I think it comes down to the fact that you have a fair amount of missing values, so the marginal odds ratios over day are strongly influenced by the slopes at each day.  For instance, for out=3, at Day=0 you have 18 observations, while at Day=1 and Day=2 there are almost 3 times as many (in the fifty-ish range).  For out=2, the dropout rate from Day 1 to Day 7 is nearly 86%.  So, even though you are not explicitly modeling the repeated nature, it is affecting what is going on.  By aggregating over time, the trend due to inlab becomes confounded with the day of measurement. Since you are using pseudolikelihood, adding a repeated effect (although it has to go in as a G side effect) can be done as below: proc glimmix data = vdata  noclprint method = rmpl; class id out day; model out (ref = '1') = lnlab*day / dist = multinomial link = glogit solution ; random intercept /subject = id type = un group = out; random day/ type=ar(1) subject=id; run; Two parts to check here--first, will it even run, and second, is it still sensitive to the specification of the reference group.  Note that the repeated nature is modeled as an autoregressive process (AR(1)), which assumes equal spacing in time.  If Day is representing a "visit" (to borrow from clinical biostats) such that these are not equally spaced in time, you may wish to consider other structures. Steve Denham ... View more