I have a theoretical question involving the use of population level data in a latent class/latent transition analysis. I am working with data that describes state level legislation that deals with drug monitoring programs at pharmacies and doctors’ offices. For example, one of the binary variables has a Yes/No possible response to the question, “Are monitoring programs required to report to law enforcement agencies”. I have data across 14 years, for all states.
I’ve been using Mplus to estimate a variety of LCA and LTA models for this data, with the end goal of identifying underlying latent classes in the states related to patterns in legislation, and which classes may be used for subsequent analyses as predictors. I have 51 rows of data(50 states + DC) which is generally considered to be a smaller than recommended sample size for these types of analysis, however, this is actually complete population level data, where the population in these analysis will always be constrained to 51 measurement units.
Can some of the small sample size concerns be ignored when the LCA/LTA analyses are being performed on population data? Are standard errors of any use when considering the population data? And, finally, should I be using this type of inferential method when using population data?
It depends on the inference you are making. If you want to make an inference about the LTA distribution for a year that you observed then the SE is 0. If you are making an inference for what will happen in the future then those standard errors are relevant. You have the full population of states conditional on the year but not unconditional. When Mplus 8.1 is released you can run a dynamic LCA on this data. http://statmodel.com/download/DLCA.pdf