cchien posted on Friday, November 12, 2010 - 12:49 pm
Hi, I am doing the EFA with ESTIMATOR = ML for 15 binary variables, and have a couple of questions:
First, why is there no eigenvalue showing in the output when I define all variables as categorical variables in Mplus? I use to define them as continuous variables, and eigenvalues appear in the output. Is there any explanation?
Second, sometimes the information criterion section and the chi-square test of fit section are not shown in the output. If both of them are gone, I guess the iteration may not reach convergence, but I found sometimes only one of them is not in the output, and the other parts (such as factor loadings, factor structure, estimated residual variance) appear normally. I checked the warning message, and it only shows "STANDARD ERRORS COULD NOT BE COMPUTED IN EXPLORATORY FACTOR ANALYSIS WITH 6 FACTOR(S)", but I don't think the calculation of AIC, BIC, likelihood, and DF need standard errors. Hence, what's the real problem when this situation happened?
My program is as simple as those showing in Mplus guidebook:
VARIABLE: NAMES ARE VAR1-VAR15; USEVARIABLES ARE VAR1-VAR15; CATEGORICAL ARE VAR1-VAR15; ANALYSIS: TYPE = EFA 1 6; ROTATION = QUARTIMIN; ESTIMATOR = ML;
With ML and categorical, means variances, and covariances are not sufficient statistics for model estimation. There is no covariance or correlation matrix to base eigenvalues on for this model. This is also why chi-square and related fit statistics are not available.
If a model does not converge, you will receive a message that states this. If a model converges, you will obtain the information criteria. If you don't, please send the output and your license number to firstname.lastname@example.org.
Hi, ML is default in conducting EFA under normal theory (NT). In Mplus guide, you said you have four estimators for continuous indicators; ML, maximum likelihood with robust standard errors and chi-square (MLM, MLMV)and ULS. which one is recommended if indicators are non-normal? Also, what the difference between MLM and MLR? I have read some articles claiming that MLR under non-normality is better than GLS, WLS and ADF?