Multinomial Distribution using GLIMMIX

[clg2]

I have a response variable with six categories, 1 - 6. 

SAS Version: TS1M6

I am analyzing the data using GLIMMIX: 

PROC GLIMMIX ; 

CLASS rep group size ;

MODEL size = group / DIST=MULT ; 

RANDOM rep ; 

OUTPUT OUT=new PRED(ILINK)=ps ; 

SIZE has discrete values 1, 2, 3, 4, 5, 6. 

When I previously used this code, the new SAS DATA set would have _LEVEL_ to identify which of the six levels of the categorical response variable the predicted cumulative probabilIty (ps)  was associated with. Now, I am no longer getting the _LEVEL_ variable; just the observed value of SIZE.  I want to make certain I align the correct value of SIZE with its cumulative probability for each group value. 

Any assistance would be appreciated. 

clg

[FreelanceReinh]

Hello @clg2 and welcome to the SAS Support Communities (as a first-time poster)!

Sorry to see that your post hasn't been answered yet. Not sure if I can be of much help, but running your code using my SAS 9.4M5 (SAS/STAT 14.3) and this very crude input dataset HAVE

data have;
call streaminit(27182818);
do i=1 to 100;
  group=rand('bern',0.5);
  rep=rand('bern',0.5);
  size=rand('integer',6);
  output;
end;
run;

creates a dataset NEW which does contain variable _LEVEL_ with values 1 through 5 for each value of variable i, hence 500 observations in total. However, if I add the OBSCAT option to the OUTPUT statement

OUTPUT OUT=new1 PRED(ILINK)=ps / obscat;

the output dataset NEW1 is only a subset of NEW: comprised of those 75 observations where SIZE=_LEVEL_, and variable _LEVEL_ is omitted (as it would be redundant). This looks like what you describe.

As far as I see, the documentation of the OUTPUT statement has not changed between SAS/STAT 14.3 and 15.1. So I wouldn't know why your SAS 9.4M6 should work as if OBSCAT was the default. Do you get variable _LEVEL_ if you use the above HAVE dataset? If so, there might be a data-dependent issue and you may want to share a portion of your data (possibly with confidential values replaced with mock-up data).

[clg2]

Thank you for your reply. The error I made was embarrasingly simple. I needed to sort the categorical responses within each of the levels of the CLASS variable. Bad oversight on my part.