Conditional Separable Effects

Researchers are often interested in treatment effects on outcomes that only defined conditional a post-treatment event status. For example, study of the effect different cancer treatments quality life at end follow-up, individuals who die during is undefined. In these settings, naive contrast variable not an average causal effect, even randomized experiment. Therefore principal stratum those would have same value regardless treatment, such as always survivors truncation by death setting, advocated for inference. While this well contrast, it hard to justify relevant scientists, patients or policy makers, and cannot be identified without relying unfalsifiable assumptions. Here we formulate alternative estimands, separable effects, natural interpretation under assumptions can falsified We provide identification results introduce estimators, including doubly robust estimator derived from nonparametric influence function. As illustration, estimate chemotherapies with prostate cancer, using data clinical trial.