gms | German Medical Science

Introduction: The Cox proportional hazards (PH) model is the standard tool for the analysis of survival time data. However, some effects may vary in time, which means that the PH assumption is violated, leading to poor model fit and wrong conclusions.

Subject and methods: To model non-proportional hazards a new procedure, MFPT, has been proposed (Sauerbrei et al. 2006). It is an extension of the multivariable fractional polynomial (MFP) approach combining selection of influential variables, determination of a sensible functional relationship for continuous variables and modelling of time-varying effects.

To investigate time-varying effects, we will discuss central issues of several methods and compare results obtained by MFPT to those received from alternative approaches. By analysing a specific data set we will consider reduced rank models (Perperoglou et al. 2006; Stat Med 25:2831-45), an approach based on fractional polynomials (Berger et al. 2003; Stat Med 22:1163-80), a semiparametric approach based on cumulative regression functions (Scheike and Martinussen 2004; Scand J Stat 31:51-62) and parametric models based on natural cubic splines (Royston and Parmar 2002; Stat Med 21:2175-97). For none of the approaches sufficient knowledge about properties exists.

Results: In the Rotterdam breast cancer series (N=2982) with 20 years follow-up and ten potential explanatory variables, the MFPT approach selects a model with eight variables, of which two have a time-varying effect. The alternative approaches require to pre-specify a 'baseline PH' model, in which time-varying effects can be investigated. There are several differences of results between the approaches.

Discussion: The approaches are very different in selecting a multivariable model with time-varying effects. In our data set they select different models. Investigations of their properties and comparisons are required.