Statistical Learning Approaches for Environmental Mixtures Studies with Survival Outcomes and Their Application to the Strong Heart Study

• Activity Number: 431 - Contributed Poster Presentations: Section on Statistics and the Environment
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Melanie Nicole Mayer* and Arce Domingo Relloso and Ana Navas-Acien and Brent Coull and Marianthi-Anna Kioumourtzoglou and Linda Valeri
• Companies: Columbia University and Columbia University and Columbia University and Harvard University and Columbia University and Columbia University
• Keywords: Survival analysis; Environmental mixtures analysis; Modeling

Abstract

Environmental mixtures studies consist of multiple continuous and correlated exposures. We compared statistical learning methods in terms of bias and efficiency when evaluating individual exposure and complete mixture effects on survival outcomes. We considered Cox proportional hazards (PH) with and without penalized splines and Cox Elastic Net. Additionally, we applied discrete-time survival analysis approach to Gaussian Process Regression (GPR), Multivariate Adaptive Regression Splines (MARS), and Bayesian Additive Regression Trees (BART). We conducted simulations under several real-world scenarios and evaluated the effect of mixtures on CVD in the Strong Heart Study (SHS) cohort. In simulations where PH assumption held, BART and MARS’s estimates of the hazard ratio (HR) for an IQR change in the mixture had high bias and variance, resulting in higher RMSEs (0.3-5.2) compared to the other methods (0.1-0.9). When it was violated, RMSEs were comparable to other methods (0.2- 0.3), other than GPR (0.1), while achieving higher coverage (>76%). In the SHS analysis, all methods found significant effects. However, MARS, BART and GPR estimated larger HRs with wider confidence intervals.