Title: Supervised-component based Cox regression
Authors: Theo Simac - Universite de Montpellier (France)
Thomas Verron - SEITA - ITG (France)
Xavier Bry - Universite Montpellier (France) [presenting]
Abstract: In survival analysis with high-dimensional regressors, the Cox Proportional Hazard Model (CPHM) encounters instability problems owing to regressor-collinearity. Its regularisation is therefore needed. Two family of methods currently perform regularisation of regression models: penalty-based methods such as Ridge and Lasso, and component-based models such as PLS-type regression models. Only the latter enable exploratory analysis of predictive structures by making components lean on the regressors' strong correlation structures. We propose a new way to take into account the structural relevance of components within the estimation procedure of the Cox Model. Our algorithm, called SCCoxR (for Supervised-Component based Cox Regression), is tested on simulated data, and then, applied to life-history data of HIV-positive Thai subjects, in order to model the age at disclosure of their serologic status.