Accident prediction models for signalized intersections

• Main Author: Feng, Shirley
• Language: English
• Published: 2009
• Online Access: http://hdl.handle.net/2429/7851

This thesis describes the development of accident prediction models for signalized intersections in the Greater Vancouver Regional District (GVRD). The traffic and road-related factors which appeared to underlie the occurrence of accidents are examined and models which explain, in a statistical sense, the generation of accidents as a function of these factors are developed. Recognizing the statistical and practical shortcomings associated with the use of the Conventional Linear Regression approach to develop accident prediction models, it was decided to utilize the Generalized Linear Regression Models (GLIM) approach. This approach addresses and overcomes the error structure problems that are associated with the conventional linear regression theory and allows for the use of nonlinear relationships in the model. In addition, the safety predictions obtained from the GLIM models can be refined using the Empirical Bayes' approach to provide, more accurate, site-specific safety estimates. The use of the complementary Empirical Bayes approach can significantly reduce the regression to the mean bias that are inherent in observed accident counts. The study made use of sample accident, traffic and intersection design data corresponding to signalized intersections located in the Greater Vancouver Region. The accident data set contained 67 urban intersections from the City of Richmond and 72 urban intersections from the City of Vancouver giving a total of 139 intersections. Three different types of models were developed: (1) models relating the total number of accidents to traffic volume; (2) models relating accidents of a specific type to traffic volume; and (3) models incorporating other geometric design variables such as the existence of left turn lanes, right turn lanes, pedestrian crossings, etc. The goodness of fit of the models was evaluated using two statistics: the Scaled Deviance (SD) and the Pearson x2 statistics. The overall fit of the models was adequate. Three applications of the GLIM models and the Empirical Bayes refinement process were described. The first related to the identification of accident prone locations. The second related to the before and after safety analysis and the third to safety planning. The usefulness of the GLIM model estimates in accounting for the randomness inherent in the accident occurrence process and the regression to the mean bias was documented and discussed.