M5 model tree based predictive modeling of road accidents on non-urban sections of highways in India

• Published in: Accident analysis and prevention 2016-11, Vol.96, p.108-117
• Main Authors: Singh, Gyanendra, Sachdeva, S.N., Pal, Mahesh
• Format: Article
• Language: English
• Subjects: Accidents; Accidents, Traffic - statistics & numerical data; Data points; Environment Design; Highways; Humans; India; Linear Models; M5 model tree; Mathematical models; Modelling; Models, Theoretical; Motor Vehicles - statistics & numerical data; Predictive modeling; Random effect negative binomial model; Regression; Risk Assessment; Road safety; Roads; Rural Population; Safety - statistics & numerical data; Trees
• ISSN: 0001-4575; 1879-2057
• DOI: 10.1016/j.aap.2016.08.004

•The potential of M5 model tree and Random effect Negative Binomial model for accident prediction on highway was examined.•M5 model tree and RENB models perform equally well.•M5 model tree provides simple linear equations at terminal nodes to predict accidents.•To make roads safer, proper design of minor accesses and service roads and reduction in dispersion of speeds is desired. This work examines the application of M5 model tree and conventionally used fixed/random effect negative binomial (FENB/RENB) regression models for accident prediction on non-urban sections of highway in Haryana (India). Road accident data for a period of 2–6 years on different sections of 8 National and State Highways in Haryana was collected from police records. Data related to road geometry, traffic and road environment related variables was collected through field studies. Total two hundred and twenty two data points were gathered by dividing highways into sections with certain uniform geometric characteristics. For prediction of accident frequencies using fifteen input parameters, two modeling approaches: FENB/RENB regression and M5 model tree were used. Results suggest that both models perform comparably well in terms of correlation coefficient and root mean square error values. M5 model tree provides simple linear equations that are easy to interpret and provide better insight, indicating that this approach can effectively be used as an alternative to RENB approach if the sole purpose is to predict motor vehicle crashes. Sensitivity analysis using M5 model tree also suggests that its results reflect the physical conditions. Both models clearly indicate that to improve safety on Indian highways minor accesses to the highways need to be properly designed and controlled, the service roads to be made functional and dispersion of speeds is to be brought down.