Holly Burke posted on Thursday, September 24, 2009 - 2:48 pm
I need to conduct an EFA with categorical and continuous factor indicators and then output the factor scores. I have missing and clustered data(I plan to specify CLUSTER = subject id under the VARIABLE command). I was told to obtain polychoric correlations, followed by least squares model fitting and the Crawford-Ferguson oblique oblimin rotation. I am trying to find these options in Mplus.
I will have a combination of categorical and continuous variables, what will be the default estimator? Should I specify "Estimator = WLS" under the Analysis command? Or, do you suggest WLSM and WLSMV or some other estimator?
Regarding the rotation, there are CF rotations, but none that specifically say CF - oblimin. Which of Mplus's rotations would be closest to "the Crawford-Ferguson oblique oblimin rotation"? How would you suggest I specify it?
I have missing data. If I specify "TYPE = EFA 1 5 MISSING;" under Analysis will this handle my missing data? If so, how? Does it use ML modeling to account for the missing data?
After I conduct the EFA then I will conduct an EFA within a CFA framework to get the factor scores. How do I account for my missing and clustered data in this CFA? Thank you.
The weighted least sqaures estimator uses polychoric correlations. This is what you want. You can use either WLS, WLSM, or WLSMV. Depending on the model, WLSMV or WLSM is the default for a combination of categorical and continuous dependent variables.
The QUARTIMIN rotation is the same as OBLIMIN (OBLIQUE 0). You can use either. See the ROTATION option.
If you want factor scores for EFA, you need to use the new ESEM feature that is described in the Version 5.1 Language and Examples Addendum which are on the website with the user's guide.
TYPE=MISSING is the default now. It is ML under MAR.
If you have clustered data, you can take this into account using multilevel modeling along with other complex survey features. See Chapter 9 of the user's guide and the section on complex survey data.
Holly Burke posted on Friday, September 25, 2009 - 11:02 am
Dear Linda, Thank you for your help. I tried running an EFA with the cluster option under VARIABLE and the COMPLEX option after TYPE (see relevant code below), but I got an error saying that COMPLEX can not be used with EFA. I also got an error saying that something is wrong with my rotation statement. Do you see what I am doing wrong?
VARIABLE: MISSING = .; CLUSTER = ID;
ANALYSIS: TYPE = COMPLEX EFA 1 5 MISSING; ROTATION = QUARTIMIN; PLOT: TYPE = PLOT3; SAVEDATA: FILE = FA1OUT.TXT;
Finally, (assuming I can eventually run the above EFA!), I read through the entire Addendums, but did not see a section on ESEM. I was able to find the slides on EFA within a CFA framework on the website. Is there any problem with using the latter approach to get the factor scores? Thank you!
David Kosson posted on Wednesday, November 18, 2009 - 10:58 am
I am running an EFA with Version 5.1. When I ran it with quartimin rotation, I recognized all the fit indices (CFI, TLI, RMSEA) and know how to interpret them. However, when i specified the varimax rotation, the output included only the RMSEA and the Root Mean Square Residual (or RMSR). Although this RMSR seems related to the SRMR or WRMR, I am guessing it is not the same. I am wondering if I can interpret it the same way. That is, does a value below .05 suggest good fit? Does a value below .08 suggest acceptable fit? And is there a reference for using the RMSR in this way that you can provide?
We give a less complete set of fit indices with our old rotations. But RMSEA is defined the same in both cases. RMSR and SRMR for EFA are the same and can use the same cutoff. See Technical Appendix 5 formulas 128, 129, and related text.
Daiwon Lee posted on Wednesday, February 03, 2010 - 8:51 am
Hello professors, I recently bought 5.21 to work on my thesis. I have a couple of fundamental questions regarding EFA output. 1) what is minimum rotation function value and how do I interpret it? 2) what is factor determinacies and how do I interpret it? Thank you.
Should I be concerned if I get the exact same coefficients in the outputs for an EFA model when (1) accounting for complex sampling structure (e.g., cluster = family id) and (2) when ignoring the potential dependencies that might result from nesting within families?