I don't know why you don't use the ml missing data estimation method. Simulations have found that multiple imputation with a sufficient number of imputed data sets yield nearly identical estimates to those obtained using the maximum likely (FIML) method. This could solve your problem. If you are concerned about the estimates you could compare the coefficients to those you are using presently to see if you get much difference.
David J. ------------------------------------------------------------------------------ David R. Johnson Professor of Sociology, Human Development and Family Studies, and Demography Department of Sociology 713 Oswald Tower The Pennsylvania State University University Park, PA 16802 814-865-9564 [email protected] ------------------------------------------------------------------------------- ----- Original Message ----- From: "Sean Hurley" <[email protected]> To: [email protected] Sent: Wednesday, October 31, 2012 1:11:36 PM Subject: Chi-square scaling correction factor under MI Hi all - I am working on some path analyses (in MPlus) using MLR estimation (because of some non-normality in our data and because we need to include sampling weights). We also have missing data, which I'd prefer to deal with using multiple imputation. I'm running into a problem doing some moderator analyses (e.g., by sex) on some of the paths in the model. The approach I'm taking is to do a multiple group analysis (e.g., males vs. females) and testing the Satorra-Bentler corrected difference in chi-square statistics for the models in which the path of interest is constrained be invariant between the two groups vs. models in which the path is not constrained. The problem is that in order to compute the Satorra-Bentler corrected test of differences in chi-square, I need the scaling correction factor for the chi-square statistic which is not included in the output of multiple imputation analyses. A message from Linda Muthen on the Mplus blog (from 2006, if I remember correctly) indicated that nobody has worked out how to compute the scaling correction factor under multiple imputation, and that it's not as straight-forward as simply averaging them across the results for the individual imputations. So I was wondering if you know a way to get the scaling correction factor in this context (I'm open to using other software if necessary), or have any suggestions for different approaches I might take for testing moderation and/or handling missing data. I've tried FIML estimation using auxiliary variables, but I'm having convergence problems when I include all of the auxiliary variables I want to include in these multiple-group analyses (and the models take FOREVER to run, which makes tweaking and adjusting cumbersome). In case it matters, all of my variables are continuous except for the grouping variables. Thanks for any insights you might have! -Sean
