Message/Author 

V X posted on Friday, February 26, 2010  4:32 pm



Dear Dr Muthen, I am wondering what is MLF ? Mplus User's Guide only provided very limited information. Would you provide more reference with regard to this estimation method? How does it differ from MLR ? When would you consider to use MLF rather than MLR? Thank you. 


MLF is a common approach in statistics to computing SEs and is defined in equation (168) of the Mplus Technical Appendices ("through Version 2") on the web site. It uses the sums of products of firstorder derivatives. It is a simpler estimator of SEs than ML and MLR because they also use approximations of secondorder derivatives. In large samples the 3 methods are equivalent. When samples are not large it depends on the situation (distributional violations, nonindependence) which is best. MLR is often preferrable given its robustness. In some cases, MLR SEs cannot be computed in which case Mplus switches to MLF. I am not aware of studies providing a comprehensive comparison of the 3 methods for the large sets of models that Mplus offers. 


Dear Dr. Muthen, I am running the mixture model for murder rates at the county level. My sample size is quite large (N=2700). I use the ML estimator, but only 1class model has executed normally with ML estimator. Starting with 2class model, I am getting a warning about
a nonpositive Fisher information matrix, and the results are presented for the MLF estimator. In your earlier post you said that in large samples, Ml, MLR and MLF are equivalent. Do you think my sample size is large enough for accepting the results based on MLF, or I need to try to modify my model to get a solution based on ML or MLR estimators? Thank you. Arina. 


The MLF estimator should be fine at this sample size. 


Dear Dr. Muthen, I was running a series of mixture models with varying number of classes for the rates of female aggregated assault at the county level simultaneously. The 1, 2, 3 and 5 and 6classs models terminated normally, but the 4class model did not. The warning said the model could not converge. Should I disregard the 5 and 6class solutions before I modify a 4class model? 


I would try more random starts for the 4class model, for example STARTS = 400 100; I would not disregard the 5 and 6class solutions. 


Dear Linda, I tried more random starts for the 4class model. I went up to 1000, and I also increased the number of miterations. All that did not help. 


Try more like 2000 500 or 5000 1000. If you continue to have problems, send the output and your license number to support@statmodel.com. 

Back to top 