Re: st: AW: estimation of series of OLS regressions based on t-values from previous regression

[Steven Samuels <sjhsamuels@earthlink.net>]

---



And, of course, throwing 200 variables into a stepwise algorithm  
presumes that they should enter as single, untransformed, effects.  
If, on the contrary, there are important interactions or, for  
continuous predictors, necessary transformations, SW will not find  
them, because it does not look.

How fortunate that a knowledgeable expert can supply 12 candidate  
predictors! From a Bayesian perspective, that is good prior  
information.  An additional benefit of starting with a  small number  
of  predictors: there is a reasonable chance that the final model(s)  
will be believable.

-Steve


On Jan 30, 2009, at 2:14 PM, Nick Cox wrote:

> In addition to the very serious methodological issues quite rightly
> raised, I'm wondering what kind of performance this would be  
> expected to
> produce. If I throw 200 noise predictors at a response, I expect to  
> get
> a pretty good R-square, for example. (They shouldn't create much
> difficulty over multicollinearity....) Conversely, if 200 sensible
> predictors aren't enough, why should one expect 200 more to do much
> better?
>
> Calibrating any procedure against what happens with stochastic garbage
> would seem essential.
>
> P.S. anyone contemplating stepwise who hasn't read Frank Harrell,
> Regression modeling strategies, Springer, New York 2001, should  
> seek it
> out straight away.
>
> Nick
> n.j.cox@durham.ac.uk
>
> Maarten buis
> Sent: 30 January 2009 16:55
> To: statalist@hsphsun2.harvard.edu
> Subject: RE: st: AW: estimation of series of OLS regressions based on
> t-values from previous regression
>
> --- sdm1 <sdm1@york.ac.uk> wrote:
> > I'm afraid that it will have to be done (considerably) more than
> > once!  If anyone could offer an idea of how to get started with a
> > program, that would be much appreciated.
>
> You seem to be aware that applying this method on that number of
> variables will mean that your results cannot in any way be generalized
> outside your data. Just out of curiosity, could you give a bit of
> substantive background behind your problem, showing that
> generalizability is not of interest in your case? It would be nice to
> have a real life example of an exception to the rule that
> stepwise/datamining/datasnooping is evil.
*
*   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/