SERIOUS HUMAN CONSEQUENCES IN SHIPPING ACCIDENTS: TEMPORAL DISCRIMINATION AND CALIBRATION USING TSB MARSIS DATA, 1995-2026

Abstract

Maritime risk management requires separating accident occurrence from consequence severity conditional on an accident. This study used the Transportation Safety Board of Canada public Marine Safety Information System (MARSIS) to examine serious human consequences among 9,491 reportable shipping accidents from 1995 through July 2026; 293 (3.09%) involved death, a missing person, or serious injury. Among accident types, the all-period occurrence-level logistic model showed the largest adjusted association for capsize (odds ratio, OR 10.36, 95% confidence interval, CI 6.76-15.88), followed by sinking/loss/abandonment (OR 2.74, 95% CI 1.79-4.19); moderate/adverse/ice sea state had OR 2.25 (95% CI 1.65-3.08). A historical model developed on 6,923 accidents from 1995-2014 retained discrimination in 2,568 later accidents (area under the receiver operating characteristic curve (AUC) 0.826; precision-recall AUC (PR-AUC) 0.237) but under predicted absolute risk: 93 events were observed versus 41.9 expected (observed/expected ratio 2.22). Connected-component resampling supported these conclusions. Post hoc timespecification checks predicted 61.3 events without calendar year and 67.7 with a linear year term, indicating that extrapolation of the original spline contributed to under prediction while not fully explaining it. The full historical model also materially outperformed an accident-type-only benchmark for AUC, PR-AUC, and Brier score. Expandingorigin refits using more contemporary data improved average calibration in later forward blocks. The results support transparent retrospective consequence-risk stratification, but not a fixed operational probability threshold: contemporary refitting, prospective evaluation, and a defined prediction time are required before deployment.