Session: Advances in Causal Modeling of Longitudinal Intervention Effects (Society for Prevention Research 21st Annual Meeting)

This symposium will discuss recent developments in causal modeling in the longitudinal intervention context. A natural question that arises in longitudinal intervention trials is if there are distinctive strata of individuals who exhibit heterogeneous longitudinal outcome and how differently these strata benefit from the intervention. The problem with this inference is that the strata membership is completely unknown (latent) and individuals cannot be observed under both intervention and control conditions. This seemingly intractable identification problem is solved via empirical model fitting in latent class type analyses such as growth mixture modeling (GMM). The main drawback of this approach is that we identify both the latent strata and the strata-specific treatment effects heavily relying on empirical model fitting, and therefore the resulting inference on intervention effects may be largely influenced by model specifications, deviations from parametric assumptions, and the presence/absence of auxiliary information.

In this symposium, we present alternative ways of identifying causally interpretable longitudinal intervention effects. In particular, our interest is in reducing the dependency on parametric assumptions and in facilitating diverse and rigorous sensitivity analysis and diagnostics methods, which we believe will significantly strengthen the inferential practice in the context of longitudinal intervention trials. The three papers in this symposium commonly utilize the idea of principal stratification (PS), which is one of the causal modeling approaches that have been specifically developed to achieve causal inference taking into account latent discrete subpopulations that are formulated based on potential values of posttreatment outcomes. Despite their conceptual similarity, potential benefits from integrating the latent class type and PS approaches have not been examined much. This session is intended to carefully assess the utility of PS in causal modeling of longitudinal intervention effects. 

The first paper (Wang) compares the widely used one-step GMM and the relatively new two-step hybrid GMM. Advantages and disadvantages of each approach and potential benefits of jointly using the two approaches will be discussed. 

The second paper (Jo) focuses on the use of two-step GMM as a practical tool for longitudinal causal inference. Various possibilities of identifying strata-specific intervention effects by combining the strengths of GMM and PS will be demonstrated. 

In the third paper (Lin), times of sequential event are explicitly modeled as outcome. A PS framework is utilized in identifying differential effects of treatment on time to event (quitting smoking) and the gap time (quitting to relapse).