Statistical Issues in Development of a Predictive Model for Survival in Chronic Lymphocytic Leukemia

• Activity Number: 91
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Minya Pu*+ and Hongying Li and Lei Bao and Loki Natarajan and Laura Rassenti and Thomas Kipps and Karen Messer
• Companies: University of California, San Diego, Moores UCSD Cancer Center and University of California, San Diego, Moores UCSD Cancer Center and University of California, San Diego, Moores UCSD Cancer Center and University of California San Diego and University of California, San Diego, Moores UCSD Cancer Center and University of California, San Diego, Moores UCSD Cancer Center and University of California San Diego
• Keywords: Model selection ; model comparison ; Bayesian model averaging ; prediction error ; integrated brier score ; LASSO

Abstract

Survival time after diagnosis for CLL patients varies tremendously, so its prediction is of great clinical interest. Our goal was to develop a robust prognostic model using data from 2886 patients in the large CLL Research Consortium data base. A moderately large number of predictors were available, including genomic measures and laboratory values. We first randomly split the patients into training and validation data sets, and built the model using training data only. We considered several model section techniques, including LASSO with cross-validated partial likelihood to obtain the best regularization parameter. However, we noticed that the order of variable entry into a model was very sensitive to data set changes, so we sought to summarize this variability using 1000 bootstrapped samples. Our final analysis used Bayesian model averaging, which searches over the whole model space. Based on BIC, 15 best Cox models were produced and were compared using an Efron-Type 0.632 prediction error based on the integrated brier scores. Our final model has smaller prediction error than using other popular methods such as random survival forests.