Nonparametric Survival Analysis with Dirichlet Processes Mixtures and Heteroskedastic Bayesian Additive Regression Trees

• Activity Number: 20 - Bayesian Additive Regression Trees: Making an Impact
• Type: Topic Contributed
• Date/Time: Sunday, July 28, 2019 : 2:00 PM to 3:50 PM
• Sponsor: International Society for Bayesian Analysis (ISBA)
• Author(s): Rodney Sparapani* and Robert McCulloch and Matthew Pratola and Brent R. Logan and Prakash Laud
• Companies: Medical College of Wisconsin and Arizona State University and and Medical College of Wisconsin and Medical College of Wisconsin
• Keywords: black-box models; ensemble predictive modeling; machine learning

Abstract

Bayesian Additive Regression Trees (BART) is a nonparametric machine learning method for continuous, dichotomous, categorical and time-to-event outcomes. However, survival analysis with BART currently presents some challenges. Two current approaches each have their pros and cons. Our discrete time approach is free of precarious restrictive assumptions such as proportional hazards and Accelerated Failure Time (AFT), but it becomes increasingly computationally demanding as the sample size increases. Alternatively, a Dirichlet Process Mixture approach is computationally friendly, but it suffers from the AFT assumption. Therefore, we propose to further nonparametrically enhance this latter approach via heteroskedastic BART which will remove the restrictive AFT assumption while maintaining its desirable computational properties.