Hi Statalisters-
I didn't hear back the other day re: my message titled "two-stage
multinomial logit, logit" (which follows), but I've been thinking more
about my model, and thought perhaps a trivariate probit would work.
Here's the original question I posted:
I'd like to run a multi-stage model, and have looked at bivariate probit
(biprobit) but I'm not sure it's working for me.
I have a first equation that predicts treatment choice (3 possibilities),
A, A+B or C. So this could be estimated using a multinomial logit. But, my
second equation is a logit model whose outcome is an exit (failure)
conditional on treatment choice.
I tried using biprobit :
biprobit (A = X)
(exit = A B Y), cluster(id)
where X and Y represent covariate matrices with some overlap.
I have multiple observations per individual and use the cluster option.
This doesn't seem to be quite right though because A, A+B and C are unique
treatment choices even though A and A+B overlap.
--------------
Would a trivariate probit work where I have
equation 1: probability of treatment A
equation 2: probability of receiving B (would always be 0 for those getting
treatment C)
equation 3: probability of exit with A and B on the right-hand side
?
Are there any trivariate probit experts who might be able to help? The
problem is the second equation because I never see B for patients getting
C. Is there a way to trick the model into integrating to 1 for the 2nd
equation for those getting C?
