Additive Bayesian Networks for Statistical Modeling of Driving Risk in Road Safety

Additive Bayesian Networks (ABNs) are graphical models that extend generalized linear models to multivariate settings by representing the full joint distribution of multiple dependent variables. Although ABNs have been widely applied in biological and epidemiological research, their use in road-safety studies remains largely unexplored. This paper presents one of the first applications of ABNs to human-driven crash data, demonstrating their suitability for modelling complex dependency structures in driving-risk analysis. Using 10,045 crash reports fromMontgomery County (Maryland, USA), we develop a complete ABN workflow comprising prior specification, score-cache structure learning, Laplace-basedmarginal likelihood computation, and robustness assessment through extensive Markov Chain Monte Carlo (MCMC) exploration. The final model identifies both direct and indirect associations among behavioral (distraction, substance use), environmental (weather, surface conditions), and contextual (speed limit) factors associated with crash severity. Results illustrate how ABNs can provide an interpretable graphical representation of conditional dependencies among crashrelated variables and can complement regression-based approaches by explicitly representing both direct and indirect statistical associations. This study provides a reproducible methodological template for ABN modelling and illustrates its potential for broader application in safety-critical and multivariate domains.