관심 있는 분의 많은 참여 부탁드립니다.
○ 주제 : Causal inference with positive-unlabeled treatments
○ 연사 : 이유진 교수 (Department of Biostatistics, Brown University)
○ 일시: 2026년 04월 22일(수) 오전 11시 - 12시
○ 장소: 서울대학교 보건대학원 221동 202호
○ 방식: 현장 참여
○ 문의: museum03@snu.ac.kr

Abstract: Treatment information in observational studies is often incompletely ascertained. In many comparative effectiveness studies using data not originally collected for research purposes, only a subset of treated units is identified, providing asymmetric treatment information. This setting corresponds to positive-unlabeled (PU) data in machine learning literature, consisting of a conservatively defined treatment group ("positives'') and a contaminated control group that may include both treated and untreated units. Although PU learning algorithms provide predicted treatment probabilities analogous to propensity scores, their integration into principled causal inference frameworks without ascertained treatment variable remains underdeveloped.
In this work, we establish identification conditions and propose estimators of the average treatment effect (ATE) based on PU-augmented propensity scores under different labeling assumptions. We also examine theoretical and practical challenges that arise when incorporating prediction-based outputs into causal estimators, particularly when the assumptions underlying causal inference and predictive modeling are not fully aligned. We illustrate the proposed framework using the data from the Future of Families & Child Wellbeing study to study the effect of smoking during pregnancy on low birth weight.