删失数据下事件持续时间多因素生存分析模型

Multivariate Survival Analysis Models for Incident Duration With Censored Data

摘要:

交通事件持续时间是事故管理研究中最重要的指数之一，应用半参数比例风险模型和参数加速失效模型分析了多因素事件持续时间及其危险因子之间的关系.利用浙江省某高速公路3年内采集的事件数据，通过对交通事件各状态概率分布函数的显著性分析，比较了比例风险模型和加速失效模型对协变量选择及生存率对共同协变量的敏感性.基于Cox比例风险模型和对数罗吉斯蒂加速失效模型的参数估计表明，有6个显著性的协变量入选各自估计的生存函数，包括报警时段、事故类型、报警方式、占用车道数、受伤人数和当场亡人.对于最显著示性变量（当场亡人），持续时间生存概率估计曲线表明，加速失效模型对于该变量更敏感，且Cox比例风险模型更适合于短时持续时间的预测.此外，生存模型可以基于事故报告信息预测持续时间大小的概率，有益于事故预后措施的实施及为紧急救援提供决策参考.

Abstract:

Traffic incident duration is one of the most important indexes in accident management studies. The semi parametric proportional hazards model (Cox model) and parametric accelerated failure time model (AFT model) were employed for analysis of the association between multivariate incident durations and risk factors. With the three year traffic incidents data collected from some freeways in Zhejiang Province, the Cox model and AFT model with covariates selected to remain in the model and survival probability of sensitivity to common covariates were illustrated and compared. Based on the Cox and the Log logistic AFT models, the parameter estimates showed that six significant covariates were selected to remain in the estimated survival functions, including accident time, accident type, number of lanes occupied, number of the injured and accident of fatality. With regard to the most significant indicator variable (accident of fatality), the estimates of the survival probability curves for duration demonstrate that the AFT model give a higher sensitivity than the Cox model, and the Cox model is more appropriate for short time incident duration prediction. Moreover, survival analysis models can predict duration based on incident report and benefit rescue countermeasure and emergency aid decision after accidents.