CMStatistics 2015: Start Registration
View Submission - CMStatistics
B0154
Title: Boosting multi-state models Authors:  Thomas Kneib - University of Goettingen (Germany) [presenting]
Holger Reulen - University of Goettingen (Germany)
Abstract: Multi-state models describe the evolution of discrete phenomena (such as the health state of individuals) in continuous time with the goal of analysing the transition intensities between the different states. A common model specification relies on a Cox-type structure with unspecified hazard rate and multiplicative effects of the covariates. We present a functional gradient descent boosting approach that allows us to implement model choice and variable selection in an automated fashion in multi-state models. The approach relies on a stratified Cox model representation that has the particular advantage to allow for the inclusion of cross-transition effects, i.e. effects that are common to multiple transition types. The automatic variable selection framework offered by functional gradient descent boosting now also allows to automatically detect which effects can be combined across the transition types.