Re: st: How to assign class membership in latent growth curve (SEM)

[Clyde B Schechter <clyde.schechter@einstein.yu.edu>]

Mark Ward wrote:

"I am attempting to develop a latent growth curve model (within the SEM
framework) in order to examine weight(Kg) trajectories for children. I
have 3 time points (birth, 9 months, and 3 years). For now I am using
a sub-set of my data with no covariates. My thinking here is to begin
with as simple a model as possible and gradually build on this.

I have built a model using a sub-set (n=200) of my data:

sem (Intercept@1 -> weightBirth) (Intercept@1 -> weight9Month)
(Intercept@1 -> WEIGHT3Y)(Slope@0 -> weightBirth) (Slope@1 ->
weight9Month) (Slope@3 -> WEIGHT3Y), covstruct(_lexogenous, diagonal)
latent(Intercept Slope ) cov( Intercept*Slope) nocapslatent

The output is
...[output omitted]...


A number of articles that have employed a similar approach, albeit
usually with Mplus, suggest there will normally be three classes
(stable, steady increase, and elevated). My question is how do I (1)
identify the latent classes(trajectories) and (2) assign individuals
to the latent classes?

A couple of things.  I'm pretty sure that latent growth mixture modeling cannot be done in Stata's SEM.  That is because the model requires a categorical latent variable representing the underlying classes, and as far as I know, Stata's SEM does not support categorical latent variables.  Even if it did, you can't expect three classes to spontaneously emerge from the analysis: even in MPlus, you have to specify the number of levels for the latent categorical variable.  (You can try different specifications for this and compare them with likelihood ratio tests, but I don't know of any software that automatically selects the "right" number of levels.)  Also note that in any latent growth mixture model, there is no "assignment" of cases to trajectories.  You can get prior and posterior probabilities of membership in each trajectory, but then if you want to "reify" the classification, you have to do that by some rule, for example, the class with the highest posterior probability.

All of that said, if you want to tackle this problem within Stata, you can come very close using  Partha Deb's -fmm- command (-ssc install fmm-).  The model it would estimate for this data is not exactly the same as what you would get doing a latent growth mixture model in MPlus, but it is quite similar and might serve just as well, depending on your purposes and which parameters of the model are most important to you.  (And it will also handle inclusion of gestational age at birth as suggested by Austin Nichols' reply to your original post.)

Hope this helps.

Clyde Schechter
Department of Family & Social Medicine
Albert Einstein College of Medicine
Bronx, NY, USA



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/