Statistical modeling of freight train derailments due to track-geometry defects: National trends and segment-level analysis

This study analyzes geometry-related freight train derailments at two scales using publicly available data. At the national level, a Negative Binomial model characterizes temporal trends in derailment rates by track class. Derailment rates per train-mile declined by approximately 5.5% per year between 2000 and 2024. Per-mile rates also declined as regulatory track class increases. At the segment level, a Zero-Inflated Poisson framework characterizes derailment frequency as a function of operational and environmental conditions. The higher train speed and traffic volume (MGT) are associated with lower derailment rates, which are interpreted as proxies for unobserved factors such as track quality, inspection frequency, and maintenance regime, rather than as direct causal effects. The segment-level model ranks segments by relative derailment risk without reliable absolute probability estimates. Modest predictive performance reflects three structural features of segment-level derailment prediction: available predictors being indirect proxies for the mechanisms causing geometry defects; conditional analysis on segments having experienced at least one derailment of any cause; and geometry-related derailments are statistically rare events. Operationally deployable risk prediction would require comprehensive exposure data for the full main-track network, segment-level inspection and maintenance records, and precursor event data such as near-misses and incidents below the federal reporting threshold.