Relative survival: comparison of regressive models and advice for the user.

BACKGROUND Relative survival is a method of analysis of failure-time data used to estimate the net survival. Cancer registries frequently use this method. The main regressive models are the Hakulinen and Tenkanen model, and the Esteve et al. model, which are easily used in practice thanks to their specific software (SURV and RELSURV, respectively). An assessment of the behaviour of the models is made, with the aim of giving advice for users of lifetime data in practice. METHODS Simulations were done by respecting, then violating, the basic hypothesis supporting the theoretical foundation of these two proportional hazard models (independence of the death and censor process, proportionality of risks). For each simulation, 100 files of either 100, 1,000, or 10,000 individuals were generated to assess the behaviour of the model. RESULTS Moderate censor rates, with or without proportionality assumption, lead to the use of the Hakulinen and Tenkanen model, especially for studies with little information. Non-proportionality of risks in the Hakulinen and Tenkanen model could be tested and analysed. If assumptions underlying the models are respected, the Esteve et al. model seems to be more precise. DISCUSSION The choice of a model in practice depends on its performance, and on the user's knowledge of statistics and computer science. Non-proportionality of risks is common in cancer registries. In theory, non-proportionality of risks could be taken into account for both relative survival models but, for the moment, it is feasible in routine only for the Hakulinen and Tenkanen model. Characteristics of the software should also be taken into account for routine relative survival analyses.