I am attempting to apply sampling weights that are available in a large scale national data in an LCA model. Further, I am doing the LCA model on a subpopulation of the larger dataset. I read on another thread that I'd ideally like to use the "subpopulation" option so that the weights sum up to the right total, but this is not available with type=mixture. Do you have any suggestions for how I can proceed?
Also, as suggested in this thread, I ran an LCA on my sub population and multiplied the posterior probabilities by the weight variable saved in the save data file and have two questions:
1) The weight variable in the original data file ranges from 535.1 to 19923.12 and the weight variable in the save data file ranges from .502 to 4.337. Why is my weight variable in the save data file different than the weight variable in the original data file?
2) I multiplied the posterior probabilities for each latent class by the weight variable in the save data file. I then summed across these multiplied columns and the value is equal to my original sample size. I though this sum would be more representative of the national population for the sub sample. Am I thinking about this incorrectly?
Thank you for the response. I have been running LCA's with and without sampling weights and am not seeing much of a difference in the fit statistics and/or the estimated class probabilities. I recently read a paper by Vermunt (2007) where he described sampling weights using pseudo maximum likelihood estimation. That is what Mplus does, right? Weights the frequencies before estimating the latent classes? I'm trying to understand and justify why the weights don't make a difference in my LCA results since it is a national data set and I expected some sort of change when using them. Thanks in advance.
Yes, Mplus uses pseudo maximum likelihood with weights.
Not seeing differences in the parameter estimates probably means that the groups for which the sampling weights differ considerably from one are not that different with respect to the model parameters.