LCGA or GMM: which suits better? PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
 Kiki van Broekhoven posted on Tuesday, April 18, 2017 - 5:24 am

I have carried out a LCGA with a continuous outcome for depressive symptoms, and found a 3-class solution. My main interest is in identifying groups of women who display a similar course of depressive symptoms over time.

However, the plots of the estimated means and observed individual values show quite a lot of variation regarding intercept, slope, and quadratic effect for the individual growth curves.
Therefore, a colleague of mine proposed that I try a GMM solution, thus with free estimation of variances and covariances. However, I am now reading the book by Nagin on group-based modeling and start to doubt about what I should use - LCGA or GMM, as GMM really alters the conception of "group".

Do you have any useful suggestions for me?
 Kiki van Broekhoven posted on Tuesday, April 18, 2017 - 5:28 am
As an additional remark:

I am currently trying to run this GMM solution with 3 classes. This started at 12.17 PM and is still running (it is now 2.27 PM, thus it has been running for more than 2 hours and is still not ready).

Is such lengthy computation time normal?
 Bengt O. Muthen posted on Tuesday, April 18, 2017 - 6:11 pm
On your first post, I have described my arguments that counter Nagin's arguments in the article:

In short, I don't think this choice should be a philosophical matter but a matter of statistical choice based on fit such as BIC.

On your second post, without seeing your run, I can only say that a long run time may be due to many factors such as

a slow computer

a large sample

non-normal outcomes resulting in many dimensions of integration (see TECH8 screen printing)

TECH8 will give you an idea of how long it will take.

If this doesn't help, send input and data to Support along with your license number.
 Kiki van Broekhoven posted on Wednesday, April 19, 2017 - 4:19 am
Thank you for the article, that was very helpful.
I have a very new and fast computer so that cannot be the problem; my sample size is 1428.

How do I know whether I have non-normal outcomes? (I am sorry, I do not know exactly what this means)

Yesterday I kept the model running for 9 hours and then it looked like it was kind of "stuck"; nothing happened anymore. However, I did not get any error messages. Should I just let such a model continue to run? I don't know whether it's normal that it takes this long?
 Kiki van Broekhoven posted on Wednesday, April 19, 2017 - 4:30 am
As a related question (I am sorry if it is a stupid question but I am relatively new to Mplus)

If I do happen to obtain results for the GMM models in the foreseeable future: do I have to start over the process of selecting the number of classes, based on BIC, BLRT, LMR-LRT etc, as I have done for the LCGA model? Or could I say something like: the LCGA pointed to a 3-class solution, and model fit increased (i.e. BIC decreased) when we performed a GMM, thus we chose the 3-class GMM model?

Thank you in advance.

- And another question (maybe this also is very stupid):
When I perform a LCGA, and then I remove the line i-q@0, does this make it a GMM with equal variances for the trajectory growth factors?
And then when I add
i s q;
for each class, does that make it a GMM with class-specific variances

i.e., are both of these models to be called GMMs?

Thank you in advance.
 Bengt O. Muthen posted on Wednesday, April 19, 2017 - 11:59 am
Answered via Support.
 Kiki van Broekhoven posted on Friday, April 21, 2017 - 7:51 am
Thank you!
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message