Modeling frailty as a function of observed covariates

In survival analysis, frailty models are potential choices for modeling unexplained heterogeneity in a population, which exists due to missing covariate information or to differential survival patterns among members of a population. Typically, in these models, the frailty term, which is a random effect, is unconditional on the observed covariates. In our model, we allow the frailty effect to be modulated by the observed covariates. In this way, the frailty effect is no longer rendered separate from the covariates, allowing the model to capture the frailty effect as function of unobserved as well as observed information. We demonstrate this model on a set of subjects in the Framingham Heart Study who had atrial fibrillation events and who were followed forward in time for the development of stroke. As assessed via performance measures, our model performs better on this data than the other models considered. It also captures unique hazard configurations not produced by the other models.