Bayesian Nonparametric Inference on Restricted Mean Survival Time with Adjustments for Covariates

• Activity Number: 208 - SPEED: Bayesian Methods and Social Statistics Part 2
• Type: Contributed
• Date/Time: Monday, August 8, 2022 : 2:00 PM to 2:45 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Ruizhe Chen* and Sanjib Basu and Qian Shi
• Companies: University of Illinois Chicago and Biostatistics, University of Illinois Chicago and Mayo Clinic
• Keywords: Dirichlet Process ; Restricted Mean Survival Times; Covariates Adjustment; Survival Analysis; Bayesian nonparametrics; Stick-Breaking Priors

Abstract

In survival analysis, Restricted mean survival time (RMST) is a function defined as the mean survival time, for either a subject or an entire study population, measured from randomization to a time point of clinical interest tau. The RMST function is easy to interpret and free of the proportional hazards assumption compared to the hazard ratio measure under a Cox model. We have developed a Bayesian nonparametric framework to model RMST functions as a mixture of covariates-dependent kernel RMST functions with covariates-dependent mixture weights by assigning either logistic or probit stick breaking priors on the parameter space. Subject-level RMST function is modeled as predictor-dependent infinite mixtures of Weibull or Gamma distributions, which allow an analytic form of the RMST function. We compare the performance of the proposed Bayesian non-parametric approach with an existing Frequentist approach that estimates RMST with adjustment for covariates. We evaluate their performances for estimating both subject-level RMSTs and average treatment effect measured by average conditional RMST difference between two treatment groups under a RCT setting.