I'm estimating a growth mixture model with 4 classes, continuous indicators, and linear, quadratic and cubic growth factors (to mimic TRAJ). I started with growth factor variances and covariances fixed at 0. That converged fine. Next I freed the growth factors. That converged fine. The last step I am attempting is to allow Psi and Theta to vary among the classes, and I cannot get it to converge. I've tried a number of different starting values including the invariant value estimates. Sometimes the problem is with a non-invertible covariance matrix for one class. Other times there is a non-positive definite Fisher-info matrix, and a third error message is that the loglikelihood increased at a particular iteration. There are a number of negative variances in the interim estimates for either the indicators or the growth factors or both in one or more classes. Any suggestions?
As I understand it, you were able to get the TRAJ model to converge. And when you tried to move away from the TRAJ model and free the variances and covariances of the growth factors (psi), that was fine. But when you started allowing psi and theta to be unequal across classes, you ran into problems.
Freeing all of the elements of psi and theta across classes often leads to instability. I suggest the following steps:
1. Free the variance of the intercept growth factor across classes -- start with one of the extreme classes first.
2. Free the variances of theta across time and classes -- start with the one of the extreme classes first.
3. Go on to the other growth factor variances and covariances.
4. Same thing for theta.
It is particularly difficult to allow the variances of the quadratic and cubic growth factors to be different across all classes.
Anonymous posted on Monday, May 02, 2005 - 1:56 am
Which command allows Psi to vary across classes? I checked the manual and I am sure it is in there somewhere but I could not find it. I am using a linear 2 class latent growth mixture model and regress i and s on a time-invariant covariate (sex). It appears that by default the residual variances between groups are constrained to be equal, however, I am interested in the moderating effect of class membership on the relationship between i (s respectively) and a covariate. Here is my Model statement: %OVERALL% i s | t1@0t2@1t3@2t4@3t5@4t6@5; i s ON SEXX; c#1-c#2 ON SEXX; THANK YOU!!!