Stata vs Mplus for LPA?

[Eunhye Kang]

Dear researchers,

I am trying to conduct Latent Profile Analysis but wonder conducting it with Stata doesn't get me into any trouble.
Mplus is mostly used for LPA but I wonder if it is okay to run it with Stata15.
Do you know what strengths Mplus has for LPA over Stata?
I appreciate your comments in advance.

[Weiwen Ng]

This is a bit difficult for me to answer as I don’t have MPlus. I have a general sense of its capabilities, though. In general, you can argue to reviewers that Stata is a well-known software package. It should not be a problem per se to use it.

One shortfall that Stata has is that it doesn’t have the bootstrap likelihood ratio test for model selection. This is difficult to program as well, so I’m not aware of anyone who’s implemented it in Stata (the Penn State U plug-in does implement it, but it only accepts binary and ordinal indicators). I would simply use the BIC to select models in Stata.

Reviewers may prefer that you use many sets of random starting values and report how many % of the runs converged at the global maximum. Stata can natively do the first bit, but it doesn’t natively report the latter bit. I present a solution in this post.

when you perform LPA with Gaussian indicators, you need to be aware that Stata’s default setting fits a very restrictive model (I.e. it has strict assumptions that I think are unrealistic). The issue is outlined and illustrated here. I have suggested that you either explore various model options, or you simply use the least restrictive set of assumptions. I don’t know what Mplus’s current default settings are, so I can’t say if it is worse or better than Stata in this regard. I think I’ve read one 2009 paper that used MPlus and didn’t seem to have used the restrictive model structures, so I would just generally beware of this trap.

last, I think Stata is likely slower than MPlus. This is something you just have to live with.