Development of Analysis Methods Using Recent Data (2012)

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.

64 C h a p t e r 6 Summary It may be helpful to recall the primary goal of this project: to develop and test analytic methods that could then be applied to data produced by the SHRP 2 Safety field studies. The gen- eral approach was first to identify interesting research prob- lems and then attempt to solve them using data similar to what is expected from the field studies. Chapters 3 to 5 illus- trated how Bayesian statistical methods can be used to fit mod- els to trajectory data, the parameters of these models being related to interesting event features. Although the empirical results presented in Chapters 3 to 5 offer some tantalizing suggestions regarding the relationships between crashes and conflicts, and the degree to which noncrash events provide information about crashes, the project team’s position is that the number of events analyzed is too small to justify drawing general conclusions. Such conclusions, of course, are what one hopes will result from the field studies. In this chapter, then, the discussion is restricted to methodological issues. In the work plan submitted at the end of Phase 1, three research problems were identified on which progress was needed. The first of these was identification of an appropriate class of structural models describing how crash and near- crash events developed, together with analytic tools for fitting these models to data expected from the vehicle- and site- based field studies. The second problem involved counter- factual screening of supposed near-crash events to determine their similarity to crashes, and in the Phase 1 report, a theo- retical result was developed, which indicated that those near- crash events that are most like crashes would have evasive actions more extreme than those used in crashes. The third problem involved developing plausible models of how driv- ers select evasive actions as functions of the situations in which they find themselves. In the Phase 1 report, it was indi- cated that given an adequately large sample of crash and near-crash events for which estimates of background condi- tions and driver actions were available, it should be possible to conduct exploratory modeling of evasive action selection using regression-type models. Solutions to the second and third research problems are contingent on a solution to the first, so the bulk of the effort during Phase 2 was devoted to structural modeling of crash and near-crash events. Chapter 2 of this report described a class of models that characterized driver behavior as a sequence of discrete changes in acceleration and illustrated how an ordi- nary differential equation taking this acceleration sequence as an input, together with initial speeds and positions, could be integrated to give a predicted trajectory for the vehicle’s motion. It was then illustrated how the parameters describing a driv- er’s acceleration sequence and the initial conditions could be estimated from vehicle trajectory data. It was also shown how the identified model and estimates could be used to address the second research problem by computing the probability a crash would have resulted, other things being equal, as a func- tion of a range of counterfactual evasive actions. Chapter 3 of this report described the application of the project team’s meth- ods to seven rear-ending crash and near-crash events obtained from the 100-car vehicle-based field study. Chapter 4 described application to six rear-ending crash and near-crash events using site-based video data obtained from the MTO, while Chapter 5 described a pilot application to an intersection angle conflict using site-based Doppler shift data obtained from the CICAS system. Conclusions Statistical analyses of crash frequency data can identify reli- able associations between crash experience and roadway or driver features but are of limited value in discovering how crashes occur. The first conclusion is that in situations where the direction of travel is roughly constant, trajectory-based reconstruction of crash-related events, where trajectory data are used to fit parsimonious models of driver behavior, is feasible using both vehicle-based and site-based data. The Conclusions and Recommendations

65 Table 6.1. Estimated Maximum Evasive Deceleration for Following Vehicles in 10 Events Analyzed in Chapters 3 and 4 Event ID Deceleration (ft/s2) Event Type 100-Car Vehicle-Based Data 99540 -12.6 Rear-end crash 104119 -9.5 Rear-end near crash 73082 -18.3 Rear-end near crash 104851 -21.8 Rear-end near crash 104283 -16.2 Rear-end near crash 60289 -10.7 Rear-end near crash I-94 Site-Based Data Oct 8 1600: vehicles 1 and 2 -8.0 Rear-end near crash Oct 8 1600: vehicles 2 and 3 -10.2 Rear-end crash Oct 13: vehicles 1 and 2 -10.5 Rear-end near crash Oct 13: vehicles 2 and 3 -12.9 Rear-end crash product of such a reconstruction is a set of estimates of when and by how much drivers changed their acceleration and the background conditions associated with these changes. These estimates can in turn be used to produce estimates of driver reaction times, following distances, and selected/rejected gaps. Bayes estimates, especially estimates of posterior prob- ability distributions, can be obtained using Markov Chain Monte Carlo simulation. This approach is especially helpful in studying crash-related events involving two or more vehicles, where information on the behavior of drivers in noninstru- mented vehicles is required. One goal of traffic safety research is identifying causes of traffic crashes. The notion of cause has historically been rather difficult to pin down, but a good case can be made that a com- mon core exists in what is meant by cause in statistical estima- tion of crash-reduction factors, in reconstruction of individual crashes, and in simulation of crash events. This common core is a counterfactual definition of cause (1). As indicated in Chapter 1, a counterfactual component is also present in working definitions of traffic conflict and near crash. Given a simple structural model of an event along with Bayes esti- mates of the posterior distribution for the model’s param- eters, it is possible to quantify the degree to which a near crash could have been a crash by perturbing a driver’s evasive action and computing the probability that a crash results. The second conclusion is that it is possible to extend the methods of counterfactual analysis to more complicated structural models involving differential equations. The Phase 1 report presented a theoretical argument to the effect that for a near-crash event to be similar to a crash event, the near-crash event should have an evasive action more extreme than that in the crash event. If crashes tend to involve extreme evasive actions, then this would imply that those near crashes that are most similar to crashes would tend to be less frequent than crashes. Although a conclusive test of this prediction would require more data than were available in this study, Table 6.1 hints that this might not be as big a problem as feared. The maximum evasive deceleration observed for the crash events was about -12.9 ft/s2, while the maximum evasive deceleration observed for a near-crash event was about -21.8 ft/s2, with the second-most extreme successful evasive deceleration being about -18.3 ft/s2. The third conclusion is that, at least for rear-ending events, there is some limited evidence that the distributions of evasive actions for crashes and near crashes share some overlap, so that it should be possible to find near-crash events that are similar in other respects to crashes. Vehicle-based data configurations have definite limits regard ing information provided about noninstrumented vehi- cles involved in multivehicle crashes or near crashes. Chapter 4 illustrated how site-based video with manual extraction of vehicle trajectory data can support structural modeling and counterfactual analysis of multivehicle events, but clearly this is not feasible for processing very large numbers of events that might be expected from a longitudinal study. The CICAS data-collection system, based on radar and LIDAR units, col- lects and processes large amounts of vehicle trajectory data at intersections. The current CICAS configuration and architec- ture is designed to provide available gap information to minor- approach drivers rather than to process and analyze data on crashes and conflicts. With some technical modifications,

66 however, this system has the potential to provide data needed for structural modeling of crash and near-crash events, at least at lower-volume intersections. Conclusions are only as reliable as the data upon which they are based, and as described in Chapter 3, in many cases, the vehicle-based data from the 100-car study showed incom- pleteness, inconsistencies, or errors that limited the ability to use them. These included cases where the forward radar data were missing or corrupted, where the speedometer data were clearly in error, and where there were marked differences in the vehicle trajectory as implied by the accelerometer measures and as implied by the speedometer and heading measures. It was also true for all cases studied that the GPS coordinates were not sufficiently refined to determine vehicle trajectories. It is recognized that the 100-car study was a pilot effort and that detailed quantitative reconstruction of events was probably not one of the study’s objectives. Nonetheless, the final con- clusion here is that the usefulness of the data produced by the SHRP 2 vehicle-based field study will be strongly dependent on the ability to calibrate and maintain the data-collection systems. recommendations for Future Work All but one of the crash and near-crash events analyzed so far show essentially straight-line trajectories for the involved vehicles. Although the project team has developed a prelimi- nary model code that allows for two-directional trajectories, it was not possible within the time and resource constraints to bring this to the degree of maturity achieved for straight- line events. Since two-directional motion occurs when a driver uses swerving as an evasive action and when a left- turning driver selects a gap, the first recommendation is for developing and testing trajectory estimation tools that handle two-directional trajectories. In the reconstruction of crash and near-crash events, driver inputs such as acceleration rates, reaction times, and follow- ing distances can be treated as exogenous quantities to be esti- mated, and the issue of how drivers select evasive actions does not arise. When attempting to include crash events in a micro- scopic traffic simulation, however, plausible models that close the feedback loop between existing conditions and driver actions are necessary. The second recommendation is for conducting research on this issue using the data from the SHRP 2 field studies. The Phase 1 report pointed out that in some cases the resid- uals obtained after fitting a trajectory model showed serial cor- relation. When serial correlation is present but unaccounted for, the standard errors and confidence intervals associated with parameter estimates suggest more precision for those esti- mates than is justified. That is, although in all the analyzed cases the trajectory models that were fit give reasonable descrip- tions of data, there may be greater uncertainty in the parameter estimates than have been so far acknowledged. The third rec- ommendation is that this project’s model estimation methods be enhanced to allow for possible serial correlation. The project’s experience with vehicle-based data indicated that except for car-following events where both vehicles remain primarily in the same lane, the quantitative information available about driver behavior in noninstrumented vehicles is essentially nonexistent. Compiling data on gap-selection and other intersection-related events will then require a dif- ferent data setup. The fourth recommendation, then, is that the vehicle-based study be complemented with site-based research. During the course of this project, a recurring factor was the unknown or uncertain influence of the measurement method or hardware, as well as the postcollection filtering, on the avail- able data. In all future studies, consistency and transparency in data collection and processing methodologies are paramount. Past efforts naturally were affected by the objectives and pri- orities of the projects funding, so that future dissemination of data may have received less emphasis. Considering that the pri- mary objective of SHRP 2 Safety Project S01 was to examine the usability of existing data, it is clear that the broader utility of data in a project be recognized. Therefore, the final recom- mendation is that, beginning with SHRP 2 and similar feder- ally funded projects, a clause be added to each RFP to make sure that the data-collection setup, postcollection processing, and storage and availability of information are clearly described in the final report. reference 1. Davis, G. Towards a Unified Approach to Causal Analysis in Traffic Safety Using Structural Causal Models. In Transportation and Traffic Theory in the 21st Century (M. A. P. Taylor, ed.), Elsevier Science, Ltd., Oxford, United Kingdom, 2002, pp. 247–266.