Gomes Marcos José Timbó Lima, Cunto Flávio, da Silva Alan Ricardo
Department of Transportation Engineering, Federal University of Ceará, Brazil.
Department of Statistics, University of Brasilia, Brazil.
Accid Anal Prev. 2017 Sep;106:254-261. doi: 10.1016/j.aap.2017.06.011. Epub 2017 Jun 22.
Generalized Linear Models (GLM) with negative binomial distribution for errors, have been widely used to estimate safety at the level of transportation planning. The limited ability of this technique to take spatial effects into account can be overcome through the use of local models from spatial regression techniques, such as Geographically Weighted Poisson Regression (GWPR). Although GWPR is a system that deals with spatial dependency and heterogeneity and has already been used in some road safety studies at the planning level, it fails to account for the possible overdispersion that can be found in the observations on road-traffic crashes. Two approaches were adopted for the Geographically Weighted Negative Binomial Regression (GWNBR) model to allow discrete data to be modeled in a non-stationary form and to take note of the overdispersion of the data: the first examines the constant overdispersion for all the traffic zones and the second includes the variable for each spatial unit. This research conducts a comparative analysis between non-spatial global crash prediction models and spatial local GWPR and GWNBR at the level of traffic zones in Fortaleza/Brazil. A geographic database of 126 traffic zones was compiled from the available data on exposure, network characteristics, socioeconomic factors and land use. The models were calibrated by using the frequency of injury crashes as a dependent variable and the results showed that GWPR and GWNBR achieved a better performance than GLM for the average residuals and likelihood as well as reducing the spatial autocorrelation of the residuals, and the GWNBR model was more able to capture the spatial heterogeneity of the crash frequency.
具有负二项分布误差的广义线性模型（GLM）已被广泛用于在交通规划层面评估安全性。通过使用空间回归技术中的局部模型，如地理加权泊松回归（GWPR），可以克服该技术在考虑空间效应方面的有限能力。尽管GWPR是一种处理空间依赖性和异质性的系统，并且已经在一些规划层面的道路安全研究中得到应用，但它未能考虑到道路交通碰撞观测中可能存在的过度离散问题。地理加权负二项回归（GWNBR）模型采用了两种方法，以使离散数据能够以非平稳形式建模，并注意到数据的过度离散：第一种方法研究所有交通区域的恒定过度离散，第二种方法为每个空间单元纳入变量。本研究在巴西福塔莱萨市交通区域层面，对非空间全局碰撞预测模型与空间局部GWPR和GWNBR进行了比较分析。根据关于暴露、网络特征、社会经济因素和土地利用的现有数据，编制了一个包含126个交通区域的地理数据库。以受伤碰撞频率作为因变量对模型进行校准，结果表明，GWPR和GWNBR在平均残差和似然性方面比GLM表现更好，同时降低了残差的空间自相关性，并且GWNBR模型更能够捕捉碰撞频率的空间异质性。
