Jeph,
The prediction is not conditional on having one or more positive
outcomes but on having _exactly_ one positive outcome per group. I
don't know how useful this is in practice as what you really want is
the unconditional probability.
As far as I know the only alternatives are including the fixed effects
manually in -logit- which is a bad idea when the group sizes are small
or, alternatively, using a random effects logit. The latter approach
assumes, of course, that the regressors are uncorrelated with the
unobserved heterogeneity which may not be the case.
Perhaps someone else have any better ideas...
Arne
On 30/11/2007, Jeph Herrin <[email protected]> wrote:
> Arne,
>
> I knew it was conditional, but since I have only
> a couple of small groups (n=2) with no positive outcome,
> I did not see why it would change from 45% to 3%.
>
> Moreover, the thread below indicates that -clogit-
> should be the better model here, but how can I get
> useful predictions out of it?
>
>
> thanks,
> Jeph
>
>
> Arne Risa Hole wrote:
> > Hi Jeph,
> >
> > When you use -predict- with the -pc1- option after -clogit- you get
> > the predicted probability of a positive outcome _conditional_ on there
> > being a single positive outcome in each group. I don't believe it is
> > possible to calculate the unconditional probability since the fixed
> > effects are not actually estimated. An alternative would be including
> > the fixed effects manually in the -logit- model, but this leads to
> > biased coefficient estimates unless the group sizes are large. See the
> > thread below for a very useful discussion:
> >
> > http://www.stata.com/statalist/archive/2007-10/msg00935.html
> >
> > Hope this helps.
> >
> > Arne
> >
> > On 29/11/2007, Jeph Herrin <[email protected]> wrote:
> >> I am surprised by the results I am getting from
> >> -predict- after -clogit- and wonder if I am missing
> >> something.
> >>
> >>
> >> . clogit endpoint age female, strata(site)
> >> [output suppressed]
> >> . predict prob1
> >> (option pc1 assumed; conditional probability for single outcome within
> >> group)
> >> . sum prob endpoint
> >>
> >> Variable | Obs Mean Std. Dev. Min Max
> >> -------------+--------------------------------------------------------
> >> prob1 | 630 .0365079 .0776009 .0053902 1
> >> endpoint | 630 .4539683 .4982722 0 1
> >>
> >>
> >> Why is the predicted so far off from the actual? Compare with the logit
> >> model:
> >>
> >> . logit endpoint age female
> >> [output omitted]
> >> . predict prob2
> >> (option pr assumed; Pr(endpoint))
> >> . sum prob2 endpoint
> >>
> >> Variable | Obs Mean Std. Dev. Min Max
> >> -------------+--------------------------------------------------------
> >> prob2 | 630 .4539683 .0251972 .4150936 .4819036
> >> endpoint | 630 .4539683 .4982722 0 1
> >>
> >>
> >> Am I using -predict- incorrectly after -clogit-? Is it a coincidence
> >> that -invlogit(prob1)- *does* look like the actual.
> >>
> >> . gen invlogpr1=invlogit(prob1)
> >> . sum invlogpr1
> >>
> >> Variable | Obs Mean Std. Dev. Min Max
> >> -------------+--------------------------------------------------------
> >> invlogpr1 | 630 .5090418 .018543 .5013475 .7310586
> >>
> >>
> >> Thanks for any insights.
> >>
> >> Jeph
> >>
> >>
> >>
> >>
> >> *
> >> * For searches and help try:
> >> * http://www.stata.com/support/faqs/res/findit.html
> >> * http://www.stata.com/support/statalist/faq
> >> * http://www.ats.ucla.edu/stat/stata/
> >>
> > *
> > * For searches and help try:
> > * http://www.stata.com/support/faqs/res/findit.html
> > * http://www.stata.com/support/statalist/faq
> > * http://www.ats.ucla.edu/stat/stata/
> >
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
