Model Selection and Inference in Developing a Predictive Model for Survival

• Activity Number: 510
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Hongying Li*+ and Minya Pu and Lei Bao and Loki Natarajan and Karen Messer
• Companies: University of California, San Diego and University of California, San Diego and University of California, San Diego and University of California, San Diego and University of California, San Diego
• Keywords: Model Selection ; LASSO ; Stability Selection ; Prediction Error ; Bootstrap Smoothing

Abstract

It is of great clinic interest to develop statistical models to predict overall survival or recurrence risk in cancer patients after diagnosis as this can facilitate treatment decisions. To do this we often collect many relevant predictor variables including clinical variables, genomic measures and laboratory values in addition to the overall survival or recurrence follow-up measure. We then then fit a cox regression model based on a subset of selected relevant variables. Many statistical methods have been developed for such variable selection and model development. In this talk, we will review the model selection framework and particularly discuss a stability selection procedure based on LASSO and compare it with some other methods such as the original LASSO using AIC or BIC, cross validation methods, and Bayesian model averaging. We will also discuss the inference after model selection based on a smoothed bootstrap method. We will demonstrate this utilizing a large cohort of cancer patient data from real life.