Abstract
Cubic spline smoothing of hazard rate functions is evaluated through a simulation study. The smoothing algorithm requires unsmoothed time-point estimates of a hazard rate, variances of the estimators, and a smoothing parameter. Two unsmoothed estimators were compared (Kaplan-Meier and Nelson based) as well as variations in the number of time-point estimates input to the algorithm. A cross-validated likelihood approach automated the selection of the smoothing parameter and the number of time-point estimates. The results indicated that, for a simple hazard shape, a wide range of smoothing parameter values and number of time-points will yield mean squared errors not much larger than parametric maximum likelihood estimators. However, for peaked hazards, it seems advisable to use the cross-validated likelihood approach in order to avoid oversmoothing.

This publication has 19 references indexed in Scilit: