Prediction of Occurrence and Severity of Run-off-Roadway Crashes on Rural Two-Lane Roadways Using Bayesian Networks

Abstract

Run-off-roadway (ROR) crashes are among the most common crash types on rural two-lane roadways. Current methodologies to predict their occurrence and severity by considering conditional nature and interactions between independent variables require complex mathematical procedures. This study employs Bayesian networks (BNs), a non-functional form graphical model, to determine factors associated with the occurrence and severity of ROR crashes. The study used five-year (2014–2018) crash data collected from 397 randomly selected road segments within Texas. Out of 397 segments, 279 did not experience ROR crashes. The first BN model used all 397 segments and explored factors associated with occurrences of ROR crashes. The second BN model used the remaining 118 segments that involved ROR crashes and focused on factors associated with different crash types (guardrail [GR], overturning [OT], and fixed object [FO] crashes) and their associated severity levels. Study results revealed that the presence of horizontal curves and utility poles within the clear zone on the road individually increased the chance of ROR crashes by about 35%. Moreover, FO crashes resulted in 36% more fatal and injury crashes than GR crashes, which showed the effectiveness of guardrails in reducing severity. This study also explored the combined influence of variables on ROR crash occurrence and severity, as well as the interrelation between several independent variables. The proposed methodology can be used to evaluate the effectiveness of countermeasures.