compareFit and categorical models

[Galla Placidia]

I have two nested models with categorical measures. I fit them using WLSMV, for polychoric correlations. Can I use compareFit from the semTools package to compare the fits? And if not, how can I compare the models.

Thanks.

[Alex Schoemann]

Yes, this should work. One important note, the compareFit function isn't doing anything magical. It is simply gathering up fit indices from several models and putting them in the same place. You could do the same thing with a piece of paper if you wanted (and the anova command to compare nested models). One other caveat, it looks like the compareFit function grabs the DWLS fit indices, not the robust fit indices when using WLSMV.

Alex

[Galla Placidia]

Thanks.

OK --- so what are the pros and cons of using DWLS vs the robust fit?

And does the anova method work for categorical data? I seem to recall reading somewhere that the ML tests, and those based on the ML, do not apply in that situation.

[Alex Schoemann]

I don't feel qualified to address your first question. As for the second. The anova method does work with categorical data. When comparing models estimated with WLSMV, the chi square difference test is not simply the difference between the two observed chi square statistics, but it is calculable. For details see: http://www.statmodel.com/download/webnotes/webnote10.pdf

Alex

[yrosseel]

> OK --- so what are the pros and cons of using DWLS vs the robust fit?

Estimates are identical. Inference is superior when using the robust
standard errors and the robust (scaled) test statistic.

Reference:

Beauducel, A., & Herzberg, P. Y. (2006). On the performance of maximum
likelihood versus means and variance adjusted weighted least squares
estimation in CFA. Structural Equation Modeling, 13(2), 186-203.


Yves.

[f1r3t...@gmail.com]

In terms of other fit indices (CFI, RMSEA) would the robust values also be preferable to the DWLS values in this context?

 

Thanks

 

Peter

 

[yrosseel]

On 04/10/2014 11:15 AM, f1r3t...@gmail.com wrote:
> In terms of other fit indices (CFI, RMSEA) would the robust values also
> be preferable to the DWLS values in this context?

There is in fact little literature on this (AFAIK). But I would be
inclined to say "yes".

Yves.

[f1r3t...@gmail.com]

Thanks for your help