Mixture Modeling

Hi All!
If you have been directed to this post, you stated that you were interested in, or are currently using mixture models in your analyses. Please use this forum to connect with each other either in person or online... It would be great if we could share knowledge and advice with each other!

Hi all! I am just branching out into mixture modeling, particularly LPA, and am having some fun issues related to the fact that I have 15,000 participants in my sample. Anyone else ever try to run these types of models on huge datasets and want to share their insights? All the information criteria are telling me I need 11+ classes and that seems a little ridiculous to me!

Hi Leanne, Ideally these models should be run in large samples. So your sample size is a definite plus. Can you give us a little more information. I'm interested to know, what type of data are you modeling? What are the observed variables? Dimensional scales or categorical items? Somewhat separately, what do the 11 classes look like? Are they gradations of severity? Or are they suggestive of qualitative differences?
Thanks, 
a

Thanks Aidan! I wasn't sure if there are issues of power (or I guess more likely being over powered) to be concerned about). The models I'm running have five continuous observed variables, which are different behavioral and academic outcomes measured in kindergarten. At this point I'm just running the LPA - no predictors, distal outcomes, anything like that, but I do have sample weights and clustering in the data too. After 4-5 classes, the additional profiles don't really look qualitatively different, just slightly above/below existing classes. The other issue is that when I add more classes, some of the groups are really tiny, and so if my goal is to identify profiles of interest that other people could potential replicate, I'm not sure how relevant these tiny groups are. So my real question - am I justified in going with the more parsimonious model even if all the empirical criteria say add more classes?

I would say yes, interpretability remains king here as well. There is a lot that could be said about any given data set or analysis, but ultimately, yes, interpretability should be your guide. I would recommend the Wiley Blue Book by Collins and Lanza (2010). Linda and Stephanie really are the guiding lights when it comes to navigating the shoals of LCA. 

A