Calibration and Transferability of Accident Prediction Models for Urban Intersections

Accident prediction models, also known as safety performance functions, have several important uses in modern day safety analysis. Unfortunately, calibration of these models is not straightforward. This paper documents a research effort that demonstrates the complexity of calibrating these models for urban intersections. These complexities relate to the specification of the functional form, the accommodation of the peculiarities of accident data, and to the transferability of models to other jurisdictions. Toronto data are used to estimate models for 3and 4-legged signalized and unsignalized intersections. Then the performance of these models is compared to models for Vancouver and California that were recalibrated for Toronto using a procedure recently proposed for the application in the Interactive Highway Safety Design Model (IHSDM). The results of this transferability test are mixed, suggesting that a single calibration factor as is currently specified in the IHSDM procedure may be inappropriate and that a disaggregation by traffic volume might be preferable.