Temporal heterogeneity in traffic crash delays: causal inference from multi-scale time factors and sample-wise structural decomposition

Traffic crashes remain a critical global public safety concern, exhibiting complex heterogeneity across multiple temporal scales that limits the timeliness and precision of current traffic safety management strategies. Modeling the underlying causation mechanisms across temporal scales presents several challenges, including weak sequential patterns and nonlinear interactions among temporal features. To address these issues, this research categorizes crash data into a macro-scale (annual seasonality) and micro-scale (daily peak intervals and daytime/nighttime variation), and proposes a Multi-Channel Feature Correlation Transformer (MCFformer) to systematically model the influence of multi-scale temporal factors on crash-induced delay. The model introduces a Multi-Channel Feature Correlation (MCFC) mechanism to capture inter-scale couplings, a Recurrent Retention Attention (RRA) module to enhance cross-sample nonlinear dependency modeling, and an attention-based causality interpretation approach to derive the dynamic contribution of each factor. Experimental results demonstrate that MCFformer significantly outperforms mainstream models (e.g., XGBoost and CatBoost) in crash delay regression tasks, achieving a 44% improvement in predictive accuracy and over 26% reduction in RMSE under weakly sequential, non-stationary conditions. Further analysis reveals that environmental factors exhibit higher influence during peak intervals and daytime/nighttime periods, with an average contribution of 55.6%, compared to 51.1% for built environment features. Conversely, in the seasonal dimension, built environment factors contribute 49% on average, exceeding that of environmental factors (44.6%). These findings highlight the scale-sensitive and structurally heterogeneous nature of crash causation, and validate the effectiveness of multi-scale temporal modeling. The proposed framework offers both predictive performance and interpretability, providing theoretical insights and practical guidance for the development of dynamic and refined traffic safety interventions.