Modeling pedestrian crash severity in St. Louis using random forests and mixed logit techniques

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI--COLUMBIA AT REQUEST OF AUTHOR.] The pedestrian-vehicle crashes in the City of St. Louis is a major concern due to the high magnitude of injuries and fatalities compared to Missouri's statewide collisions. Thus, there is a need to identify and prioritize the correlates of injury severities both in the collision data and the neighborhood environment surrounding every crash location. This research: (1) employed a comprehensive data collection effort to link the neighborhood environment with the crash data using the ArcGIS package; (2) gathered the built environment data inside three separate buffers with radius of 500 ft, 1000 ft, and 1500 ft from the collision; (3) screened the most significant variables using random forests and stepwise logistic regression; and (4) modeled three binary mixed logit models for the aforementioned buffers and a single model for the police data. The goodness of fit for both training and validations sets suggest that the 500 ft is superior compared to the other models. The results elucidate that the evening off-peak time crash is the most significant random variable which increases the odds of severe crashes (disabling and fatal) by 28.023%. The rest of the variables follow the order as follows: network connectivity (21.241%); speed involvement (12.767%); morning peak time (7.963%); elderly pedestrians beyond 65 years old (6.527%); old pedestrians from 45 to 65 years old (6.342%); male drivers ( 4.232%); and AADT (3.529%) . This research carries out the first effort to identify the correlates of injury severities in the City of St. Louis and is a milestone toward identifying and mitigating the potential risks of the pedestrian-vehicle collisions in this city.