Variable Bandwidth Kernel Hazard Estimators

Variable kernel hazard estimators are considered in the case, where the bandwidth is allowed to depend on the exposure. Simulations show, that when the exposure varies substantially, then this can improve the performance of the basic kernel smoother con-siderable. A two-stage approach for kernel function smoothing of hazard functions is also suggested. The pilot estimator is a fixed bandwidth kernel estimator. The second and final estimator is a variable bandwidth kernel estimator, where the bandwidthis determined by Abramson's square root law, see Abramson (1982). This approach reduces the theoretical bias compared to the fixed andwidth kernel estimator, but simulations show that the exposure dependent kernel hazard estimators have better small sample performance.