Analyzing influencing factors of crash injury severity incorporating FARS data

Abstract

The purpose of this study is to deeply analyze the influencing factors of drivers’ traffic accident casualties and reduce the occurrence of casualties. From the FARS database of the National Highway Safety Administration (NHTSA), 93248 traffic accident data were extracted as analysis samples. On this basis, the Bayesian network and multinomial logit model are established. The constructed model was tested from the perspective of model prediction accuracy and variables importance. Bayesian networks are used to analyze the interrelationships among influencing factors, and multinomial logit models are used to compare and evaluate the impact of different variables on the injury severity under different circumstances. The results show that: the prediction accuracy of the Bayesian network model and multinomial logit model is 64.57% and 65.97%, respectively. The Bayesian network reference analyses indicate that injury severity is affected by the crash factors, and there are various interactions between the various factors. The multinomial logit model analyses indicate that the factors including drivers’ age, female driver, rural roads, drunk driving, drug driving, crash time, side collision accident, etc. could significantly increase injury severity. Airbags are more effective in reducing fatal crash than injury crash. The probability of accidents caused by drug driving drivers is greater than drunk driving, drunk driving drivers are 1.79 times and 2.34 times more likely to suffer an injury severity and fatal injury severity in crashes as compared to a no injury severity, respectively, and drug driving is 1.93 times and 2.6 times, respectively. Seat belts may avoid 92.2% of fatalities. Roadside guardrail reduces the incidence of fatal crash better than injury crash. Fatal injuries severity and injury severity are 1.124 times and 1.141 times more likely to occur during the 0\u200a:\u200a00 to 6\u200a:\u200a00 as compared to no injuries, respectively, etc.