A comparative analytical framework for modeling road fatalities with count regression techniques

• Benchmarked 6 count models on 20,448 Thai road accidents in 2021. • Five-fold CV, Friedman and Vuong tests confirmed HNB’s superiority. • Two-part HNB separates fatal risk and counts for clearer interpretation. • Uncovered Thai misperception that fatalities peak during festive seasons. • Findings guide targeted policies for high-risk month, road, vehicle, and crash types. Regions with persistently high road traffic fatality (RTF) rates remain a major global safety concern. Using a nationwide dataset of 20,448 road accidents recorded in Thailand in 2021, this study compares 6 count regression models including Poisson, Negative Binomial (NB), Zero-Inflated Poisson (ZIP), Zero-Inflated Negative Binomial (ZINB), Hurdle Poisson (HP), and Hurdle Negative Binomial (HNB) to identify the best-fitting model for explaining RTF counts. Each model was evaluated using five-fold cross-validation across seven performance criteria, including AIC, BIC, Pearson’s χ 2 , dispersion, pseudo-R 2 , MAE, and RMSE. The HNB model achieved the lowest AIC (11,557.17) and BIC (11,819.05), the dispersion statistic closest to unity (1.0460), and the highest pseudo-R 2 (0.1479), confirming its superior goodness-of-fit and parsimony. The two-part HNB model separately estimated the likelihood of fatalities and the expected number of deaths given a fatal crash. In the zero part, festive months showed lower odds of fatality, while national highways and northeastern regions exhibited increased fatal risks. In the count part, fatal severity rose on highways and among larger vehicles such as cars and pickups. These findings reveal that road fatalities are driven by temporal, environmental, and vehicular factors. The study underscores the importance of robust count models for evidence-based road safety policy and targeted interventions.