Efficient Bayesian nonparametric hazard regression

We model the log-cumulative baseline hazard for the Cox model via Bayesian, monotonic P-splines. This approach permits fast computation, accounting for arbitrary censorship and the inclusion of nonparametric effects. We leverage the computational efficiency to simplify effect interpretation for metric and non-metric variables by combining the restricted mean survival time approach with partial dependence plots. This allows effect interpretation in terms of survival times. Monte Carlo simulations indicate that the proposed methods work well. We illustrate our approach using a large data set of real estate data advertisements.