Sensitivity Analysis via the Proportion of Unmeasured Confounding

In observational studies, identification of ATEs is generally achieved by assuming that the correct set confounders has been measured and properly included in relevant models. Because this assumption both strong untestable, a sensitivity analysis should be performed. Common approaches include modeling bias directly or varying propensity scores to probe effects potential unmeasured confounder. article, we take novel approach whereby parameter “proportion confounding”: proportion units for whom treatment not as good randomized even after conditioning on observed covariates. We consider different assumptions probability unit being unconfounded. each case, derive sharp bounds average effect function propose nonparametric estimators allow flexible covariate adjustment. also introduce one-number summary study’s robustness number confounded units. Finally, explore finite-sample properties via simulation, apply methods an database used assess right heart catheterization. Supplementary materials article are available online.