socrates posted on Thursday, February 05, 2009 - 4:08 pm
When using GMM in my lonigtudinal dataset, I always run into convergence problems. However, when fixing the variances and covariances of the latent growth parameters (intercept, slope, and quadratic term) at zero, a LCGA solution with three latent classes turned out to be optimal. In a further step, I would now like to examine predictors of individual differences in development over time by regressing the latent growth parameters on a set of potential predictors in each of the classes (as it is often done in GMM). Is this allowed although I fixed the growth parameter (co-)variances at zero in my LCGA?
With LCGA you would let the covariates predict the probability of class membership, not the growth factors. When you use LCGA the growth factor variances are zero so that there is no variability to predict in the growth factors. Your non-converging GMMs might benefit from having growth factor variance for say only the intercept, not all growth factors.