Crash frequency modeling for signalized intersections in a high-density urban road network

• Correlation of crashes is studied using Moran׳s I and Lagrange Multiplier tests.
• We propose a novel hierarchical conditional autoregressive model for crash count.
• We use a cross-validation test to compare the predictive performance of models.

Conventional crash frequency models rely on an assumption of independence among observed crashes. However, this assumption is frequently proved false by spatially related crash observations, particularly for intersection crashes observed in high-density road networks. Crash frequency models that ignore the hierarchy and spatial correlation of closely spaced intersections can lead to biased estimations. As a follow-up to our previous paper (Xie et al., 2013), this study aims to address this issue by introducing an improved crash frequency model. Data for 195 signalized intersections along 22 corridors in the urban areas of Shanghai was collected. Moran׳s I statistic of the crash data confirmed the spatial dependence of crash occurrence among the neighboring intersections. Moreover, Lagrange Multiplier test was performed and it suggested that the spatial dependence should be captured in the model error term. A hierarchical model incorporating a conditional autoregressive (CAR) effect term for the spatial correlation was developed in the Bayesian framework. A deviance information criterion (DIC) and cross-validation test were used for model selection and comparison. The results showed that the proposed model outperformed traditional models in terms of the overall goodness of fit and predictive performance. In addition, the significance of the corridor-specific random effect and CAR effect revealed strong evidence for the presence of heterogeneity across corridors and spatial correlation among intersections.