Code and Derived Results for Cross-State Transferability and Recalibration of Highway–Rail Grade Crossing Collision Models

This repository contains research code, derived results, fitted model coefficients, calibration summaries, and figure-generation scripts supporting the manuscript Transferability and recalibration of highway-rail grade crossing collision models across U.S. states: implications for safety screening, prepared for Accident Analysis & Prevention. The study evaluates highway–rail grade crossing collision-count prediction models across 44 U.S. states, using Federal Railroad Administration (FRA) historical crossing inventory and collision records. The reported cohort includes 189,982 eligible crossings and 7,248 linked collisions during the 2021–2024 evaluation period. The deposited scripts support Poisson and negative-binomial modeling, geographic holdout evaluation, local recalibration, shrinkage-based adjustments, collision-count diagnostics, and safety-screening comparisons. Reproducibility scope: The primary analysis requires locally prepared historical crossing inventory and collision extracts, which are not redistributed. The deposited code does not automatically reconstruct these extracts from raw FRA downloads. The 200 training-refit and rolling-origin analyses are included as stored outputs; their underlying model runs are not repeated by the primary scripts. Figures derived from those stored outputs can be regenerated. The repository provides public source references, input documentation, software dependencies, numerical outputs, and instructions for using the available scripts. The findings concern prediction and safety screening, not demonstrated reductions in collisions or injuries.