Abstract
Police enforcement constitutes a key element of traffic safety management efforts, yet existing studies have largely relied on system‐wide and temporally aggregated analyses, limiting insight into spatial heterogeneity and dynamic enforcement effects. This study develops an integrated framework to examine the dynamic relationship between police enforcement and crashes at fine spatial and temporal resolutions, with the objective of characterizing spatial heterogeneity, temporal dynamic, and persistence of enforcement effects. The framework integrates resolution diagnostics, dynamic time‐series models, and deep‐learning verification. Effective sample size and impulse‐response stability metrics are used to identify spatial and temporal resolutions that balance analytical precision with inferential reliability. Vector autoregression and structural vector autoregression models with impulse response functions quantify contemporaneous and lagged enforcement effects, while deep‐learning models, including encoder–decoder architectures, transformers, and graph neural networks, provide independent directional verification under flexible nonlinear specifications. The results identified a consistently short‐lived inverse citation–crash relationship across all segments, with substantial spatial heterogeneity and generally stronger enforcement responses in urban segments than in rural segments. The consistency between deep‐learning sensitivity measures and impulse response estimates supports the robustness of these findings. The results provide actionable and practical insights into spatially and temporally targeted enforcement strategies, supporting data‐driven police resource allocation to improve freeway safety.