Crash Prediction Modeling for Horizontal Curves on Two-Lane, Two-Way Rural Highways Based on Consistency and Self-Explaining Characteristics Using Zero-Truncated Data

Consistency and self-explaining characteristics play important roles in road safety performance, especially at rural highway curves. This study aims to take into account the effect of several contributing factors associated with consistency and self-explaining. The scope is limited to the roads based on the crash frequency that occurred at horizontal curves of Two-Lane, Two-Way Rural Highways (TLTWRHs). The main contribution is to simultaneously consider the traffic, geometry, consistency, and self-explaining variables, as novel parameters, for a set of 224 selected horizontal curves of TLTWRHs in Iran. The curves with existing at least one fatal crash in their history for the three-year period, from 2018 to 2020, have been selected as a case study. The collected data was zero-truncated and under-dispersion. The modeling process was carried out using the Poisson, Zero-Truncated Poisson (ZTP), and Conway-Maxwell-Poisson (COM-Poisson) regression models followed by analyzing the results. The results showed that the COM-Poisson regression model could effectively be used in the case of under-dispersed zero-truncated crash data and demonstrated that there are strong relationships between the crash frequency and the consistency variables: the ratio of curve radius to the average radius of the adjacent curves (as the alignment consistency variable) and the difference between expected and existing superelevation (as the vehicle stability consistency variable). Furthermore, the findings indicated enhancing the Field of View (FOV), as one of the self-explaining characteristics of the roads, is an effective low-cost approach for improving road safety on TLTWRH horizontal curves, compared to the other measures. Moreover, the results confirmed that constructing TLTWRH self-explaining horizontal curves is four times more effective than improving its consistency in terms of crash reduction meanwhile the curve self-explaining is 33% more effective than the superelevation improvement of TLTWRH curves.