I am estimating a four factor model with CFA. My 23 observed variables are likert-type and is multivariate non-normal with missing. What estimator should I use(MLR, MLF, WLSM, WLSMV....)? When I used WLSM or WLSMV, the relative fit indices were good(CFI=.966 or .90, TLI=.962 or .96) but the absolute fit index were not so good(RMSEA=.086 for both). In case of MLM or MLMV, the relative fit indices were not so good(CFI=.83 or .838, TLI=.81 for both) but the absolute fit index were good(RMSEA=.054 for both). I wonder my variables should be regarded as continuous or ordered categorical.
If your Likert variables have floor or ceiling effects which it sounds like they have, then you should use categorical data modeling using either weighted least squares or maximum likelihood estimation. The way to specify this is to put the dependent variables on the CATEGORICAL list. I suspect they should be treated as categorical based on the difference in the results by treating them as categorical versus continuous.
You should not obtain very different results with different estimators as far as parameter estimates go. Standard errors and fit statistics will be different and could result in slightly different patterns of significance. For information about the estimators implemented in Mplus, see Technical Appendices 4 and 8 and references therein and pages 482-485 of the user's guide. Also, see the general literature on maximum likelihood and weighted least squares.
Shani Ofrat posted on Thursday, March 10, 2016 - 10:44 am
Hello, I am attempting to use the MLR estimator in a multiple groups CFA, with binary observed variables and continuous latent variables. The dataset has weights and stratification, so type=complex. When I try to run the model with default WLSMV, it works well and the model has good fit, but there are no differences (difftest; CFI/TLI/RMSEA fit stats) between models with constrained and unconstrained regressions. I believe this is because the sample size is so large that difftest is overpowered, and the regressions are too small a component of the model to produce significant change in fit. As such, I would like to use MLR to get AIC and BIC for a more sensitive measure of change to model fit, but MLR will not run in my current model. Output error is: *** ERROR in ANALYSIS command ALGORITHM=INTEGRATION is not available for multiple group analysis. Try using the KNOWNCLASS option for TYPE=MIXTURE. Question: Can I run MLR in type=complex multiple groups CFA? How must I change the defaults to get it to run?
You need to use TYPE=MIXTURE, the CLASSES option, and the KNOWNCLASS option. When classes are known, this is the same as multiple group analysis.
Sunmi Seo posted on Monday, November 13, 2017 - 8:29 pm
I will run multiple group CFA. But, I am confused to choose a proper estimator. My data has non-multivariate normality and no missing data. All variables are continuous variables. I have 2 groups (e.g., male vs. female). I wonder which estimator is the best fit. In this case, is MLR the best?