Causal Methods for Mediation and Interaction

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): New causal methods are proposed for understanding randomized behavioral interventions in randomized mental health trials with mediation and interaction analyses. Motivated by specific hypotheses from four mental health studies, the two primary aims are: 1) to aid the explanation of the mechanism of action of these interventions through post-randomization factors such as provider adherence to treatment guidelines and social support (i.e., mediation with possible interactions); and 2) to help identify subgroups of patients based on post-randomization factors such as medical comorbidities and non-study treatments across which randomized intervention effects vary. An additional third aim addresses dissemination. The proposed methods relax a strong assumption of current mediation/interaction methods that assumes no unmeasured confounding for the mediating or interaction factors (sequential ignorability). We make other assumptions involving more parametric models with sensitivity analyses based on different modeling approaches. For the mechanism of action goal, we propose extensions of structural mean models (SMM) for estimating prescriptive and natural direct effects and natural indirect effects. Natural effects are appropriate for the effectiveness research represented by the studies of interest as they provide a theoretical basis for indirect effects and accommodate interactions between baseline interventions and post-baseline behavioral and process factors. We also will focus on developing optimally efficient weights that improve precision without assuming sequential ignorability. For the post-randomization stratification goal, latent principal strata or subgroups are identified under the principal stratification (PS) approach, using post-randomization factors and randomized baseline intervention information. The baseline intervention effect in strata corresponding to constant post-randomization factor levels regardless of randomization arm represent prescribed direct effects. We extend both the SMM and PS approaches to multiple nested post-randomization factors (e.g., physician prescription and patient medication behavior), longitudinal outcomes, binary outcomes, and cases where post-randomization outcomes influence post-randomization adherence behavior. The above methods and standard mediation/interaction procedures and their assumptions will be evaluated and compared with simulations and analyses to answer the specific hypotheses from four studies of interest.