Propensity Score Methods for Assessing Causal Mediation in Cluster-Randomized Substance Use Prevention Interventions

Substance use prevention interventions are often designed to affect mediators, such as substance use norms, which in turn are hypothesized to reduce substance use. When evaluating the effectiveness of interventions, the causal effect of the intervention through a mediator may be of interest. Estimating this causal effect is complicated by the fact that, unlike the treatment condition, individuals cannot be randomly assigned to levels of the mediator, because the mediator is an outcome of the intervention. In the absence of randomization, propensity score methods can be used to estimate the causal effect of the mediator (measured at the individual level), yet two complications may arise this context. First, interventions are often group randomized and/or administered at the school or classroom level and second; mediators are often continuous rather than categorical variables.

In this paper, we seek to build on previous methodological work that has been done regarding propensity score modeling in a multilevel context. However, given our interest in continuous mediators, we extend previous work that has been done in the context of binary propensity score modeling to the continuous case, with a focus on propensity score weighting. Specifically, we seek to identify the optimal form of the propensity score model, given the multilevel nature of the intervention. We will present a simulation study comparing propensity score models for a continuous mediator using a multi-level model, a single-level model with fixed effects, and a single-level model without fixed effects. Additionally, we will discuss implementation challenges and logistics, including how to assess balance with propensity score weights for a continuous exposure, and whether balance should be obtained within clusters or across clusters.  Finally, we will present an application in which we investigated the causal mediation effect of HealthWise South Africa, a prevention intervention randomly assigned and administered at the school-level, on substance use.

Given the prevalence of group-based designs as well as the importance of understanding mediational pathways in prevention research, methodological developments in the field of causal mediation are widely applicable to the field of prevention science.