The use of multilevel models for the prediction of road accident outcomes

Abstract

An important problem in road traffic accident research is the resolution of the magnitude by which individual accident characteristics affect the risk of fatality for each person involved. This article introduces the potential of a recently developed form of regression models, known as multilevel models, for quantifying the various influences on casualty outcomes. The application of multilevel models is illustrated by the analysis of the predictors of outcome amongst over 16,000 fatally and seriously injured casualties involved in accidents between 1985 and 1996 in Norway. Risk of fatality was found to be associated with casualty age and sex, as well as the type of vehicles involved, the characteristics of the impact, the attributes of the road section on which it took place, the time of day, and whether alcohol was suspected. After accounting for these factors, the multilevel analysis showed that 16% of unexplained variation in casualty outcomes was between accidents, whilst ∼1% was associated with the area of Norway in which each incident occurred. The benefits of using multilevel models to analyse accident data are discussed along with the limitations of traditional regression modelling approaches.