topic Re: Common reference category for interaction in PROC GLIMMIX in SAS Programming

Common reference category for interaction in PROC GLIMMIX

njgrubic — Tue, 10 Nov 2020 14:44:17 GMT

I am conducting an interaction analysis in PROC GLIMMIX and would like to present my interaction analysis with a single reference category. The interaction is between two categorical variables (2-level and 5-level variable). I have only been able to build obtain an output displaying ORs with separate reference groups. See tables below - I am attempting to display results similar to Table 6 with a single reference group.

proc glimmix data=cares2 method=quad;
class group outcome (ref="No") intervention(ref="No") area(ref="small") gender(ref="Female") /;
model outcome = intervention area gender age intervention*area / dist=binary link=logit ddfm=bw solution oddsratio;
random intercept / subject=group;
lsmeans intervention*area / slicediff=area adjust=bon oddsratio cl ilink;
run;

Re: Common reference category for interaction in PROC GLIMMIX

Reeza — Tue, 10 Nov 2020 15:58:17 GMT

You'll need to manually double code that and you can get an OR table that compares the different levels.

One quick way:

Asbestos_Smoker = catx(" - ", var1, var2);

Then use the asbestos_smoker variable in your model instead of the interaction term and specify your combined reference level as desired.

Re: Common reference category for interaction in PROC GLIMMIX

njgrubic — Thu, 12 Nov 2020 02:03:53 GMT

Thanks Reeza. I have attached an image of part of the Differences of intervention*area LSM output adjusted for multiple comparisons below.

Do the 'pre-underscored' variables (i.e. _intervention, _area) indicate the reference category that is being used to calculate the OR? In this case, would I be able to obtain these comparisons by locating the appropriate comparisons in this table, as opposed to creating a new variable?

Re: Common reference category for interaction in PROC GLIMMIX

Reeza — Fri, 13 Nov 2020 20:26:50 GMT

No, but I do think you could if you used a custom ODDS ratio statement but you'll have to manually specify it. Recoding the variable is the simplest option as the output won't be ambiguous in any manner.