A0156
Title: A Bayesian nonparameteric approach to competing risks
Authors: Antonio Lijoi - Bocconi University (Italy) [presenting]
Claudio Del Sole - Bocconi University (Italy)
Igor Pruenster - Bocconi University (Italy)
Abstract: Competing risks arise in several problems in survival analysis. In such a framework, data consists of survival times and an associated cause of death. A new class of priors is displayed for the transition probabilities, which are used in a multi-state modeling approach to competing risks. The proposed specification is obtained through a suitable transformation of a vector of discrete hierarchical random measures. These have been successfully applied in several areas (density estimation, clustering, prediction in species sampling problems, inference with hidden Markov models, etc.), while their uses in survival analysis are very limited. A new strategy that leads to closed-form expressions for marginal, posterior and predictive distributions is illustrated and is based on two main tools: (i) A set of latent variables that naturally arise from the data-generating distribution and can be seen as marks associated with the atoms; (ii) An identity for suitably defined moment measures. The undertaken approach allows for evaluating estimates of the (cumulative) incidence and survival functions and of the so-called prediction curve, which is related to future causes of death.
