The Super Learner for Estimating Nonlinear Associations in the Cox Regression Model

• Activity Number: 377
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Elizabeth Malloy*+ and Philip Gautier and Cynthia Cook and Melissa K. Bergeron
• Companies: American University and Purdue University and American University and Freddie Mac
• Keywords: Cox regression ; cross validation ; machine linearning ; smoothing parameter ; smoothing splines ; super learner

Abstract

We demonstrates the use of the super learner algorithm to estimate a nonlinear exposure-response relationship based on penalized splines in the Cox regression model. Two types of super learners are examined. A discrete super learner, constructed by choosing amongst a set of candidate models, defined to be the class of penalized splines over different levels of the smoothing parameter, and a continuous super learner, constructed by combining estimates across the candidates. For the discrete super learner we compare different loss functions to choose the optimal model from the candidates, with two cross-validated partial-likelihood-based loss functions demonstrating lower bias in a simulation study. Unconstrained and constrained continuous super learners were also compared in the simulations with a Lasso constraint generally demonstrating better properties in the simulations.