Poisson and negative binomial regression techniques were used as a means to predict accidents on the basis of highway geometrics and traffic related factors. For a particular highway section the overall number of accidents was predicted using both Poisson and negative binomial distributions. The predictions were then compared with actual accident statistics. Both methods use a log-linear function to ensure that accident prediction is always non-negative. The primary data sources were the Washington State Department of Transportation's Transportation and Planning Support database for geometric and traffic information and the Washington State Patrol's accident database for accident information. The results suggested that horizontal curvature, daily traffic, speed, number of lanes, and tangent length between curves are significantly correlated with accident occurrence. The results indicated that if accident data are dispersed relative to the mean, negative binomial regression is the most appropriate method of analysis.
Washington State Transportation Center (TRAC)
Accidents, Alternatives analysis, Annual average daily traffic, Curvature, Forecasting, Geometric design, Mathematical prediction, Poisson distributions, Regression analysis, Statistics, Traffic accidents, Traffic data, Traffic lanes, Traffic speed.