I recently attended the Baltimore conference on MLM and am attempting SEM with the Add Health Data set. I am in the preliminary stages and have not added the multi-levels as of yet. I have decreased the background variables I am controlling for and have only included the variables of primary interest. I have TWO categorical DVs. I am getting the following error message: THE ESTIMATED COVARIANCE MATRIX COULD NOT BE INVERTED. COMPUTATION COULD NOT BE COMPLETED IN ITERATION 1. CHANGE YOUR MODEL AND/OR STARTING VALUES. I reviewed the discussions on error messages but each time I found this message it seemed the response was to send the data to Linda. Thank you for your help. Sue
I made an initial error but didn't know how to retract my post. I'm sure I'll still end up contacting support, but not yet. Thanks
Qilong Yuan posted on Thursday, August 12, 2010 - 8:54 pm
I have the same problem. I am running a CFA and I have 6 factors in my model and had to reduce the integration points to 4 since my computer does not have enough memory. I am using a MLR estimator and use the results from WLSMV as starting values.
My understanding is fixing factor loadings does not help, so I fixed the factor correlations but the model stopped at iteration 1. Is there any other parameters that I can change (I mean change the default)? I just need the fit indices (like AIC BIC).
WLSMV gives probit coefficients for factor loadings. The default for maximum likelihood is logistic regression so your starting values are not going to be in the right ballpark. You can use LINK=PROBIT with maximum likelihood to get probit coefficients.
You should free all factor loadings and fix the factor variances to one. It may be that the first factor loading that is fixed at one as the default is not close to one and this is causing problems. Otherwise, send the full output and your license number to firstname.lastname@example.org.
Qilong Yuan posted on Thursday, August 19, 2010 - 6:24 pm
Thank you very much for your reply. I have four follow up questions:
1. For some models, if I fix all factor correlations the model actually converges in the end. However, if I only specify the starting values then the model won't converge. Why fixing factor correlations helps?
2. Is there a way that I can get the fit indices (AIC, BIC etc) without using numerical integration?
3. Will it help if I am able to use more integration points? Right now I am using 4.
4. I got different fit indices using logit and probit link function. Why is that?
1. If it works when all factor indicators are free and factor correlations are fixed at one, this means that the first factor indicator is most likely not estimated close to one which is the value it is fixed at as the default. This suggests choosing another factor indicator to fix at one.
2. You get these whenever you use maximum likelihood estimation. It is not necessary to have numerical integration.