Bayesian spatial modeling to incorporate unmeasured information at road segment levels with the INLA approach: A methodological advancement of estimating crash modification factors

Abstract

Estimating safety effectiveness of roadway improvements and countermeasures, using cross-sectional models, generally requires large amounts of data such as road geometric and traffic-related characteristics at road segment levels. These models do not consider all confounding crash contributory factors such as driving culture and environmental conditions at the segment level due to a lack of readily available data. This may result in inaccurate models representing actual conditions at road segment levels, followed by erroneous estimations of safety effectiveness. To minimize the effect of not including such variables, this study develops a new methodology to estimate safety effectiveness of roadway countermeasures, based on generalized linear mixed models, assuming zero-inflated Poisson distribution for the response, and adjusting for spatial autocorrelation using the spatial random effect. The Bayesian approach, with Integrated Nested Laplace Approximation, was used to make inference on this model with computational efficiency. Results showed that incorporating a spatial random effect into the models provided better model fit than non-spatial models; hence, estimated safety effectiveness based on such models is more accurate. The proposed approach is a methodological advancement in traffic safety, which allows evaluation of safety effectiveness or roadway improvements when data are not readily available.

Keywords

• Traffic safety
• Roadway countermeasures
• Crash modification factors
• Spatial random models
• Hierarchical Bayesian models