Influence of segment length on the fitness of multivariate crash prediction models applied to a Brazilian multilane highway

Abstract

Road safety modeling enables the development of crash prediction models and the investigation of which factors contribute to crash occurrence. Developing multivariate response models is also valuable, but such models are currently under-exploited. Machine learning techniques, especially artificial neural networks (ANN), have been presented as possible alternatives. Furthermore, selecting a proper roadway segmentation is one of the first tasks in the standard crash modeling workflow. However, this is a challenging task, especially in terms of choosing a segment length. This article presents a study of the influence of segment length on the development of multivariate response models (i.e., three response variables: property damage only crashes, injured victims crashes, and fatal crashes). The models use ANN for a road segment of a Brazilian divided multilane highway. The highway to be modeled was divided into segments with 10 different fixed lengths. The model characterization included geometric and operational data available for the years from 2011 to 2017. The models were evaluated in terms of errors and by residual plot analysis. The 5-km segment of the northbound carriageway and the 4.5-km segment of the southbound carriageway presented the smallest errors and the highest values of R2. The residual analyses confirmed the trend to improve the model with the greater segment lengths. This was clear by the residues' distribution around zero, except for the output “Fatal crashes”. The better performance of the longer segments models was expected because these models aggregate more crashes into one segment. The reduction of no crash observations also facilitated the improvement of the models' goodness-of-fit. The use of ANNs also revealed its potential value. However, it is still important to seek strategies to deal with the excess of zeros in fatal crashes; a problem that also occurs in the traditional statistical modeling process.