A Nonparametric Method for Predicting Survival Probabilities

Public programs often use statistical profiling to assess the risk that applicants will become long-term dependent on the program. The literature uses linear probability models and (Cox) proportional hazard models to predict duration outcomes. These either focus on one threshold duration or impose proportionality. In this paper we propose a nonparametric weighted survivor prediction method where the weights depend on the distance in characteristics between individuals. A simulation study shows that an Epanechnikov weighting function with a small bandwidth gives the best predictions while the choice of distance function is less important for the performance of the weighted survivor prediction method. This yields predictions that are slightly better than Cox survivor function predictions. In an empirical application concerning the outflow to work from unemployment insu rance, we do not find that the weighting method outperforms Cox survivor function predictions.