Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Semiparametric Proximal Causal Inference with Invalid Proxies
Authors: Myeonghun Yu, Eric Tchetgen Tchetgen, Xu Shi,
Presenting Author: Myeonghun Yu*
Proximal causal inference, introduced by Miao, Geng, and Tchetgen Tchetgen [Biometrika 105 (2018) 987–993], has garnered significant attention for estimating causal effects in the presence of unmeasured confounders through the use of proxy variables. This framework has found applications in diverse fields such as longitudinal data analysis, mediation analysis, and survival analysis. However, current approaches critically assume complete knowledge of proxy validity, rendering identification results invalid when some candidate proxies are not valid. To address this limitation, we propose a novel identification framework that ensures validity as long as at least some candidate treatment-inducing proxies are valid, even without identifying the exact subset of valid proxies. Building on this framework, we develop the semiparametric theory for the average treatment effect in the presence of invalid treatment-inducing proxies and establish properties of doubly robust and locally efficient estimators. Extensive simulations validate the proposed framework, demonstrating its practical utility.