Effects coding versus dummy coding in with two-level categorical variables in PROC GLIMMIX

[cmo5]

Hi Paige, 

Thank you for the quick and helpful response. It looks like the LSMEANS statement doesn't work in multivariate models. I plan to use the class statement for the dichotomous independent variables. It seems that the ODDSRATIO statement may give the same results for contrasting males vs. females and ascending limb vs. descending limb. The ESTIMATE statement might be another option, though I'm not certain on the specifics on this yet.

The reason I'm interested in the intercepts is that I'm hoping to statistically examine the distribution of responses across levels of the DV (distance participants were willing to drive while intoxicated, which has 4 levels: 0 miles, 1 mile, 3 miles, 10 miles). For example, when blood alcohol level is descending (when limb = 1, descending) how much further are participants willing to drive? That is, after accounting for the IVs in the model, is the "typical participant" willing to drive 3 or more miles on the descending limb, but only 0 or 1 mile on the ascending limb? If I center on the limb I'm interested in and center other variables at their means, can the intercepts be interpreted this way in a GLMM? I realize that the cumulative logit model is calculating the effect across all levels of the DV, but is there a way to distinguish between levels of the DV (i.e., proportion of participants at a certain level of the DV or below is more than the other levels of the DV). 

This SAS paper (https://support.sas.com/resources/papers/proceedings15/3430-2015.pdf) makes me think that I can interpret the intercepts, but I'm still a still a little uncertain whether I should be doing this, which levels of the DV the intercepts would be distinguishing, if so, and whether the p values associated with the intercepts are informative in drawing conclusions about what the "typical participant" (all IVs centered) did. 

[PaigeMiller]

It looks like the LSMEANS statement doesn't work in multivariate models

I'm not sure why you say this or what you mean, but as far as I am concerned, this is not a correct statement.

Bottom line: if you want to interpret the model, LSMEANS for categorical variables is a better solution than anything else you have mentioned. If you want to make predictions from a model, then you need the intercepts and regression coefficients. Yes, this is a bit of oversimplification, but you ought to start here, and only if this simplified approach doesn't provide what you need should you go to more complicated approaches.

[cmo5]

Hi Paige,

Thanks for the reply. Here's my edited syntax and the error that I received in GLIMMIX when specifying LSMEANS.

proc glimmix data=three method=laplace empirical=mbn noitprint noclprint ;
	where PacketN>2;
	class Subj Female AscendLimb;
	model DriveDist (order=data)= Female cBase_AIDattitudes AscendLimb / 
                DIST=MULTINOMIAL LINK=CLOGIT SOLUTION CL ddfm=bw 
		oddsratio(DIFF=LAST LABEL);
	random intercept / sub=Subj TYPE=VC;
	lsmeans Female AscendLimb / ilink ;
	covtest / WALD;
run;

[cmo5]

In my previous post, I meant to say that LSMEANS does not work in "multinomial" rather than "multivariate" models.