Multinomial models for clustered data using GLIMMIX

Melk — Fri, 22 Nov 2019 23:44:02 GMT

I have clustered data (right and left hand - not all patients contribute both) and I am trying to run a multinomial model in GLIMMIX. My categorical outcome is actually ordinal in nature (5 categories self rating, about 400 observations total), but I am also okay with it being treated as nominal depending on better model fit. My questions are:

1) How do I check the proportional odds assumption using GLIMMIX?

2) I am running the dist=multinomial with both the clogit and glogit links but I cannot run a model without some sort of a error or note in the log. I either get "NOTE: At least one element of the (projected) gradient is greater than 1e-3." or "NOTE: Estimated G matrix is not positive definite." singly or both together when a tweak the model using various options such as "method=laplace, nloptions MAXITER=5000 gconv=0." I was hoping someone could provide some general advice on what else I can do to fix the errors?

Thank you.

Re: Multinomial models for clustered data using GLIMMIX

Rajesh3 — Wed, 27 Nov 2019 19:10:47 GMT

I'm not sure about the test for the proportional odds assumption in GLIMMIX.

But the error that you are getting is because your data is sparse and is not sufficient for the model you are specifying. Possible solutions would be to reduce the number of covariates or try to make your outcome binary. You can run a simple proc freq on your outcome stratified by cluster to get a general idea on how sparse your data is.

I hope that helps!

Thanks,

Rajesh.

topic Multinomial models for clustered data using GLIMMIX in Statistical Procedures

Multinomial models for clustered data using GLIMMIX

Re: Multinomial models for clustered data using GLIMMIX