An accelerated hierarchical Bayesian crash frequency model with accommodation of spatiotemporal interactions

Area of Science:

• Transportation Science
• Statistical Modeling
• Urban Planning

Background:

• Spatial and temporal correlations in crash data are well-studied.
• The interaction between spatial and temporal factors in crash frequency modeling remains underexplored.

Purpose of the Study:

• To propose a novel Bayesian spatiotemporal interaction (BSTI) approach for crash frequency modeling.
• To enhance the accuracy and interpretability of traffic safety analytics.
• To address the understudied interaction between spatial and temporal crash determinants.

Main Methods:

• Developed a Bayesian spatiotemporal interaction (BSTI) model using integrated nested Laplace approximation (INLA) for efficient estimation.
• Utilized hexagonal units in Manhattan, NYC, to integrate crash, transportation, land use, and demo-economic data (2013-2019).
• Compared various spatiotemporal models to identify the best-performing approach.

Main Results:

• The BSTI model with Type II interaction (structured temporal effect interacting with unstructured spatial effect) demonstrated superior goodness-of-fit.
• This model effectively reduced residual dependencies and identified spatial effects as the primary source of unobserved heterogeneity.
• The BSTI Type II model achieved the lowest predictive error when validated with the most recent year's data.

Conclusions:

• The proposed BSTI approach offers a significant advancement in traffic safety analytics.
• It provides high prediction accuracy and computational efficiency.
• The model maintains interpretability regarding contributing factors and unobserved heterogeneity.