topic Re: Different parameter estimates between proc logistic, proc glimmix, and proc genmod in New SAS User

Different parameter estimates between proc logistic, proc glimmix, and proc genmod

AXR — Wed, 26 Feb 2020 18:06:17 GMT

Hello,

I'm testing out some logistic regression models and I am comparing results across proc logistic, proc glimmix, and proc genmod. I wanted to first look at models without specifying any random effects. With these fixed effects models, I anticipate that the coefficient estimates would be the same across procedures; however proc logistic produces substantially different coefficient estimates for the fixed effects (proc glimmix and genmod produce the same coefficient estimates). Do you know what would be causing this? I've pasted example code below.

Addendum: I've noticed that the different parameter estimates are specifically for variables in the class statement (I am specifying the same reference category across models).

proc logistic data = dat desc;
	class  var1 (ref = "0");
	model outcome = var1 var2 var3;
run;

proc glimmix data = dat;
	class  var1 (ref = "0");
	model outcome = var1 var2 var3/solution dist=bin link=logit;
run;

proc genmod data = dat;
	class  var1 (ref = "0");
	model outcome = var1 var2 var3/dist=bin link=logit;
run;

Parameter estimates output

Re: Different parameter estimates between proc logistic, proc glimmix, and proc genmod

PaigeMiller — Wed, 26 Feb 2020 17:43:01 GMT

Oh come on, show us the output!

And you say

(I am specifying the same reference category across models)

but the code you show doesn't do any such thing.

Re: Different parameter estimates between proc logistic, proc glimmix, and proc genmod

AXR — Wed, 26 Feb 2020 18:06:54 GMT

updated the post.

Re: Different parameter estimates between proc logistic, proc glimmix, and proc genmod

PaigeMiller — Wed, 26 Feb 2020 18:16:15 GMT

The GLIMMIX and GENMOD estimates match. The LOGISTIC estimates are the exact same model, just parameterized differently. If you had shown the Intercept from LOGISTIC, we could all see these are the same models. These models give the exact same predictions.

Instead of looking at parameter estimates for CLASS variable levels, you ought to be using LSMEANS, which I suspect will provide the same LSMEANS for each of the three PROCs.