事故嚴重度之多層次風險因子分析

Abstract

以往交通事故嚴重度分析大多以一般線性模式 (Generalized Linear Models)分析，而忽略事故資料階層性結構之問題；然而，在相同路口發生之事故，由於有共同的車流量、交通號誌設計、及交通管制設施等路口特性因子，事故的嚴重度結果可能有相似性。為了要從事故資料及路口調查資料中取得更多資訊，若能進一步加入路口特性因子以比較造成事故嚴重度之差異與原因，將有助於提出更為有效的交通安全改善策略。 由於事故嚴重度除了受到個體風險因子 (如當事人特性、當事人狀態及事故特性)，總體風險因子(如路口特性)可能也會影響事故嚴重度差異，並造成對個體風險因子及事故嚴重度關係之影響，本研究採用階層線性模式 (Hierarchical Linear Model)發展事故嚴重度模式分析。此外，為確認以階層線性模式發展事故嚴重度模式可獲得之模式改善效果，本研究以過去文獻中進行事故嚴重度分析常採用之多項羅吉特模式及次序普羅比模式，作為單層次模式分析之方法，以比較單層次模式及多層次模式分析事故嚴重度模式之績效表現。 本研究蒐集台北市2012年及2013年之交通事故資料及路口調查資料，納入模式分析之資料共計53個號誌化路口，2,132件事故資料，將事故嚴重程度分為死亡或重傷、中度受傷、輕度受傷、及財損四類，並將2,132件事故以事故發生之路口分成53群，利用路口特性因子作為階層線性模式之上層因子，個體事故特性因子作為階層線性模式之下層因子。先以多項羅吉特模式及次序普羅比模式建構單層次之事故嚴重度分析模式，並比較其差異性，再建構多項羅吉特模式及次序普羅比模式之階層線性模式，以比較單層次模式及多層次模式之差異性及模式績效表現。 研究結果發現當事人中有高齡者、酒駕、超速、碰撞型態為路口交叉撞及路口中有順向公車道，會提高事故嚴重度，而事故為大小車輛碰撞及路口號誌週期長度則被發現會降低事故嚴重度。在跨層級解釋變數的交互作用效果中發現，在超速事故中，車道數差與死亡或重傷的事故嚴重度有正向影響，平均車道數則與死亡或重傷的事故嚴重度有負向影響；在路口交叉撞的事故中則被發現，有快慢車道分隔島的路口事故會傾向有較低機率的死亡或重傷事故發生，路口各方向車道數差距較大的路口事故則傾向於有較高的死亡或重傷事故發生。且以單層次模式及多層次模式分析可獲得相似的事故嚴重度風險因子正負影響結果，但以多層次模式分析可獲得較佳的模式績效表現。

Generalized linear models has been widely adopted in crash injury severity analysis, but ignored the hierarchical structures within accident data. However, the accidents happened at the same intersection are supposed to have common intersection factors (such as traffic, signal design, and facilities), indicates that the severity outcomes of these accidents are similar. In order to get more information from crash data and intersection investigation data, modeled crash injury severity analysis with intersection factors to compare the differences and causes of crash injury severity will help make more efficient traffic safety improvement strategy. Since individual risk factors (such as parties properties, parties states, and accident characteristics) affect crash severity, general risk factors (such as intersection properties) may affect crash severity also moderate the relationship between individual risk factors and severity, this study developed hierarchical linear models to analyze crash injury severity. Moreover, in order to make sure that hierarchical linear models provide better performance, this study also developed severity model with multinomial logit model and ordered probit model (which widely adopted in crash injury severity analysis) as single-level models to compare model performance with multi-level models. In total, 2132 accidents that occurred at 53 signalized intersections during 2012-2013 in Taipei City are collected. Injury severity is the dependent variable that is killed/serious injury, moderate injury, minor injury, and property damage only. 2132 accidents are grouped to 53 group according the occurred intersection, as a result, we have intersection factor as the upper level factor and individual crash factor as lower level factor in hierarchical linear models. This study build crash severity models with multinomial logit model, ordered probit model, multinomial logit hierarchical model, and ordered probit hierarchical model for the purpose of discussing the differences of the factors effects and the performances. The model results indicate that elderly, alcohol and speeding involved the crash, angle collision, and same direction bus lane existed at the intersection will increase the injury severity. Moreover, four-wheeled vehicle involved the crash and the signal cycle length will decrease the injury severity. Besides, there has the similar sign in the results of the single-level models and multi-level models, but multi-level models provide better performance. Also, the difference of number of lanes has positive moderating effects on the relationship between speeding and injury severity, while average number of lanes has negative moderating effects on the relationship between speeding and injury severity. At the intersections having traffic island for separating slow cars has negative moderating effects on the relationship between angle collision and injury severity, while having traffic island for separating slow cars has positive moderating effects on the relationship between angle collision and injury severity.