A multivariate spatial approach to model crash counts by injury severity

Area of Science:

• Transportation Engineering
• Spatial Statistics
• Urban Planning

Background:

• Conventional traffic safety models often assume crash independence, which is unrealistic in urban environments.
• Spatial autocorrelation and correlations between different crash types are critical factors in understanding road safety.
• Existing models struggle to simultaneously address these multivariate spatial dependencies.

Purpose of the Study:

• To develop and validate a novel multivariate conditional autoregressive (MVCAR) model for traffic crash data analysis.
• To account for spatial autocorrelation among neighboring geographic units and correlations across various crash types.
• To assess the performance of the MVCAR model against conventional methods in a densely populated urban setting.

Main Methods:

• Development of a multivariate conditional autoregressive (MVCAR) model.
• Application of the model to crash, transportation, land use, and demo-economic data in Manhattan, NYC.
• Utilizing census tracts as the primary geographic units for analysis.
• Employing Moran's I tests to assess spatial autocorrelation.
• Comparison of model performance using the Deviance Information Criterion (DIC).

Main Results:

• The MVCAR model successfully captured multivariate spatial autocorrelation among different crash types.
• Moran's I tests confirmed the model's ability to account for spatial dependencies.
• The MVCAR model demonstrated superior performance with the lowest DIC value compared to other models.
• Unobserved heterogeneity in crash data was primarily attributed to spatial factors.

Conclusions:

• The proposed MVCAR model provides a robust framework for analyzing multivariate spatial crash data.
• Spatial factors significantly influence crash risk, exhibiting a shared geographical pattern across different injury severities.
• This approach enhances the accuracy of traffic safety modeling in complex urban environments.