Freeway Accident Prediction Using Quantile Regression

With the rapid growth in highway mileage and the traffic volume, traffic accident becomes more and more serious and brings graveness consequences to human life and wealth. Therefore, in-depth analysis and forecasting traffic accidents, which has important practical implications for road safety evaluation, planning and decision making. For this purpose, this paper proposed freeway traffic accident prediction method using Quantile Regression.In contrast, in the study of accident prediction method, China started late. Especially in the regression model, there is a certain gap with abroad both the depth and breadth of the research, and need to be further deepened. Firstly, analyses the common characteristics in traffic accident data (Over-dispersion, Under-dispersion, Low sample-mean and small sample size) and the problems (Under-reporting, Time-varying variables, Endogenous variables, Omitted-variables) need to consider. The basic principle and the advantages and disadvantages of regression models, neural networks, time series, and gray prediction method were introduced. It provided a reference for the selecting of accident prediction method. Next, from the perspective of road environment factors, a detailed investigation of the road alignment (Horizontal Alignment, Vertical Alignment and the combination), special sections (bridges, tunnels) and traffic environment (traffic volume, traffic composition, operating speed) impact on traffic safety was made, in order to determine the accident prediction variables.Finally, based on the basic principles of quantile regression, this paper presented the method for quantile regression application in the count data, and proposed two indicators (marginal effect, sensitivity) to explain the variables. Based on the series of parameters estimation of quantile regression, this paper proposed two accidents prediction method:Historical data and probabilistic methods, and compared with the negative binomial regression method.The numerical case study shows the superiority of quantile regression in traffic accident prediction. In the process of analysis, quantile regression is more comprehensive for parameters, marginal effect and sensitivity compared to the negative binomial regression. In addition, in the accidents prediction, the accuracy of the historical data method up to 74%, RMSE is 0.618. Two prediction methods based on quantile regression is better than negative binomial regression model in the accuracy and error.