Inference for Survival Prediction in the High-Dimensional Setting

• Activity Number: 596
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Jennifer Sinnott*+ and Tianxi Cai
• Companies: Harvard University and Harvard University
• Keywords: high dimensional ; survival function ; prediction error ; resampling ; shrinkage ; prostate cancer

Abstract

For a risk prediction model to provide clinical utility, it is crucial that it deliver both a prediction for a new patient's risk and an honest assessment of the error inherent in that prediction. When the number of predictors is small, classical methods can capture prediction error; however, when the predictors are high dimensional, estimation of the prediction error can be challenging. In this high dimensional setting, we investigate inference on the survival function estimated using a model, such as the Cox model, under shrinkage. A shrinkage method can be chosen to give nice theoretical properties including asymptotic normality and asymptotically perfect variable selection; nevertheless, in finite samples, the estimated conditional survival distribution can be difficult to approximate using either asymptotic results, which can underestimate the variability, or the standard bootstrap, which may yield overly-conservative standard error estimates. We propose an adaptation of perturbation resampling designed to improve estimation of the error in survival prediction. We demonstrate our method in a study relating a large panel of tissue biomarkers to prostate cancer progression.