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.
4C h a p t e r 2 The goal of the Safety Project S01 studies was to develop and demonstrate analytical methods that could be used with nat- uralistic driving data and to formulate additional research questions. As discussed in Appendices A and B, these research questions, together with those developed in Safety Project S05, were used to prepare a set of global research questions. S01 studies provided examples of the different analytical approaches that can be used with naturalistic driving data. Each of the individual S01 studies is reviewed below to pro- vide context for the recommendations of the present study. The S02 team also considered information on data type and quality in formulating the sample work plans to demonstrate the types of information that will be necessary in a successful S08 proposal. Four contractors contributed to S01: the University of Minnesota, the University of Michigan Transportation Research Institute (UMTRI), the Pennsylvania State Uni- versity Transportation Institute (PTI), and the Iowa State University Center for Transportation Research and Education (CTRE). These institutions developed and tested analytical methods for existing data sets that are similar to but much smaller than the data set to be developed under SHRP 2. University of Minnesota The University of Minnesota study focused on crashes and near crashes involving more than one vehicle. Researchers used data from the Virginia Tech Transportation Institute (VTTI) 100-car study and from two University of Minnesota projects: the Minnesota Traffic Observatory and the Coop- erative Intersection Collision Avoidance Systems project. The study produced three main outcomes for examining car-following and gap-selection crashes: it (1) identified an appropriate class of structural models for crash and near- crash events and the analytic tools needed for fitting these models to the data, (2) performed a counterfactual screening of near-crash events to determine their similarity to crashes, and (3) developed models of evasive actions that drivers take as a function of the situation. The researchers introduced a Bayesian approach to micro- scopic modeling of crash and near-crash events that accounted for vehicle kinematics, trajectories, and driver evasive actions. Several methodological conclusions resulted from the study. The first was that trajectory-based reconstruction of crashes and near crashes is feasible using vehicle and site data when the direction of travel is roughly constant. Examining crash and near-crash events from the perspective of a second vehicle is possible using a two-directional trajectories approach. The second conclusion was that research on the feedback loop between existing conditions and driver actions is both rec- ommended and necessary for an accurate microscopic traffic simulation. The third conclusion was that model estimation methods should be enhanced for serial correlation, which is critical for appropriately sized standard errors and confidence intervals for parameter estimates. The findings of this study demonstrated that site- and vehicle-based data can provide complementary results. Finally, the authors recommended that all aspects of such studies, including data collection setup, postcollection processing, and the storage and availability of data, be well documented to ensure the usability of the study data by future researchers. University of Michigan transportation research Institute UMTRI research focused on capturing the associations among highway factors, crashes, and driving behavior for road-departure crashes. The underlying hypothesis was that connections could be drawn between variations in continu- ous driving behavior (as seen in normal driving) and the dis- crete crash events recorded in crash databases. This focus was addressed using spatially referenced databases with geographic information system (GIS) tools and the concept of disturbed Review of Analytical Methods Proposed in SHRP 2 Safety Project S01
5 Departure and Curve Warning FOT. The goal of the project was to explore structural paradigms in order to better ana- lyze naturalistic driving data. Analysis of naturalistic driving data is inherently complex as a result of the obscure inter- actions between physical infrastructure and human factors. PTI explored how to address this complexity by using linear regressions, count regressions, and categorical and hierarchical Bayesian models tested at both the driver and the event level. For analyses at the driver level, drivers were treated individu- ally and also grouped by gender. Three events were analyzed for the event models: crashes, near crashes, and critical inci- dents. In addition, kinematic models were constructed using the kinematic data from UMTRI. PTI researchers considered regression models from both a frequentist and a Bayesian perspective to describe the benefits and limitations of each technique in the analysis of naturalis- tic driving data. A negative binomial model was constructed to evaluate the relationship between driver characteristics (e.g., gender, education level) and the probability of a ROR event. This is appropriate when events (crashes) are over- dispersed (i.e., the variance is larger than the mean). Zero- inflated Poisson or negative binomial models better account for the abundance of zeros in the data sets. A logistic model was constructed to evaluate dynamic and static driver factors (e.g., distraction, Dula Dangerous Driving Index score) and environmental factors (e.g., surface condition). It has been suggested that hierarchical models can capture driver differences over time and space, depending on how the data are clustered. The ability to model such driver dif- ferences would allow the classification of static and dynamic driver parameters to be treated as random effects. It has also been suggested that naturalistic driving data analysis would benefit from the use of hierarchical models when the param- eters are largely unknown. The frequentist approach may be hindered by sample-size limitations, however, and while driver, event, and context variables are known in the data, the relationships between these variables in crash modeling have not been well examined (Shankar et al. 2008). PTI researchers found that even with large data sets, rig- orous application of Poisson, negative binomial and zero- inflated Poisson, zero-inflated negative binomial, and other count regressions were needed. Main effects alone were insuf- ficient to generate consistent model parameters, and they pro- duced reduced goodness-of-fit statistics. Although including explanatory variables improved the fit of some data, the over- dispersion parameter in negative binomial models warrants further study. With naturalistic driving data, it is important that models integrate kinematic data along with event, driver, and context attributes. Hierarchical models offer some specific advan- tages given their flexibility and the relaxation of assumptions of probability distributions for dependent and independent control. Disturbed control was defined as any interruption or delay in the process of driver perception, recognition, judg- ment or decision making, and action. Bayesian multivariate generalized models, seemingly unrelated regressions (SURs), and extreme value theory were used to test this association. Naturalistic driving data from the UMTRI Field Operational Test (FOT), highway data from the enhanced Highway Per- formance Maintenance System (HPMS), and Michigan crash data were included in the analysis. The Bayesian and SUR models were applied to road- departure crashes that occurred on the right side and to three potential surrogates for this type of event: right-lane devia- tion, right-lane-departure warning, and time-to-right-edge crossing (TTEC). The analysis suggested that TTEC was the best surrogate, right-lane departure warning was an interme- diate surrogate, and right-lane deviation was the weakest of the three surrogates for right-sided road-departure crashes. Extreme value analysis was used to model rare events that lie outside the range of available observations. UMTRI research- ers modeled the TTEC variable on a specific roadway and determined that this variable may also be a good surrogate from an extreme value analysis perspective, a finding that warrants further investigation. They also found that the yaw rate error might be a better surrogate than TTEC for right- hand roadway departures. The yaw rate error generates a smooth, continuous, and differentiable data series even when a lane boundary crossing occurs. In contrast, TTEC includes discontinuities whenever the vehicle crosses the lane boundary. Yaw rate error strongly correlates with rapid steering interventions by the driver and may be a useful pre- dictor of degraded or ineffective lane keeping. The analysis of SUR models and the yaw rate error pro- vided evidence that disturbed control is a fruitful perspective for interpreting naturalistic driving data. In addition, these analyses demonstrated that roadway departure surrogates could be useful for future naturalistic driving studies. The researchers demonstrated that exposure should be based on instrumented vehicle traversals of directional road segments. The researchers noted that refinements to roadway data are needed; for example, HPMS roadway data are not directional and lack adequate curve information. They also noted that many innovative analysis methods can potentially be devel- oped to link crash, roadway, and naturalistic driving data. pennsylvania State University transportation Institute PTI examined the relationship between various precrash events and identified methodological paradigms that can be used to answer research questions specific to roadway departures. PTI used run-off-road (ROR) events from the VTTI 100-car study and data from the UMTRI Roadway
6they differ in ways that are considered unimportant for their analysis. Data mining methods were used to explore associations and patterns in events such as event sequencing, clustering, and association with future events. CTRE researchers constructed classification and regression trees to find associations between environmental, roadway, and driver factors for lane depar- tures in the UMTRI FOT data. The main advantage to data mining is that it can be used to identify relationships in the data that were not apparent a priori or that may not be found using methods that focus on linear relationships and simple combinations of predictor variables. Data mining can also use automated processes to evaluate large amounts of data. However, because many researchers are not familiar with data mining, it may be difficult to interpret results in a manner that will allow practitioners to incorporate the information into decision-making models, such as a model for comparing the costs and benefits of a particular countermeasure. Summary of Safety project S01 The results of the four studies undertaken in Project S01 pro- vide a basis for understanding what research questions can be posed using naturalistic driving data and how different analytical methods can be applied to these data. The studies demonstrate the use of Bayesian methods, logistic and SUR models, kinematic models, and microscopic event models. These very different approaches underscore the wide range of methods that can be used with naturalistic driving data. Identifying the most appropriate approach depends on the research question and the sampling plan selected for the data. Importantly, the different approaches govern both the types of questions that can be asked and the validity of the result- ing answers. The studies show that the approach and variables considered in the analysis can be complementary but also pro- duce potentially conflicting outcomes. These reports should serve as illustrative examples for S08 proposers and help guide them toward selecting the most appropriate approach from the wide range of available analytical techniques. variables. PTI also found evidence of driver adaptation to tech- nology, including changes in driver behavior with and without warning systems. Models need to take into account how driver behaviors change over time and in response to technology. Overall, the PTI results confirm that the simple statistical approaches commonly used in experimental and epidemio- logical studies fail to address the complexities of naturalistic driving data. Such failures in managing complex data can result in erroneous conclusions. Iowa State University Center for transportation research and education CTRE studies focused on lane-departure and ROR crashes. They defined the following crash surrogates for these types of crashes: nondeparture lateral drifts, nonconflict road depar- tures, and road-departure conflicts. The effectiveness of these surrogates depends on the potential hazard (i.e., what the vehicle will strike if the driver does not recover control), as well as the time to collision, time to lane departure, and time to hazard (e.g., oncoming vehicles, adverse slopes, and fixed objects). Vehicle kinematic signatures were used to define events (e.g., lateral acceleration, forward acceleration, and speed). The CTRE team developed three approaches to answering lane departure research questions using natural- istic driving data: data mining, logistic regressions and odds ratios, and timeâseries analysis. Using the three approaches, CTRE researchers estimated the odds of lane departures on the basis of a series of indepen- dent factors. For example, left- and right-side lane departures, respectively, were 10.9 and 29.2 times more likely to occur on curves with a very small radius (less than 200 meters) than on a tangent section. This analysis shows that to appropriately apply logistic regression to the full SHRP 2 NDS, the data must be aggregated to identify event and nonevent driving periods. This aggregation process is problematic with nat- uralistic driving data because all situations are unique and are only identified as being similar when researchers decide
