Crash prediction based on random effect negative binomial model considering data heterogeneity

In order to improve traffic safety, a large amount of works focusing on crash prediction and identifying factors contributing to crashes. However, the ignorance of data unobserved heterogeneity in some traditional models may lead to biased parameter estimation and erroneous inferences. To investigate the relationship between crash and the potential contributing factors, the crash data occurred in 3-year survey period on Interstate highways in Washington, including 134 fatal crashes, 13936 injury crashes, and 34,084 property damage only (PDO) crashes were collected. A data quality control method based on sensitivity analysis is used to determine the road segments. Then a negative binomial (NB) model and a random negative binomial (RENB) model are constructed to predict crash number. The inverse stepwise procedure was applied to examine the significance of explanatory variable. The horizontal alignment type, speed limit, visibility, road surface condition, and AADT are identified as significant factors on the crash. In the comparison, four standard errors are designed as indicators, and the results show that the errors of RENB model are lower than that of NB model. The comparing results illustrate that the RENB model outperforms the NB model in crash number prediction and safety service level prediction (C) 2019 Elsevier B.V. All rights reserved.