
Message/Author 


I'm trying to estimate a 3class GMM. I have set quad var to zero but want to allow i & s var to differ across classes. I specify random starts with starting values. C#1 needs the s var to be fixed @0, and s var are est separately for c#2 & c#3. When I order the class starting values to put the largest class last and reestimate, the means "move", but the fixed s var "stays" with the c# it was fixed for. Now, the new model gives me the same "not pos definite" warning for the class that used to be c#1, and the new c#1 has the s var fixed at zero. How can I fix the var for a class allowing other class var to differ, and still get a stable solution? Here are example starting values: MODEL: %OVERALL% i s q  sa0@0 sa1@1 sa2@2 sa3@3 sa4@4 sa5@5 sa6@6 ; q@0 ; %c#1% [ i*40 s*6 q*0.6 ] ; i*56 s@0 ; %c#2% [ i*46 s*0.5 q*0.2 ] ; i*75 s*2.9 ; %c#3% [ i*37 s*3 q*0.2 ] ; i*30 s*0 ; In the last run, c#3 moved to c#1 and had a fixed s var, and c#1 moved to c#3 and gave the "not positive definite" warning. Thanks, Bruce 


Can't you do a run with the s variance free and request SVALUES in the Output command to save the estimates. Then you do a new run with Starts=0 and  if you like  switch the classes around to get the largest class last (the class logits won't be right, but you can refigure those if need be), and fix the s variance in the class that you want it fixed. 


Thanks Bengt  Again, forgetting the basics! I've been using the final estimates to fix the starting values for the next run, but the result from your suggestion was much more informative. I got starting values, but the result from the 3class solution produced a third class that was only 3% of the sample, and (1) the best three LL nonreplicated with 800 random starts, (2) 29 parameters fixed to avoid a singular information matrix, and (3) nonpos def cov matrices for two classes involving i in C#1, and q in C#3. From this, I'm guessing that I'm trying to squeeze water from a stone, and the program is telling me in a more fundamental way (than the BIC, VLMR, and BLRT), that 3 classes won't fit the data better than 2 classes without using very samplespecific fixed parameters and starting values. Does that sound right? Bruce 


Letting growth factor variances vary across classes can be dicey. A good way to approach it is to first go by the default which holds the variances class invariant, then look at the plots for estimated classspecific mean growth curves versus observed curves for people most likely to be in each class. That can show if one class seems to have less/more observed variability than the other classes. Then go and free the variance(s), such as for the intercept growth factor, for that class. 


Thanks, Bengt! I'm responding via support email, because I can't get my note to fit in 1.5K!  bac 

Back to top 

