I am doing an EFA on a set of 12 continous indicator varialbes, but some of them are obviously non-normal. Among the four choices, (ML, MLM, MLMV,ULS), which can be used and which is the best choice? Thanks.
Both MLM and MLMV are robust to non-normality. MLM seems to perform better in the limited simulations that we have done.
Dave Flora posted on Tuesday, November 29, 2011 - 12:05 pm
Is MLM estimation with EFA essentially an application of the Satorra-Bentler scaled chi-square and robust standard errors to the EFA model? Is there any methodological literature I can cite regarding the use of MLM in an EFA context? A reviewer recently challenged my claim that the Satorra-Bentler procedure can be applied to EFA using specialized software; Mplus' MLM is what I had in mind.
In EFA settings the model fit and chi-square are obtained for the unrotated solution (a CFA model). The rotation of the parameter estimates does not affect fit. Thus there are no chi-square issues that arise from EFA.
Dave Flora posted on Wednesday, November 30, 2011 - 6:44 am
Thanks Tihomir. What about robust standard errors with EFA?
The asymptotic distribution of the rotated solutions is based on the asymptotic distribution of the correlation matrix. If you use robust standard errors for the correlation matrix you get robust standard errors for the rotated solution.
AT Jothees posted on Sunday, February 26, 2017 - 1:03 pm
This is a very important discussion. However, I need more clarification on which estimator to use for non-normally distributed data.
I am running CFA with 7 continuous indicators and four likert item (ordered categorical indicators). I assume that most of my variables are not normally distribute. Therefore, please suggest which estimate I should use (ML,WLS,ADF).
I am new to mplus and also performing factor analysis for first time. So, please bare with my ignorance.