The purpose of this study was to establish empirical relationships between truck accidents and highway geometric design. First, statistical frameworks based on Poisson and negative binomial regression models were proposed. Preliminary models were then developed using accidents and road inventory data from the Highway Safety Information System (HSIS). Three roadway classes were considered in the model development: rural Interstate, urban Interstate and freeway, and rural two-lane undivided arterial. The maximum likelihood method was used for estimation of model parameters. Information criterion, asymptotic t-statistic, and goodness-of-fit test statistics were employed to evaluate the estimated models. The model results based on data from one of the HSIS States--Utah, were used for analysis and for suggesting areas in which the quality and quantity of the existing HSIS data can be enhanced to improve the developed models.

Despite the limitations in existing Utah data, some encouraging preliminary relationships were developed for horizontal curvature, length of curve, vertical grade, length of grade, shoulder width, number of lanes, and annual average daily traffic (AADT) per lane (a surrogate measure for vehicle flow density). Goodness-of-fit test statistics indicated that extra variations (or overdispersion) existed in the data over the developed Poisson models for all three roadway classes. Subsequent analyses suggested that a future study can be performed to enhance the predictive power of these preliminary models by including detailed truck exposure information, e.g., time-of-day, truck type, and weather conditions, by considering more explanatory variables, such as roadside design and superelevation, and by reducing the sampling errors of vehicle exposure data (both AADT and truck percentages).