Title: Cox regression with Potts-driven latent clusters
Authors: Alejandro Murua - University of Montreal (Canada) [presenting]
Danae Martinez-Vargas - Universite de Montreal (Canada)
Abstract: A Bayesian nonparametric survival regression model with latent partitions is considered. The goal is to predict survival, and to cluster survival patients within the context of building prognosis systems. We propose the Potts clustering model as a prior on the covariates space so as to drive cluster formation on individuals and/or Tumor-Node-Metastasis stage system patient blocks. For any given partition, our model assumes an interval-wise Weibull distribution for the baseline hazard rate. The number of intervals is unknown. It is estimated with a lasso-type penalty given by a sequential double exponential prior. Estimation and inference are done with the aid of MCMC. To simplify the computations, we use the Laplace's approximation method to estimate some constants, and to propose parameter updates within MCMC. We illustrate the methodology with an application to cancer survival.