Abstract
Although crashes involving hazardous materials (HAZMAT) are rare events compared with other types of traffic crashes, they often cause tremendous loss of life and property, as well as severe hazards to the environment and public safety. Using five-year (2013-2017) crash data (N = 1610) from the Highway Safety Information System database, a two-step machine learning-based approach was proposed to investigate and quantify the statistical relationship between three HAZMAT crash severity outcomes (fatal and severe injury, injury, and no injury) and contributing factors, including the driver, road, vehicle, crash, and environmental characteristics. Random forest ranked the importance of risk factors, and then Bayesian networks were developed to provide probabilistic inference. The results show that fatal and severe HAZMAT crashes are closely associated with younger drivers (age less than 25), driver fatigue, violation, distraction, special roadway locations (such as intersections, ramps, and bridges), higher speed limits (over 66 mph), midnight and early morning (12:00-5:59 a.m.), head-on crashes, more than four vehicles, fire/explosion/spill, poor lighting conditions, and adverse weather conditions. The overall prediction accuracy of 85.8% suggests the effectiveness of the proposed method. This study extends machine learning applications in a HAZMAT crash analysis, which would help develop targeted countermeasures and strategies to improve HAZMAT road transportation safety.