multinomial logit estimate using proc logistic or catmod

[Chaoran]

Hello all, I'm a newcomer and have a question on how to do multinomial logit estimate with either proc logistic or catmod. Here is my question.

Let's say we are looking at events of 5 mutually exclusive outcomes: 1, 2, 3, 4, 5. We want to estimate probability models for each of them, subject to the constraint of all of their probabilities summing up to 1. Something like the following would do:

PROC LOGISTIC DATA=myData;

     CLASS outcome;

     MODEL model = x1 x2 x3  / LINK = GLOGIT;

     RUN;

That would estimate a set of coeffs for each of the first 4 outcomes, but all for the same set of explanatory variables x1, x2, x3. Now if I really need to use different set of regressors for different outcomes (e.g. x1 x2 for outcome 1, x2 x3 for outcome 2, etc). How would I do that? Would appreciate any suggestions.

[bncoxuk]

PROC LOGISTIC DATA=myData;

     CLASS outcome;

     MODEL model (REFERENCE ='5') x1 x2 x3  / LINK = GLOGIT;

RUN;

My experience is that:

To run multinomial logit model, PROC LOGISTIC can handle large-scale analysis with many variables and cases. PROC CATMOD stop running if the model is too complex with too many cases. PROC CATMOD is quicker to get the results if the model is not too complex and the dataset has manageable size.