Comparing non-nested SEM wlsmv PreviousNext
Mplus Discussion > Structural Equation Modeling >
 GRAZIA DE ANGELIS posted on Friday, September 30, 2016 - 6:27 pm
Dear Drs. Muthen,

I'm testing two alternative non-nested full SEM:

-in the first model, A affects C directly and indirectly through B;

-in the second model, A affects B directly and indirectly through C.

As B is a categorical dicothomous variable, WLSMV is the estimator.

I have two questions:

1- Is it possible to use a categorical variable as a mediator?

2- How can I compare the two alternative models? I cannot use the DIFFTEST, because the two models are not nested, and in the output I do not have indices such as Akaike, because of the estimator, I suppose.

Thanks a lot for your help
 Bengt O. Muthen posted on Saturday, October 01, 2016 - 10:58 am
1) Yes. See our Mediation Analysis webpage and the heading

Counterfactual causal effects for mediation modeling

2) Both models are just-identified/saturated and cannot therefore be rejected by overall model fit tests. Typically, the causal order of variables cannot be determined by statistical analysis.
 GRAZIA DE ANGELIS posted on Saturday, October 01, 2016 - 5:07 pm
Dear Dr. Muthen,

thanks a lot for your reply. I would like to ask some clarifications

- could I use MLR as estimator, instead of WLSMV?

- regarding your answer to my second question, I would like to add some specifications. In my previous post I tried to be as concise as possible, however my models are much more complex than I described.

I have 3 independent continous latent variables (each one of them has two indicators);

another continous latent variable (with two indicators) that is the mediator in the first model, and the independent variable in the second, alternative model;

a dichotomous categorical observed variable (i.e. belonging to group1 or group2), that is the independent variable in the first model, and the mediator in the second, alternative model;

finally, I have three observed covariates.

Thus each model is correctly overidentified, with 38 degrees of freedom. Does this means that I coul use some test to compare the two models?

Fit indices were very similar, chi-squared were just a bit higher in the alternative model

I apologize if I have not correctly understood your reply, I am new in using MPLUS.

Thanks for your help
 Bengt O. Muthen posted on Monday, October 03, 2016 - 6:44 am
You can use MLR instead of WLSMV but because you have categorical variable you won't get an overall fit index. And the directions of influence from one variable to another can still not be determined statistically.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message