事故分析除了事故頻次(crash frequency)之探討外,嚴重程度(severity)及碰撞型態(collision type)也同等重要。以往研究大多先依嚴重程度及碰撞型態事故分別統計事故頻次,再分別建立單變量頻次模式,忽略事故類別問可能存在關聯性,而導致估計偏誤。基此,本研究建立多變量廣義卡瓦松(Multivariate Generalized Poisson; MVGP)模式,同時模化不同類別事故頻次,並計算事故類別閉關聯性及資料離散程度。本文以中山高速公路為例,結果證明多變量模式確較單變量模式為佳。另由關聯鋒、數知,A1與A2事故類別具有高度正相關,而A2與A3事故問則呈中度正相關,Al與A3事故幾近無關。追撞、同向擦撞與其他通撞三者問皆為高度正相關;撞護欄與追撞、同向擦撞呈現中度正相關,另與其他碰撞型態則為低度負相關。此外,碰撞型態關聯程度高於嚴重程度,車種則為碰撞型態與嚴重程度共同肇因。依據模式推估結果,本文也據以研提相關對策。
In addition to crash frequency analysis, crash severity and collision type are another two equally important factors that have to be assessed so as to propose effective countermeasures accordingly. Since crash severity and collision type are categorical, most of previous studies independently developed crash frequency models by severity levels and collisions types, ignoring the potential correlation among categorical levels and leading to biased estimation and lower prediction accuracy. To accommodate the potential correlation and data dispersion, this paper attempts to develop integrated models for freeway crashes across severity levels and collision types, respectively, by using the Multivariate Generalized Poisson (MVGP) model. A case study on the crash data of Taiwan NO.1 Freeway shows that the MVGP model performs significantly better than the Univariate Generalized Poisson (UVGP) model in terms of model prediction accuracy. According to estimated correlation coefficients, for crashes with severe injury levels which are Al (fatal crashes) and A2 (injury crashes) exhibit highly positive correlation. A2 and A3 (property damage only crashes) have medium positive correlation. However, Al and A3 are nearly uncorrelated. As to collision type, rear-end, sideswipe and other collision types are highly positively correlated, while collisions with roadside barriers have medium positive correlation with rear-end and sideswipe collisions, but low positive correlation with other collision types. Additionally, it is also found that the correlation effects among collision types are higher than those among severity levels. Moreover, traffic composition is identified as the common factor contributing to both collision types and severity levels. Finally corresponding countermeasures are then proposed based on the estimated parameters and elasticity of significantly tested variables.