Applications of Generalized Additive and Bayesian Hierarchical Models for Areal Safety Analysis: Case Study of an Urban Multimodal Transportation System in Chicago, Illinois

Abstract

Areal crash modeling has gained increased attention in the past decade because of initiatives to incorporate safety performance–based decision making in transportation planning. Particularly in urban multimodal transportation systems, safety outcomes may be influenced by long-term planning decisions at the area and network levels. How multimodal facilities are layered and prioritized eventually affects conflicts that may result between modes, influencing expected crash frequencies and severities for various road user types. The emphasis on areal crash modeling has opened the door for various innovative statistical methods, applied to explain factors that contribute to crashes, as well as to address issues that arise in spatially aggregated count data. These issues include spatial autocorrelation, ecological fallacy, and the modifiable areal unit problem. Previous studies used generalized linear models with fixed and random effects to address these issues, while the Bayesian framework has become a dominant approach in areal crash analysis—particularly at the county level—in the past decade. This paper explores an alternative frequentist approach to areal crash modeling with generalized additive models with smooth functions across the location and compares these models to negative binomial and Bayesian hierarchical models. The results, based on data from Chicago, Illinois, show that generalized additive models can account for spatial autocorrelation in the data, particularly when autocorrelation is lower and more data are available to reduce the number of potentially omitted variables.