Meta-analysis of crashes on two-lane rural highways using Bayesian Networks and Random Forest

The study used data for crashes that occurred on two-lane rural highways in Granada, Spain, over a four-year period (2016–2019). The obtained dataset consisted of 1,291 crashes. The data was used to identify the variables that influence crash injury severity. In this study, a hybrid meta-learning methodology based on Bayesian Networks (BNs) and Random Forest (RF) was used. To the authors’ knowledge, these methods have never been combined in a single analysis. BNs identified the main factors influencing crash injury severity and the categories of these factors that are most important for each injury severity. These factors were ranked in importance from highest to lowest using RF. It was found that the following variables significantly increase the risk of a fatality in a traffic crash: driver age, crash type, time of day, main crash cause according to police assessment, and lane width. In addition, it was found that the following categories of these factors were associated with a higher risk of fatality: driver age over 45, rollover crash type, nighttime lighting conditions, police reported driver distractions, and lane width of less than 3.25 meters. It became evident that using a hybrid meta-learning methodology based on BNs and RF was helpful in identifying the most significant factors, which influence the negative outcome of crashes. Local criteria that may evaluate the traffic safety of present and future highways should be developed and put into use in order to improve the level of traffic safety in two-lane rural highways, reducing the severity of injuries caused by traffic crashes and improving traffic safety.