Georgia Road Safety Screening: Crash Data, Model Comparisons, Temporal Validation, and Geographic Transferability — Reproducibility Package

This repository provides the research datasets, analysis scripts, model evaluation outputs, and supporting reproducibility materials associated with the manuscript “Do Complex Crash Prediction Models Improve Road Safety Screening? Evidence from Temporal Validation and Geographic Transferability,” prepared for submission to Accident Analysis & Prevention. The study investigates whether complex machine-learning models improve roadway safety-screening decisions compared with conventional approaches, including historical crash-based rankings, negative-binomial safety performance functions, and empirical Bayes methods. Using crash and roadway data from Georgia, United States, the research evaluates screening performance under temporal validation and geographic transferability. The study includes historical crash observations from 2017–2019 and an additional evaluation period covering available 2024–2026 records. The analytical framework compares conventional statistical approaches with XGBoost and crash-history-aware XGBoost models. Performance is assessed through roadway safety prioritization metrics, including crash capture under constrained screening budgets, with attention to model fairness, uncertainty, and transferability across geographic areas and time periods. The repository is intended to facilitate transparent reporting, independent verification, and reproducibility of the study's analytical methods and findings. Supporting materials document data preparation, model evaluation, validation procedures, and the generation of research outputs, as available in the deposited package. Associated manuscript: Do Complex Crash Prediction Models Improve Road Safety Screening? Evidence from Temporal Validation and Geographic Transferability. Target journal: Accident Analysis & Prevention (Elsevier). Geographic coverage: Georgia, United States. Study periods: 2017–2019 and available 2024–2026 observations. Repository version: 1.0.0. DOI: https://doi.org/10.5281/zenodo.23959534 Data provenance and reuse: Users should consult the repository documentation for original data sources, licensing and redistribution conditions, software dependencies, and reproduction instructions.