Direct Spatial Correlation in Crash Frequency Models: Estimation of the Effective Range

Abstract

Spatial correlation has shown to be present in highway safety data, yet the distance at which sites should be considered correlated is largely unknown. The purpose of this research is to explore the effect of direct spatial correlation structures on crash frequency models at the road segment level and compare them to spatial conditional models. By using direct spatial correlation structures the “effective range” (i.e., the distance at which there is no lingering spatial correlation) is estimated. Full Bayes hierarchical models with direct spatial correlation effects and conditional autoregressive (CAR) spatial effects are estimated. The model of crash, traffic, and roadway inventory data from Pennsylvania shows an average effective range of around 168 m which is in line with previous findings from conditional models. The direct spatial correlation model has a better goodness of fit than the random effects and the CAR model. Furthermore, the proportion of variation in the data explained by the spatial correlation term is almost the same for both spatial models. The standard deviations in coefficient estimates are slightly lower for the direct spatial correlation model compared to the random effects model but significantly lower than the CAR model.

Keywords:

• Bayesian statistics
• direct spatial correlation
• effective range
• CAR
• crash data