A1904
Title: A general framework for incorporating identification uncertainty in individualized treatment rules
Authors: Muxuan Liang - The University of Texas MD Anderson Cancer Center (United States) [presenting]
Abstract: Estimating individualized treatment rules (ITRs) from observational data or clinical trials with non-adherence is challenging due to potential unmeasured confounding bias. Instrumental variables (IV) can offer partial identification of the possible values for conditional average treatment effects (CATEs). However, optimal treatment decisions under partially identified CATEs can be uncertain, and current literature fails to inform such uncertainty in treatment decisions. A deferred treatment option is adopted and a novel class of optimal ITRs called Enhanced IV-optimal ITRs with deferred option (EIV-ITRs) is developed to guide treatment decisions and account for identification uncertainty. The deferred option extends beyond the original treatment options, identifying patients susceptible to identification uncertainty and allowing for the collection of additional information or physicians' guidance to aid in optimal decisions. To estimate the EIV-ITR, a weighted classification framework is developed with a modified hinge loss function, where the weights are nonsmooth transformations of nuisance parameters. An augmented empirical risk minimization approach is further proposed to estimate the EIV-ITRs, achieving a fast convergence rate even if the nuisance parameters are estimated using nonparametric or machine learning methods.
