Crash Frequency Analysis Using Hurdle Models with Random Effects Considering Short-Term Panel Data.

Feng Chen,Lin Yang,Xiaoxiang Ma,Suren Chen

doi:10.3390/ijerph13111043

Abstract

Random effect panel data hurdle models are established to research the daily crash frequency on a mountainous section of highway I-70 in Colorado. Road Weather Information System (RWIS) real-time traffic and weather and road surface conditions are merged into the models incorporating road characteristics. The random effect hurdle negative binomial (REHNB) model is developed to study the daily crash frequency along with three other competing models. The proposed model considers the serial correlation of observations, the unbalanced panel-data structure, and dominating zeroes. Based on several statistical tests, the REHNB model is identified as the most appropriate one among four candidate models for a typical mountainous highway. The results show that: (1) the presence of over-dispersion in the short-term crash frequency data is due to both excess zeros and unobserved heterogeneity in the crash data; and (2) the REHNB model is suitable for this type of data. Moreover, time-varying variables including weather conditions, road surface conditions and traffic conditions are found to play importation roles in crash frequency. Besides the methodological advancements, the proposed technology bears great potential for engineering applications to develop short-term crash frequency models by utilizing detailed data from field monitoring data such as RWIS, which is becoming more accessible around the world.

Highlights

Traffic crashes cause a lot of occupancy injury and serious congestion on the road systems around the world
In light of the above discussion, this study explores the development of the random effect hurdle models for crash frequency prediction on fine time scales for the first time
The results show that many factors significantly influence the crash frequency on I-70, including time-varying variables and site-varying variables

Summary

Introduction

Traffic crashes cause a lot of occupancy injury and serious congestion on the road systems around the world. Lord and Mannering [1] as well as Mannering and Bhat [2], some methodology challenges encountered in current crash frequency studies were summarized, including temporal and spatial correlation, time-varying explanatory variables, and omitted-variables bias. Most of the existing crash frequency modeling methods use aggregated data in extended time scales (e.g., yearly or monthly), instead of fine time scales (e.g., hourly, daily) containing detailed time-varying information. The extended scales and aggregated variables lead to some limitations as summarized in references [1,2]. Some important explanatory variables in crash frequency models sometimes change quickly over time, such as weather, road surface conditions and traffic flow. By adopting extended time scales, some critical information over time of those important influencing variables Public Health 2016, 13, 1043; doi:10.3390/ijerph13111043 www.mdpi.com/journal/ijerph

Methods

Results

Conclusion