I am fitting a repeated measures model using proc mixed, data are from randomized parallel clinical trials with three treatment groups, patients are followed for a few weeks with measurement at each week after baseline. A big percentage of the patients discontiued before the end of the study (dropped out at different visits). Below is the model I try to fit:
Proc mixed;
Class treatment visit patientid;
Model change = baseline_score treatment visit treatment*visit /s;
Repeated visit / subject=patientid type=un;
lsmeans treatment*visit;
I am very puzzled that the lsmeans by visit and treatment from this model fitting are quite different from the raw means, with almost all (except one visit for one group) bigger than the raw means, which is odd. Could someone offer any insight how is this possible (because of the inbalance re discontinuation)? Or something wrong with my proc mixed model? I thought this is a pretty straightforward model.
Any response will be appreciated.
Best,
John