I'm running a CFA of survey response items with two variables on one factor and three variables on another, and I'm correlating them. I'm often getting errors on the factor with two variables. Here is my error:
WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR AN OBSERVED VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO OBSERVED VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO OBSERVED VARIABLES. CHECK THE RESULTS SECTION FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE HI4R.
The only possibility is a negative variance, so I set HI4R@0. Then I get this error:
*** ERROR in MODEL command Variances for categorical outcomes can only be specified using PARAMETERIZATION=THETA with estimators WLS, WLSM, or WLSMV. Variance given for: HI4R
I set the parameterization to theta, and I get more errors. Any suggestions?? Thanks!
Residual variances for categorical variables are not parameters in a cross-sectional model. They can be specified only in multiple group and multiple time points models. I'm assuming that you see the negative residual variance with R-square. Here it is computed as a remainder. You need to change your model.
Xu, Man posted on Wednesday, April 10, 2013 - 3:57 am
I have run into the same situation myself. I am comparing several nested models. It is the bi-factor model that gave this warning for one item.
Would it be unacceptable to ignore this warning? The model was converged and all estimates were given, apart from r-square of this item, of course.
Also, when I revert to theta parameterization, this warning went away but a different one pops up: MINIMIZATION FAILED WHILE COMPUTING FACTOR SCORES FOR THE FOLLOWING OBSERVATION(S) : 126 FOR VARIABLE GHQ0899
This message cannot be ignored. Please send the output with the not positive difference message and your license number to email@example.com. Changing to the Theta parametrization is not a solution.
Xu, Man posted on Wednesday, April 10, 2013 - 8:34 am
Thank you very much. I have sent relevant information to you at this email address.
It seems that this is at least related to sample size. Another point is that I need to look at the results based on listwise complete data including some observed external predictors. This substantially reduces sample size and probably leads to problems in the estimation of the latent model.
In this situation, would it be a reasonable compromise if I export factor scores from the FIML analysis model based on complete sample, then defined the complete sample on the factor scores and external predictors?