Estimating the Causal Effect of Latent Class Membership on a Distal Outcome: A Step-By-Step Primer

Introduction. Modern approaches to latent class analysis (LCA) with distal outcomes can provide insight into the potential risk conferred by class membership. However, as with any standard regression-based approach, coefficients describing the association between latent class membership and a distal outcome are correlational. Propensity score techniques are being used with increasing frequency in prevention science to strengthen causal inferences with observational data. These methods can be directly extended to studies where the exposure of interest is a latent class variable. We present a step-by-step primer on integrating inverse propensity weights and LCA with distal outcomes in the context of estimating the causal effect of adolescent substance use latent class membership on adult outcomes. 

Methods. We focus on a new approach to LCA with distal outcomes that is straightforward to implement: classify-analyze with improved posterior probabilities. We present the following steps to apply inverse propensity weighting to LCA with distal outcomes: (1) select LCA model; (2) obtain posterior probabilities; (3) estimate propensity scores, check overlap; (4) calculate weights, check balance; (5) apply weights and fit analysis model. Although some of these steps may seem straightforward to scientists familiar with LCA or inverse propensity weighting, unique challenges arise with their combination, including model identification with many confounders and model congeniality when estimating posterior probabilities and propensity scores. 

Results. Data from the National Longitudinal Study of Adolescent and Adult Health (n=1315) were used to identify four classes of adolescent substance use based on cigarette use, regular cigarette use, alcohol use, binge drinking, drunkenness, and marijuana use: Non-Users (42%), Smokers (13%), Drinkers (19%), and Heavy Users (26%). Inverse propensity weights were used to adjust for twenty-three confounders at the individual, family, peer, school, and neighborhood levels. There was a significant causal effect of adolescent latent class membership on a variety of adult outcomes, including regular cigarette use, binge drinking, marijuana use, and depression, even after adjusting for confounding.   

Conclusion. Understanding the consequences of latent class membership is often of great interest to prevention scientists; these consequences dictate the significance of unique behavioral profiles in terms of their relevance to health outcomes. Although recent methodological advances now enable researchers to estimate associations between latent class membership and distal outcomes, the integration of modern causal inference methods will permit a greater understanding of the role that complex behavior profiles play in public health.