面板数据模型分析及交通事故预测

Panel data models for analysis and prediction of crash count

构建了一种基于面板数据的空间非集计模型——负二项面板模型,包括混合效应、固定效应和随机效应3种类型,以同性质路段为研究单元,选取道路线形、交通特性、土地利用和降雨量等影响因素,利用事故率比例指标IRR,分析和预测未受伤事故、受伤事故、死亡事故和事故总数等4种类型事故,并通过F检验和Hausman检验以及对数似然值和离差信息准则DIC,对比分析3种类型模型的拟合效果。发现协变量对各类型交通事故的影响作用和统计显著性不尽相同,如限速每增加1.609km/h(1mile/h),未受伤事故、受伤事故、事故总数分别减少3.89%、2.24%和2.79%,而死亡事故增加6.38%。研究结果表明:负二项面板固定效应模型比混合模型和随机效应模型更优,另外越严重的事故,模型拟合效果越好。

Spatially disaggregated panel data models using negative binomial process,including pooled model,fixed-effects model and random-effect model,were developed.The contributing factors,such as roadway geometrics,traffic characteristics,land usage and rainfall were collected.Incidence Rate Ratio（IRR）and homogeneous segments were employed for analysis and prediction of crash count,which contains Property-Damage-Only（PDO）crashes,injury crashes,fatal crashes,and total crashes.Meanwhile,F test and Hausman test,log-likelihood value and Deviance Information Criterion（DIC）were used to compare the performance of the three types of models.It is found that the effects and statistical significance of covariates on the four types of crashes are not the same.For example,if the speed limit is increased by 1.609km/h,the PDO rate,injury crash rate and totalcrash rate are reduced by 3.89%,2.24% and 2.79%,respectively;however,the fatal crash rate is increased by 6.38%.Results show that the fixed-effects panel negative binomial model surpasses the pooled model and random-effect model,and the modeling is even better for crash count associated with more severe injuries and fatalities.