Non-overlap Average Treatment Effect Bounds

Abstract: The average treatment effect (ATE), the mean difference in potential outcomes under treatment and control, is a canonical causal effect. Overlap, which says that all subjects have non-zero probability of either treatment status, is necessary to identify and estimate the ATE. When overlap fails, the standard solution is to change the estimand, and target a trimmed effect in a subpopulation satisfying overlap; however, this no longer addresses the original goal of estimating the ATE. When the outcome is bounded, we demonstrate that this compromise is unnecessary. We derive non-overlap bounds: partial identification bounds on the ATE that do not require overlap. They are the sum of a trimmed effect within the overlap subpopulation and worst-case bounds on the ATE in the non-overlap subpopulation. Non-overlap bounds have width proportional to the size of the non-overlap subpopulation, making them informative when overlap violations are limited -- a common scenario in practice. Since the bounds are non-smooth functionals, we derive smooth approximations of them that contain the ATE but can be estimated using debiased estimators leveraging semiparametric efficiency theory. Specifically, we propose a Targeted Minimum Loss-Based estimator that is $\sqrt{n}$-consistent and asymptotically normal under nonparametric assumptions on the propensity score and outcome regression. We then show how to obtain a uniformly valid confidence set across all trimming and smoothing parameters with the multiplier bootstrap. This allows researchers to consider many parameters, choose the tightest confidence interval, and still attain valid coverage. We demonstrate via simulations that non-overlap bound estimators can detect non-zero ATEs with higher power than traditional doubly-robust point estimators. We illustrate our method by estimating the ATE of right heart catheterization on mortality.