An integrated data-driven and causality crash severity model

Crash injury severity models are used to assess accident impacts, yet predictive models often lack causal insight, while causal models may compromise accuracy. Despite societal concerns over severe crashes, their rarity limits data-driven analysis. Therefore, we propose an integrated approach combining Bayesian neural networks with causal inference to enhance prediction performance and interpretability. This method identifies direct and indirect causal links, offering deeper insights into how various factors contribute to accident severity and their underlying mechanisms. It also improves our understanding of complex interactions between variables, resulting in more interpretable predictions for single- and multi-vehicle crashes in China. Our findings demonstrate that our model outperforms traditional methods, revealing key causal relationships and the impact of contributing factors. Furthermore, Bayesian network graphs illustrate the effects of these factors on injury severity probabilities across different crash types. These insights can inform traffic safety interventions and strategies to reduce accident severity.