Thakali, L., Fu, L., & Chen, T. (2016). Comparing Crash Prediction Techniques for Ranking of Sites in a Network Screening Process. Proceedings of the 26th Canadian Association of Road Safety Professionals (CARSP) Conference, Halifax, NS, June 5-8, 2016.
Network screening, a process for an effective and efficient management of road safety programs, relies on crash prediction techniques to quantify the relative risks of given sites. The two most commonly used statistical approaches are- cross-sectional model-based approach and Empirical Bayes (EB) approach as they are known for reducing regression-to-mean bias problem of a simple crash history-based method. Meanwhile, relatively the EB approach is known to be a robust as it accounts for a site-specific risk level while still incorporating the risk estimates obtained from a cross-sectional model. Common to both the approaches is they are relatively convenient to apply and easy to interpret due to a defined mathematical equation used to relate crashes and the potential explanatory variables. However, pre-specification of such relations are challenging, as the true cause-effects are not known in advance. One approach is to use trial and error process to select for the final relation. Nonetheless, potential misspecification may still remain which consequently could result in an inaccurate list of crash hotspots in a network screening process. As an alternative to model-based approach, this study applies kernel regression (KR), which is a data-driven nonparametric method. In addition, the KR method is extended in a similar framework of EB approach to account for site-specific risk levels. All these techniques when applied in a case study using crash data from Highway 401 of Ontario, Canada showed that some deviations exist between the methods, particularly when applied in the ranking of sites in a network screening process.