Predicting Construction Safety Violation Recurrence Using Longitudinal Occupational Safety and Health Administration Enforcement Data

Construction safety prediction commonly addresses accidents or current hazards, whereas the timing of repeated violations after case resolution is less clearly characterized. This retrospective longitudinal cohort prediction study used public U.S. Occupational Safety and Health Administration enforcement records to construct 741,436 employer–standard observations from 2012 to 2025, including 16,002 observed same-standard recurrences. Follow-up began at a resolution-based anchor and was right-censored at five years or the administrative study end. CatBoost-Cox, XGBoost-Cox, and XGBoost-AFT combined inspection context, resolution information, pre-index history, and regulatory hierarchy. Models were trained on 2012–2018 anchors, validated in 2019, and compared in a 2020 benchmark that informed model selection. The selected 63-predictor XGBoost-Cox model achieved a 2020 C-index of 0.8062 (95% CI 0.7908–0.8200) and 3-year AUC of 0.8128 (0.7953–0.8290); recent AUC was 0.7988 (0.7884–0.8091). A separate highway holdout yielded AUC 0.7281 (0.6463–0.8080) among calendar-complete observations. A sensitivity analysis censoring development outcomes at 31 December 2019 retained AUCs of 0.8105 and 0.7948 in the two later cohorts. Section-level encoding provided a consistent predictive increment. The highest-risk decile contained 42.61% and 36.45% of recorded 3-year events, although absolute-risk calibration deteriorated. The contribution is a resolution-anchored framework linking standard-specific recurrence, unequal follow-up, and construction hazard history. A predictor-level audit could not confirm historical availability of the extracted inputs, leaving temporal leakage unresolved. The framework supports retrospective assessment; prospective use requires verified inputs, recalibration, and independent validation. The endpoint remains inspection-detected recurrence rather than unobserved hazards or a demonstrated preventive effect.