Modeling spatial nonstationary and overdispersed crash data: Development and comparative analysis of global and geographically weighted regression models applied to macrolevel injury crash data

• Published in: Journal of transportation safety & security 2021-09, Vol.13 (9), p.1000-1024
• Main Authors: Soroori, Emad, Mohammadzadeh Moghaddam, Abolfazl, Salehi, Mahdi
• Format: Article
• Language: English
• Subjects: Agricultural land; Bus stops; Case studies; Comparative analysis; Crashes; Developed countries; Developing countries; Extreme values; geographically weighted negative binomial regression; geographically weighted Poisson regression; Goodness of fit; Heterogeneity; Injuries; Kernel functions; Land use; LDCs; macrolevel safety analysis; overdispersion; Parameter estimation; Poisson density functions; Public transportation; Regression analysis; Regression models; Safety; Spatial data; Spatial heterogeneity; spatial nonstationary; traditional models; Transportation models; Transportation planning; Transportation safety; transportation safety planning
• ISSN: 1943-9962; 1943-9970
• DOI: 10.1080/19439962.2020.1712671

There is a wide body of research, especially in developed countries case studies, conducted to quantify safety impacts of transportation strategies. These models can be broadly divided into global and local types. Contrary to consistency assumption of global parameters, parameters' variation over the space, which stems from heterogeneity, can be considered by the local models. Geographically weighted regression (GWR) is a proper technique which is employed to locally estimate the parameters. Geographically weighted Poisson regression (GWPR) is one of the most prevailing count models to be exercised in transportation safety planning. Overdispersion as another consequence of heterogeneity, however, is not addressed by GWPR. Therefore, this study is centered on three targets. First, crash models are developed in a developing country case study. Second, overdispersed and spatial nonstationary data is scrutinized by applying and investigating the performance of geographically weighted negative binomial regression with two aspects of global and spatially varying overdispersion parameters (i.e., GWNBRg and GWNBR). Also, the explanatory performance and prediction accuracy are compared with GWPR, traditional Poisson and negative binomial regressions as two widely used classes of the global models. Finally, the fluctuating safety effects of spatial-unit-related variables at the macrolevel are examined. To this end, geographically coded data on injury crashes, road characteristics, socioeconomic and demographic factors, public transit and land use were collected for 253 traffic zones as the macrolevel spatial units from Mashhad city. The findings indicated that GWR family outperforms traditional regressions in capturing spatial heterogeneity when goodness of fit criteria is compared. Also, GWNBR revealed better performance in terms of AICc and lesser sensitivity to extreme values by developing smoother kernel function of geographical weights. Based on our findings, however, GWNBRs cannot predict other data sets as well as GWPR and researchers are recommended to use GWPR rather than GWNBR. The results also demonstrated that major arterials, ramps, number of intersections, bus stops and employees as well as commercial, residential, and agricultural land uses have considerable impact on injury crashes. In addition, implications regarding the detection of multicollinearity among explanatory variables based on condition index were noted.