Jen posted on Tuesday, November 09, 2010 - 1:50 pm
I'm using GMM w/ non-linear data: 4 pts, 1 pre- & 3 post-intervention. Change after the intervent. wears off. I'd like to use GMM to ID subgroups responding differently.
I've noted some pubs using LCGA don't address intercepts & slopes at all; fit stats are used to decide on # of classes & comparisons made based on membership. Do you consider this an acceptable approach w/ LCGA &/or GMM?
I tried 2 things:
1) A discontinuous model– i s1 | a@-1b@0c@0d@0; i@0; s1@0; i s2 | a@0b@0c@1d@3; s2@0; (LCGA.) W/ only 2 pts for the 1st slope, the var for s1 will always be set to 0, but freeing the var of i & s2 would make this GMM. Is anything wrong w/ this approach? Would adding a quad to the 2nd part of the model (w/ var=0) hurt?
2) A cubic model is not IDed w/ 4 pts, but is anything wrong w/ fitting a cubic model & fixing var=0? I'm not interested in using the vars within classes as predictors, just class membership.
If I'm only interested in IDing classes & not description, does it matter if the predicted trajectories are good fits?
If you are looking for intervention effects I think it is awkward to use LCGA - you can't pinpoint the intervention effect. With GMM the intervention effect is on the slope of the within-class development which is a clear concept.
Looks like you find a big change from the 1st to the 2nd point. Piecewise can be good, but with only 4 time points, each piece has only 2 and cannot support both an intercept variance and a slope variance. Perhaps one quadratic model is ok.