Message/Author 


Dear Bengt or Linda, We conducted a multilevel MIMIC model with kids nested within families and also a single level model randomly selecting one child. A reviewer is asking us to report an assessment of multivariate normality. I ran Tech 13 for a one class model but we are using missing data. Is there a way to get a Mardia's coefficient for missing data? Thank you in advance 


No, I don't think so. But why don't you instead analyze with ML versus MLR and see how different the SEs and the chi2 test of model fit are  there you would see the real effect of lack of multivariate normality. 


Thank you for this suggestion. For a typical FIML ML model would you consider a chi2 test of model improvement for MLR a general good practice step to take in choosing MLR over ML? Or mainly check if the ML model has poor fit. I assumed that among the consequences of multivariate kurtosis for ML there would be poor model fit, reviewer preferred a test. 


You can't do a difference test between ML and MLR. You should just look at the two tests and see if there is a difference in fit and also look at the standard errors. If they are different, this suggests that nonnormality exists. 


Dear Dr. K. Muthen/ O. Muthen, I have checked my dataset for multivariate normality and I have found out that there are positive skewness and kurtosis in all variables. I tried to logtransform my dataset using stata. Then I tried to work with my LGC model with both datasets. But: 1. the logtransformed dataset did perform worse than the nontransformed interms of the chisquare and the other model fit indices 2. when I compared the ML and MLR for the nontransformed dataset, I can see that there is a massive difference in the chisquare values( MLR=589 and ML=1064; pvalues <0.0001) which according to your explanations above is the sign of significant effect of the nonnormality So, how do I deal with such situation? Can I simply pick the nontransformed data under MLR? I have also "Scaling Correction Factor 1.805 for MLR" in my output. What is this telling me? Regards. 


I would use the nontransformed data and MLR. The scaling correction factor is used only for difference testing. It is related to nonnormality. 


Thank you very much. 

Back to top 