depc_gpf30 close_g120, group(ticker_id) gradient hessian
trace showstep showtolerance
http://pastebin.com/xgcud5QF

3) clogit depc_gpf30 close_g30, group(ticker_id) gradient hessian
trace showstep showtolerance
http://pastebin.com/d22GhCVF

These also include the gradient, hessian and showtolerance options.

This is real price data, I also verified it several times. There
appear to be no mistakes in the data.

Do you have an explanation why close_g100 would fail whereas close_g30
does not? If you look at the summary statistics you'll see that the
close_g30 variable and close_g5 etc. are actually much more skewed and
have higher variations.

About the "bad answer", I don't think the answers given with the
xtdate nodouble option were that bad, because the predictions were
correct a considerable amount of times.

If any more information is needed please ask.. I'm very much willing
to solve this situation.

Tom

On Tue, Apr 2, 2013 at 8:06 PM, JVerkuilen (Gmail)
<jvverkuilen@gmail.com> wrote:
> It sounds like you're running into some kind of weird near but not
> complete perfect prediction. If you look at the log-likelihoods in the
> model they are frequently diverging:
>
>> From A:
>> log likelihood = -8.99e+307
>> (initial step bad)
>
>> From A and E (different likelihoods of course):
>> log likelihood = -235698.21
>> (backed up)
>
>> From B, C and E:
>> log likelihood =    -1.#INF
>> (initial step bad)
>
> So these fail simply on the face of as the first one is essentially
> diverged to -Infinity and the third has as well. The middle is bad,
> too, as "backed up" means that the step wasn't an improvement.
>
> The IV that you want to use has some truly wicked properties. It is
> quite possibly the most skewed variable I've seen in my career....
>
> If you just run with the offending variables does it fail? When I say
> "fail" I don't mean does it run to completion because that's a very
> low standard, but do you get a ludicrous log-likelihood such as the
> one there? Is the Hessian and parameter correlation matrix similarly
> silly? I suspect it must be regardless of whether the model converges
> or not.
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/