Comparing Three Approaches to Latent Class Analysis with Distal Outcomes: A Scientific Summary

Prevention scientists increasingly are using latent class analysis (LCA) to identify subgroups of individuals based on unique patterns of behavior, risk exposure, mental health symptoms, and other characteristics. Now that LCA has proven to be a powerful and intuitive tool for studying heterogeneity in risk behaviors and related constructs, new methods are needed to address the next generation of complex questions about how subgroup membership is embedded in developmental pathways. These questions are often concerned with how subgroup membership is linked to later outcomes; for example, do patterns of early risk exposure during childhood predict later binge drinking during adolescence? Addressing questions using LCA with a distal outcome poses interesting methodological challenges; during the past two years the literature has included a rapidly increasing number of publications proposing competing approaches to address these methodological challenges. This talk presents a scientific summary of the three state-of-the-art approaches to LCA with distal outcomes, in which we sort out important differences in terminology that have made penetration of this literature difficult. We will focus on the simplest case of a latent class predictor and an observed distal outcome. In the talks that follow in this symposium, this simple model will be expanded so that control variables, moderators, and propensity scores can be added to address the complex developmental questions posed by prevention scientists. 

The approaches we synthesize fall into three general categories based on: (1) weighting by the classification error; (2) Bayes’ Theorem; and (3) improvements to the posterior probabilities. Each approach has been shown to work well under certain conditions in recently published simulation studies. To date, however, there has been no comprehensive overview summarizing the approaches and their assumptions or integration of “take-home messages” across simulation studies. To further complicate matters, not all approaches are implemented in all LCA software packages and the availability of high-quality standard errors depends on the combination of approach and software package selected. In order to lay the foundation for recent advances in LCA with distal outcomes presented in the other talks of this symposium, we will (a) describe the three approaches to LCA with distal outcomes, (b) clarify the assumptions of each approach, (c) integrate “take-home messages” of published simulation studies, and (d) summarize available software options. Our goal in this talk to is to help scientists make informed decisions about which approach to choose given their research questions, data, latent class model quality, outcome model complexity, and available software.