Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Conditional probabilities with one fixed effect at a constant


From   Michael Ingre <[email protected]>
To   [email protected]
Subject   Re: st: Conditional probabilities with one fixed effect at a constant
Date   Wed, 7 Jul 2004 20:56:33 +0200

Buzz Burhans have made several suggestions to help me with my problem. I appreciate it a lot. It has inspired me to think in new ways and come closer to a solution.

I now think I have the solution. The one presented earlier is probably completely wrong. I will not go into much detail but, there might be other gllammers out there that might be interested.

If I want to predict probabilities with a fixed effect removed (or at a constant) this is what I believe needs to be done:

1. estimate the gllamm logit model
2. predict the random effects with -gllapred u , u-
4. set the random effect of the effect that should be constant to 0 - replace um3 = 0 -
3. predict conditional probabilities with -gllapred mu , mu us(um)-
4. calculate how much the fixed effect that should be removed adds to the model.
5. remove the probabilities (at 4) from the predicted probabilities in -mu-

If anyone has any comments I appreciate it.

In particular I'm not quite sure how to calculate the probabilities to remove. If it was a linear regression model it would be a pice of cake. But I'm not used to work with logit estimates. I guess I have to convert everything to log-odds before adding and/or subtracting and then convert back to probabilities.

A comment on the previous solution:

My best guess is that -gllapred- does not use subject specific information from the estimation sample at all when predicting responses from observations outside the estimation sample. Instead it seems like -gllapred- uses the estimated variances and covariances of the random effects as well as the responses in the new sample to simulate a plausible solution with subject specific probabilities.

This is quite different from -predict-, -gllapred- also uses the responses when predicting in a new sample. And if I'm not mistaken, it completely invalidates the solution suggested earlier. It makes no sense at all.

Michael

*
* 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/




© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index