You are here


Primary tabs

Bachir A. FAWAZ



University of Liverpool


Civil Engineering/




# Pages






The observed “migration” of traffic accidents from treated sites to their neighbours has been the subject of considerable controversy in recent years.  Some authors have argued that the observed migration represents a real change in the underlying mean accident frequencies at the sites involved, while others have argued that it is merely a statistical illusion arising, like regression-to-mean, due to bias by selection.  The problem remains unresolved essentially because of the lack of a sufficiently detailed and extensive set of data against which the alternative theories can be rigourously tested.

The main objective of the research described in this thesis was to resolve the controversy by establishing a suitable data set and identifying a method by which changes in underlying mean accident frequencies could be reliably determined.

Detailed accident, traffic and treatment data were obtained for a network of some 280km of highway over an 11 year study period.  A particular feature of this data was that all engineering measures applied to the network were included, not just accident remedial measures. This ensured that the effects of individual schemes could be isolated and that data could be obtained for sites unaffected by treatment (at which any apparent migration of accidents would be attributable to statistical effects).  Variants of two methods were developed to estimate underlying mean accident frequencies: an Empirical Bayes approach, which gives an estimate based on the observed accident frequencies at the study site and the mean accident frequency at similar sites; and a regression approach in which a regression equation is fitted to accident frequencies at untreated sites for two separate time periods.

Both methods were shown to be capable of providing reasonably reliable estimates for a group of sites.  It is demonstrated that an illusion of accident migration can arise due to regression-to-mean effects, but there was some evidence to suggest that a genuine migration of “accident risk” does occur in some circumstances.  Problems of “overlap” of treated sites reduced the sample size in the curent study to some 200 treated sites.  Assuming that an average increase of 0.5 accidents in 3 years over the 500m sections of the network immediately adjacent to the treated site would constitute a migration effect of practical significance, it was concluded that a ten fold increase in the sample size would be required to ensure that an increase of this magnitude would be statistically significant at the five percent level.