Forum: open-discussion

RE: CFA with ordinal variables [ Reply ]
By: Michael Hunter on 2014-06-16 16:46

Hi Jeffrey,

If the same model runs fine with one estimator (MLR) but not with another (WLSMV), then it doesn't sound like a specification error to me. It's important for this check that the same model is compared where the only difference is the estimator.

If it is an estimator issue with WLSMV it might be missing data related. I've heard from Mike Neale that WLSMV can give very poor estimates under certain missing data conditions that can be common in behavior genetics work.

Beyond that my general advice for debugging this kind of problem is to start with somethings very basic that you know will work. Try that. Then slowly add pieces, iteratively checking and running, until the problem occurs. Then the newest change is what caused the problem.

Cheers,
Mike Hunter

CFA with ordinal variables [ Reply ]
By: Jeffrey Henry on 2014-06-13 22:40

[forum:41105]

bifactor.R (7) downloads

Dear all,

Dear Mr Beasley,

I am running into an odd problem with some data I am
analyzing via R (lavaan). I am testing a general-specific (or bifactor) CFA
model with one general factor + 3 specific factors, on a twin sample
(grouping=MZ;DZ). Each item is assessed twice (i.e., for each twin (twin 1, twin 2)) and is rated on a 4-point scale. When I treat the data as
continuous (MLR), the model runs smoothly and provides results coherent with the literature and what is expected from twin models. However, when I treat the data as ordinal (estimator="WLSMV"), inter-factor covariances (i.e., twin1-twin2 covariance for each factor) are > 1. This is unexpected as I set the mean to 0 and variance to 1 for these factors so that, to my understanding, they should behave as correlations (which they do in the MLR model and other models I have tested). Also, the estimates are superior to one and the std.all are equal to the estimates. Please see attached syntax.

In your opinion, is there something wrong with the specifications? Or, is it something normal in threshold models? Or identification issue or estimation problem?

Thanks in advance,

Jeffrey Henry