Understanding Interaction Effects in PROC GLIMMIX

[njgrubic]

I am running a two-level multivariate logistic regression model in PROC GLIMMIX. My relationship of interest of the association between "intervention" and "outcome". My clustering variable is "level2". My moderating variable is "class" and I am adjusting the model for "a", "b", "c", and "d" confounders. All variables are categorical.

I am interested in identifying the effect if "class" on the relationship between "intervention" and "outcome". Most of the interaction terms in the Solutions for Fixed Effects tables are significant (p<0.001), although the type III tests of fixed effects shows a non-significant p-value for the overall "intervention*class" term (p=0.3881). The intervention*class Least Squares Means are all significant (<0.001). I am not exactly sure how to interpret these tables and could use some advice.

I am then provided with the Simple Effect Comparisons of intervention*class Least Squares Means By class. From my understanding, these are the effect estimates of the relationship of interest, stratified by my class levels. Is this correct? Also, is there a way to compare these effects apart from just observing to see if the ORs overlap?

Thank-you!

proc glimmix data=cares2;
class level2 outcome (ref="No") intervention(ref="No") a(ref="No") b(ref="No") c(ref="No") d(ref="No");
model neurooutcome = intervention a b c d intervention*class / dist=binary link=logit ddfm=bw solution oddsratio;
random intercept / subject=level2;
lsmeans intervention*class / slicediff=class oddsratio cl ilink;
run;

[PaigeMiller]

Show us the output you are looking at. You can make a screen capture and then include it in your reply by clicking on the "camera icon". Do not attach files.

[njgrubic · Accepted solution]

Thank-you Paige. I have attached the three table outputs I am confused about.

   

[PaigeMiller]

Most of the interaction terms in the Solutions for Fixed Effects tables are significant (p<0.001), although the type III tests of fixed effects shows a non-significant p-value for the overall "intervention*class" term (p=0.3881)

The solutions output is testing if the coefficient is zero. They are not, they are statistically different from zero, because the p value is less than 0.05.

The Type III test is testing if all of the coefficients are equal. They are equal, because the p value is greater than 0.05. So no difference between the coefficients.

I am then provided with the Simple Effect Comparisons of intervention*class Least Squares Means By class. From my understanding, these are the effect estimates of the relationship of interest, stratified by my class levels. Is this correct?

You are testing to see if "yes" is diffferent than "no" for FIRST and then for SECOND and so on, is that what you want?

Also, is there a way to compare these effects apart from just observing to see if the ORs overlap?

Since the Type III effect of the interaction in the model is not significant, I would assume this means that these are not different.