Hello Linda! I'm estimating GMMs with a sample that includes about 70% dyads (patients and caregivers). To accommodate the dependency within dyads, I use dyad as a clustering unit with COMPLEX (following an earlier recommendation from you). I'm testing whether the latent growth classes differ, regressed on group (pt vs cg) as a primary question. I have been estimating the i s & q for each latent class separately, of course, and estimating i and s variance within classes, but regressing i s & q on group under %OVERALL% as a secondary question instead of within the latent classes. Does that seem sensible, or should I include the tests of the growth parameters by group within each latent class? Thanks for any comments and guidance! Bruce
To use CLUSTER=DYAD; your data must be in the long format. When have data in the long format, patients and caregivers cannot be compared. They are assumed to be statistically equivalent. You should put your data in the wide format where you can then compare patients and caregivers. The multivariate nature of the data takes into account the non-independence of observations. See the Topic 7 course handout starting at Slide 158 and the same topic in the Topic 8 course handout.
Thanks very much for your pointers, Linda! Topic 8 did the trick and clarified what I needed to do to build the models.
I must have been too confusing in my note, or I'm missing something. For these 3-level hierarchical data (occasions within persons within dyads), I wanted to test whether patients and caregivers had different change trajectories in general, and if there were different latent growth classes, whether (1) the classes differed in their compositions of pts and cg, and if so, what growth parameters they differed on. The Topic 8 example (slides 142ff) gave me the clues I needed for my models. In this case (and hence your previous suggestion), only 65% of the cases are in dyads, so the multilevel approach didn't work. I used TYPE=COMPLEX only to accommodate nonindependence in those dyads. However, there was no desire to test whether there were differences betw cg and pts within dyads. Best, Bruce