Crash Prediction Modeling Using a Spatial Semi-Local Model: A Case Study of Mashhad, Iran

The rapid expansion of road construction and ever-increasing growth of urbanization have led to increased number of vehicles. Achieving the safe trips without personal harm or property damage has always been the concern of safety specialists. Over the last few years, the safety researchers attempted to develop the innovative methodologies to explore the crash affecting factors and obtain practical models with high prediction power. Generalized Linear Models (GLMs) with Poisson or Negative Binomial distribution of errors are known as the most common techniques to investigate the relationship between crashes and the related factors. Such models assume the dependent variable (e.g. crash frequency or crash rates) to be statistically independent. However; spatial traffic accidents have the tendency to be dependent, a phenomenon known as spatial autocorrelation. Values over distance are more or less similar than expected for randomly associated observations. This study aims to develop a series of crash prediction models based on Traffic Analysis Zone (TAZ)-level crashes and contributing associated factors using GLMs and spatial semi-local Poisson-Gamma-CAR model. Trip generation variables will be employed as the surrogate variable for land use and demographic characteristics in models in addition to network variables and traffic volume. The significant Moran’s I as the spatial autocorrelation indicator performing on crash frequencies grouped in 253 TAZs in Mashhad, Iran and the same analysis for residuals of all models proved the reliability of spatial model over conventional GLMs. The spatial model also indicated an improvement in model performance as indicated by the set of goodness-of-fit criteria. The results of local analysis can provide a predictive tool at the planning-level which can be applied on different travel demand policies to evaluate their traffic safety impacts.