Detecting Distributional Treatment Responders with False Discovery Rate Control

Abstract:

In this paper,
we introduce causal responder detection (CARD), a method for distributional responder analysis that identifies treated subjects whose outcomes significantly depart from the control response distribution while controlling the false discovery rate (FDR) marginally over the tested treated population. CARD builds on the AdaDetect framework and, in randomized settings, inherits finite-sample FDR control under the exchangeability conditions required by AdaDetect when coupled with the Benjamini–Hochberg procedure. For observational studies, we propose a propensity score–adjusted extension whose validity is asymptotic and depends on ignorability, overlap, and adequate propensity-score estimation. Simulation studies and real-data applications demonstrate that CARD effectively detects distributional treatment responders with high power across a range of heterogeneous distributional treatment-effect scenarios.

Submission Type: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission: To make the revision easy to evaluate, we briefly summarize the main changes made in response to the reviews. First, we revised the title, abstract, introduction, inferential-goal section, and discussion to consistently frame CARD as detecting distributional responders, rather than certifying unit-level individual treatment effects (see revised title; Abstract; Introduction; Section 1.2; and Discussion). Second, we moved the marginal-versus-conditional error-control discussion into the main text and added an illustrative example showing that conditional Type I error can be poor even when marginal control is preserved (Section 3.4, “Marginal vs. conditional error control”). Third, we strengthened the distinction between randomized and observational settings, emphasizing that CARD has finite-sample FDR control in RCTs, while the observational extension is asymptotic and depends on ignorability, overlap, and adequate propensity-score estimation (Abstract; Section 3.1; Section 3.3; Discussion). Fourth, we expanded the empirical characterization by adding component ablations, observational overlap/propensity sensitivity analyses, and a derandomization analysis for knockoff sampling (Appendices E, F, and G). Finally, we moved a condensed exploratory semaglutide real-world data example into the main text, added a computational-cost discussion, and added a broader-impact discussion (Section 4.3; Appendix D; Discussion; Broader Impact Statement).

Assigned Action Editor: ~Arash_Mehrjou1

Submission Number: 7717