Stephan posted on Thursday, May 29, 2008 - 6:24 pm
Hello, I'm running a multiple group analysis where the data set contains a grouping variable (1/2). The analysis worked well and also mean difference tests showed group differences among 4 latent variables. Now I had the idea to confirm group membership by computing a CFA mixed model as it is shown in ex 7.17. However, the fit criteria are reasonable but the classification does not match (correlation ~.4) real class membership. My two questions:
1) Did I choose the wrong approach? How is it possible to increase the overall classification accuracy? 2) ex 7.17 says (…) %c#1% [f*1] “the factor mean varies across the classes”
Is this simply done by [f*1]? And with multiple factors will this be done through %c#1% [f1*1] [f2*1] [f3*1] (…) I’ asking because 7.20 deals with 2 LV and there it is stated “[f1*1 f2]”
You may come closer if you let the factor covariance matrix vary across classes, not only the factor means which is the default. But there is no guarantee that the same known classes will be recovered - another class formation may fit better. So the attempt to "confirm" the known groups may not be achievable and I am not sure why it would be desirable.
Dear Dr. Muthen, thanks for your response. Why? I'm just comparing/try to get familiar with various classification approaches using LCA and Yarnold's ODA. It might be tough to confirm group membership due to unmeasured effects but it's interesting.
My angle was that if the known groups are say males and females, it may well be that some males are more similar to the average female on these outcomes and vice versa. That is, the known groups may not represent the most or all of the heterogeneity. But as an excercise in understanding the methods it could be useful.