LPA profiles as DV in multinomial logistic regression

[Lisa-Marie Kraus]

Hi everyone,

I aim to run LPA with the end goal of using the derived latent profiles as DVs in a multinomial logistic regression analysis (and describe the profiles) to identify class predictors by using different independent variables.

I already established that a four-profile solution is the best fit (with AIC/BIC),

Now I want to save the outcome of this command

gsem (attitutde behaviour <-_cons), lclass(C 4)

in a new variable where 1=profile1, 2=profile2, 3=profile3 and 4=profile4.

I am aware of the "predict" command, which I have encountered in another thread, but since I am new to Stata I am not sure how to use it correctly.

In the end, I want to be able to understand why these different profiles emerge. For instance, by saying something like "females are more likely to belong to profil1 than males" or "older people are less likely to belong to profile3 than younger people"

Thank you for you help!

Best,
Lisa-Marie

[Weiwen Ng]

Welcome to Statalist. This is a good question. If you used the predict post-estimation command, you would actually wind up with the probability that each observation falls into each of the k latent classes in whatever your final model was. What would you do then? That doesn’t work with multinomial logistic regression.

you might think that, well, maybe I assign people to the class they’re most likely to belong to, I.e. modal class assignment. You’d then wind up with a DV you could feed into the multinomial regression. However, in some sense, this is wrong. You are ignoring the uncertainty in the estimates of the probability of belonging to each latent class. This is not good.

the better alternative is latent class regression.

https://www.statalist.org/forums/for...ass-regression

[Lisa-Marie Kraus]

Dear Weiwen,

Thanks a lot for your help! Much appreciated.

Best,
Lisa-Marie

[Weiwen Ng]

Dear Weiwen,

Thanks a lot for your help! Much appreciated.

Best,
Lisa-Marie

Actually, here's a pro tip that didn't address in the other post. You should enumerate the latent profile model first. Once you've settled on the final model, you can extract the model's parameters as start values, then specify your latent class regression model with those start values. If you don't do this, you're not guaranteed to converge on the same latent classes in your regression model.

gsem (attitutde behaviour <-_cons), lclass(C 4)
estimates store class4
matrix b = e(b)
gsem (attitutde behaviour <-_cons) (C <- i.female age), lclass(C 4) from(b)

Because multinomial logistic coefficients can be a bit hard to interpret, I'd suggest using the margins command (as detailed in the post) to show what you want to show.