Time-aware crash-severity modelling: STATS19 and CAS

Abstract

v1.3.0: Corrected analysis and verified clean reconstruction Release date: 2026-09-07 This version supplies the corrected computational scope for the Applied Sciences study. It supersedes earlier results for the current analysis without moving old tags, overwriting historical scientific artifacts or changing the input archives. Corrected scope STATS19 uses 15 features, excluding special_conditions_at_site and carriageway_hazards, with six split-local model sets and eleven evaluation cohorts. Models are Dummy, unweighted/weighted multinomial Logistic and weighted LightGBM. Random-reference LightGBM selection uses each split's own training/validation records, correcting overlap in the historical shared-selection procedure. The fixed-parameter exclusion-2020 sensitivity is retained. Ordered Logit, matched-subset comparisons and tree-count extension are omitted from the corrected primary scope; their audit history remains available. CAS retains its independently audited temporal workflow and post-hoc urban/ sparse-speed removals, followed by corrected split-local random references. The V2 catalog selects 20 compact result tables. CAS H2 exports only nine temporal-test contrasts; 90 superseded random contrasts remain in the preserved V1 history, not the current CAS H2 export. No scientific source result was changed by this export-only correction. Validation Author-run reconstruction in two new Windows/Python 3.13.15 virtual environments completed all 22 STATS19 and 25 CAS stages. All 20 selected V2 result comparisons passed with the frozen tolerances (atol=1e-10, rtol=1e-8); exact fields were compared exactly. No fitted models, processed data, predictions or SHAP arrays were copied. Nine existing raw snapshot files passed size/SHA-256 verification; they were reused rather than freshly downloaded in this run. There were 159 passing applicable unit tests and seven passing standalone CAS audits. The two original historical D14 exact-provenance assertions remain separate documented exceptions, not relabeled as passing tests. The independent bounded historical check passed. See docs/FINAL_REPRODUCTION_VALIDATION.md and logs/final_analysis/final_reproduction_acceptance.json for evidence and scope. This is author-run same-machine validation, not third-party or cross-platform validation. Manuscript layout was not part of this acceptance. The release preserves the original bytes of newly sealed protocols, result tables and acceptance records through explicit Git attributes, preventing automatic line-ending conversion from invalidating their recorded hashes. Reproduce this version Use Python 3.13 from the extracted release source directory: python -m pip install --requirement requirements-lock.txt python download_and_verify_data.py download --dataset all python download_and_verify_data.py verify --dataset all python run_final_analysis.py --stage reproduce --dataset all --workspace ../final-reconstruction To continue an interrupted run, repeat the last command with --resume. --stage check is an author-side audit requiring existing analytical artifacts; it is not the first command for a fresh clone. See FINAL_ANALYSIS.md. Interpretation and archive boundaries Corrections were made after historical outcomes were known; they are not a new untouched-test experiment or preregistration. H1 remains metric-dependent; Macro-F1 gains do not establish uniformly safer LightGBM performance. Failed safety constraints, the 1200-round search limit, label-recording changes, unverified spatial generalization and dataset-specific SHAP designs remain explicit limitations. SHAP is descriptive, not causal. Raw records, models and record-level outputs are excluded from the software archive. Fixed data archives are unchanged: STATS19: https://doi.org/10.5281/zenodo.22290566 (OGL v3.0). CAS: https://doi.org/10.5281/zenodo.22296725 (CC BY 4.0). The software concept DOI is https://doi.org/10.5281/zenodo.22231696. The version-specific DOI will be added to the GitHub release description and the main-branch citation metadata after Zenodo archives this tag. The tag itself will not be moved for this metadata-only backfill.