Abstract

Survival analysis has received significant attention in recent decades, particularly in estimating the cure rate - the proportion of subjects who will never experience an event. Mixtures have been the established parametric model for estimating the cure rate and associated survival curves. In this talk, we introduce a novel cure model based on phase-type distributions, which are a generalization of mixtures and rely on a Markovian latent path. The proposed model incorporates a hidden Markov jump process, enabling a proportion of subjects to be immune at inception and another proportion to become immune at later time points, resulting in an overall cure rate. In addition, we present a unified approach to regression on both the cure rate and the distribution of susceptibles. The added parameters provide the model with enhanced flexibility and interpretability compared with traditional cure models. Expectation-maximization algorithms are derived to estimate all models, which are then exemplified with synthetic and real-life data. For insurance applications, this approach allows insurers to assess risk and calculate premiums more accurately, for instance, by differentiating between those likely to recover and those with chronic conditions in health insurance. Other applications include lapse or credit risk modeling. 

Speaker

Jorge Yslas (University of Liverpool)