Modelling trends in road accident frequency - Bayesian inference for rates with uncertain exposure

Abstract

Several thousand people die as a result of a road accident each year in Great Britain and the trend in the number of fatal accidents is monitored closely to understand increases and reductions in the number of deaths. Results from analysis of these data directly influence Government road safety policy and ensure theintroduction of effective safety interventions across the country. Overall accident numbers are important, but when disaggregating into various characteristics, accident risk (defined as the number of accidents relative to an exposure measure) is a better comparator. The exposure measure used most commonly for accident rate analysis is traffic flow which can be disaggregated into vehicle types, road type, and year. Here we want to assess the accident risk across different car types and car ages, and therefore alternative exposure sources are required. We disaggregate exposure to a further extent than possible with currently available data in order to take the increased variability within these new factors into account.

Exposure data sources are mainly based on sample surveys and therefore have some associated uncertainty, however previous accident risk analysis has not, in general, taken this into account. For an explicit way to include this uncertainty we use a Bayesian analysis to combine three sources of exposure using a log-Normal model with model priors representing our uncertainty in each data source.
Using further Bayesian models, we propagate this uncertainty through to accident rates and accident severity, determining important factors and inter- relationships between factors to identify key features affecting accident trends,and we make the first exploration of the effect of the recent recession on road accidents.

More information

Identifiers

ORCID for Jonathan J. Forster: orcid.org/0000-0002-7867-3411

Catalogue record

Export record