An empirical evaluation of multivariate spatial crash frequency models

Area of Science:

• Transportation Engineering
• Traffic Safety Analysis
• Statistical Modeling

Background:

• Roadway crash analysis commonly uses spatial or temporal models.
• A comprehensive comparison of spatial weight matrices and temporal treatments in spatiotemporal crash models is lacking.
• Existing studies often do not evaluate the combined impact of various spatial and temporal specifications.

Purpose of the Study:

• To compare the crash estimation performance of different Full Bayesian (FB) multivariate spatiotemporal models.
• To evaluate various spatial weight matrices (adjacency- and distance-based) combined with different temporal treatments.
• To identify the optimal model structure for analyzing pedestrian and bicyclist crashes.

Main Methods:

• Developed three groups of FB multivariate spatiotemporal crash models using pedestrian and bicyclist crash data from 58 California counties over eight years.
• Group 1: Unstructured error and spatially structured conditional autoregressive (CAR) term.
• Group 2: Added a linear time trend; Group 3: Incorporated space-time interaction. Evaluated 17 models per group using various criteria.

Main Results:

• The pure-distance model D0.5 generally performed best across all groups based on training and test errors.
• Models incorporating space-time interaction (Group 3) showed superior performance compared to other temporal treatments.
• Distance-based spatial weight matrices were mostly superior to adjacency-based ones, though with higher variability.

Conclusions:

• The optimal choice of spatial weight matrix and temporal treatment significantly impacts crash estimation accuracy.
• Incorporating space-time interaction generally improves model performance for pedestrian and bicyclist crash analysis.
• Distance-based models, particularly D0.5, are recommended, but careful selection is advised due to performance variability.