In this series of papers, we have described a selection of statistical methods used for the initial analysis of survival time data (Clark et al, 2003), and introduced a selection of more advanced methods to deal with the situation where several factors impact on the survival process (Bradburn et al, 2003). The latter paper focused on proportional hazards (PH) and accelerated failure time (AFT) models, and we continue the series by demonstrating the application of these models in more detail. Whereas the focus of the previous paper was to outline the purpose and interpretation of statistical models for survival analysis, we concentrate here on approaches with which to undertake the actual modelling process. In other words, the aim of this paper is to promote the correct use of the models that have been suggested for the analysis of survival data. When used inappropriately, statistical models may give rise to misleading conclusions. Checking that a given model is an appropriate representation of the data is therefore an important step. Unfortunately, this is a complicated exercise, and one that has formed the subject of entire books. Here, we aim to present an overview of some of the major issues involved, and to provide general guidance when developing and applying a statistical model. We start by presenting approaches that can be used to ensure that the correct factors have been chosen. Following this, we describe some approaches that will help decide whether the statistical model adequately reflects the survivor patterns observed. Lastly, we describe methods to establish the validity of any assumptions the modelling process makes. We will illustrate each using the two example datasets (a lung cancer trial and an ovarian cancer dataset) that were introduced in the previous papers (Bradburn et al, 2003; Clark et al, 2003). The covariates that we consider here are fixed, that is, known at baseline or entry to the study. The handling of covariates that change values over time (e.g. white blood cell count as measured at different time points) will be described in the subsequent paper in this series. The power (and indeed in some cases validity) of a survival analysis is related to the number of events rather than the number of participants. Simulation work has suggested that at least 10 events need to be observed for each covariate considered, and anything less will lead to problems, for example, the regression coefficients become biased (Peduzzi et al, 1995). In the ovarian study, there were 550 deaths and 11 covariates for the five prognostic factors, implying 50 events per covariate. In the liver cancer trial with 114 events, a full model of 11 covariates has approximately 10 events per covariate. For prospective studies, several books (e.g. Machin et al, 1998) and software packages (e.g. nQuery, power and precision) are available to assist the calculation of adequate sample sizes, and many general purpose statistical packages also perform such calculations. This selection of scenarios is far from exhaustive, and in practice a study may combine all of the above types. The ovarian study is a combination of (b) and (c). We illustrate some straightforward approaches to the choice of covariates in the two example datasets used in this paper. In the final paper of this series, we will outline the rationale behind semiautomated methods (together with their limitations) and give further advice on hands-on modelling. In this analysis, all the covariates were included yielding the model presented in Table 2. Advanced FIGO stage, higher grade, presence of ascites and increased age all impaired survival to varying degrees. The mucinous and serous histology types had a better prognosis, and undifferentiated and mixed mesodermal a lesser one. No grade–histology interactions were included in the final model, either due to insufficient numbers of patients to allow meaningful modelling (e.g. clear cell, mixed mesodermal, adenocarcinoma or undifferentiated), or for statistical insignificance (the remainder). In fact no second-order interaction or nonlinearity was detected. If this model were to be used for the purpose of predicting future survival patterns, it is appropriate to ensure that the effect sizes are robust. One approach is to use bootstrap sampling, which involves randomly resampling the data and fitting the model to these modified datasets (Clark and Altman, 2002). These produce a series of effect sizes that should be similar to those derived from the original data if the model is sufficiently stable, and indeed do so here. Regardless of which type of model is fitted and how the variables are selected to be in the model, it is important to evaluate how well the model represents the data. A survival model is adequate if it represents the survival patterns in the data to an acceptable degree. This aspect of a model is known as goodness of fit. For example, if a given group of patients have a poor (or good) prognosis, then the model should predict this group to have that outcome. In practice, the issues in choosing the most appropriate type of model and the most appropriate covariates are heavily related, and the adequacy of a model may be assessed in several ways. In this section, we discuss methods to verify fit that are common across all survival models, before describing approaches specific to different model types. We will use the ovarian database example to demonstrate these checks. If the shape of the disease hazard is known to be different from that of a particular distribution, then the data should not be analysed with this parametric model. For example, consider the hazard for overall survival after cancer diagnosis. The hazard is rarely constant, thus ruling out an Exponential distribution. In some cases, the hazard rises sharply (due to treatment deaths) before tailing off, which would also rule out the Weibull. An informal assessment of a parametric model's appropriateness...