253 – Contributed Oral Poster Presentations: Biometrics
Validating a Cox Proportional-Hazards Model
David Elashoff
University of California at Los Angeles
Tristan Grogan
University of California at Los Angeles
While several methods have been published on validating standard logistic or linear models, much less material exists on validating time-to-event models, such as the Cox proportional-hazards model. During the course of this research, an investigation of four different strategies for validating the Cox model was carried out, utilizing data from a prostate cancer study. These validation techniques are especially important for biomarker studies to aid in combating the effect of selection bias, and would strengthen the results and credibility of any study. This paper will present four performance measures for assessing whether a model has evidence for being validated or not-comparing model coefficients between training and test data sets, assessing Harrell's c-index between the training and test models, running a cross-validation technique, and comparing recurrence risk predictions between the training and test models.
