Estimating heterogeneous survival treatment effect in observational data using machine learning

• DOI: 10.1002/sim.9090
• Publication Date: 2021-06-11T07:33:08Z
• Authors: Liangyuan Hu; Jiayi Ji; Fan Li
• Keywords: Male; FOS: Computer and information sciences; Models, Statistical; Bayes Theorem; Machine Learning (stat.ML); Statistics - Applications; 01 natural sciences; 3. Good health; Causality; Machine Learning; Statistics - Machine Learning; Humans; Computer Simulation; Applications (stat.AP); 0101 mathematics

Methods for estimating heterogeneous treatment effect in observational data have largely focused on continuous or binary outcomes, and been relatively less vetted with survival outcomes. Using flexible machine learning methods the counterfactual framework is a promising approach to address challenges due complex individual characteristics, which treatments need be tailored. To evaluate operating characteristics of recent estimation heterogeneity inform better practice, we carry out comprehensive simulation study presenting wide range settings describing confounded effects varying degrees covariate overlap. Our results suggest that nonparametric Bayesian Additive Regression Trees within accelerated failure time model (AFT-BART-NP) consistently yields best performance, terms bias, precision, expected regret. Moreover, credible interval estimators from AFT-BART-NP provide close nominal frequentist coverage when overlap at least moderate. Including nonparametrically estimated propensity score as an additional fixed formulation can further improve its efficiency coverage. Finally, demonstrate application causal through case examining two radiotherapy approaches localized high-risk prostate cancer.