Smart Transportation Key Laboratory of Hunan Province, School of Traffic and Transportation Engineering, Central South University, Changsha, 410075, China.
School of Transportation Engineering, Changsha University of Science and Technology, Changsha, 410205, China.
Accid Anal Prev. 2020 Dec;148:105833. doi: 10.1016/j.aap.2020.105833. Epub 2020 Oct 22.
In recent years, globally quantile-based model (e.g. quantile regression) and spatially conditional mean models (e.g. geographically weighted regression) have been widely and commonly employed in macro-level safety analysis. The former ones assume that the model coefficients are fixed over space, while the latter ones only represent the entire distribution of variable effects by a single concentrated trend. However, the influence of crash related factors on the distribution of crash frequency is observed to vary over space and across different quantiles. Therefore, a geographically weighted Poisson quantile regression (GWPQR) model is employed to investigate the spatial heterogeneity of variable effects crossing different quantiles. Five categories, including exposure, socio-economic, transportation, network and land use were selected to estimate the spatial effects on crash frequency. In the case study, vehicle related crashes collected in New York City were used to validate the predicted performance of the proposed models. The results show that the GWPQR outperforms the NB, QR and GWNBR for modeling the skewed distribution, reconstructing the crash distribution and capturing the unobserved spatial heterogeneity. Additionally, the significant coefficients are further used to classify all 21 variables into key, important and general parts. Then we discuss how these factors affects the regional crashes over space and distribution of crash frequency. This study confirms that the influencing factors have varying effects on different quantiles of distribution and on different regions, which could be helpful to provide support for making safety countermeasures and policies at urban regional level.
近年来，基于全球分位数的模型（例如分位数回归）和空间条件均值模型（例如地理加权回归）已在宏观安全分析中得到广泛应用。前者假设模型系数在空间上是固定的，而后者仅通过单一集中趋势来表示变量效应的整个分布。然而，观察到与碰撞相关的因素对碰撞频率分布的影响在空间上和不同分位数之间是不同的。因此，采用地理加权泊松分位数回归（GWPQR）模型来研究变量效应在不同分位数上的空间异质性。选择了五类，包括暴露、社会经济、交通、网络和土地利用，以估计对碰撞频率的空间效应。在案例研究中，使用纽约市收集的与车辆相关的碰撞来验证所提出模型的预测性能。结果表明，GWPQR 在对偏态分布进行建模、重建碰撞分布和捕获未观察到的空间异质性方面优于 NB、QR 和 GWNBR。此外，显著系数进一步用于将所有 21 个变量分为关键、重要和一般部分。然后，我们讨论了这些因素如何在空间上影响区域碰撞和碰撞频率分布。本研究证实，影响因素对分布的不同分位数和不同区域的影响不同，这有助于为城市区域层面的安全措施和政策提供支持。
