A0549
Title: A general framework for incorporating identification uncertainty in individualized treatment rules
Authors: Muxuan Liang - University of Florida (United States) [presenting]
Yingqi Zhao - Fred Hutchinson Cancer Research Center (United States)
Ting Ye - University of Washington (United States)
Abstract: Estimating individualized treatment rules (ITRs) from observational data or clinical trials with non-adherence is challenging due to possible unmeasured confounding bias. Partial identification approaches using an instrumental variable (IV) provide characterizations on possible values of the conditional average treatment effects (CATEs). A new class of `optimal' ITRs is developed to guide treatment decisions when the CATEs are only partially identified. A novel value function is defined, allowing a reject option in treatment decisions under partial identification, and that value function is used to define a class of IV-optimal ITRs with a reject option. The reject option informs those susceptible to identification uncertainty and allows the use of alternative ITRs derived from other studies or outcomes for these patients. To estimate the IV-optimal ITRs with a reject option, a weighted classification framework is developed with a modified hinge loss function, where the weights are non-smooth transformations of nuisance parameters. Simulations and real data analysis are conducted to demonstrate the superiority of the developed framework and estimation procedure.
