A0717
Title: Bayesian mixture of accelerated-failure-time experts model
Authors: Elham Mirfarah - University of St Andrews (United Kingdom) [presenting]
Abstract: Exploring the relationship between survival time and the covariates of interest is a challenging topic in biomedical studies. The classical accelerated failure time (AFT) model is often more flexible, powerful, and interpretable than the Cox proportional hazards model if the underlying assumptions (distributional and homoscedasticity) are met. However, real-world data often exhibit heteroscedasticity, which compromises the robustness of the classical AFT model. A parametric approach to addressing non-homogeneous datasets is the mixture-of-experts (MoE) models. The MoE is an extension of the finite mixture model wherein the mixing proportion varies for each observation. A Bayesian analysis is presented for censored survival time data, employing a broad class of distributions (scale mixture of normal) for the error term in the AFT model. A weakly informative prior structure is proposed for the parameters, and the corresponding posterior distributions are demonstrated to be proper. By leveraging the Ultimate Polya-Gamma data-augmentation method, gating parameters are efficiently sampled, and cluster memberships are allocated for data subgroups. The effectiveness of the proposal is illustrated through synthetic studies and a real data example.
