WLSMV estimation in stata

[Rune Lomholt]

Hello everyone

I am working with confirmatory factor analysis in statas SEM module. I am concerned with non-normality in my data, since all indicators of latent variables af categorial variables (likert scaled).

Is it possible to use WLSMV estimator for CFA in stata?

[Jane Burke]

I believe that ADF is the option for the weighted least squares estimator.

[Erik Ruzek]

You can use ADF, but it is not equivalent to WLSMV. Unfortunately, StataCorp has not been interested in adding a WLSMV estimator to sem, and you must instead use gsem,(or irt) for factor models in which the items are binary, ordered categories, non-ordered categories, etc. As much as I love Stata, for measurement type modeling it is very hard to beat Mplus.

[Weiwen Ng]

You can use ADF, but it is not equivalent to WLSMV. Unfortunately, StataCorp has not been interested in adding a WLSMV estimator to sem, and you must instead use gsem,(or irt) for factor models in which the items are binary, ordered categories, non-ordered categories, etc. As much as I love Stata, for measurement type modeling it is very hard to beat Mplus.

Is it not true that ADF is still a weighted least squares estimator, even though it’s not weighted least squares mean and variance (WLSMV)? If so, is WLS considered significantly inferior to WLSMV, to the point where reviewers would or should reject something estimated with WLS?

I am genuinely asking here. I suspect the answer is that WLS isn’t so inferior that it should be laughed out of the room, but I don’t know, as I’m not a traditional SEM practitioner.

[Christopher Bratt]

ADF is not recommended (based on simulation studies).

With access to Mplus, I would suggest using the Stata addon -runmplus- by Rick Jones.

Without access to Mplus, lavaan in R is a better option and has easy syntax.
Here on estimators: https://lavaan.ugent.be/tutorial/est.html

[Erik Ruzek]

To add a bit to Christopher's point, ADF, which is more or less WLS, has all sorts of problems with small sample sizes (with 500 being considered pretty small). See the Flora and Curran paper from 2004 and the Forero et al. paper from 2006. The other thing to consider with Stata's ADF in sem is that it doesn't truly treat the items as ordinal, meaning that you get no information about the item thresholds.