The influence of categorizing survival time on parameter estimates in a Cox model

With longer follow-up times, the proportional hazards assumption is questionable in the Cox model. Cox suggested to include an interaction between a covariate and a function of time. To estimate such a function in Stata, a substantial enlargement of the data is required. This may cause severe computational problems. We will consider categorizing survival time, which raises issues as to the number of cutpoints, their position, the increased number of ties, and the loss of information, to handle this problem. Sauerbrei et al. (2007) proposed a new selection procedure to model potential time-varying effects. They investigate a large dataset (N = 2982) with 20 years follow-up, for which the Stata command stsplit creates about 2.2 million records. Categorizing the data in 6-month intervals gives 35,747 records. We will systematically investigate the influence of the length of categorization intervals and the four methods of handling ties in Stata. The results of our categorization approach are promising, showing a sensible way to handle time-varying effects even in simulation studies. References: Sauerbrei, W., Royston, P. and Look, M. (2007). A new proposal for multivariable modelling of time-varying effects in survival data based on fractional polynomial time-transformation. (Biometrical Journal, in press)