Change and safety: decision-making from data

The UK rail industry collects and analyses incident data to assess the risks experienced by passengers, the railway workforce and members of the public. The current analysis mainly compares current and historic incident rates and, for each type of incident, looks at the railway as a whole. However, changes to reduce risk are often made locally, at least at first, so a modelling approach is needed that is able to analyse local risk. In this paper we present a form of model that is able to make local risk estimates from incident data, using a case study - boarding and alighting incidents at stations. Using a Bayesian network (BN), we analyse the incident data with expert judgements about causal factors. The BN cannot be directly leant from data because the dataset contains no entries for the overwhelming majority of cases where the railway is used without incident. Instead, the incident data is analysed in the context of a model of the use of the railway, so that the prevalence of the causal factors in normal use can be compared with that in the incident data. The usage model is derived from multiple data sets; the use of a BN allows this to be done with approximations. Finally, we show how the model allows local risk to be estimated and consider the wider applicability of the techniques described.