Accident prediction model development: signalized intersections

An intuitive methodology of developing injury, property-damage only (PDO) and fatal accident models for signalized intersections based on the Traffic Accident Surveillance and Analysis System (TASAS) in California is illustrated. A fairly new grouping and classifying technique called Classification and Regression Trees (CART) was used as a building block for developing prediction models. The proposed methodology includes a 3-Level prediction procedure with a "tree" structure for easy interpretation and applications. It also includes an adjustment procedure for different reporting levels of PDO accidents in different police jurisdictions. Applications of accident prediction models are also discussed in detail. Macroscopic-type models for injury, and PDO accidents per year were derived, and the following factors were found to be significant: traffic intensity, proportion of cross street traffic, intersection type, signal type, number of lanes, and left-turn arrangements. Relevant factors for fatal accidents are traffic intensity, intersection type, and design speed. Based on the results, it is also apparent that the models derived from the proposed methodology and TASAS provide more intuition and flexibility than the existing models used in California and some other models derived from site observations and accident record systems (A).