multinomial logistic comparing A vs (B+C) and B vs C in single model

[emaguin]

Part 1. Suppose a nominal variable with three categories (A,B,C). Proc Logistic, as does GLMMIX and others procs, solve that. However, the PI insists that in CATMOD an analysis can be constructed to compare A vs (B+C) and B vs C in a single proc statement. However, he doesn't recall how. Is this possible? Please, details.

Part 2. The covariate list will include continuous variables that, as I read the documentation, catmod doesn't work well with. Can the analysis be done in logistic or glmmix? Details please. 

Part 3. The dataset is multilevel, occasions within persons, so if the analysis can done in sas, can random effects be included? Again, details please.

Thank you, Gene Maguin

[PaigeMiller]

1. This variable that has A B C as levels, is it a predictor variable or a response variable?

2. GLM or GLIMMIX can handle a mix of continuous and class predictor variables.

3. Both GLM and GLIMMIX allow random effects.

[emaguin]

Although I thought it was, my question wasn't clear. Apologies.

Response variable.

[PaigeMiller]

I think PROC LOGISTIC and PROC GLIMMIX allow three-level response variables and allows both continuous and categorical predictor variables.

I believe there are ways built into these procedures to compare A to B+C and then also B to C, but I have never done this.

[emaguin]

Thanks, Paige. Yes, both logistic and glmmix will execute multinomial models. I've already done both. Thank you for saying that you hadn't done what I am asking for. Now, a harder question. You know the system. Where do I go next and how do I get there.

Gene Maguin

[StatDave]

The two logit response functions that you have defined are known as continuation ratio logits. The continuation ratio logit model is one type of model used for modeling an ordinal response and is mentioned in this note about the types of logistic models available in SAS. The model can be fit in PROC CATMOD, as you mentioned, using weighted least squares. A version of it can also be fit by maximum likelihood estimation in PROC LOGISTIC by fitting the two logits separately. Both are discussed and illustrated in this note. Since you have continuous predictors, the approach using separate binary logit models is probably the better approach. Since ordinary binary logistic models are used in that approach, you could instead use a PROC GLIMMIX to include random effects.

It should also be possible to fit the overall model in a single step by defining the model in PROC NLMIXED which also can incorporate random effects. This is probably the best and most general approach. However, an example of doing this is not available. You can see an example of how NLMIXED can similarly be used to fit a cumulative logit model in this note which shows the NLMIXED code in the Addendum at the end.

[emaguin]

StatDave, Thank you for your reply and it looks like this fits what I was told by the PI with respect to catmod. I think this is the answer to my question but I want to hold the topic open so that i can come back to it if I need to, which is probably likely. Gene Maguin