2. It is the final value that the EM algorithm ends with. However this is not an identified parameter. Use the output option H1SE to determine what parameters are identifiable. You should get an entry of 999 for non-identified parameters in the "STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS" section.
yang posted on Thursday, September 08, 2011 - 2:21 pm
Thanks for your response. Just wanna make sure I got it...
1. By FIML I mean maximizing likelihood computed from individual data...not the Wishart likelihood computed via the covariance matrix which is produced by EM algorithm. Is that correct?
2. The EM algorithm you're talking about is only for estimating covariance matrix in the presence of missing data (i.e. if I delete SAMPSTAT from the input file, the algorithm won't be invoked)...not for estimating the model that I specify in the MODEL statement. Is that correct?
1. The estimator is the maximum likelihood estimator.
Ting Dai posted on Monday, April 22, 2019 - 9:37 am
Similar to Yang's post on 9/7/2011, I have a CFA model (Model A) whose input var-cov matrix has about 40% of 0 covariance coverage. Mplus still is able to estimate the model for me, and generate a full estimated var-cov matrix. But my CFA Model B with about 60% of 0 covariance coverage fails. The error message is "THE COVARIANCE COVERAGE FALLS BELOW THE SPECIFIED LIMIT." My questions are: 1. What is Mplus's covariance coverage limit? 2. Can the covariance coverage limit be specified to be lower, so that Model B would be able to run and for me to obtain its estimated var-cov matrix? 3. In the current situation, are Model A's estimated covariance matrix results all trustworthy, especially those cells with 0 input covariance? 4. In general, is there a rule-of-thumb % of 0 covariance coverage below which you would not advise to handle with FIML (even though the missing mechanism is MAR or MCAR)?
3. That depends on the situation - I would have to see the full output (send to Support along with license number). Perhaps you have missing by design given your exact zero coverages - if so, that can be better handled by multiple-group analysis.
4. Chapter 10 in our RMA book discusses this and shows examples where you can go wrong with 60% coverage.