Murali Kuchibhotla <[email protected]>:
A better question might be: Is there a way to correct for this in the
real world? Probably not, in most senses. Think of it as a
logit/probit regressing "survival" on the k variables in X and a
training dummy T and the k interactions T*X, so you have at least k+1
endogenous variables measuring the effects of training. You could
switch to an -ivprobit- model if you had a large number of
instruments, but that seems unlikely (plus, I think you would want an
-ivcloglog- command which does not exist). One way forward is to
reweight the samples so they look identical on observables, using
propensity scores, as described in
http://www.stata-journal.com/article.html?article=st0136 and elsewhere
(the sample will look identical in the sense that a -hotelling- test
would fail to reject the null that the mean of X is the same in the
training group and the non-training group--the distributions may
differ in other ways). But this corrects only for selection on
observables, not unobservable differences--for the latter, you need
training to be randomly assigned, for a start. Or you need a very
convincing natural experiment.
On Wed, Sep 10, 2008 at 11:22 PM, Murali Kuchibhotla
<[email protected]> wrote:
>
> Hello,
> I am interested in estimating separate Cox regression models for 2
> groups of individuals- trainees and non-trainees. The problem is that the
> training decision is endogenously determined, so that the differences
> between the 2 sets of parameter estimates may well be driven by the
> differences in the unobservable characteristics between the 2 groups of
> respondents. Is there a way to correct for this in Stata?
>
> Murali Kuchibhotla
*
* 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/
