RE: st: Penalized/shrinkage estimators for probi

[Charles Vellutini <charles.vellutini@ecopa.com>]

Thanks Stas
You are right, for some reason -findit plogit- does not work. For those interested, a -net from http://www.homepages.ucl.ac.uk/~ucakgam/stata- will lead you to the plogit package written by Gareth Ambler. 
It is also listed on Stata Corp's website: http://www.stata.com/links/resources2.html

On scaling: yes but my understanding is that penalized estimators normally work on standardized data, that is, columnwise centered on mean zero and  with variance one (both the response Y and the exogenous Xs). And no, I did not find anything in the literature specifically on penalized probit estimators. The closest reference I have is the Clarke and al. book "Principles and theory for data mining and machine learning", chapter 10 and particularly section 10.3.4.2. While logit is explicitly covered, probit is not. But since the probit model is clearly a generalized additive model with probit as the link function, I don't see why these techniques could not be implemented. 

Does anyone know of a reference on probit penalized estimators? 

Thanks,
Charles
-----Message d'origine-----
De : owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] De la part de Stas Kolenikov
Envoyé : vendredi 12 août 2011 16:43
À : statalist@hsphsun2.harvard.edu
Objet : Re: st: Penalized/shrinkage estimators for probi

-findit plogit- does not return anything by that name. It may have been written and circulates in your organization, but for the rest of Stata world, it is as good as non-existent. Since the coefficients in limited dependent variable models are identified up to a scale (or at least that's how econometricians view it, and I think it is best to know of all potential pitfalls even if your discipline does not believe in them), imposing lasso-type conditions on probit coefficients may lead to strange and inconsistent results (although a solid reference would certainly convince me otherwise :)).

Most of the time, the choice between logit and probit models is a matter of personal preference (although in case of -mlogit- and -mprobit-, there huge differences, of course).

On Fri, Aug 12, 2011 at 4:03 AM, Charles Vellutini <charles.vellutini@ecopa.com> wrote:
> Hi all,
>
> I need to estimate a probit model starting with a very large number of candidate explanatory variables -this is essentially a data mining, a-theoretical context. I therefore need a robust approach to variable selection.
>
> I am aware of the plogit package, which implements penalized estimators for the logistic model (including lasso), but I did not find similar estimators for the probit model. Is anyone aware of one?
>
> Failing that, what about using the plogit package to do the variable selection bit and then run a regular probit on the selected model?
>
> Thanks
>
> Charles Vellutini
> charles.vellutini@ecopa.com
> ECOPA
>
>
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>



--
Stas Kolenikov, also found at http://stas.kolenikov.name Small print: I use this email account for mailing lists only.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/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/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/