Many researchers have developed predictive models of crashes based on the safety scores of commercial truck companies, but these studies have been based on aggregated data at the truck company level—evaluating the total crashes and violations per company over a period of time. This level of aggregation obscures critical information. Here, a new approach to organizing non-aggregated data is taken, and logistic regression and random forest models are applied to non-aggregated FMCSA roadside inspection, violation, and crash data at the specific vehicle level. Resampling methods are used to improve model performance where there are relatively few events of interest—crashes. These results point not to specific “unsafe” drivers, but rather, patterns of unsafe behaviors or conditions that predict roadway crashes. Working toward reducing these behaviors systematically could save lives on US highways.

许多研究人员已经基于商用卡车公司的安全评分开发了事故的预测模型，但是这些研究是基于卡车公司级别的汇总数据-评估了一段时间内每个公司的总事故和违规情况。这种聚合级别掩盖了关键信息。在这里，采用了一种组织非聚合数据的新方法，并将逻辑回归和随机森林模型应用于特定车辆级别的非聚合FMCSA路边检查，违规和碰撞数据。重采样方法用于改善感兴趣的事件相对较少（崩溃）的模型性能。这些结果并非指向特定的“不安全”驾驶员，而是指向道路交通事故的不安全行为或状况的模式。