MONDAY, November 14, 2005
Time: 1:00 - 2:00 PM
Constant Hall 2099

Title: "Smooth" Inference for Survival Functions With Arbitrarily Censored Data

Kirsten Doehler
Department of Statistics
North Carolina State University

Standard methods for inference on survival functions with right- or interval-censored data are traditionally nonparametric, and hence impose no assumptions on the true survival distribution. We propose a new procedure for estimation of survival functions that allows a unified approach to handling different kinds of censoring that is based on the premise that, if one is willing to make mild smoothness assumptions on the underlying true survival distribution, efficiency gains and computational advantages over nonparametric methods may be possible. The approach assumes that the survival distribution has a "smooth" density, which is approximated by the so-called seminonparametric (SNP) density. The SNP has a flexible "parametric" representation that admits a convenient expression for the likelihood and allows it to capture arbitrary shapes through choice of a tuning parameter, which may be carried out based on standard criteria such as AIC and BIC. We describe the approach and its implementation and validate its performance in empirical studies. Using right- and interval-censored data from popular parametric models, we compare survival curves generated from the application of the SNP density with the corresponding nonparametric survival curves. We also develop a test statistic where each survival curve is estimated from application of the SNP density.