Luke Hardcastle

Piecewise Deterministic Markov Processes for transdimensional sampling from flexible Bayesian survival models

Flexible survival models have seen increasing popularity for the estimation of mean survival in the presence of a high degree of administrative censoring where survival curves need to be extrapolated beyond final observed event times. This increased flexibility, however, often introduces challenging model selection problems that have limited their wider application. In this talk I will focus on two such models, the polyhazard model and the piecewise exponential model. We introduce new prior structures that allow for the joint inference of parameters and structural quantities. Posterior sampling is achieved using bespoke MCMC schemes based on Piecewise Deterministic Markov Processes that utilise and extend existing methods for these samplers to target transdimensional posterior distributions. This is a joint work with Samuel Livingstone and Gianluca Baio.