The Effect of Hazard Assumptions on Split Selection Criteria and Predictive Error in Survival Trees

• Activity Number: 466
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Risk Analysis
• Author(s): Brian White*+
• Companies: University of South Florida
• Keywords: Bootstrap ; Brier Score ; Data Mining ; Prediction Error ; Survival Analysis ; Survival Trees

Abstract

Many survival analysis techniques assume constant and proportional hazards although for many real data sets, these assumptions do not hold. Survival trees is one such technique commonly used to assign risks to individuals based on their personal covariates. We evaluate split selection criteria for survival trees when these assumptions are violated and propose new criteria that do not rely on such assumptions. We generate multiple simulated data sets under various scenarios to evaluate the predictive accuracy for the proposed techniques and compare them to commonly used techniques. In medical applications, a typical goal might be to define risk groups of patients at the time of treatment thus we employ our methods to a real data set from a clinical study of patients treated for breast cancer. We find that the prediction error of survival trees depends on the choice of split selection criterion as well as the measure used to estimate it. Our results are affected by the degree of censoring, nonconstancy, and nonproportionality of the hazards, thus the best choice may depend on characteristics of the data set in addition to any available a priori assumptions regarding the hazards.