Sandra Nitz posted on Tuesday, August 02, 2011 - 3:40 pm
Hi, I'd like to do a twolevel-CFA with 6 latent Variables measured by 40 items in 45 classes. For analyzing my data I follow a protocoll including 6 steps based on Muthen (1994). In step 1 it is advised to explore the data using the (total) sample covariance matrix: a) trying the hypothesized model ignoring the clustering (type=general) and b) checking for changes in SEs and chi-square by using type = complex. Step 1a converged and resulted in satisfying model fits (taking into account that the total sample covariance matrix was used). However, in step 1b I get the following error message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS -0.256D-15. PROBLEM INVOLVING PARAMETER 45. THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER. As step 1b is intended to check for changes in SEs, I think I cannot use the estimated SEs due to the error. Nevertheless, the model converged and showed satisfying model fits (better model fits compared to previous unclustered cfa). I am not sure how to deal with this error message and would be happy to hear your suggestions. Sandra
We give that warning to make people aware of the fact that they have more parameters than independent pieces of information. The effect of this on standard errors has not been studied as far as I know. To know how it would affect your study would require a simulation study.