Re: st: FW: help on variable selection problem

[Joerg Luedicke <joerg.luedicke@gmail.com>]

On Fri, Jun 10, 2011 at 3:40 PM, Lachenbruch, Peter
<Peter.Lachenbruch@oregonstate.edu> wrote:
> This is not especially a Stata question, but it is driven by an analysis issue...
>
> A student is trying to analyze data from a national survey (no weights needed).  She has 26 variables plus 10 years of data.  There are about 1,000,000 observations.  With this many observations, everything is significantly different from 0.  She's using mlogit (predicting medical care expenses), so she'd like to cut down the number of 'important' predictors.  I have thought of several options: backward stepwise  (not available with mlogit); look at effect size and insist it be larger than 0.05 - again not available since there are four categories of the response variable; use a Bonferroni inequality on the coefficients and insist on a low p-value to begin with - e.g. try for a size of 0.01 adjusting for 25 tests, so p must be less than 0.0004.  The issue seems to be the huge sample size pushing everything to significance.
> Does anybody have any ideas?
>

Some $0.02:

1) "She's using mlogit (predicting medical care expenses), so she'd
like to cut down the number of 'important' predictors."

I do not quite understand the logic here. Let's say you have 25
variables, all significant. Now you remove 15 and the remaining 10 in
the model are all significant. What would you gain by that? (BTW are
"medical care expenses" not at least measured on ordinal scale?)

2)  "I have thought of several options: backward stepwise"

This is usually problematic, see:
http://www.stata.com/support/faqs/stat/stepwise.html

3) "adjusting for 25 tests"

I cannot really see how this is a multiple comparisons problem. But
even if you would do an adjustment like that, it would not really help
(see the point below)

4) "The issue seems to be the huge sample size pushing everything to
significance."

That is why you should look at effect sizes first and care less about
p-values. Just see if the predictor's contributions are of substantive
size. For example, if you find an odds ratio for women of 1.01, you
can conclude that there is not much difference across gender,
regardless of whether this is "significant" or not. Even if it was
significant at p<0.0004 or whatever other level you would chose, that
would not change anything.

J.

*
*   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/