multinomial regression with repeated measures using PROC CATMOD

[levender8622]

I am trying to fit a multinomial regression with repeated measures, and specifically, I need GEE with sandwich estimator but not random effects model (so PROC GLIMMIX does not work for me). It seems PROC CATMOD is the only possible solution. However when I tried the following codes:

proc catmod data = sjeur;
direct MET_vigday;
response logits;
model obesity = MET_vigday/param=reference;

weight wt;
repeated sjlid $/_response_=sjlid;
run; 

The model is only for testing purpose so I only have one continuous independent variable.

However I got the following error:

ERROR: For the standard response function, _Response_ may be defined only in terms of the dependent
variables, and no factor names should appear in the REPEATED statement.

I tried various ways of incorporating _response_ in model statement but always had errors. Could someone help me with proper coding? Thank you!

[StatDave · Accepted solution]

Use PROC GEE, not PROC CATMOD. PROC CATMOD is rarely needed since other procedures, especially the LOGISTIC, GENMOD, and GEE procedures, have much of its capabilities and are easier to use. In PROC GEE, specify the DIST=MULT and the LINK= option with the appropriate link (CLOGIT if the response is ordinal, GLOGIT if it is nominal) in the MODEL statement. Also specify the REPEATED statement and the variable that defines your repeated clusters in its SUBJECT= option.

[levender8622]

Thank you so much! I tried PROC GEE and it worked perfectly. And as a test, I used PROC GEE on binary outcome and compared to PROC GENMOD and results were identical. Sorry I couldn't get back to you earlier, as I was sick for a few days and now I'm well again.