Freeway Crash Frequency Modeling under the Time-of-Day Distribution
|關鍵字:||事故頻次;時間分布;負二項迴歸;多項羅吉特;集群分析;多變量卜瓦松迴歸;Crash frequency;Time-of-day distribution;Negative binomial regression;Multinomial logit model;Cluster analysis;Multivariate Poisson regression|
The key factors explaining the spatial and temporal distribution of crash frequency are essential for proposing corresponding countermeasures. However, most of previous studies only focus on the key factors contributing to the spatial distribution, while rather few studies further examine the time-of-day distribution of crash frequency. Based on this, this study aims to not only identify the spatial key factors, but also to examine those affecting the time-of-day distribution of crash frequency. To do so, three freeway crash frequency models under the time-of-day distribution are developed, estimated and compared in this study. Model 1 uses of count models, including Poisson regression (PO), negative binomial regression (NB) and generalized Poisson regression (GPO), to explain the spatial distribution of crash frequency and uses of a ratio model, i.e. multinomial logit model (MNL), to determine the time-of-day distribution probability. Model 2 combines the abovementioned count models and a clustering model which classified freeway segments into different clusters according to their time-of-day distribution of crash frequency. The average time-of-day distributing pattern of each cluster is then use to represent the distribution of freeway segments which belong to the cluster. Model 3 is a multivariate count model by treating crash frequencies by time-of-day periods as target variables and two formulations, multivariate Poisson regression (MPO) and multivariate generalized Poisson regression (MGPO), are attempted to exhibit the spatiotemporal distribution of crash counts simultaneously. The abovementioned count and ratio models are then developed by considering the explanatory variables, including geometrics, facilities, environment condition, and traffic characteristics. Crash datasets of Taiwan Freeway No.1 in 2005 and 2006 are used to estimate and validate the models, respectively. The performances of three models are measured in terms of the Adjusted Mean Absolute Percentage Error (Adj-MAPE) and the Root-Mean-Square Error (RMSE). Four time-of-day periods, Morning (23~07), Afternoon (07~14), Evening (14~20) and Night (20~23), are formed according to the crash frequency distribution. The results show that the NB model performs better than the GPO and PO models, which is then adopted for the univariate count model in both Models 1 and 2. In terms of Adj-MAPE and RMSE, Model 2 performs best, followed by Model 1 and Model 3. According to the estimated parameters in the NB model, four variables of the maximum downward slope, the Clothoid curve value, the number of speeding cameras, and the percentage of heavy trucks exhibit significant negative effects on crash frequency, while the curvature rate, the adjacent to metropolitan and the traffic volume of small vehicles have significant positive effects on crash frequency. Corresponding countermeasures are then proposed. It is interesting to note that according to the clustering results, the freeway segments located in the non-metropolitan area (i.e. Cluster 2) tend to have higher crash frequency in the night (20~23) and morning (23~07) while those located near the system interchange (i.e. Cluster 4) tend to have higher crash frequency in the afternoon (07~14) and do not located near the system interchange (i.e. Cluster 3) tend to have higher crash frequency at night(14~20), suggesting the time-of-day distributions of crash frequency of different segments remarkably differ. Different countermeasures should be proposed for different segments. Keywords：Crash frequency, Time-of-day distribution, Negative binomial regression, Multinomial logit model, Cluster analysis, Multivariate Poisson regression.