

Singular inf matrix probs for gmm w/b... 

Message/Author 


Hi Linda  I'm estimating a GMM with MLR for pain occurrence as a binary variable across 7 months. I'm getting conceptually very interesting results, but in trying to finalize the 3class and especially the 4class solutions, I keep getting errors that result from (at least partly), classes that keep shifting in order. One problem is that the L and Q slope estimates are sometimes astoundingly big (and I don't know how they are scaled). For example, in the most recent problematic model, the problem class has S= 539.098 and Q= 269.345 and the program fixed Q (threshold is 1.723). But what is that S?!? But S is large for another class also (8.117). Basically, the primary problem class has no pain at baseline, 100% with pain at month 2, then a steep decline in the probability of pain over months 37. Any help with these estimates and how to keep the classes in order would be great! Thanks, Bruce 


Please send the output and your license number to support@statmodel.com. 


Thanks, Linda  So the 2nd class in my 4C model starts at close to a P=0, then goes to P=1 at T2, then returns to P=0 by T4. They represent a meaningful class. (1) How can I model the group? I can't fix the S and Q to e(logit) = infinity or minus infinity. (2) Can I estimate a mixture model for these data with a piecewise approach? If so, what would the model look like, taking off from i s q  bpn1m0@0 bpn1m1@1 bpn1m2@2 bpn1m3@3 bpn1m4@4 bpn1m5@5 bpn1m6@6 ; I specified a model with the 2nd piece beginning at T4 and got interesting est & plots. i1 s1 q1  bpn1m0@0 bpn1m1@1 bpn1m2@2 bpn1m3@3 bpn1m4@4 bpn1m5@5 bpn1m6@6 ; i2 s2 q2  bpn1m0@0 bpn1m1@0 bpn1m2@0 bpn1m3@1 bpn1m4@2 bpn1m5@3 bpn1m6@4 ; Before I encourage a migraine, does this look right? Thanks! Bruce 


You should not do anything. You should simply describe the class. 

Back to top 

