事故鏈與因果分析

Abstract

分析事故因果關係為改善交通事故、提升交通安全的重要方法之一。本研究的目的在利用交通事故資料庫，憑藉先進之方法論與逐漸成熟之電腦計算能力，從事故鏈的觀點有效挖掘事故發生影響因子以及事故因果關係。研究中以粗略集合理論作為從橫斷面事故資料有效取得事故鏈之方法，該理論的優點在於可同時控制眾多影響變數，反應事故發生為眾多因子交互作用的本質；粗略集合理論規則的產生為比較事故個體差異的結果，可有效避免總計誤差在資料推論時可能造成之謬誤。 本研究以事故鏈為核心概念進行三項研究：首先，藉由系統性地導入不同組合之條件屬性，分析粗略集合理論解釋事故資料之能力，以及粗略集合規則解釋事故鏈之有效性。接著以粗略集合規則對應之事故發生頻率為指標進行資料分群，以事故鏈的觀點分析事故資料異質性。最後以成對比較粗略集合規則的方式，分析事故情境變動對事故後果可能之影響，藉以挖掘可能之事故因果關係。 本研究利用內政部警政署之事故資料庫，針對台灣地區小客車單一車輛事故進行實證分析。研究結果發現：粗略集合理論的上下界近似、近似精度、近似品質、規則產生數以及判中率，為比較不同事故種類發生過程的有效指標，粗略集合規則並可幫助研究者了解事故發生情境。在單一車輛事故中，衝撞道路設施為可預測性較高的事故種類，其與撞建築物、衝出路外與翻車事故之發生過程可能類似。另外，經常發生與稀少發生之事故型態特性確有明顯差異；前者為過去研究中常被指稱為高風險之駕駛族群，後者則與不良之駕駛環境連結。過去常以改善道路環境作為增進交通安全的方式，此等手段雖可有效降低中、低發生頻率情境之事故，但良好之道路環境可能間接鼓勵高風險駕駛人提高行駛速度。研究並發現事故的發生並非由單一因素造成，而是由一連串不利因素組合而成。若能在事故鏈中移除部分不利因素，有可能改變事故後果、降低事故嚴重度。 本研究以事故鏈為核心所主張之分析概念與架構，提供一個更貼近事故發生本質的分析方法，其中並對交通安全研究中常見之總計誤差、資料異質性、干擾因子等議題，進行深入探討。本研究所提之分析架構，可根據研究者手中資料完整性、對分析對象的了解程度，在事故鏈的大架構下作相對應之延伸。

Analyzing accident causality has been one of the many ways to enhance traffic safety. The objective of this research was to explore contributing factors and accident causality by utilizing crash databases with mature methodologies and powerful computational powers from chain perspective. Rough sets theory was adopted in this research to obtain accident chains from cross-sectional databases. This theory is advantageous due to its ability to simultaneously control numerous factors, which reflect the fact that the occurrence of accidents results from complex interactions of many contributing factors. The other advantage is that rough set rules are generated by comparing the individual differences, which would partially alleviate the issue of aggregation bias. Three studies were conducted based on the concept of accident chains. The first study was to assess the ability of rough sets theory in explaining the underlying process of accident occurrence and in demonstrating accident chains by systematically loading combinations of condition attributes into rough sets. Second, the issue of data heterogeneity was examined from chain perspective by grouping accidents with the occurring frequency of rules. Finally, accident causality was addressed by comparing individual rules in pairs. Taiwan's crash databases were adopted in the empirical study, where single auto-vehicle (SAV) accidents were chosen as the subject to analysis. It was found that lower/upper approximation, accuracy of approximation, quality of approximation, number of generated rules, and hit rates could effectively address the differences between accident types. The occurrence of crashes with facility may follow similar paths and is more predictable; these crashes have some similarities between the crashes with architecture, with facility, off-road and rollover types. Moreover, significantly different features were shown between frequently repeated and sparsely unique rules. The former rules linked to the characteristics of high-risk drivers shown in past studies while the latter was connected with poor road conditions. Providing better road environment has been considered as an effective way to improve traffic safety; however, better roads could encourage high-risk drivers to raise their driving speeds. Furthermore, instead of one single factor the combinations of unfavorable factors were found to be the causes leading to fatal accidents. If one or several undesired factors were removed from the chain, accident severity might be reduced. The proposed approaches in the research provide a way to analyze accidents closer to the essence of accident occurrence. Meanwhile, these approaches also provide alternative ways to alleviate issues often seen in safety research such as aggregation bias, heterogeneity of accident data, confounding factors, and so on. These approaches can be expanded based on analysts' on-hand data and their understanding of target subjects.