Bayesian machine learning for causal inference with multiple treatments and multilevel survival data

• Activity Number: 357 - Cross-Cutting Research in Causal Inference and Survival Analysis
• Type: Invited
• Date/Time: Thursday, August 12, 2021 : 12:00 PM to 1:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Liangyuan Hu* and Jiayi Ji and Joseph Hogan
• Companies: Icahn School of Medicine and Icahn School of Medicine and Brown University
• Keywords: Causal inference; machine learning ; sensitivity analysis ; survival outcome ; clustered data ; Bayesian analysis

Abstract

Despite numerous recent advances in causal inference, the literature for handling data with multiple treatments and multilevel censored survival outcomes is sparse. Here we develop a way to use Bayesian Additive Regression Trees, a high performance machine learning modeling technique, into a causal inference framework for clustered survival data with multiple treatments. This will provide a substantial level of modeling flexibility for a data structure for which few off-the-shelf causal inference methods are available. We further develop a flexible and interpretable sensitivity analysis framework to handle the assumption of no unmeasured confounding, respecting the multilevel survival data structure. Our approach addresses unmeasured confounding at both cluster- and individual-level and incorporates uncertainty about unidentified model components formally into the analysis. The operating characteristics of our proposed method are examined via an extensive simulation. We demonstrate the developed methods via a case study evaluating the survival effects of three popular types of treatments for high-risk localized prostate cancer using the national cancer database.