LCA - mix of categorical and continuous indicators

[Anna Manzoni]

Dear Stata Users.
Is there a way in Stata to include a mix of categorical and continuous indicators (i.e. observed variables) in Latent Class(/Later Profile) Analysis?

Say I have 3 indicators: education, income, occupation.
If there were all categorical I would do something like (for example for a model with 3 latent classes):

gsem (education income occupation<-,mlogit) ///
, nocapslatent logit lclass(C 3) ///
startvalues(randomid, draws(15) seed(123321)) ///
em(iter(5)) ///
nodvheader ///
nonrtolerance
estat lcprob


But what if occupation is a continuous scale?

Any suggestion is greatly appreciated.
Thanks
Anna

[Weiwen Ng]

Easily.

gsem (education income <-,mlogit) (occupation <- _cons, family(gaussian)) ///
, nocapslatent lclass(C 3) ///
startvalues(randomid, draws(15) seed(123321)) ///
em(iter(5)) ///
nodvheader ///
nonrtolerance
estat lcprob

Some minor cautions. You're treating education and income as un-ordered categorical indicators, but it seems more likely that they're coded as ordered categorical. If so, think about using ologit. I suspect the same is true of occupation, and it is probably best practice not to treat an ordered categorical variable as a continuous one.

I note that you are using the nonrtolerance option. You presumably have seen the related discussions about that option on the forum. Please do note that it's highly recommended to re-fit your final model without the nonrtolerance option - and by "highly recommended," I mean that I would personally do this.

Last, you're using 15 draws for start values. As you increase the number of latent classes, I'd increase the number of random draws. I'd read Kathryn Masyn's chapter on latent class/profile analysis as cited in our SEM manual. I think that she recommends as many as 100 random draws. In my experience, this might not be necessary with 3-4 classes. However, I'd definitely recommend it starting with 5 latent classes.