Identifying and modelling prognostic factors with censored data

Abstract

A major issue in the analysis of diseases is the identification and assessment of prognostic factors relevant to the development of the illness. Statistical analyses within the proportional hazards framework suffer from a lack of flexibility due to stringent model assumptions such as additivity and time-constancy of effects. In this paper we use tree-based models and varying coefficient models to allow for detectability of prognostic factors with possibly non-additive, non-linear and time-varying impact on disease development. Questions concerning model and smoothing parameter selection are addressed. An analysis of a data set of breast cancer patients demonstrates the ability of these methods to reveal additional insight into the disease influencing mechanisms.