Ordinal Multinomial modelling with random effects

SergioP — Thu, 08 Nov 2018 15:48:14 GMT

Dear all,

I have a set of evaluation scores (from 1=worst quality to 5=best quality) given by N=25 subjects to each of M=90 videos varying by content, resolution and compression. The contents are six video games and the subjects are individuals of similar age and playing experience. The main interest is in the fixed effects of resolution and compression with respect to the population of games and individuals. Thus I would like to perform a GEE marginal analysis with 1-nested log odds ratios, but I understand that is not available for the multinomial. I ended up using the following repeated statement

repeated subject=Subject* Game / logor=EXCH;

which is not what I was looking for but probably gives similar answer.

The full code is

proc gee data=scores;
class Subject Game Resolution;
model Score = Game Resolution Game*Resolution sqrtComp logComp logComp*Game logComp*Resolution/ dist=mult type3;
repeated subject=subject*Game/logor=Exch;

output out=out(keep= config Game Resolution Compression Pred) Predicted=Pred;
ods output GEEFitCriteria=QIC2;
run;

Is there any other way?

Would glimmix with method=Laplace helping in any way?

Thank you

Sergio Pezzulli

Re: Ordinal Multinomial modelling with random effects

Rick_SAS — Sat, 10 Nov 2018 21:52:28 GMT

I am not an expert in this area, but I recently read a SAS Global Forum paper about using PROC GLMMIX to perform GEE modeling. The author, Kathleen Kiernan, is an expert who works in SAS Technical Support:

"Insights into Using the GLIMMIX Procedure
to Model Categorical Outcomes with Random Effects"

Pages 10-11 discuss the difference between GEE models in GENMOD or GEE versus GLIMMIX. Example 3 is an example with multinomial response.

topic Ordinal Multinomial modelling with random effects in Statistical Procedures

Ordinal Multinomial modelling with random effects

Re: Ordinal Multinomial modelling with random effects