Ein Bayesianisches Modell für Ereignisszeitanalyse

Two sources of heterogeneity are often overlooked in the applied survival literature. On the one hand, time-varying hazard contributions of explanatory variables cannot be captured in the widely used Cox proportional hazard model. To this end, this paper investigates a dynamic survival model within a Bayesian framework. Such a specification allows parameters to gradually evolve over time, thus accounting for time-varying effects. On the other hand, unobserved heterogeneity across (a potentially large number of) groups is often ignored, leading to invalid estimators. This paper makes accounting for such effects feasible for even large numbers of groups through a shared factor model, which picks up unexplained covariance in the error term. This paper uses the triple gamma prior in the to detect which parameters should be included in the model and which should be allowed to vary over time. Finally, an R package which makes the routine easily available is introduced.