Capturing Road-Level Heterogeneity in Crash Severity on Two-Lane Rural Highways: A Multilevel Mixed-Effects Approach

Accurately modeling crash severity on rural two-lane roads is essential for effective safety management, yet standard single level approaches often overlook unobserved heterogeneity across road segments. In this study, we analyze 19 956 crash records from 99 rural roads in Iran during recent four years incorporating crash level predictors such as driver age, education, gender, lighting and pavement conditions, along with road level covariates like annual average daily traffic, heavy-vehicle share and terrain slope. We compare three binary logistic frameworks: a single level generalized linear model, a multilevel model with a random intercept capturing latent road level effects (intraclass correlation = 21 %), and a multilevel model with random coefficients that allows key predictor effects to vary by road. The random coefficient model achieves the best fit in terms of deviance, AIC and BIC, and substantially improves predictive performance: classification accuracy rises from 0.62 to 0.71, recall from 0.32 to 0.63, and AUC from 0.570 to 0.775. Results from 200 simulation runs reveal notable variability in slopes for pavement and lighting variables, underscoring how local context influences crash risk. Overall, our findings demonstrate that flexible multilevel modeling not only enhances prediction accuracy but also yields context-specific insights to guide targeted safety interventions on rural road networks.