Are estimates of crash modification factors mis-specified?

Transportation planners and traffic engineers are using crash modification factors to evaluate how changes in road geometry and design features can reduce crashes. Crash modification factors are typically estimated based on segmenting links on a highway and associating with geometric features. This allows statistical methods to be applied to the data. Concurrently there is a stream of research that relies on spatial units of analysis to examine crashes; these typically use broad features of the road network combined with socio-economic and demographic factors that are associated with crashes. In this paper, we examine whether omission of these spatial factors in a link-based model results in mis-specified models, in particular, omitted variable bias. Our results suggest that there is no change in coefficient signs, but that there is a reduction in the magnitude of estimates, suggesting that omitted variable bias exists. The sign of spatial variables differ substantially when combined with a link-based model. We also find substantial variability in coefficient estimates, and discuss the implications of these results for the use of crash modification factors in cost-benefit analysis of road safety projects.