I thought about it a bunch last night, and I think I have a better articulation of my hesitations.

SEM is a better tool for just about everything. The @quantitudepod guys are right on that.
1/6 https://twitter.com/KMKing_Psych/status/1350306463831375874

The challenge, in my experience, is when you want to take advantage of several of its strengths at once, you rapidly run into models that aren't estimable.

2/6

Modeling latent factors with ordinal indicators for multiple scales? Sure.

Interactions among latent variables? Sure.

Non-linear outcomes? Sure.

Multiple mediators, that are also moderated? Sure.

Account for missing data using FIML or imputation? Sure.

3/6

Do that all at once?

THERE IS NOT ENOUGH MEMORY SPACE TO RUN THE PROGRAM ON THE CURRENT INPUT FILE.

So you start having to make sacrifices. That's the reality of using SEM in many situations.

4/6

You end up using observed variables instead of latent.

You ignore the zero-inflated nature of your count, and use a robust estimator instead.

You make sacrifices in some parts of your model so you can get the damn thing to converge.
5/6

Of course, this isn't just limited to SEM.

Any model is about sacrificing, making strict (untenable) assumptions in one place so you can make nuanced assumptions somewhere else.

Which approach you choose depends on where you want to make those sacrifices.
6/6