A Comparative Analysis on Performance of Severe Crash Prediction Methods

Abstract

The objective of this paper is to compare the performance and tradeoffs between two alternative analysis methods for developing crash prediction models for severe crashes: a direct estimation of severe crashes using frequency models, and an indirect but popular approach of combining frequency of total crashes models and some form of severity distribution functions (SDFs). The researchers conducted a comprehensive comparison of these modeling methods to illustrate the strengths and weaknesses of each alternative, and to inform future research that intends to develop such models. An examination of the theoretical characteristics of the modeling approach is presented and discussed. The performance of the two modeling alternatives is compared using two different datasets. The results of those comparisons showed very similar performances by both techniques. Finally, a sensitivity analysis is presented to explore how the performance of these techniques vary by degree of dispersion and observed correlation levels of total and severe injury crashes (KAB; injury scale in which K = fatal [killed], A = incapacitating injury, B = nonincapacitating injury) with potential explanatory variables. The results from these analyses tended to favor the use of SDFs in combination with total crashes safety performance functions (SPFs), as the prediction tended to show reduced dispersion under most conditions. However, performance of the KAB SPF model outperformed the combination of SDF and SPF for total crashes when KAB and non-KAB crashes had a common predictor but with effects in opposite directions.