Vehicle and time specific crash modelling on selected rural highway curves using geometric and speed parameters: A transformed linear regression approach

Abstract

This study develops vehicle and time-specific crash rate prediction models for rural highway curves using high-resolution geometric and speed data. A 30km segment of State Highway-1 in Karnataka, India, encompassing 32 horizontal curves, served as the study site. Detailed data collection included 10 years of crash records, traffic volume count, LiDAR-based geometric features, and spot speeds recorded from laser speed cameras. Distinct models were built for motorized two-wheelers (MTW), passenger cars (CAR), heavy commercial vehicles (HCV), and for both daytime and nighttime conditions. The study offers a novel contribution by incorporating nighttime crash rate modelling rarely addressed due to challenges in data availability, and by developing disaggregated models for multiple vehicle classes. A backward stepwise regression (BSR) approach with square root transformation was employed, ensuring model transparency and interpretability. Sight-distance deficiency consistently emerged as the most influential predictor of crash rate, highlighting the critical role of visibility on curved segments. Validation through Leave One Out Cross Validation (LOOCV) confirmed acceptable predictive performance (R² = 0.43-0.80), with residuals exhibiting normal distribution. The findings underscore the importance of curve geometry and visibility in crash risk and provide actionable insights for design audits and safety interventions on rural highways.