Crash severity modeling in transportation systems

Abstract

Modeling crash severity is an important component of reasoning about the issues that may affect highway safety. A better understanding of the factors underlying crash severity can be used to reduce the degree of crash severity injury, locate road hazardous sites, and adopt suitable countermeasures. In order to provide insights on the mechanism and behavior of the crash severity injury, a variety of statistical approaches have been utilized to model the relationship between crash severity and potential risk factors. Many of the traditional approaches for analyzing crash severity are limited in that they are based on the assumption that all observations are independent of each other. However, given the reality of vehicle movement in networked systems, the assumption of independence of crash incidence is not likely valid. For instance, spatial and temporal autocorrelations are important sources of dependency among observations that may bias estimates if not considered in the modeling process. Moreover, there are other aspects of vehicular travel that may influence crash severity that have not been explored in traditional analysis approaches. One such aspect is the roadway visibility that is available to a driver at a given time that can impact their ability to react to changing traffic conditions, a characteristics known as sight distance. Accounting for characteristics such as sight distance in crash severity modeling involve moving beyond statistical analysis and modeling the complex geospatial relationships between the driver and the surrounding landscape. To address these limitations of traditional approaches to crash severity modeling, this dissertation first details a framework for detecting temporal and spatial autocorrelation in crash data. An approach for evaluating the sight distance available to drivers along roadways is then proposed. Finally, a crash severity model is developed based upon a multinomial logistic regression approach that incorporates the available sight distance and spatial autocorrelation as potential risk factors, in addition to a wide range of other factors related to road geometry, traffic volume, driver's behavior, environment, and vehicles. To demonstrate the characteristics of the proposed model, an analysis of vehicular crashes (years 2013-2015) along the I-70 corridor in the state of Missouri (MO) and on roadways in Boone County MO is conducted. To assess existing stopping sight distance and decision sight distance on multilane highways, a geographic information system (GIS)-based viewshed analysis is developed to identify the locations that do not conform to AASHTO (2011) criteria regarding stopping and decision sight distances, which could then be used as potential risk factors in crash prediction. Moreover, this method provides a new technique for estimating passing sight distance along two-lane highways, and locating the passing zones and no-passing zones. In order to detect the existence of temporal autocorrelation and whether it's significant in crash data, this dissertation employs the Durbin-Watson (DW) test, the Breusch-Godfrey (LM) test, and the Ljung-Box Q (LBQ) test, and then describes the removal of any significant amount of temporal autocorrelation from crash data using the differencing procedure, and the Cochrane-Orcutt method. To assess whether vehicle crashes are spatially clustered, dispersed, or random, the Moran's I and Getis-Ord Gi* statistics are used as measures of spatial autocorrelation among vehicle incidents. To incorporate spatial autocorrelation in crash severity modeling, the use of the Gi* statistic as a potential risk factor is also explored. The results provide firm evidence on the importance of accounting for spatial and temporal autocorrelation, and sight distance in modeling traffic crash data.