George Howe posted on Tuesday, July 26, 2011 - 10:27 am
I've created 50 imputed data sets using MPLUS, and have been running some structural equation models. I'd like to do nested model comparisons to test some equality constraints, but am not sure this is acceptable. In addition I've been using MLR, and no correction factor is provided, so it looks like Satorra-Bentler is not possible? (I haven't seen anything on this in the imputation literature, but am still trying to run down Craig Ender's book; references would be nice if you know of any).
Currently the only likelihood ratio test available with imputations is the ML estimator (for testing unrestricted against a structural model).
You can easily however use the Wald test for your testing purposes with imputation. Look at page 618 "Model Test" command.
George Howe posted on Tuesday, July 26, 2011 - 12:57 pm
Thanks, Tihomir, I'd wondered about that. Worked fine!
shaun goh posted on Saturday, March 09, 2013 - 9:30 pm
I'm wondering how to compare two nested models from 2 multiple imputed data sets in a multi-group, LGC setting.
I am interested in group differences of trajectories - differences in the means of the latent factors of the growth curves between 2 groups. I was going to constrain the means to equality and employ chi-sq testing, but am unsure how to conduct likelihood testing with multiple imputed data. Is there a way to do so in Mplus (i.e. by "Model Test" or by hand-calculation of chi-sq differences?)
You can use MODEL TEST. Difference testing using chi-square or the loglikelihood has not yet been developed for multiple imputation.
shaun goh posted on Sunday, March 10, 2013 - 4:21 pm
Thanks Linda, I will utilise MODEL TEST.
I was wondering, if difference testing using chi-square/loglikelihood/fit statistics have not yet been developed for multiple imputation, how does MODEL TEST run a valid test for equivalence between two nested models in the multiple imputation context?
I am currently working with multiply imputed data sets and want to compare models. I used the model test command (Wald tests) as were recommended in this thread, and this worked fine. However, I wonder whether there is a publication (or something like that) that I can refer to (cite) to argue that traditional chi-square difference tests are not yet implemented when using multiple imputation?