基于有序Logit与Probit模型的交通事故严重性影响因素分析

Analysis for the influential factors of the accident severity based on the ordinal Logit and Probit models

为准确分析影响交通事故严重性的各项因素,引入有序选择模型中的Ordinal Logit模型和Ordinal Probit模型,研究驾驶员、车辆、环境、管理因素与事故严重程度之间的耦合关系,利用计量经济学理论和统计学方法对模型进行显著性检验。并选择美国北卡罗来纳州2010年至2014年385个翻车事故样本进行严重性影响因素分析,所得回归模型参数估计均符合Wald检验（p≤0.0001）和似然比检验（p≤0.0001）,模型的拟合度也都符合Pearson检验（p=0.7976）、偏差检验（p=0.6006）和信息准则检验,模型预测精度指标值均大于0.7。所得两个模型中变量＂安全带＂（p≤0.001）、＂路面状况＂（p=0.0071）、＂道路线形＂（p=0.0077）、＂路面类型＂（p=0.0251）都对翻车事故的严重性具有显著影响,表明有序Logit模型和有序Probit模型均适用于分析和揭示影响交通事故严重性的各个因素。

The present paper takes it as its purpose to make an analysis of the influential factors of the accident severity based on the ordinal Logit and Probit models. For the analysis purpose,we have chosen all the 385 rollover accidents that took place in North Carolina,USA,during the period of the five years from 2010 to2014 so as to identify and determine the influential factors on the severity of such accidents. However,to clarify the underlying severity propensity for the rollover accidents,it is necessary to take two different statistical approaches,that is,the Ordinal Logit（ ORL） regression and that of the Ordinal Probit（ ORP） to determine the causative and resultative relations between the resultant factors pertinent to the drivers,vehicles,roadway features,the environment and management strategies both quantitatively and qualitatively,in addition to calibrating the parameters both of ORL and ORP models via the maximum probability judgment through screening of the independent variables concerned. And,then,the tests of parallel line factors assumption and likelihood ratio can be done to examine the model validity,the relevant factors and the goodness-of-fit criteria so as to enhance the model parsimoniousness and screen out the key variables. As a result,the ORL regression results show that the following four variables are of great significance in the final parsimonious model： ＂ the seatbelt＂（ p ≤ 0. 0001）,＂the road line condition ＂（ p =0. 0059）, ＂ the road category or classification＂（ p = 0. 0226）,as well as the ＂road surface maintenance status-in-situ＂（ p =0. 010 3）. What we have gained by using the model tends to be of the closest correlation between the seatbelt and the severity of injury,instead of whether to wear or not the seatbelts,merely.And,in contrast,negative effect can be detected by the lane curvature,gravel-covering,and friendliness of the road surface on the severity of the rollover accidents. All in all,the ORP regression results demonstrate that the final model encompasses the significant variables as shown below： ＂the seatbelt wearing situation＂（ p ≤0. 0001）,＂the road line condition＂（ p = 0. 0077）,＂the road classification category＂（ p = 0. 0251）,and＂the road surface status-in-situ＂（ p = 0. 0071）,for,only in the significance level of each variable can the other results of ORL and ORP be made highly similar. The modeling results also show that the ordinal discrete choice models are valid to reveal the factors influential on the resultant severity of the rollover accidents and imply the methodological necessity for the roadway safety consequence to take the distinctive modeling perspectives for a more grounded and informed comprehension of the safety issues. Thus,the analysis can only be made fully informative for the roadway management agencies to assess the roadway design schemes by taking the safety countermeasures against such rollover accidents（ e. g.,guardrail installation） and boosting the public education campaign to promote the safety belt usage and the safety driving behaviors in the case of adverse weathers.