dear all, I'm trying to run a two part growth mixture model. I did a lcga for the binary part first.
1.) 3 or 4 class solutions fitted better than a one class or two class solution, but I get bad average post probabilities for the 3 or 4 classes. If I have more than two groups, at least on group has a averagepost-prob of 60-70 (diagonal of matrix). What is an acceptable value? I thought of 80!? Are there any chances to isolate more than two meaningful and valide groups in a lggmm of boh parts? Is it worth it to proceed?
2.) I went to joint LGMM of both parts. I used the means of the growth factors lcga in the following lgmm as starting values, but it didn't converge.
After nearly 2 hrs, I get the following messages: THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO A NON-ZERO DERIVATIVE OF THE OBSERVED-DATA LOGLIKELIHOOD.
THE MCONVERGENCE CRITERION OF THE EM ALGORITHM IS NOT FULFILLED. CHECK YOUR STARTING VALUES OR INCREASE THE NUMBER OF MITERATIONS. ESTIMATES CANNOT BE TRUSTED. THE LOGLIKELIHOOD DERIVATIVE FOR PARAMETER 20 IS -0.55772741D-01.
Any suggestions with regard to model convergence, additional starting values, may be variances? I was inspired by the recent Hser et al. article (2007). I think they used one latent class for both model parts. Is it significantly more computational demanding to use 2 latent class variables for both parts separately?
1. I would say any entropy value is acceptable in the sense that entropy does not have to do with model fit or the best choice of number of classes. But to be useful for classification, I would think you want higher numbers.
2. Just increase to say miter=1000; as suggested in the error message.
2. I will try but I'm nt very optimistic that it works.
I'm very interested in isolating a "zero-group" (structural zeroes). How do I specify such a group? I've heard of threshold = 15 for the binary part and zero mean for intercept and slope of the continuous part. Do I have to add zero means of the growth parameters for the binary part or is threshold = 15 sufficient?
o.k., I have given it up, it is far too complex and miles away from convergence. I focus on an alternative, namely to isolate groups within the entire model (both parts) on the basis of LCGA, which is often done by people who use SAS (group based approach, Nagin). I know the disadvantages (too many groups may be isolated by LCGA) and it is rough and dirty but may be sufficient for me. Entropy looks very good.
1.)Is LCGA compatible with two-part in general? I only have read an article by Bengt, using LCGA on the binary part (but not on both parts simultanously).
Two-part modeling requires that the two parts correlate in some way. Generally this is through the growth factors. If you do an LCGA, the growth factor variances are fixed at zero and there can be no growth factor covariances between the two processes. LCGA can be done with a two-part model if the latent class variable influences both parts. The following paper on two-part factor mixture may be of interest. It can be downloaded from our website.
# Kim, Y.K. & Muthén, B. (2007). Two-part factor mixture modeling: Application to an aggressive behavior measurement instrument.
o.k., your first statement was the reason for asking this question :-).
"LCGA can be done with a two-part model if the latent class variable influences both parts."
Sorry, this was a missunderstanding. Bengt used LCGA on the separated binary part to get a clue about the number of groups in general. This was followed by LGMM on both parts. Now I'm separating groups on the basis of both parts simultanously with one class variable in my "Variable" command. Your answer implies, that this seems o.k., albeit having no correlated parts within the classes.
thank you again, as always :-) So I have to use the means not adjusted for covariates for the visual plot in my paper, if this is not possible with imputation. I want to calculate confidence intervals for these means. Using "cinterval" only gives me cintervals for my model but not for the estimated means. I there any way to get them out of mplus?
I noticed that the current version 5.2 contains improvements concerning type=imputation. Has anything changed regarding means available in this kind of analysis? This would, for instance, alleviate the exploration of LGMM conditioned on imputed covariates, since sometimes the mean trajectories of the classes look different as compared to unconditional models. So far, the only chance to have a closer look on this issue is with non-imputed conditional models.