Extending safety performance functions with connected vehicle data using a random-parameter negative binomial Lindley model with spatial effect.
Abstract
Severe crashes remain a critical roadway safety concern, yet traditional Safety Performance Functions (SPFs) primarily rely on traffic exposure and roadway characteristics, often overlooking driving behavior. Leveraging high-resolution connected vehicle (CV) trajectory data, this study derives behavioral measures, including hard braking, hard acceleration, radial acceleration, lane changes, operating speeds, and speed variability, and integrates them into an extended SPF framework for segment-level crash frequency modeling. Using data from the Lake-Sumter Metropolitan Planning Organization region in Florida, Bayesian random parameter negative binomial-Lindley models with spatial effects are estimated separately for two-lane and multilane segments for both severe (KA) and total (KABCO) crashes. Six alternative spatial weight structures, accounting for segment adjacency, length, and distance decay, are compared, with the second-order adjacency structure providing the best overall performance. Results indicate that CV-based behavioral variables improve model fit and predictive performance compared to traditional SPFs. Among the behavioral variables, radial acceleration on two-lane segments and lane changes on multilane segments are positively associated with KA and KABCO crash frequencies, respectively. Elasticity analysis identifies traffic volume (AADT) as the most influential factor across facilities, while over-speeding proportion (0-10%) and speed standard deviation show strong effects for two-lane KA and KABCO crashes, respectively, and speed standard deviation remains particularly influential for multilane KABCO crashes. By integrating CV data into SPF modeling, the proposed framework advances behavior-informed roadway safety analysis and offers practical insights for proactive safety management and prioritization.