I'm doing GMM analysis with 5 imputed datasets. When I do the analysis for each imputed dataset separately, I have similar class membership results across 5 imputed datasets. But I run the analysis for all 5 imputed datasets with "type=imputation" option, the class membership results are substantially different from any of the results from a single imputed dataset. Please let me know how this might happen, and how MPLUS produces combined results for a class membership based on multiply imputed datasets. Thank you!
I'm running a 4-class GMM using multiple imputation (20 datasets), and I'm running into some inconsistencies when I save class membership and class probabilities. I've used the procedure I've seen on this discussion board in the past (i.e., set up a model with all values fixed to the parameters from the final model and save data on one of the imputed data sets).
The problem is that the class membership assignments do not correspond perfectly with the classes assigned to cases based on most likely membership across imputed datasets (close, but off by a few cases). I reasoned that this might be a function of saving class probabilities and membership from just a single data set. So, I saved data from all of my imputed data sets and calculated average class probabilities myself -- but these don't match either!!
Now I am questioning how the class assignments are calculated to arrive at the numbers shown in the output. The output reads "CLASSIFICATION OF INDIVIDUALS BASED ON THEIR MOST LIKELY LATENT CLASS MEMBERSHIP", and I assumed this was a function of posterior probabilities across imputations. Am I interpreting this wrong? Is there a way to save class assignment from multiple imputation and get the same classification as printed in the output?
I am running LCA on 5 imputed datasets using 4 categorical variables with 3 levels each. I have also included 3 covariates in the model. After running the model using type=imputation I then fix the parameters to the average estimates and run the analysis again on one single dataset in order to get the class probabilities. When I do this the classes are different in size to those from the results averaged across the imputed datasets. My model input looks like this: