高速公路事故頻次、嚴重度及碰撞型態整合迴歸模式之比較

Abstract

在分析事故的時候，事故頻次(總共的事故件數)、事故嚴重度(例如死亡、受傷、財物損失)、碰撞型態(例如追撞、對撞、側撞)是三個很關鍵的指標變數，雖然過去有許多的研究透過不同的統計方法將這三個指標變數以個別的方式或是聯合的方式建構模式，但是，大部分建構的模式卻忽略在不同類別(例如事故嚴重度或是碰撞型態)的事故件數之間的潛在相關性，這將會導致估計效率較低以及推估參數的偏誤，因此，為了要表達不同類別的事故件數之間的相關程度，本研究使用多變量廣義卜瓦松迴歸來同時建構事故頻次、嚴重度及碰撞型態的模式，以便找出影響不同受傷嚴重度或是不同碰撞型態的事故件數之顯著變數，然而，為了避免產生太多的指標變數，本研究分別構建事故頻次及嚴重度模式與事故頻次及碰撞型態模式。

本研究收集2005年在台灣的國道一號的事故資料，並將其用於推估模式，其中這些資料主要有三個來源：(1)事故資料庫; (2)幾何設計的文件; 以及(3)交通資料庫。其中事故資料庫由國道公路警察局所提供，並包含許多的事故資訊，例如事故嚴重度、碰撞型態、事故的時間與地點、事故發生的總車輛數與車輛型態，至於幾何資料則使用官方的國道繪製圖予以數值化，其中包含車道數、坡度、曲率、克羅梭曲線參數。由於台灣的國道一號全長373.3公里，並且有63個交流道，因此，為了方便模式之推估，本研究以兩個相鄰交流道之間的路段來作為一個分析樣本，並分別考慮南向與北向，總共可得到124個分析樣本，最後，本研究將每個路段上的總事故件數依照不同的受傷嚴重度分為A1事故、A2事故、A3事故，另外也有依照不同的碰撞型態分為同向擦撞事故、追撞事故、撞護欄事故、其他碰撞型態的事故。

除了使用多變量廣義卜瓦松迴歸之外，本研究也採用另外兩種的研究方法建構模式，包含單變量廣義卜瓦松迴歸以及誤差項的多項廣義卜瓦松，並分別將這三種建模方法進行推估與比較，結果顯示多變量廣義卜瓦松有最佳的模式配適度與預測準確度，根據多變量廣義卜瓦松模式的估計結果，下坡度、車道數、交流道兩端均鄰都會區對事故嚴重度以及碰撞型態皆有相當高的影響程度，表示必須針對這三個影響因子來研擬改善之對策。

Crash frequency (number of crashes), severity (e.g. fatality, injury, and property damage only) and collision type (e.g. rear-end, head-on, swipe…) are three key target variables in accident analysis. Numerous studies attempted to separately or jointly model the these three target variables with various statistic methods. However, most of models ignored potential correlation of crash frequency with other variable, which led to low estimation efficiency and biased estimated parameters. In order to better capture the correlation of crash frequency with other variables, this study aims to use of three models to simultaneously model crash frequency, severity and collision type. It is hoped to identify the key variables contributing to crash frequency by severity or crash frequency by collision type. To avoid too many target variables being generated, however, severity and collision type are separately integrated with crash frequency.

The accident dataset for Taiwan’s No. 1 Freeway in 2005 was collected and used to estimate our proposed models. Data were drawn from three sources: (1) the accident database; (2) the geometric documents; (3) the traffic database. The accident database, maintained by the National Highway Police Bureau (NHPB), contains accident information, such as severity, collision type, location and time of accident, and number and types of vehicle involved. The geometric data were digitalized from the official freeway construction drawings, including number of lanes, slope, curvature degree, and clothoid curve value. Taiwan’s No. 1 Freeway is north-south bounded 373.3 km long with 63 interchanges. To facilitate model estimation, each study segment is formed by two adjacent interchanges and north-south-bound directions are treated altogether; therefore, a total of 124 analytical samples are obtained. The proposed models consider three levels of crash severity: fatality (A1), injury (A2) and property damage only (A3) and four types of collision: sideswipe, rear-end, collision barrier, and others.

For comparison, three modeling approaches are estimated and compared: Multivariate Generalized Poisson (MGP), Univariate Generalized Poisson (UGP), and Multinomial Generalized Poisson with Error Components (EMGP). The estimation results show that the MGP model performs the best in terms of goodness of fit and prediction accuracy. According to the MGP model estimation results, downward slope, number of lanes, and adjacent to metropolitan have relatively high effect on severity and collision type. Some safety strategies are proposed accordingly.