GLIMMIX Solutions vs. Type III for Fixed Effects

[JonKetchup]

I am creating a GEE in PROC GLIMMIX (I asked about this model yesterday), and I am having trouble understanding the difference in output produced by the Solutions for Fixed Effects and Type III Tests of Fixed Effects.

My model has 4 fixed effects and 1 interaction. All fixed effects are binary including the outcome. I am most interested in location, sex, and their interaction - var3 and var4 are just variables I need to control for. I would like the odds ratios with confidence intervals for all fixed effects including location and sex. I ran the following estimate statements to do so. The estimates I ran are not equal to those from the Solutions table, and the p-values match that of the Type III Tests rather than the Solutions table.

Here are my questions:

1) Are the odds ratios for each fixed effect the exponentiated estimates in the Solutions table or the exponentiated estimates I ran?

2) What is the difference between the p-values in the Solutions table compared to the Type III Tests?

Appreciate your help.

proc GLIMMIX data = DATA ORDER=INTERNAL EMPIRICAL=root;

class ID Cluster Location(REF='1') Sex(REF='1') Var3(ref='1') Var4(ref='1'); model OUTCOME(event='1')= Location|Sex Var3 Var4

     / dist=bin link=logit covb solution;
ESTIMATE 'Location' Location1 -1 / exp cl;

ESTIMATE 'Sex' Sex 1 -1 / exp cl;

RANDOM _residual_ / SUBJECT=Cluster TYPE=cs vcorr solution;

run;

[StatDave]

Just like in an ordinary regression model where, if there is an interaction, then the effect of one variable depends on the level of the other and you don't look for overall effects of either variable, the same is true here for a logistic model and odds ratios. If your interaction is significant, then you should only be looking at the effects of one variable within each level of the other. You did this using the DIFF and ODDSRATIO options in the LSMEANS statement in your previous post. There is no need for ESTIMATE statements. 

[JonKetchup]

Appreciate the quick response. As you said, I outputted all the estimates of both variables within each level of the other. However, the confidence intervals of the fixed effects are of clinical significance. 

[StatDave · Accepted solution]

A significant interaction implies that the main effects of each variable on its own do not exist and it would be incorrect to report such. If you assume there is no interaction and remove it from the model, then it would be reasonable to report the main effects. Again, this is accepted statistical practice regarding interaction in any type of model.