Extending Latent Class Analysis with Distal Outcomes: Recommendations for Adding Control Variables and Moderators

Introduction. Latent class analysis (LCA) and latent profile analysis (LPA) have been applied widely to identify multivariate profiles. As prevention scientists have adopted these approaches, more sophisticated questions arise about how subgroup membership is linked to later outcomes, and how this link may be affected by other variables. We will present results from a simulation study that allow us to make recommendations about best practices for adding control variables and moderators in LCA and LPA with distal outcomes. 

Methods. We conducted a Monte Carlo simulation study that implements a modern classify-analyze procedure that improves the posterior probabilities in order to investigate the optimal way to integrate control variables and moderators when predicting a distal outcome from latent class membership. The simulation study compares five models for improving the posterior probabilities under different conditions of measurement quality and relative class size. Based on an empirical study of the effects of personality profile membership on frequent heavy drinking, data for the simulation study were generated using “true models” with five latent profiles. We considered true models with a control variable present but no interaction between the control variable and the latent profile variable, and true models with a control variable present and an interaction between the control variable and the latent profile variable. 

Results. One thousand replicate datasets for each cell of the simulation study were analyzed to compare five ways of improving the posterior probabilities for classify-analyze. Results suggest that when there is moderation, posterior probability improvements must include information about the moderator, but that when controlling for additional variables, performance is only modestly improved when their information is included. We summarize the results of this simulation study in the context of other published studies in order to make recommendations on extending existing approaches to LCA and LPA with distal outcomes. 

Conclusion. As state-of-the-art approaches to LCA and LPA with distal outcomes are more widely adopted by prevention scientists, these approaches will need to accommodate more complex models that include control variables and moderators. The recommendations made here will broaden the set of questions that can be addressed by prevention scientists. This is critical so that researchers can better isolate the effect of latent class membership on an outcome, and also so that heterogeneity across groups in the association between latent class membership and a distal outcome can be studied.