(with Anton Strezhnev)
Matching methods are widely used to reduce the dependence of causal inferences on modeling assumptions, but their application has been mostly limited to the overall effect of a single treatment. Direct effect analyses, which estimate the effect of a treatment not due to some mediator, have become increasingly popular in the social sciences for a wide variety of inferential targets, including understanding the causal mechanisms of a treatment. Standard matching analyses, however, are not directly applicable to direct effect analyses because of their tendency to induce post-treatment bias, and so almost all applications are dependent on the correct specification of several models. In this paper, we propose a novel two-step matching approach to estimating direct effects, telescope matching, that reduces model dependence without inducing post-treatment bias. This method uses matching with replacement to impute missing counterfactual outcomes in a flexible manner and relies on regression models to correct for bias induced by imperfect matches. We show in simulations that our approach is more robust to misspecification of these regression models than non-matching estimators. We derive the asymptotic properties of this estimator and provide a consistent estimator for its variance. Finally, we apply this approach to estimating the direct effect of a job training program on long-term mental health not due to employment and show that it can generate substantively different inferences than standard approaches.