SEM and latent class analysis

[Niels Henrik Bruun]

Hi
Are there anyone who have tried to do Latent Class Analysis based on SEM in Stata?
According to what I can read it is almost the same as example example 28g (IRT as SEM) in the SEM documentation except for that the latent class is multinomial instead of normal.
But I can't figure out how to specify the last part.
Maybe because it isn't possible in the current setup?

Looking forward to hear from you

[wbuchanan]

UPenn's methodology center has an LCA plugin put together and it might be possible to fit an LCA with gllamm but those are the only ways that I am aware of.

[Niels Henrik Bruun]

Thank you for help.
I've tried the Upenn's LCA plugin and it works fine.
I have not yet tried the gllamm solution.
I was mainly courious about what SEM could do and what it could not.

[wbuchanan]

Ah. The SEM stuff at the moment seems to be solely focused on continuous latent variables. Depending on the complexity, there were also a few user-written finite mixture model packages that would provide some functionality for estimating latent classes - or some approximation thereof. I know a few people have mentioned LCA at the North American user group meetings, but would guess there might not be a high enough demand for it to be implemented natively in Stata.

[Niels Henrik Bruun]

Thank you very much.

[Brian Flaherty]

Hello,
This post showed up when I was searching for information about gllamm and I thought I'd chime in.

Just started using Stata 15 (I hoped it would make problems I was having in Stata 14 disappear, but alas no.) and it does fit LCA with categorical indicators. I've wanted to try it and your question spurred me on to it. I've just run a couple examples and I think it will take me some getting used to, but so far it looks good. I'm totally unfamiliar with Stata's sem and gsem syntax, but here is an example of syntax I created navigating the menus:

gsem (cig100lf cigdlymo cig30use cig30av <- _cons), family(bernoulli) link(logit) lclass(A 2) lcinvariant(none)

Converged to a reasonable solution and post-estimation produced fit statistics and marginal LC probabilities. (That's as far as I got.) Also, I might totally be missing something, since I just started playing with it.

Also wanted to mention that the Methodology Center that produces the LCA plugin is at Penn State, not UPenn.

Have a good night and weekend.

Brian

[Niels Henrik Bruun]

Hi Brian
Thank you for noticing me