Is "FIML" for lavaan models still unsupported?

[follhim]

Question says all. I was hoping to know if FIML is still unsupported in bain.

Thank you.

[cjvanlissa]

Where did you get the impression that you can't use FIML? Bain is run on the estimated lavaan model, and lavaan uses FIML by default, so I don't see any obstacle!

[follhim]

It says here under "use bain with lavaan object" that users should not use fiml since it will bias the model.

In my own checks, I see that the BF changes drastically between when I do and don't use FIML.

[cjvanlissa]

I see, I didn't know - Herbert must have written that.

I see the rationale: bain doesn't know that the data are partially imputed, and so the standard errors are too small.

On the other hand, list wise imputation also biases the estimates.

The vignette refers to a method for multiple imputation, which should be the best of both worlds: best practice for imputation, and works with bain. I'd recommend that over FIML.

If you insist on FIML, you csn estimate your model with FIML, extract the parameter estimates and their covariance matrices, and apply bain.default() to those. Supply the sample size you'd get for the complete data, or do a sensitivity analysis of the effect of applying diffďerent sample size penalties in between the complete data and imputed sample size.

You can compare this method to the existing solution with multiple imputation.

[follhim]

Super helpful! If you do end up adding functionality for FIML, please let me know, as that is (as you are likely aware) the most popular/easiest option.

Thank you

[cjvanlissa]

The problem is not that it's unsupported, but that you shouldn't use it, because it will bias the result.

If you set the sample size argument n in bain.default to the number of complete observations, that should somewhat attenuate this bias.