I am considering ways to empirically tease out individuals for whom a given model DOES NOT fit well and looking for meaningful patterns and predictors of misfit. I'm wondering in what ways a mixture modeling approach might differ from more traditional treatment of residuals. Does the conditional independence assumption of mixture modeling capture any systematic patterns and correlates of misfit, rendering further analysis of the residuals unnecessary? Or might there be an advantage to calculating individual-level residuals for a given model (even a mixture model) and subjecting these residuals to further analysis?
bmuthen posted on Thursday, January 22, 2004 - 10:06 pm
I think you are on to an interesting topic that will probably expand in the near future. Individual-level residuals would certainly be of interest also with mixtures and there is some new, unpublished work in that area by the PSMG group. It is also true that mixtures have traditionally been used to capture misfitting individuals in terms of outliers forming their own class. Much more can be done along these lines I would think.