Treatment Heterogeneity with Survival Outcomes
Yizhe Xu, Nikolaos Ignatiadis, Erik Sverdrup, Scott Fleming, Stefan Wager, Nigam H. Shah
Abstract
This chapter is accompanied by survlearners, a package that provides well-documented implementations of the conditional average treatment effects (CATE) estimation strategies described in this work, to allow easy use of recommendations as well as reproduction of numerical study. It builds on the metalearners literature and provides concrete guidance for their usage in estimating treatment effect heterogeneity from RCT data with right-censored survival outcomes. Prior works have introduced metalearners for HTE estimation and provided comprehensive tutorials and simulation benchmarks that explicate their usage. Several researchers have developed machine learning based methods for estimating HTEs with data that have time-to-event outcomes. Metalearners are specific meta-algorithms that leverage predictive models to solve the causal task of estimating treatment heterogeneity. The strength of treatment heterogeneity may also influence how easy or difficult it is to estimate CATEs. An ubiquitous challenge in working with survival outcomes is the presence of right-censoring.