Hi, I am a bit confused about the difference between GMM and LCGA. I am reading your article (Muthen & Muthen, 2000) and am starting to understand the difference, but could use a definition directly from you.
GMM and GGMM refer to continuous outcomes. For categorical outcomes, only LCGA is available in the current version of Mplus. LCGA is latent class growth analysis that allows for mixtures with no within-class variation.
Mplus Version 3 is unique in making that analysis possible using ML with numerical integration over the within-class random effects (continuous latent variables). You request analysis type=mixture and algorithm=integration. Note that mixture modeling with categorical outcomes is harder than with continuous outcomes and you want to use aws few random effects as possible, e.g. only making the intercept random.
LCGA has no within class variance of the growth factors. If there is no within class variance of the growth factors, GMM and LCGA should look the same. Variation being the same could be no variation or all classes with a lot of variation.
EFried posted on Wednesday, February 08, 2012 - 7:21 am
I'm running 10 GMMs, one with a continuous outcome variable and nine with categorical outcome variables.
From the videos on the website, the Wickram 2008 GMM paper and also from what I read in the MPLUS 6 manual so far I got the impression that the difference between LCGA and GMM is:
However, on p.219 of the manual, chapter 8.09 you state that "when type=mixture without algorithm=integration is selected, a LCGA is carried out". This was new to me and honestly I don't find "algorithm=integration" in the GMM papers I've read so far.
The difference between LCGA and GMM is that in LCGA growth factor variances are zero and in GMM they are estimated.
TYPE=MIXTURE without ALGORITHM=INTEGRATION does not estimate growth factor variances. With ALGORITHM-INTEGRATION, it does.
EFried posted on Wednesday, February 08, 2012 - 8:48 am
Thank you for your quick reply.
(1) Is there a difference between "fixing GF variances to zero" and "not estimating GF variances"?
(2) I just ran two GMM (continuous outcome variable) a) ! integration=algorithm; i-s@0; b) ! integration=algorithm; !i-s@0; If !integration=algorithm; fixes the growth factors variances to zero, the results should be the same for both models, but they are not. I must misunderstand something, obviously.
(3) Does this have to do with categorical vs. continuous outcome variables? Is there some paragraph or paper I can read this up on? I don't find an overview about this in the manual, but it looks like the integration algorithm is only mentioned in examples for categorical outcome variables.
EFried posted on Thursday, February 09, 2012 - 10:52 am
Linda, before I burden you with the output:
do I understand correctly that ALGORITHM = INTEGRATION; is only used for models with categorical/count outcome variables, and not including it into models does the same as the model statement "i-s@0", that is fixing the variance of the intercept and slope growth factors to zero?
For TYPE=MIXTURE, leaving out ALGORITHM=INTEGRATION is the same as fixing the growth factor variances and covariances to zero. This applies to all variable types.
EFried posted on Thursday, February 09, 2012 - 1:38 pm
(1) According to the manual and your videos, type=mixture is the only viable type option for conducting GMMs - correct?
(2) Algorithm = Integration is said to be computationally very demanding (both in the manual and in a couple of forum posts I found). With N=1000, 5 measurement points, 7 time-invariant and 2 time-varying covariates, is algorithm=integration a viable option for running GMMs for me? If not, does that mean that with my model it is not possible to run GMMs, only LCGAs?
Thank you so much, this is incredibly helpful for me. Eiko
2. The computational burden with categorical outcomes is a function of the number of growth factors and the sample size. With your sample size, 2-3 growth factors (linear or quadratic growth) is not computationally demanding.
EFried posted on Thursday, February 09, 2012 - 2:04 pm
Thank you, Linda!
1. Wonderful! What is the computational burden for continuous outcomes?
2. If leaving "Algorithm = Integration" out of a type=mixture model specification, models with and without the "i-s@0;" statement should be identical models, because both should have their growth factors (co-)variances fixed to zero, right? This is not the case for me, results are drastically different. Why is that?
2. Leaving out integration and zero variances are the same. If you do not get the same results, you are doing something else. Please send the two outputs and your license number to firstname.lastname@example.org.
I have a short question: does it make any sense to test the class invariance assumption of residual variances when modeling LCGAs? In my data, I found that relaxing the invariance assumtpion of residual variance across trajectory classes resulted in a considerably better model fit based on BIC and the impact on trajectory class solutions was relatively big. However, to date, I have only heard of testing the invariance assumtpion of residual variances when GGMMs were used.
Thank you! I'm sorry, my post was formulated in a misleading way. I was interested in class-specific residual variances in growth mixture models and I wanted to know whether it is a good idea to test the need for class-specific residual variances in Nagin models (LCGA). So far, I have only heard of the potential relevance of class-specific residual variances in growth mixture models (LGMM; cf. Enders & Tofighi, 2008). In my LCGAs (used as a first step to determine possible trajectory classes) the specification of residual variances as class specific appears to have a great impact on the final solutions in my data. But I'm somewhat confused because I have never seen that anybody checked for class specific residual variances in LCGAs (and very seldom in LGMMs).
I think a good strategy with both GMM and LCGA is to go through the sequence
1. use class-invariant variances
2. plot and see which classes might need class-specific variances
3. rerun and see if BIC is better
Class-specific variances can give different class formations. I think that should be checked for as above in both model frameworks. The Nagin model also holds residual variances equal across variables within class, and that too should be investigated and relaxed.
Thank you! Btw, is there any rationale why mplus, by default, estimates time-specific residual variances in all kinds of growth models? SAS, and Proc Traj, to my knowledge, estimates time-invariant residual variances by default (as you have mentioned in your last sentence).
In the cases you consider, I believe those other programs take a univariate, two-level, long data format approach to growth modeling, whereas in Mplus you also have the option of a multivariate, single-level, wide data format approach. In the first approach it is not possible to have time-specific variances, whereas in the second approach you can allow that and test if that is needed - which it often is. We recommend the second approach when it is possible.
Hi, I have a question concerning the differences between GMM and LCGA. It is said in the User’s Guide on the page 219 that “When TYPE=MIXTURE without ALGORITHM=INTEGRATION is selected, a LCGA is carried out.” In the first GMM example (8.1) on the page 201 there is no “ALGORITHM=INTEGRATION” mentioned in the analysis command. Why is that? When I look at the LCGA example 8.11 on page 220 there is no ALGORITHM=INTEGRATION (as it is not supposed to be) but the variances are not fixed at zero either ("i-s@0"). What is the difference between these examples, 8.1 and 8.11? What makes the former an example of GMM and the latter an example of LCGA?