I am using a mixed linear model to analyze the effect of two types of treatment on symptoms. The two treatments were administered in smaller groups (clusters), and delivered by several doctors, so I need to account for clustering (ICC is substantial). I am using difference in symptom scores from baseline as my outcome measure, treatment type and baseline symptoms (to control for them) as fixed effects, and group cluster and doctor, specified in the 1st level as "subjects" as random (intercept) effects. My model is as follows:

MIXED outcome BY treatment WITH baseline_sx
/CRITERIA=CIN(95) MXITER(500) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE)   
/FIXED=treatment baseline_sx | SSTYPE(3)  
/METHOD=REML   
/PRINT=G  SOLUTION TESTCOV   
/RANDOM=INTERCEPT | SUBJECT(doctor*cluster_gr) COVTYPE(VC).

For some reason, the model chokes (gives me a warning about the Hessian matrix not being positive definite) and does not produce any estimate for my random effects. It does just fine for very data from a different outcome point -- the only difference is that some observations are available for one outcome point and not the other, but overall the number of observations is about the same. I don't think it has anything to do with covariance structure -- tried different kinds, does not seem to solve the problem. I pretty much narrowed it down to some kind of problem in one of the groups, possibly having to do with cluster variance... Has anyone encountered anything similar or have any suggestions for how to overcome this issue?