B1291
Title: An empirical comparison between gradient boosting methods and Cox's PH model for right-censored survival data
Authors: Yingwei Peng - Queen\'s University (Canada) [presenting]
Peizhi Li - Dongbei University of Finance and Economics (China)
Abstract: Gradient boosting methods become popular in recent years to analyze right-censored survival data, where Cox's proportional hazards model is widely used in statistical models. However, there are limited studies on the differences between the two approaches for right-censored survival data. The aim is to compare two boosting methods with Cox's proportional hazards model: the gradient boosting decision tree and the gradient boosting with component-wise linear models. The differences are discussed between the two boosting methods and a simulation study investigating the performance of the three methods in practice where only the main effects of covariates are included. The results show that the boosting methods outperform Cox's proportional hazards model in both the relative and absolute risk estimation in the proportional hazards model except when Cox's proportional hazards model is fully specified with nonlinear and interaction covariates effects. It indicates that the boosting methods, particularly the gradient boosting decision tree, are very competitive for right-censored survival data if complicated covariate effects exist but are unknown to the investigator. The application of the boosting methods with real data analysis is further illustrated.