Causal Effect Estimation after Propensity Score Trimming with Continuous Treatments
share
Most works in causal inference focus on binary treatments where one estimates a single treatment-versus-control effect. When treatment is continuous, one must estimate a curve representing the causal relationship between treatment and outcome (the "dose-response curve"), which makes causal inference more challenging. This work proposes estimators using efficient influence functions (EIFs) for causal dose-response curves after propensity score trimming. Trimming involves estimating causal effects among subjects with propensity scores above a threshold, which addresses positivity violations that complicate estimation. Several challenges arise with continuous treatments. First, EIFs for trimmed dose-response curves do not exist, due to a lack of pathwise differentiability induced by trimming and a continuous treatment. Second, if the trimming threshold is not prespecified and is instead a parameter that must be estimated, then estimation uncertainty in the threshold must be accounted for. To address these challenges, we target a smoothed version of the trimmed dose-response curve for which an EIF exists. We allow the trimming threshold to be a user-specified quantile of the propensity score distribution, and we construct confidence intervals which reflect uncertainty involved in threshold estimation. Our resulting EIF-based estimators exhibit doubly-robust style guarantees, with error involving products or squares of errors for the outcome regression and propensity score. Thus, our estimators can exhibit parametric convergence rates even when the outcome regression and propensity score are estimated at slower nonparametric rates with flexible estimators. These findings are validated via simulation and an application, thereby showing how to efficiently-but-flexibly estimate a dose-response curve after trimming.
READ FULL TEXT
