A New Solution to the Problem of Posttreatment Confounding in Causal Mediation Analysis

• Activity Number: 414 - Making Questions Relevant and Assumptions Realistic: New Strategies for Causal Mediation
• Type: Topic Contributed
• Date/Time: Wednesday, August 5, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Guanglei Hong* and Fan Yang and Xu Qin and Stephen Raudenbush
• Companies: University of Chicago and University of Colorado Anschutz Medical Campus and University of Pittsburgh and University of Chicago
• Keywords: Causal inference; counterfactual outcomes; direct effect; indirect effect; ratio-of-mediator-probability weighting; selection bias

Abstract

In a causal mediation analysis that decomposes the treatment effect on an outcome into a natural indirect effect and a natural direct effect, past research has generally considered it infeasible to adjust for a posttreatment confounder when there exists treatment-by-mediator interaction. This is because a posttreatment confounder observed under treatment condition t that an individual has been assigned to (which is denoted by Z(t) = z) is unobserved under the counterfactual treatment condition t’ (which is denoted by Z(t’) = z’). This study presents a novel solution to the problem, the key of which is to impute the values of the counterfactual posttreatment confounder Z(t’) that can be adjusted for together with Z(t) in ratio-of-mediator-probability weighting (RMPW). The imputed value is a function of not just observed pretreatment covariates; the imputation also captures unobserved influences on the posttreatment confounder through making use of random effects and residuals. We evaluate the performance of this new strategy through simulations and illustrate with data from the National Evaluation of Welfare-to-Work Strategies (NEWWS).