Hit and run crashes: Knowledge extraction from bicycle involved crashes using first and frugal tree

Abstract

Hit and run crash is a punishable offense. In many cases, it involves higher severity levels of the associated roadway users due to the delay of emergency help. For vulnerable road users like pedestrians and bicyclists, this issue is more gruesome. We present an analysis of the effect of crash, geometric, and environmental characteristics on the bicycle-involved hit and run crashes by using six years of Louisiana crash data with an application of fast and frugal tree (FFT) heuristics algorithm. Over 1000 bicycle-involved hit and run crashes occurred in Louisiana out of around 108,000 hit and run crashes during 2010–2015. The fatal bicycle crashes represent 10% of the total fatal hit and run crashes. Additionally, hit and run bicycle crashes represent 22% of total bicycle crashes in Louisiana. In the preliminary analysis, we provided statistical significance test of the key contributing factors for two major groups (bicycle-involved hit and run crashes, and not bicycle-involved hit and run crashes). We divided the complete dataset into two separate datasets: training data for model development, and testing data for performance evaluation. FFT identifies five major cues or variable threshold attributes that contribute significantly in predicting bicycle-involved crashes. These cues include fatal and injury crashes, right angle/turning/head on collisions, city streets/others, intersection, and residential/mixed localities. The balanced accuracy is around 76% for both training and testing data. The current model shows higher sensitivity than other complex and black box machine learning models (e.g., support vector machine, random forest). Findings of our study will provide valuable insights for hit and run bicycle crash reduction in both planning and operation levels. Keywords: Hit and run crashes, Bicycle, Fast and frugal algorithm, Heuristics, Decision tree