A1517
Title: Estimation methods of heterogeneous treatment effects extending the w-method and a-learner for multiple outcomes
Authors: Shintaro Yuki - Doshisha University (Japan) [presenting]
Kensuke Tanioka - Doshisha University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: In a two-arm trial, participants are assigned to either the treatment or control group. If the treatment's efficacy is unclear within the overall population, identifying effective subgroups is crucial. This can be done by estimating heterogeneous treatment effects (HTE). While recent advances in HTE estimation use complex models that improve accuracy, they often reduce interpretability. Although methods for single continuous or binary outcomes are well-developed, approaches for multiple outcomes are less common. Current interpretable methods, like the W-method and A-learner, estimate HTE for single outcomes but struggle with the correlation structure in multiple outcomes. The aim is to propose two interpretable HTE estimation methods for multiple continuous and binary outcomes. Multiple squared loss and multiple logistic loss functions are introduced, which are bi-convex and based on reduced-rank regression, capturing correlations among outcomes. The proposed methods extend the W-method and A-learner to handle multiple outcomes. By leveraging the bi-convexity of the loss functions, the methods express HTE by fixing one parameter and optimizing the other. Additionally, it is shown that correcting bias in the traditional A-learner for single binary outcomes enables optimal treatment selection, and subgroup interpretability is enhanced using a group lasso penalty.
