Penalized Cox Regression Models of Sparse Outcomes

• Activity Number: 452
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): G. Kolm*+ and Pan Wu and Claudine Jurkovitz and Pranav Kansara
• Companies: Christiana Care Health System and Christiana Care Health System and Christiana Care Health System and Christiana Care Health System
• Keywords: Penalized regression ; sparse outcomes ; prediction model

Abstract

A number of outcomes of interest in medical research occur infrequently in a given population (e.g., sudden cardiac death in hypertrophic cardiomyopathy patients). When there is interest in building a predictive model of the outcome and the number of potential risk factors is large, commonly used regression methods may not result in valid models. Suggested numbers of patients per event necessary for valid models range from 10 - 20 (Harrell, Jr., 2001; Peduzzi, et. al., 1995). In this study, we apply several penalty methods for time-to-event data from a sample of coronary artery disease patients undergoing percutaneous coronary intervention. The event rate in the sample is 1.55% with 20+ risk factors of interest to clinicians. We apply penalized Cox regression models for time from intervention to stent thrombosis.