A Multivariate Analysis of Crash and Naturalistic Driving Data in Relation to Highway Factors (2013)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

3Introduction This study focuses on a new challenge and a new opportunity in highway safety research. In the upcoming SHRP 2 naturalistic driving study (NDS), a large new database will be created with the potential to provide entirely new information about risk factors and highway crashes. While smaller data sets of a simi- lar nature have been created in earlier work—especially at the University of Michigan Transportation Institute (UMTRI) (LeBlanc et al. 2006) and the Virginia Tech Transportation Insti- tute (VTTI) (Dingus et al. 2006)—the efficient analysis of the SHRP 2 data will require the development of new and innova- tive methods. Although the SHRP 2 database will be the largest of its kind, the number of actual crashes seen in the study is likely to be small. Therefore, as part of the overall risk analysis, the formulation and validation of surrogates are major goals. The present study uses smaller existing data sets as testing ground for the statistical analysis of candidate surrogates. The major focus is on highway factors and the codependence of both crash events and surrogate events on these factors. The analysis to be presented relies on the integration of several sources of data: from naturalistic driving (as will be provided by the SHRP 2 Safety project), from historical crash data, and from databases of highway characteristics. These diverse data sets are related by spatial coincidence (“same highway segment,” generalizing to “same segment properties”) and in particular via reference to underlying road map data. This analysis approach comes under the heading of geographic information systems (GISs), where different layers of infor- mation are related via a suitable map-referencing system. Thus, the goal of this study is to develop a GIS-based analysis of crashes and crash surrogates related to highway variables to address many of the priority SHRP 2 Safety questions. The particular crash problem addressed in this study is that of road departure crashes. Road departure crash rates depend on multiple factors, prin- cipally those associated with human behavior and highway/ traffic conditions. Traditional analysis of crash databases can- not determine the influence of human behavior in any great detail, so the “missing information” is to be developed from naturalistic driving studies. All approaches explored in this study include some mapping or common reference for asso- ciating naturalistic driving with the occurrence of crashes, and define surrogates that typify physical mechanisms that lead to road departure crashes. Many possible events or conditions can be proposed as surrogates for crashes. The surrogates can be discrete events or continuous conditions that result in a crash in the extreme, or in a noncrash event that is necessary for crash occurrence. In this study the research team focused on devel- oping surrogates based on measures of lane-keeping control performance. Those measures started with relatively simple ones based on lane position and time to the crossing of a lane boundary but included more complex measures, such as a driver’s adjustment of the yaw angle of the vehicle to match that of the road. Two analytical methods developed in this study focused on the statistical relationship between surrogate measures of crashes and actual crashes and on formulation of exposure- based risk measures using the surrogate measures. The first approach is an extension of the traditional univariate response model for crashes to a model that treats crashes and crash surrogates as a bivariate response variable, incorporates a correlation structure between them, and can be extended to a Bayesian model. The second approach is based on extreme value theory and estimates the probability of events that are more extreme than any that have been observed. The spatially linked data also offer opportunities to examine driving behavior from different perspectives. A set of three small exploratory studies that were orthogonal to the main thrust of the project were also conducted. The first study examined the application of spatial statistics to the problem of determining if concentrations of crashes were really “hot spots” or if they could just be considered to be random groupings. The second study compared surrogate event rates in episodes of driving while on and off a cell phone. The comparison was made for the same drivers under similar conditions. The third study c h a p t e r 1

4and the extreme value model in Chapter 6. Full statisti- cal modeling is demonstrated in these sections for a set of the simpler lane-keeping surrogate measures. Chapter 7 focuses on the feasibility of defining and computing a sur- rogate based on control-oriented driving performance. The orthogonal studies are summarized in Chapter 8. The feasi- bility of transferring methods developed in this study for building spatially linked GIS databases by other researchers, with other data, in other areas is demonstrated in Chapter 9. Conclusions and implications for analysis of data from larger studies based on more extensive data sets, in partic- ular those from the SHRP 2 Safety program, are discussed in Chapter 10. Supporting literature review and technical details are in the appendices. compared the YRE of drivers through areas of lane widening and locations with uniform lane widths. Transferability of methods developed in this study was in part demonstrated by researchers from VTTI, using VTTI’s own naturalistic driving data (NDD) and highway and crash information from the state of Virginia. This report begins with a summary of hypotheses and research questions, from fundamental research hypotheses to detailed technical questions designed to validate the methods in Chapter 2. Data sources and the development of the spatially linked data used in this study are described in Chapter 3. Chapter 4 discusses crash surrogate measures. The statistical analysis methods follow with the bivariate response and Bayesian update model reported in Chapter 5