Examining non-linear built environment effects on injurious traffic collisions: A gradient boosting decision tree analysis

Abstract

IntroductionDuring rapid urbanization, the optimization of a built environment (BE) in metropolises to reduce the frequency of traffic collisions can considerably promote traffic safety. Previous studies have seldom discussed the non-linearity and threshold effects of the BE on injurious collisions, and they propose few evidence-based synergistic effects for policy development. MethodsWith the injurious collision frequency as the dependent variable, this study established a BE indicator system including travel demand, roadway design, non-motorized infrastructure, land use, traffic control and disadvantage neighborhoods as explanatory variables. The gradient boosting decision tree (GBDT) model based on machine learning was used to explore the relative importance and the non-linearity of each variable. Furthermore, bivariate partial dependence plots were applied to reflect the synergistic effects of intersection density and other high-contribution BE variables on injurious collision frequency. Results(1) The riding volume, intersection density, and distance from CBDs are the top three contributing variables, which should attract more attention when planning. (2) Several thresholds were found: controling the width of arterial roads within 25 m, limiting the average vehicle speed in 36–38 km/h, setting subway stations within 1 km away from key city locations, adjusting the land use mix below 0.6, regulating the density of floating population around 3000 people/km2and setting the greenway density at 2 km/km2are all beneficial to reduce the injurious collision frequency. (3) Retaining the above thresholds from univariate PDPs, the synergistic effects not only causes high collision frequency, but also improves the risk growth rate. ConclusionAiming at the variables with high contribution rates and exerting these threshold effects in transportation plans will effectively reduce the occurrence of injurious collisions.