Proc glimmix dist=multinomial link=logit

[tlse31]

I have two programs which give me contradictory results for a calculation of the number of subjects per simulation.

Here enclosed, the SAS program perform with the 2 GLIMMIX models:

• Model1 with 4 groups (A B C P)
• Model2 with 2 groups (A P) only

The Results of the "A -P" comparisons are the opposite M1: "A-P" p<0.001 / M2: "A-P" p=0.971.

Thanks in advance for your help.

[SteveDenham]

If you do a cross-tabulation of the data in ZZ using PROC FREQ, you will see that when you restrict the dataset to just A and P categories you have quasi-complete separation.- the max for A is yvalue 3 with one obs, the min for P is yvalue3 with one obs.  When you fit all the levels, this isn't a problem, but when you restrict like this, the estimation technique results in a poor fit.  What I would suggest is using PROC SURVEYLOGISTIC (since you have no random effects).  Somehow, it gives reasonable results for these data.

SteveDenham 

[tlse31]

Thank you for your proposal. We suspected that the near-complete separation was the cause of the problem, but we were surprised by the lack of warning from SAS. We are going to test the surveylogistic proc.
Thank you again for your support.