Warning GMM
Message/Author
 Corinna posted on Tuesday, August 18, 2015 - 8:21 am
Dear Linda & Bengt,
I am trying to calculate a GMM with a continuous outcome variable. When I am freely estimating intercept and slope I get the following warning for i and s and every class:
WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IN CLASS 1 IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/ RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL
TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES.CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE S.

When I am fixing i and s I do not get the warning. Does this indicate that the model with free i and s is to complex and cannot converge? Should I therefore decide to fix i and s to 0?

Thank you very much.
 Linda K. Muthen posted on Tuesday, August 18, 2015 - 8:27 am
You should look at the output and see if there are negative residual variances/variances for s. Only if they are small and not significant should they be fixed at zero.
 Corinna posted on Tuesday, August 18, 2015 - 8:36 am
Yes there are negative residual variances for my outcome variable. What does that mean? Did I specify the model incorrectly?

Thank you very much.
 Linda K. Muthen posted on Tuesday, August 18, 2015 - 2:55 pm
It means the model is not appropriate for the data. You should change the model.
 Corinna posted on Wednesday, August 19, 2015 - 12:59 am
Yes, that's what I thought as well. But it only happens when I am freeing the variances of i and s or when I am using the default option (for the variances of i and s) or including a quadratic slope. So the only consequence I see is to fix the variances of i and s to 0. In this case the model converges and shows very good decision criteria (entropy, posterior class probability). Or is there another possibility?

Thank you very much.
 Linda K. Muthen posted on Wednesday, August 19, 2015 - 9:41 am
 Cristan Farmer posted on Monday, February 08, 2016 - 1:49 pm
Hello,

I am working on a dual-trajectory growth mixture model but am new to this methodology. In my dataset, there is limited coverage at the final timepoint (6). I found the best-fitting GMM for each variable and am now trying to assess the joint model of the 2-class variable 1 and 3-class variable 2. I am getting the following error:

THE ESTIMATED COVARIANCE MATRIX FOR THE Y VARIABLES IN CLASS 1 COULD NOT
BE INVERTED. PROBLEM INVOLVING VARIABLE CORAW6. COMPUTATION COULD
NOT BE COMPLETED IN ITERATION 3. CHANGE YOUR MODEL AND/OR STARTING
VALUES. THIS MAY BE DUE TO A ZERO ESTIMATED VARIANCE, THAT IS, NO WITHIN-CLASS
VARIATION FOR THE VARIABLE.

When I did the single-trajectory GMM for this variable, I had a problem with theta because of timepoint 6, but because the variance was small and negative, I set it to zero and the model ran fine. So, in this joint trajectory model, the variance of CORAW6 is fixed to zero. Can you tell me how I might find out more information about the problem? Thank you so much for your time.
 Bengt O. Muthen posted on Monday, February 08, 2016 - 6:25 pm
 Daniel Lee posted on Friday, September 30, 2016 - 7:51 am
Hi Dr. Muthen,

In a 2 class solution for my GMM, the variance for intercept and slope in class 1 and 2, and the residual variance for the observed measurements, were identical in class 1 and 2...Of note, the means for the intercept and slope terms were different by class. Nevertheless, that does not seem right since I would naturally expect the mean of int/slope to vary by class (not be equal by class). In the code, I did not include any equality constraints for these parameters, so I'm doubly confused now! Would you have any insight as to what might be occurring? Thank you!
 Bengt O. Muthen posted on Friday, September 30, 2016 - 3:46 pm
The default in mixtures is that means vary across classes and variances don't. Your second and third sentence appear to contradict each other.

You can change the defaults any way you like.