Kernel Machine Collapsing--Based Prediction for Ordinal Outcome

• Activity Number: 483
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Yuanyuan Shen*+ and Tianxi Cai
• Companies: Harvard University and Harvard University
• Keywords: ordinal outcome ; kernel machine ; collapsing

Abstract

Ordinal outcome data play an important role in medical research, since they usually contain more specified information about the measured outcome. Current methodologies for analyzing ordinal outcome data include creating dichotomies among the levels of the response to apply standard procedures for binary outcome, and using classical ordinal response models such as continuation-ratio logit model. These methods suffer from either the loss of information due to collapsing, or the difficulty in deciding between proportional model and full model. In additional to those challenges, many predictors may relate to the outcome non-linearly. Based on the continuation-ratio full model, we propose a kernel-based grouping method to determine statistically whether we should collapse neighboring regression models. Our method extends the idea of group lasso for ordinal outcome data to build up prediction model. The use of kernel transformation adequately accounts for the non-linearity. Simulation results show that our method has better prediction accuracy comparing to proportional model or full model under many cases, and well handles the non-linearity among variables.