DIFFTEST with EFA PreviousNext
Mplus Discussion > Exploratory Factor Analysis >
Message/Author
 Keith D. Wright posted on Saturday, March 28, 2015 - 6:48 pm
I am conducting an exploratory factor analysis with categorical variables

VARIABLE: NAMES = v1-v60;
CATEGORICAL = v1-v60;

ANALYSIS: TYPE = EFA 1 4

* The output indicates that chi-square value for ....WLSMV cannot be used....

I did some research and what I found indicated to use DIFFTEST....but I saw a post from Linda in 2012 which indicated this does not work with Type = EFA, please help. So, should I not use the Chi-Square results at all, and only focus on others such as RMSR value

-Thanks
 Linda K. Muthen posted on Sunday, March 29, 2015 - 9:29 am
Which two nested models do you want to test with DIFFTEST? Or are you just asking about fit for EFA?
 Keith D. Wright posted on Sunday, March 29, 2015 - 10:07 am
Fit for EFA, thanks Linda.
 Linda K. Muthen posted on Sunday, March 29, 2015 - 12:26 pm
You would look at the p-value for chi-square and all of the other fit measures.

The note relates to only difference testing of nested models.
 Claire Mitchell posted on Tuesday, March 31, 2015 - 7:07 am
Hello Dr. Muthen,

I feel my post relates to this thread - hopefully I am on the right track!

My data (questionnaire data) is categorical (ordinal; 5-point Likert scale) and non-normal. After researching the best estimator for such cases, I planned to run an EFA using the WLSMV estimator. Despite this, I realised that the DIFFTEST option is not available using the WLSMV estimator when doing an EFA. After searching the discussion boards, it was suggested that to test the model difference I could run exploratory structural equation modelling in the program.

I have a1-a7 and v1-v8 items, and want to determine if the 1-factor or 2-factor model has better fit (this is preliminary and I have other subscales of a similar fashion). Does this approach sound right, or am I going down the complete wrong track? I know it is not appropriate to use the WLSMV chi-squares for difference testing, hence my line of thinking regarding DIFFTEST. In saying that, I am not sure if my models are nested?

If ESEM is appropriate, are you able to direct me to a resource that details how to run the syntax for this? I have looked at the "Exploratory Structural Equation Modeling" paper (Asparouhov & Muthén) but I getting confused as to how to specify an EFA model within this framework.

Thanks and kind regards,

Claire
 Linda K. Muthen posted on Tuesday, March 31, 2015 - 8:26 am
You can run the 1 and 2 factors models as ESEM models and use DIFFTEST. In EFA these models are nested. See Example 5.24. If you remove the covariates and direct effects. it is an EFA.
 Claire Mitchell posted on Wednesday, April 01, 2015 - 3:09 am
Thanks for your prompt reply and advice Dr Muthen! I will look into this approach.

Claire
 Bengt O. Muthen posted on Wednesday, April 01, 2015 - 4:25 pm
UG ex 4.2 output in version 7.3 shows that EFA Difftests for different number of factors is now automatically provided in the output.
 Claire Mitchell posted on Wednesday, April 01, 2015 - 5:44 pm
Thanks again for your reply.

I have looked at the ex 4.2 output, which is similar to what I obtained when I ran my EFA. I observed that it has the '1-factor against 2-factor' model comparison under model summary information. Is this the same result I would obtain if I were able to run DIFFTEST with my EFA (using WLSMV)? I think when I initially looked at this model comparison I thought I could not interpret it, given the warning within the same output about the inappropriateness of using the chi-squares values for difference testing when using WLSMV.

If I am able to just use this information (which would be great!), would it still be correct to report the other model fit indices associated with the best model?

Kind regards,

Claire
 Bengt O. Muthen posted on Wednesday, April 01, 2015 - 5:52 pm
Q1. Yes, this output draws on Difftest behind the scenes and can therefore be used.

Q2. Yes.
 Claire Mitchell posted on Wednesday, April 01, 2015 - 6:18 pm
Thank you so much! Your help is greatly valued and appreciated. This has made my analyses a lot more simple.

Kind regards,

Claire
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: