Xu, Man posted on Tuesday, January 15, 2013 - 7:36 am
I was wondering under MPlus what the most appropriate way would be to model a set of data like this:
In a longitudinal dataset (several waves), indivduals are nested in cluters. The outcome variable is death (binary, available at individual level). The predictor is state policy, at cluster level (NOT avaible at individual level), and vary over time.
Is it appropriate to take cluster aggregate of the outcome variable (e.g., the total number of death or ratio of death in each cluster at each wave), and use the time varying cluster level variable to predict the intercept and slope of the aggregated death outcome in a growth model?
Or is it better to fit a growth model with two levels, but only with the growth model described above at the between level? In this case, the individual level model would probably just be a mean and variance structure. Would this one be more advantageous as compared to the first model (maybe in terms of missing data, and ICC)?
I guess both would have to do with Ecological Fallacy - but it could be defensible to the extent that one doesn't draw inference to the indivduals, right?
I suppose one of the advantages of the second model is also that one can include indivdual level preditors as well (such as gender, ses, education level, etc).