PROC GLIMMIX Contrasts

[mthorne]

Hi,

I am using proc glimmix to analyze count data. I notice that the pvalues for a contrast between two simple effects are the same as the pvalues for lsmeans diff for the same simple effect. If I use the contrast comparison, do I report the actual means or the lsmeans?  

[PaigeMiller]

Depending on how you did the contrast, it can indeed be the same as the LSMEANs difference. Not just the p-values should match, but the actual difference should match.

Again, depending on what you did, I would report the LSMEANs and their differences as a general rule, since I have seen only very rare situations where the contrast is actually needed. I would report Means instead of LSMEANS only if you have a completely balanced and orthogonal experiment, or you have the equivalent of a one-way ANOVA.

[mthorne]

Thanks for the reply!

This is my model.

ods graphics on; 
proc glimmix data=counts method=laplace nobound plot=studentpanel; by loc time;
class loc time herb rep;
tc2=log(c2+1);
model tc2 = herb / dist=normal link=identity;
random intercept / subject=rep;
lsmeans herb / ilink diff; 
contrast '0 vs 1' herb 1 -1 0;
contrast '0 vs 2' herb 1 0 -1;
contrast '1 vs 2' herb 0 1 -1;
*output out=allresid pred=p resid=student;
run; quit;
ods graphics off;

 The design is a balanced factorial with no missing data. Three locations, four application times, three herbicide treatments, and four replicate blocks at each location. Glimmix couldn't handle all zeros from some treatments, so I had to transform and compare to a normal distribution.

[PaigeMiller · Accepted solution]

contrast '0 vs 1' herb 1 -1 0;
contrast '0 vs 2' herb 1 0 -1;
contrast '1 vs 2' herb 0 1 -1;

There's no need to type all this out, the LSMEANS statement with the DIFF produces the same results.

[mthorne]

Great, thanks Paige!

Mark