Hello,

I am using Mplus to conduct a latent transition analysis. I have examined both regular LTA and LTA with random intercepts and found that a five-class RI-LTA model best fits the data. Each class has three indicators. I also found that a full invariance model resulted in a significantly worse model fit than a non-invariant model. I want to test for partial invariance as a visual inspection of the results suggests that the means are very similar on specific indicators within a class across time but vary somewhat for other indicators. I want to check whether constraining specific means to equal across time improves model fit.

I have a couple of questions. When I estimated the RI-LTA, I only constrained the means, but should I also constrain the variances? Also, is there a simple way to determine if any means should be constrained to be equal, such as using modification indices? If not, should I label each indicator and compare them using the model test option?

Is a RI-LTA an appropriate method, or is a regular LTA fine? I chose this approach because I am examining change and stability in psychopathology across time and believe there might be some trait-like influences on stability. I found that factor loadings for conduct problems and hyperactivity are relatively high (between .55 and .85), but factor loadings for emotional problems are lower (around .20). From what I understand, this doesn’t necessarily suggest that RI-LTA is inappropriate, only that emotional problems are less stable across time.

Any suggestions or references to papers with relevant examples and syntax would be greatly appreciated.