ARIA process-industry accident dataset: derived feature table and analysis pipeline
Description
This deposit accompanies the manuscript "What do industrial accident databases actually predict? Documentation depth, curation policy, and the limits of barrier-based machine learning in 46,394 process-industry accidents." CONTENTS - data/aria_features.csv - derived analysis table, 64,800 rows x 158 columns (147 predictors + 4 identifiers + 7 outcome/derived fields). UTF-8, comma-separated. - code/ - the complete Python analysis pipeline in execution order (feature engineering, logistic regression with barrier-block ablation, interaction tests, three-way temporal split with Platt calibration and permutation importance, France-only robustness, gradient boosting with SHAP, figures). - results/ - machine-readable JSON outputs of the reported analyses. - README.md - full provenance, column dictionary and reproduction instructions. SOURCE DATA AND ATTRIBUTION The underlying records originate from the ARIA database (Analyse, Recherche et Information sur les Accidents), maintained by BARPI, French Ministry of Ecological Transition, published as open data on data.gouv.fr (full export dated 7 July 2026; 64,800 records). Reuse of the ARIA records is permitted for non-commercial purposes with attribution to BARPI. Any reuse of this derived table must credit BARPI. ANALYSIS POPULATION AND DESIGN 46,394 process-industry (classified-installation, Seveso-type) records with valid dates. Strictly temporal validation: training up to 2015 (n = 32,937), calibration 2016-2018 (n = 3,687), held-out test from 2019 (n = 9,770). Binary outcomes: any human consequence (test base rate 12.43%) and major human consequence (test base rate 1.55%). HEADLINE FINDINGS REPRODUCED (1) Documentation drift: equipment documented in 30.7% of records before 2000 versus 84.7% after 2021, with the equipment-severity association reversing sign across periods. (2) Safety-barrier features add only +0.0060 ROC-AUC (any human consequence) and +0.0030 (major). (3) Major-consequence ROC-AUC falls from 0.8201 on the full sample to 0.5916 on French domestic events alone, indicating a curation artefact. (4) Class weighting inflates predicted probabilities roughly fourfold (Brier 0.2446 uncalibrated versus 0.1081 after Platt scaling). LICENCE Data: CC BY 4.0, subject to the upstream ARIA terms (attribution to BARPI required). Code: MIT.