Optimal stratification of survival data via Bayesian nonparametric mixtures

TL;DR

This paper introduces a Bayesian nonparametric framework for optimal data stratification in survival analysis, allowing flexible, data-driven partitioning that accounts for heterogeneity in baseline survival and predictor effects.

Contribution

It extends stratified survival models by using Bayesian nonparametrics to determine optimal strata based on the data, rather than predefined categories.

Findings

Method performs well in simulations with censored data.

Robust to right-censoring effects.

Applied successfully to AIDS research data.

Abstract

The stratified proportional hazards model represents a simple solution to account for heterogeneity within the data while keeping the multiplicative effect on the hazard function. Strata are typically defined a priori by resorting to the values taken by a categorical covariate. A general framework is proposed, which allows for the stratification of a generic accelerated life time model, including as a special case the Weibull proportional hazard model. The stratification is determined a posteriori by taking into account that strata might be characterized by different baseline survivals as well as different effects of the predictors. This is achieved by considering a Bayesian nonparametric mixture model and the posterior distribution it induces on the space of data partitions. The optimal stratification is then identified by means of the variation of information criterion and, in turn, stratum-specific inference is carried out. The performance of the proposed method and its robustness to the presence of right-censored observations are investigated by means of an extensive simulation study. A further illustration is provided by the analysis of a data set extracted from the University of Massachusetts AIDS Research Unit IMPACT Study.