I wonder which strategy would be the best to approach this research problem. I have a cross-sectional time-series dataset with 20+ time-points (years). The design is not panel; the subjects are sampled at random at every time point. I want to do a Latent Class model that would involve a time variable, since I have good reasons to think that classes are subject to change across time. If I ignore the time-aspect, I loose a meaningful change in the pattern. The question is, can I do latent class growth (?) model if the data is not panel/longitudinal. If I can, what would be the “best-fitting” MPlus example from the manual?
You need panel data to do laten class growth modeling. You should see your years as different subpopulations from which you have independent observations. That then leads to multi-group analysis, one group for each year (or you can limit yourself to some relevant years). You do LCA in these multiple groups and you can then test for measurement invariance (equality of conditional item means) and structural invariance (equality of class proportions). Multi-group LCA can be done either by using Knownclass or by having group as dummy covariates in MIMIC style.
Many thanks for your reply. It makes sense. I will follow this approach then, however, I believe that if I do the analysis for each year in a sample and then follow-up with the metric equivalence, the latter assumption won't hold and I won't be able to compare the constructs over time.