A Bayesian Survival Model for Time-Varying Coefficients and Unobserved Heterogeneity

Dynamic survival models are a flexible tool for overcoming limitations of popular methods in the field of survival analysis. While this flexibility allows them to uncover more intricate relationships between covariates and the time-to-event, it also has them running the risk of overfitting. This paper proposes a solution to this issue based on state of the art global-local shrinkage priors and shows that they are able to effectively regularize the amount of time-variation observed in the parameters. Further, a novel approach to accounting for unobserved heterogeneity in the data through a dynamic factor model is introduced. An efficient MCMC sampler is developed and made available in an accompanying R package. Finally, the method is applied to a current data set of survival times of patients with adenocarcinoma of the gastroesophageal junction.