I have 21 items scored from 0-3 and I am conducting EFAs in MPLUS with the expectation of two common factors (n = 690). I used ML (treating the items as not categorical), WLS, WLSM and WLSMV estimation methods and I noticed that the degrees of freedom for each model from one to five factors are different for the WLSMV estimator than the other methods. The 3 other methods have the following df :
It looks like WLSMV must be estimating more parameters than the other methods, but I don't see additional parameters in the output. It also seems strange that the df for my two factor model were greater than the df for my one factor model under WLSMV. Could someone explain this behavior?
The degrees of freedom for WLSMV are not calculated in the same way as for the other estimators. The formula for the degrees of freedom for WLSMV is found on page 281 of the Mplus User's Guide in formula 109.
Anonymous posted on Wednesday, June 08, 2005 - 10:19 am
Linda, I had similar question as Charles, but I could not find the formula you referred to on page 281. I have Mplus's User's Guide, Version 3.
Is there a different User's Guide for technical detail? The User's Guid I have does not see to have any technical thing besides Mplus language syntex.
The technical appendices are not included in the Version 3 Mplus User's Guide. You can find them on the website. Formula 109 is in Appendix 4.
anonymous posted on Friday, April 10, 2009 - 4:06 pm
Hello, I am attempting to conduct a dichotomous factor analysis with up to 5 factors and a few questions have come up: 1) The output states that factor determinancies could not be computed for an EFA with 4 or 5 factors b/c the model covariance matrix is not positive definite. How do you suggest I address this? 2) Also, I noticed that the default rotation is geomin (which I think is oblique)? Is this usually advisable for a dichotomous factor analysis or would orthogonal also be an option?
1. It sounds like you must be getting a negative residual variance. You may be trying to extract too many factors. If this is the case, there is no way to change it. If you would like me to look at the output, please send it and your license number to firstname.lastname@example.org.
2. The choice of an orthogonal versus an oblique rotation is available. It is your choice which to use. See the ROTATION option in the user's guide for a list of the orthogonal and oblique rotations available in Mplus.
anonymous posted on Friday, April 17, 2009 - 12:16 pm
Hi Dr. Muthen, I'm guessing that it is not advisable to interpret factors when factor determinancies could not be computed for an EFA b/c the model covariance matrix is not positive definite? Why should one avoid doing this?
The estimates produce an inadmissible model if the covariance matrix is not positive definite - for the model to make make sense this matrix has to be pos def. Try to modify the model to avoid this problem.
anonymous posted on Sunday, April 19, 2009 - 12:53 pm
Dr. Muthen, Thanks. Just to clarify, is interpretation to be avoided b/c the inadmissable solution makes the results very unreliable?
Also, could the non-positive definite matrix be due to too many variables (N=400, 42 dependent variables) or some missing data in 8 variables (maximum of 10%) or both?
Interpretation is to be avoided because the solution is analogous to having a negative variance - not an acceptable solution.
The typical reason is that you have a negative residual variance, which is often a function of trying to extract more factors than the data can provide.
anonymous posted on Tuesday, April 21, 2009 - 3:30 pm
Dr. Muthen, Thanks very much. One of my colleagues noted running the syntax of another dichotmous EFA using Mplus, version 3, and obtained the warning: THE INPUT SAMPLE CORRELATION MATRIX IS NOT POSITIVE DEFINITE.THE ESTIMATES GIVEN BELOW ARE STILL VALID.
Is this true? It seems that if the solution is inadmissable or unacceptable, then the estimates should also be unacceptable or invalid.