3

I am trying to fit a growth mixture model to a set of trajectory data. I am using a model that on a generic level is the same as this one.

I do understand how the latent intercept, linear, and quadratic variables are determined. However, I do not understand how the latent class variable is determined. Do I have to assign cases manually to different classes (as if I were doing a multi-group analysis when doing Structural Equation Modeling) or is there some automatic way of assigning cases to the classes? If so, how does this work?

Corvus
  • 4,573
  • 1
  • 27
  • 58
histelheim
  • 2,465
  • 4
  • 23
  • 40
  • 1
    Conditional on the number of classes chosen, you can estimate the posterior class membership distribution for each individual; this is typically standard output from software that fits such models. If there are large differences between the classes (with, presumably there are, if you've chosen the number of classes in a principled way, e.g. AIC) then most of the individual posterior distributions will be highly concentrated on one class, giving a basis for class assignment. – Macro Jan 28 '13 at 16:42
  • 1
    I agree with @Macro. I had a look at the paper you referenced and it seems to me that they used the data to find classes (and they tried several numbers of classes) and then looked to see if the classes found in this way corresponded to High and Low risk individuals. It doesn't look as if they used risk category as a covariate (as you might in a standard regression approach). As to assigning cases, the software gives you classification probabilities for each case, and the assignment is done on that basis. – Placidia Jan 28 '13 at 16:45
  • @Macro, when you say "chosen the number of classes in a principled way, e.g. AIC", do you mean trying several models with different number of classes and choosing the one with the best AIC? – histelheim Jan 31 '13 at 15:38

0 Answers0