Incorporating the Influence of Vehicle Mix on Crash Frequency and Severity

Abstract

The current approaches for crash frequency and severity prediction in the Highway Safety Manual (HSM) do not employ vehicle mix information. In this research effort, we build advanced alternatives to HSM methods while incorporating vehicle mix information. Two model systems—(a) multivariate Poisson-lognormal model and (b) negative binomial-ordered probit fractional split model—are estimated by incorporating vehicle mix variables. The developed model systems can also capture the influence of observed and unobserved heterogeneity of different independent variables including vehicle mix variables. We estimate the models for three facility types including Urban Arterial 4-Lane Divided segments, Rural 3-Leg STOP-Controlled and Rural 4-Leg STOP-Controlled intersections using data from four Highway Safety Information System (HSIS) states including California, Illinois, Minnesota, Washington, and three non-HSIS states including Connecticut, Florida, and Texas. For modeling crashes at each facility level, we adopt a pooled modeling technique that accounts for state-specific observed and unobserved heterogeneity in the pooled datasets. A comprehensive set of independent variables including traffic volume, vehicle mix indicators, roadway characteristics, and state-specific indicators are considered in the analysis. The model comparison exercise is conducted based on a comprehensive set of quantitative and qualitative metrics. The study highlights how different methodological approaches perform better for different facilities. The study findings also underscore how capturing the observed and unobserved impacts of vehicle mix variables improves model performance in crash frequency and severity dimensions across the facility types.