Analysis of Naturalistic Driving Study Data: Safer Glances, Driver Inattention, and Crash Risk (2014)

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.

102 C h a p t e r 9 This chapter focuses on answering the detailed research ques- tions first and then, the main research question. What are the most dangerous glances away from the road, and what are safer glances? Can risk from distracting activities (secondary tasks) be explained by glance behavior? The analysis started by replicating previous findings. The analy sis shows generally similar results that are consistent with previous findings regarding distracting activities and glance metrics. In general, distracting activities occurred frequently—much more frequently in crashes, near crashes, and baselines than impairments such as drowsiness. In line with previous naturalistic driving studies (e.g., Fitch et al. 2013; Klauer et al. 2006, 2010, 2014; Olson et al. 2009), visu- ally demanding tasks were associated with the highest risk. When considering crash and near-crash (CNC) situations combined, the results showed that the aggregate category of Portable Electronics Visual-Manual (OR 2.7, CI 1.4–5.2), and in particular, one individual activity in that category, Texting (OR 5.6, CI 2.2–14.5), had the highest odds ratios, suggesting a substantial risk. Talking/Listening on Cell Phone was found to decrease crash/near-crash risk significantly compared with not engag- ing in a phone conversation (OR 0.1, CI 0.01–0.7), represent- ing an estimated 10-fold reduction in risk compared with baseline (OR 10 if the sign of the coefficient is reversed). There were no crashes when drivers were Talking/Listening on Cell Phone. Odds ratios for more than 50 distracting activities were examined. However, many of the activities did not occur fre- quently enough to achieve statistical significance. Distracting activities do not occur as frequently as glances and thus need larger sample sizes. Other individual categories (e.g., Locating/ Reaching/Answering Cell Phone or Adjusting/Monitoring Radio) or other aggregate categories (e.g., Original Equipment or Vehicle External Distraction) were not significantly risky. To determine whether risk from distracting activities (sec- ondary tasks) can be explained by glance behavior, it was necessary to first find the most predictive glance metrics. Many Eyes-off-Path glance behavior metrics were found to be power- ful predictors of risk, much more so than the type of distracting activity (secondary task). The finding that glance behavior plays a key contributing role in crashes and near crashes is in line with existing research (e.g., Klauer et al. 2006, 2010, 2014). However, our analyses of separate glance metrics specified more strongly the benefits of using glance metrics to estimate risk. In general, the largest risk estimates were shown when crashes were analyzed separately from near crashes. Although very strong Eyes-off-Path–risk relationships were shown in separate glance metrics, the relationship between glance behavior and risk cannot be reduced to a single metric, as there is no separate metric that fully accounts for risk on its own. The relationship is analogous to accounting for discom- fort associated with heat. Temperature is a good metric that accounts for much of the variance, but including humidity would result in better predictions, as would including wind speed. Each glance metric helps inform the risk estimates. The most sensitive glance metric model was a linear com- bination of three-glance metrics as it was most predictive of crashes and near crashes. The first glance metric, Off3to1, rep- resents the proportion of time the eyes were off path from 3 seconds before until 1 second before the crash or minimum time to collision. The second glance metric is the mean dura- tion of off-path glances during the 12 seconds preceding the crash point or minTTC (mean.off). The third metric, mean uncertainty (m.uncertainty), is the mean value (in the same 12 seconds) of a composite measure based on the “uncertainty model” of the driving situation (Senders et al. 1967). However, if an individual metric is to be used, the Off3to1 metric was the second most powerful model after the linear combination of the three-glance metrics and much more powerful than the Conclusions and Recommendations

103 max.off and m.uncertainty metrics used individually. Note that the Off3to1 metric is equivalent to the metric Percent Road Center (PRC) (Victor et al. 2009). The PRC metric takes the percentage of on-road gaze, while the Off3to1 metric takes the proportion (0–1) of off-road gaze; thus, they produce directly comparable values (Off3to1 being the inverse of PRC). Thus, based on our results up to this point, the question, What are the most dangerous glances away from the road, and what are safer glances? can be answered in this way: the most dangerous and safest glances are quantified by the three- metric glance model, which combines a metric of inoppor- tune glance, mean glance duration, and a composite measure estimating the driver’s uncertainty of the driving situation. Returning to the question of whether this most sensitive glance metric model can explain the risk from distracting activ- ities, we found that it was substantially more predictive than the models based on distracting activities. Because the three- metric glance model was so superior, it might be expected to fully account for the effect of distracting activities (secondary tasks). However, the three-metric model could not be com- pared in a straightforward manner with distracting activities because the distracting activities were only coded in the 5 seconds preceding the precipitating event and 1 second after. Instead the best-performing glance model available at the precipitating event was used: the proportion of Eyes off Path in the 2 seconds overlapping the precipitating event. The aggregated category of distracting activities called Portable Electronics Visual-Manual was accounted for by the propor- tion of Eyes-off-Path metric. However, we also found that the risk-increasing effect of Texting and risk-decreasing effect of Talking/Listening on Cell Phone were not accounted for by that metric. This gives some indication that the crash/near-crash risk of neither Texting nor Talking/Listening on Cell Phone is fully determined by glances. Hence, properties of the activi- ties themselves further add to the risk. The present analysis does not further indicate what these properties might be, but cognitive load and motivational factors likely play a role in both Texting and Talking/Listening on Cell Phone. However, the influences of distracting activities might be explained by the more powerful three-metric glance model. This is clearly an interesting a topic for further research. On the basis of these results our answer to the question, Can risk from distracting activities (secondary tasks) be explained by glance behavior? is mixed and requires further research: Portable Electronics Visual-Manual was accounted for by the proportion of Eyes off Path in the 2 seconds overlapping the precipitating event, but Texting and Talking/Listening on Cell Phone were not. The finding that Talking/Listening on Cell Phone signifi- cantly reduces risk compared with not engaging in a phone conversation (OR 0.1) is in line with previous naturalistic driving studies (Olsson et al. 2009; Hickman et al. 2010) but shows an even stronger protective effect (lower OR). The present study extends these results in several ways. First, the number of crashes in Olson et al. (2009) was limited, and it has been suggested that the protective effect primarily occurs for near crashes and incidents rather than crashes. However, the present results point in the opposite direction. None of the crashes in the present study involved phone conversation and thus, increasing the proportion of crashes would be expected to reduce the odds ratio for combined crashes and near crashes even further. Second, previous studies found the protective effect only for commercial vehicle (mainly long-haul truck) drivers; the present results show that it also occurs for passenger- car drivers. Third, one criticism against the previous studies has been that the baselines were poorly matched with the crashes and near crashes. However, the present results demonstrate that the effect also occurs with a baseline sample of closely matched events. One important caveat is that the present study included only rear-end scenarios, and it is unclear to what extent these results generalize to other scenarios. The cognitive load induced by phone conversation may lead to an increased crash risk in other crash scenarios involving higher cognitive or executive functions (such as planning, decision making, novel sequences of action, or inhibiting habitual responses) rather than the more automatic reaction to a looming lead vehicle. For exam- ple, cognitive load may lead to impaired detection of red traffic lights or other types of signalized information in intersection scenarios. Indeed, one possible reason for the stronger protec- tive effect in the present study is that the protective effect is strongest in rear-end scenarios, which constitute a relatively large portion (about 30%) of all crashes. Thus, any increased risk due to cognitive load in less frequent crash scenarios may be washed out by the strong protective effect in the more fre- quently occurring rear-end scenarios. This hypothesis can be further investigated by examining the prevalence of cognitive load for different crash types in the SHRP 2 data set. Note that the protective effect of talking/listening found in the present data may not necessarily generalize to more severe crashes involving injuries and fatalities. So how can this strong protective effect for rear-end crashes be explained? One common suggestion is that the cognitive load induced by phone conversation counteracts drowsiness. This explanation seems consistent with previous findings that the effect mainly occurs for commercial vehicle drivers, in particular long-haul truck operations. However, in the present study, drowsiness was relatively rare in both baselines and crash/near-crash events. Thus, this explanation does not seem to account for the present results. Another possible explanation derives from the well- established experimental finding that phone conversation (and other cognitively loading tasks) induces a concentra- tion of gaze toward the road center. If this effect occurs in

104 naturalistic driving, the chances are greater that the eyes are on the forward path when a lead vehicle brakes, thus disabling the key mismatch mechanism behind Category 1 rear-end crashes (further discussed below). Indeed, Victor and Dozza (2011) found such a gaze-concentration effect in the 100-car study data. In the present data shown in Figure 6.5, Talking/ Listening was not associated with a general reduction in off- path glances. Thus, by contrast to the 100-car study, the present data do not lend strong support for a pure gaze-concentration explanation. However, if the glances to the phone associated with hanging up are excluded, the present data show at least a tendency for gaze concentration. Thus, subtle differences in the distraction coding may be one reason for the differences between these studies. Also, the 100-car analysis included all scenario types while the present analysis only included rear-end scenarios. Another related explanation is the task displacement hypo- thesis. Recent naturalistic studies suggest that drivers are on the phone about 10% of their driving time (Fitch et al. 2013). In our present matched baselines sample, the prevalence of Talking/Listening on Cell Phone was about 5%. Thus, because of its prevalence, phone conversation may displace or reduce engagement in other more risky activities such as texting, thus reducing the overall risk. The analysis of glance locations in Figure 6.5 clearly offers some support for this idea. A more specific version of this hypothesis is what may be called the glance displacement hypothesis. As shown in Fig- ure 6.5, the proportion of off-path glances was similar for matched baseline events with and without talking/listening coded as a distraction. Even if the phone glances are removed (as discussed above), the proportion of off-path glances in these events is still fairly high. The great majority of the glances presumably occurring during phone conversation are toward the left/right windshield and the left/right mirrors; almost no glances are to other secondary tasks (Figure 6.5). Thus, off-path glances while Talking/Listening on Cell Phone were mainly related to the driving task (e.g., road scanning and routine mirror checks while overtaking). This suggests that visual time sharing between driving-related activities is more “in pace” with the driving situation than visual time sharing with a secondary task. So, over and above a general gaze-concentration effect, replacing secondary-task glances with driving-related glances may increase the chance that the driver looks ahead at the critical moment when the lead vehi- cle brakes. In other words, driving-related glances may be less likely than secondary-task-related glances to combine with a sudden, unexpected, lead-vehicle closure. Therefore, the “per- fect mismatch” mechanism outlined in Chapter 7 is more likely to be disabled during talking/listening compared with other instances of normal driving. This hypothesized difference between driving-related and secondary-task-related glances may be further exacerbated by the fact that driving-related glances usually occur at a smaller visual eccentricity relative to the forward path than secondary-task glances. A small visual eccentricity increases the chance that a looming lead vehicle will be detected early in the peripheral field of view. This explanation appears to be the one that best fits the pres- ent data. However, more detailed analyses are needed to fur- ther examine this and other potential explanations for the protective effect. Factors other than Eyes off Path also contribute to rear-end crashes. First, factors such as age (16–17 years and 76+ years) and visibility problems (visual obstructions or rain) may influence risk. Second, the present sample contains a signifi- cant proportion of crashes and near crashes in which off-path glances were not the main contributing factor, such as when the driver looked at the road for the entire 12 seconds before the crash or only looked away at the beginning of the event. These crashes and near crashes were not systematically investigated in the present study. However, it may be that false expectations, small headways, and/or inadequately performed avoidance maneuvers may be key contributing factors. Clearly, further analysis is needed. How does the timing of lead-vehicle closing kinematics in relation to off-road glances influence crash risk? Given the present and previous findings on the importance of glance timing (e.g., Liang et al. 2012; Victor and Dozza 2011), we set out to study off-path glance timing with lead-vehicle closing kinematics and visual cues. This analysis revealed a distinct mechanism for many of the crashes. In line with Tijerina et al. (2004), we found that drivers in most cases did not look away when the lead vehicle was closing. Drivers who crashed typically looked away just before the lead vehicle started closing and did not look back until collision was unavoidable. This implies that, while drivers generally self- regulate their off-road glances based on expectations of how the situation will develop, their self-regulation is not always effective because expectations are sometimes violated. The criticality when looking back, and thus the crash risk, is largely determined by an interaction between last glance duration and the rate at which the situation changed during the glance (operationalized here in terms of inverse TTC change rate). The event outcome is also determined by the vehicle’s braking capacity and the driver’s time to react. Thus, the key mechanism behind these types of rear-end crashes (grouped as Category 1, Inopportune glance, in Fig- ure 7.8) can be understood as a perfect mismatch between last glance duration and situation change rate (in line with the general mismatch conceptualization of inattention suggested in Engström, Monk et al. 2013). The crashes grouped under Category 2, Looking away in an already critical situation,

105 followed a similar pattern. But here the driver looked away when the vehicles were already closing, often because of visibility problems that presumably impaired looming detection. The probability for the type of mismatch characterizing Category 1 and, to some extent, Category 2 crashes depends on the joint probability distributions of glance durations and situation kinematics. Since long glances are rare, many crashes occur due to the combination of a relatively short glance and a high change rate. Thus, an important finding of the present analysis is that glances that lead to crashes may not necessarily have to be long. The majority of the crashes in the present sample were associated with glances shorter than 2 seconds. This key finding motivates reconsideration of the question, What are the most dangerous glances away from the road, and what are safer glances? One way to think about which glances are safer than others is in terms of the boundary drawn in Figure 7.8. Under normal conditions (e.g., a dry road surface, normal braking capacity, normal visibility conditions), glances can be regarded as safe as long as they appear under this line, which is determined by the interaction of glance duration and kinematics change rate. Thus, the answer to the first part of the question can be reformulated like this: Dangerous glances are those during which the driver is exposed to the risk of a rapidly changing situation. This is naturally partly related to the glance dura- tion; the longer the glance, the greater the probability that the kinematics will develop in such a way that the perfect mis- match occurs. However, the second part of the equation is the natural variability in vehicle-following situation kinematics. Drivers are normally successful in controlling this variability by means of anticipation. However, as shown in the present analysis, the safety margins adopted by drivers when looking away are often insufficient to protect them from unexpected rapid changes in situation kinematics. A reformulated answer to the second part of the question can be stated as follows: An off-road glance is only perfectly safe when the safety margins adopted are sufficient to protect the driver if the situation changes rapidly during the glance. An important further question is this: What made the driver look away at that critical moment? It is likely that off-path glances further “upstream” in the chain of events leading to the crash may induce misunderstandings of the situation that influence the decision to initiate the critical last glance. These types of upstream effects of off-path glances have not been addressed in the present analysis and constitute an important area of further research. What crash severity scale is best suited for analysis of risk? The analyses to answer to this research question resulted in the formulation and proposal of two potential severity scales: Model-estimated Injury Risk (MIR) index and Model- estimated Crash Risk (MCR) index. These scales were created using mathematical simulations, applying a model of driver glance behavior to kinematics based on actual crashes and near crashes, and they represent what might have happened had the event played out according to a specific driver model. The two scales (MIR and MCR) provide continuous values and can be calculated for actual crash and near-crash events. However, further work is necessary to validate the scales. Note that the severity scales are simulated. Actual severity scales Delta Velocity (DeltaV) for crashes and minimum time to collision (minTTC) for near crashes are still the most relevant metrics when ana- lyzing actual severity (what actually happened in the event), and analyses of the SHRP 2 crashes and near crashes should use them. The main drawback with the actual severity scales is that they cannot be used to compare both crashes and near crashes. This property—the ability to compare potential severity across crashes and near crashes—is enabled by the proposed MIR and MCR scales. It is important to note that these scales are also enabled by naturalistic data. Without the detailed time- series data leading up to crashes and near crashes, the MIR and MCR scales could not be computed. How can we change glance behavior to be safer, and how do the results of this research translate into countermeasures? The findings of this research have several implications for countermeasures. Based on the general mismatch mecha- nism identified in the present analysis, countermeasures for rear-end crashes may be considered in terms of two general aims: (1) reduce the risk for the occurrence of mismatches between off-path glances and changes in situation kinematics and (2) if such mismatches still occur, maximize the chance of recovery. Human-machine interaction design, distraction guidelines, and other regulatory agency countermeasures HMI design and performance guidelines, such as the ESoP, AAM, and JAMA guidelines, and standards such as ISO (see Regan, Lee, and Young 2008) are important tools that should be used at different stages in the user-centered design process to support the safe design and evaluation of vehicle human- machine interfaces for driving-related systems. HMI design guidelines provide design specifications for installation, infor- mation presentation, interaction with displays and controls, system behavior, and information about the system. HMI performance guidelines set out the minimum level of perfor- mance that a system must meet when tested in accordance with a prescribed test method.

106 Current HMI guidelines provide a mix of design and perfor- mance guidelines. Problems associated with current guide- lines, as well as standards developed around the world, were reviewed in Regan, Lee, and Young (2008). Regan, Victor et al. (2008) indicated that the main concern with current guide- lines and standards is a lack of scientific knowledge to provide unequivocal assessment and robust compliance criteria for performance testing. Ideally, the approach to develop and implement performance standards for vehicle electronic devices would prescribe “practical, repeatable methods to measure the distracting effect of these devices and reliable benchmark levels of unacceptable performance” (Regan, Victor et al. 2008). These should be consistent for original, aftermarket, and portable devices and should not stifle product innovation. An aim of this project was to address that lack of scientific knowledge. The main recent development regarding HMI guidelines is the NHTSA Visual-Manual Driver Distraction Guidelines for In-Vehicle Electronic Devices, released in 2013. It contains a set of design guidelines and acceptance criteria for perfor- mance testing with the aim of minimizing visual-manual dis- traction and promoting nonvisual means for interaction. The evaluation methods in these guidelines are based on previous naturalistic driving study reports, such as Klauer et al. (2006) and Olson et al. (2009). In particular, they argue that in-vehicle systems which cannot be used without taking the eyes off the road for more than 2 seconds at a time are inappropriate for use while driving. The 2-second limit is included in several of the guideline performance criteria that must be met by in- vehicle systems to be considered safe to use while driving. The 2-second limit for what constitutes unsafe off-road glances is also central to the Alliance of Automobile Manufacturers design guidelines. Note that the present results are based on Lead-vehicle pre- crash scenarios and may not transfer to other precrash types. However, given these results, some observations can be made that may have relevance for distraction guidelines, as follows: • The results clearly emphasize the importance of designing interfaces that minimize the need for visual interaction. Eyes-off-Path glances are strongly associated with crash risk, near-crash risk, and the combination of both. • The results show that off-path glances leading to rear-end crashes are most often due to visual interaction with porta- ble electronic devices rather than vehicle-integrated sys- tems. Thus, efforts should focus on minimizing Eyes-off-Path glances with portable electronic devices. • The risk-reducing effect found for Talking/Listening on Cell Phone seems to support the potential for nonvisual inter- faces to enable safe interaction. Future research is needed to investigate whether the present findings generalize to other scenarios and to other nonvisual, but cognitively loading tasks, such as voice interaction. In addition, the present results emphasize the need to confirm that such interfaces are indeed nonvisual. • The results show that the majority of crashes were associ- ated with relatively short glances. This is likely due to their higher frequency; thus, based on the present results, there is a higher likelihood of mismatch with external events. These results indicate that eliminating long glances (e.g., glances above a limit of 2 seconds) will not eliminate the problem. Rather, the results clearly demonstrate that inopportune glances of normal duration with the wrong timing relative to high lead-vehicle closure rates often produce rear-end crashes. HMI design should thus also minimize occurrence of shorter glances. • The present analyses were not specifically designed to answer the question of whether or not the NHTSA 12-second Total Eyes off Road Time (TEORT) limit for distracting activities (secondary tasks) is supported. Fully addressing this issue would require coding each distracting activity from start to end. The present analysis used fixed 6-second or 12-second windows; thus, the coded data excludes cases that would vio- late the 12-second NHTSA limit. However, the present data can be examined to see if they support the general notion of the NHTSA 12-second TEORT limit, or if crashes are associ- ated with more TEORT. Several sources of results need to be examined. First, the analysis in Section 6.2 showed that odds ratios are highest for the proportion of time the eyes are off the forward path during the window from 3 seconds to 1 second before the crash point (Off3to1), but the odds ratios in the two preceding windows (5 seconds to 3 seconds, and 7 seconds to 5 seconds) also achieved statistical significance. Note that significance in those two preceding windows can be present even if they do not provide a significant contribution beyond that of the Off3to1 (the extreme case would be if these variables were perfectly correlated with Off3to1). The research team analyzed whether extended periods of eyes off the road in preceding windows might add to the risk. The results showed that there was no significant cumulative effect from the proportion of Eyes off Path in the windows preceding the Off3to1 variable. Thus, this analysis showed that risk was not associated with more Total Eyes off Path Time (TEOPT) than the 3 seconds to 1 second before the crash point. Second, the analysis of glance characteristics showed that in addition to the Off3to1 variable, the mean single glance dura- tion during the 12-second window (mean.off), and the mean level of uncertainty (m.uncertainty) during the 12-second window were together the best predictors of risk from glance data. Thus, although cumulative TEORT (in the time period before 3 seconds before the crash) was not supported as a more predictive indicator of risk, other characteristics within

107 the 12-second window (mean single glance duration and mean level of uncertainty) were supported. Third, when examining the joint probability of lead-vehicle closure change rate and eyes off road (Chapter 7), a strong relationship was shown in the mismatch mechanism. The main determinant of risk appears to be the probability of mismatch, because fast change rates and short glances are associated with crashes, as well as long glances and slower change rates. Thus, although there does not appear to be a risk contribution from an accumulation of TEORT and the risk can be pinpointed in the mismatch mechanism, it can be argued that risk increases with a total exposure from the amount of eyes off road over time (as an increase in joint probability). Exposure to the joint probability of mismatch is affected by a number of factors, such as the driver’s choice to frequently perform distracting tasks (e.g., texting), the type of traffic, and whether or not active safety systems are reducing the lead-vehicle closure change rate. These results indicate that the conclusions drawn here would not have been different if the entire lengths of secondary tasks had been coded from start to finish. Vehicle design and driving support The results provide strong support for the potential of active safety systems—such as autonomous emergency braking (AEB) systems, forward collision warning (FCW), autono- mous cruise control (ACC)—as main countermeasures for inattention-related crashes. Active safety systems provide the safety margins needed to protect the driver if the situation changes rapidly during an off-path glance by creating more time headway, issuing warnings, and actively braking. As shown in Figures 7.3 and 7.4, drivers often adopt small headways in vehicle-following situations and do not seem to adapt the headway much when looking away from the road. Figures 7.3 and 7.4 also show that drivers looking away from the road almost never crash if the time headway at the begin- ning of the last glance is larger than 2 seconds. Thus, in theory, a simple way to prevent the crashes in the present sample is to increase the time headway. The European Field Operational Test (euroFOT), an EU-funded project, demonstrated that adaptive cruise control has the potential to significantly increase the average headway adopted by drivers. In terms of the present mismatch framework, this would reduce the risk of unexpected, high closure rates. Given the present results, such an increase in headway would be expected to have strong safety benefits (as long as the situations in which the adaptive cruise control is used overlap with those in which crashes occur). As discussed in Chapter 7, the boundary suggested for safe glances (Figure 7.8) depends critically on the braking perfor- mance of the lead vehicle. In many situations, a slight differ- ence in the vehicle’s braking performance may distinguish a crash from a near crash. A short glance may be perfectly safe under normal road surface conditions but may induce a crash if the stopping distance is increased due to a wet road surface or an ineffective braking system. As mentioned above, this was precisely the mechanism behind the single crash in Figure 7.8 not assigned to any of the three categories. Thus, improving vehicle braking systems should have great potential for miti- gating the effects of inopportune glances. The same holds for AEB systems. The results on driver reactions show that drivers react to progressively faster invTTC values above 0.2 and do not react to values below 0.2. An accumulator model successfully predicted driver reactions, and this model could likely be developed to improve the activation of warnings and interventions to better correspond to driver reaction times. Systems for real-time inattention detection and mitigation have great safety potential, and development within inatten- tion monitoring is a priority (e.g., NHTSA 2010a). Inattention sensing and support is being developed to detect drowsiness and distraction. Driver inattention sensing can be used for a wide range of functions and driver feedback (for a review, see Engström and Victor 2008). Driver assistance can be provided either by information and warnings or by adapting vehicle collision avoidance functionality, depending on detection of the risky inattention mismatch situations that were identified in the present research (see Figure 7.8). The project results are directly relevant for and have the potential to greatly improve real-time algorithms used for distraction and inattention detection (e.g., NHTSA 2013). The results are very relevant for the design of forward col- lision warning (FCW) algorithms. Most FCW algorithms in use are based on a kinematic threat analysis, using some kind of time-to-collision–based threshold for when to warn the driver. Because this approach has difficulty separating late but controlled hard braking from hard braking in response to an unexpected event, nuisance alarms are common. Many researchers have therefore envisioned that the next step for FCWs in terms of increasing their hit rate (i.e., reducing the rate of nuisance warnings) would come through including in the warning algorithm assessments of driver states—that is, make it an inattention-adaptive collision warning. The present results provide an enhanced understanding of the crash-producing mechanisms targeted by FCW, which can be used to improve the sensitivity and specificity of collision warning algorithms and to enable inattention-adaptive algo- rithms. An alternative approach, which is worth pursuing, is to tune FCW to warn more exactly when the risk is greatest according to the present results. For the analyzed events, it will be possible to establish how well existing FCW algorithms would have done in terms of alerting the driver. It will also be possible to devise principles for how to further optimize FCW algorithms in terms of warnings given and warning timing,

108 based on the driver’s attention state and the traffic environ- ment. For example, in some cases with relatively low kinematic threat level, a warning might still be warranted if the driver is very distracted; inversely, an attentive driver might not need a warning even if the kinematic threat level is high (for more information, see Engström and Victor 2008). The present results also indicate a strong potential for vehicle-to-vehicle (V2V) communication-enabled FCW func- tions to prevent many of the crashes in the present sample. Video observation revealed that a common type of precrash scenario leading up to these crashes is the rapid build-up of a traffic queue in front of the lead vehicle. The traffic queue, however, is often hidden from the driver by the lead vehicle, so the driver often judges that it is safe to look away even if the threat is already present. If the driver had been alerted to the situation early by means of V2V communication with a vehi- cle in the traffic queue, many of these crashes would very likely have been prevented. Education and behavioral change As noted above, the present results indicate that a behavioral change toward longer time headway might have prevented a major portion of the present crashes. However, motivating drivers to make a sustained behavioral change is notoriously difficult. One way to address this is by means of public aware- ness campaigns. Another is behavior-based safety (BBS) pro- grams, which are relatively well established in the commercial transport domain and have proven very efficient for inducing long-term behavioral improvement (e.g., Hickman et al. 2007). Such programs have also been successfully applied with teen drivers (Carney et al. 2010). BBS programs of this type are naturally not well suited to private motorists. However, one way to influence the behavior of this group is through usage-based insurance, which is a rapidly growing business linked to the general Big Data trend. For example, points or rewards could be given for keeping longer time headways. Whatever the means for influencing drivers’ behavior, a key implication of the present research is that the adoption of safe headways should be a key target in behavioral change programs for road safety. The results identify the types of glance behaviors and contexts that are particularly dangerous and also identify safer behaviors and contexts. Thus they can be used for education, outreach, training, and licensing to teach and inform drivers how to behave in a safer manner. Road and infrastructure design Road and infrastructure design is of key importance both for preventing glance–traffic situation mismatches and for maxi- mizing the chance of recovery in case they occur. First, as discussed above, drivers’ glance allocation is strongly determined by expectations, and attentional mismatches typically occur when expectations are violated. Designing the infrastructure layout in a way that supports the development of correct expectations and minimizes the risk of misunderstand- ing situations should have a strong potential for preventing rear-end crashes. While the present research has focused on mechanisms related to last-second reaction failures, under- standing how expectations shape visual behavior and how this relates to infrastructure layout is an important topic for further research. The concept of self-explaining roads, coined by Theeuwes and Godthelp (1995), is a good starting point. Second, addressing the other side of the mismatch equa- tion, improvements in traffic control that aim to reduce dis- ruptive traffic flows have potential for reducing the prevalence of sudden, unexpected, kinematic changes that often combine with off-path glances in producing rear-end crashes. Third, as discussed above, the boundary for safe glances (Figure 7.8) depends critically on the stopping distance, which, in turn, depends on both the braking system and the road surface. Thus, by the same argument as for vehicle braking systems, improving road surfaces and their maintenance should have great potential for mitigating the effects of inopportune glances. Assessing the potential efficiency of countermeasures As described above, the present research has yielded several novel insights with respect to countermeasures for rear-end crashes. However, due to the complexity of the mechanisms involved, the potential efficiency of these countermeasures is difficult to assess. Computer simulations based on the present type of natural- istic crash and near-crash data have great potential for obtain- ing more accurate estimates of the safety benefits of crash countermeasures. In particular, the level of detail with respect to crash kinematics and driver behavior available in these data makes it possible to reconstruct the crash scenario in simula- tion and ask the question, What if things had been different? One variant of such what-if simulations (focusing on the effects of different last-glance distributions) was presented in Chapter 8. The SHRP 2 data were also used to develop driver reaction models for use in such simulations, as described in Chapter 7. The severity scales and simulation models described in Chapter 8 will also be very useful for addressing not just crash prevention but also the injury reduction potential of different countermeasures. What is the relationship between driver inattention and crash risk in Lead-Vehicle Precrash Scenarios? The answer to the main research question can be found in the general pattern of the results. They show, in line with previous

109 naturalistic driving studies, that some activity types signifi- cantly increase risk (such as Texting and Portable Electronics Visual-Manual). However, a strong significant decrease in risk was found for Talking/Listening on Cell Phone. Notably, there were no crashes while the driver was talking/listening on the phone. Three types of glance metrics showed the largest odds ratios: (1) the proportion of time the eyes were off path between 3 seconds and 1 second before the crash or minimum time to collision, (2) mean duration of off-path glances, and (3) the mean value of a composite measure estimating the driver’s uncertainty of the driving situation. Further, when the three-glance metrics were combined in a glances model, they were more predictive of crashes and near crashes than each metric individually. However, an important limit of the glances model that included the three-glance metrics is that it assumes risk is purely a function of the driver’s attention to the road (Eyes-off-Path metrics). Risk stems more precisely from both the driver’s attention to the road and the demands of the road. Analyses of the tim- ing of off-path glances with lead-vehicle closing kinematics and visual cues revealed a distinct mechanism behind most of the crashes that can be understood in terms of a perfect mis- match between last glance duration and the change rate of the lead vehicle closing. Crashes occur with short glances and high closure rates, just as crashes occur with long glances with slow closure rates. These mismatches can be understood in terms of a joint probability distribution for glance durations and closure rates where the most likely combinations will show up in a crash sample like the present one. Since long glances are rare, many crashes occur because of the combination of a rela- tively short glance and a high change rate. Another group of crashes followed a similar pattern, but in those cases the driver looked away when the vehicles were already closing, often because of visibility problems that presumably impaired looming detection. This pattern of results, or mechanism, was further confirmed in what-if simulation and modeling of reaction time. The main pattern is that lead-vehicle crashes can be under- stood as the mismatch between glance duration and the lead-vehicle closure rate. Timing matters greatly, and taken together, the analyses strongly reflect this mechanism. Limitations This study is very specific as it examined only Lead-Vehicle Precrash Scenarios (rear-end crashes and near crashes). These Lead-Vehicle Precrash Scenarios represent about 29% of crashes (Najm et al. 2003). Consideration should be given to several issues regarding generalizability and biases in the data. To what extent can the current findings be generalized to the greater population of Lead-Vehicle Precrash Scenarios (rear-end crashes and near crashes)? To our knowledge, the current data set, although not the final SHRP 2 data set, is the largest sample of Lead-Vehicle Precrash Scenarios used for risk estimation. Previous naturalistic data studies have had smaller samples of lead-vehicle crashes, and crash databases do not include driver behavior data (e.g., glance behavior) in the precrash phase. Clear indications of similarity and con- gruency with previous research are given by the replication analyses—for example, similarity in odds ratios on distract- ing activities and glance behavior at the precipitating event (Chapters 4 and 5) and similarity with the glance data in the 100-car data set (Figure 5.2). Then the question can be asked, What biases in data and findings are due to the sample of data used versus what might have been the sample if the entire SHRP 2 data set? Although an estimated 20% to 30% of the expected final data set was fully surveyed through kinematic triggers, the full data set was “surveyed” by automatic notification processes (e.g., onboard Automatic Crash Notification algorithms, incident button presses, and site reports). Arguably, the most severe or notice- able crashes were found through the automatic notification processes. Future research on the full SHRP 2 data set is needed to determine similarity and biases in the current data set. The question can also be asked, What is the representative- ness of the current crash data sample to more serious crashes? Is it possible that the distraction risks and characteristics under- lying serious rear-end crashes are not the same as those in the relatively minor crashes in the current crash data set? Figure 8.6 (and Figure 8.2) indicates that the actual DeltaVs in the crashes in the sample are substantially lower than the DeltaVs in the CDS accident statistics. This is expected, since CDS samples tow-away crashes, which are generally the more severe 40% of police-reported crashes. The SHRP 2 sample includes crashes with very low DeltaV (even near zero), which in many cases would not be severe enough to meet police-report criteria. As the present sample does not include crashes above DeltaV 8 m/s2, this is a limitation of this study, and care should be taken in drawing conclusions beyond this level. Section 7.6 suggests that more severe crashes would be expected to be less linearly organized along the negative slope in Figure 7.8, fea- turing more rare, and severe, combinations of long glances and fast kinematics. This is clearly a topic for further research. To what extent do the findings transfer to other crash types, for example, run-off-road crashes and intersection crashes? Some indication of transferability is given by the similarity with the 100-car data set (as in Figure 5.2), as the 100-car data set included all crashes, near crashes, and incidents found in that study. As discussed above, consideration of other crash types should be given, particularly with regard to the finding of reduction in risk from Talking/Listening on Cell Phone. The cognitive load induced by phone conversation may lead to an increased crash risk in other scenarios that involve higher cognitive or executive functions (such as planning,

110 decision making, novel sequences of action, or inhibiting habit- ual responses) rather than the more automatic reaction to a looming lead vehicle. For example, cognitive load may lead to impaired detection of red traffic lights or other types of sig- nalized information in intersection scenarios. In run-off-road scenarios, the present methodology, which quantifies change rate of a looming lead vehicle (invTTC), would have to be modified and developed. In particular, defining the crash point in run-off-road crashes is challenging, as time to line crossings or road edge crossings may not be as imminent to the driver as hitting an object. Further, run-off-road crashes do not typically involve an unpredictable behavioral component from another driver as Lead-Vehicle Precrash Scenarios do. To what extent are near crashes reasonable as surrogates for crashes? There were differences between crashes and near crashes in many metrics, such as eyes-off-road proportions, odds ratios, lead-vehicle-closure change rates, and driver reaction times. Near crashes are generally different than the baseline events in this sample and also generally different than crashes, emerging somewhere between the two in many met- rics. Because near crashes are more numerous than crashes, there is a clear weighting effect from using proportionally more near crashes than crashes in risk estimates combining crashes and near crashes, as can be seen in odds ratios (e.g., Figure 5.4 and Figure 6.8) and percentage of eyes off path (e.g., Figure 5.2). Combining crashes and near crashes generally dilutes the effect found in crashes but reduces the confidence intervals because of the larger sample and thus, it allows detection of more significant effects at a lower magnitude in combined crash/near-crash risk. Clearly, there is value in using near crashes, but risk magnitude estimations are lower. As more crashes become available for analyses in the SHRP 2 data set, this near-crash surrogate issue can be investigated further and perhaps modeled. One further limitation is that we cannot compare our results with a similar data set without portable electronic devices. Portable electronic device interactions, including visual-manual interactions, texting, and talking/listening on a cell phone, are present in the current data set, but we do not know the proportions and distributions of glance behaviors had they not been available in society. Thus, we do not know whether their presence displaces other interactions and glance behaviors or if they increase Eyes-off-Path glances. Although a subset or control condition could be created, selecting drivers that do not use their cell phones, it would not be representative of drivers who would choose to use them if they had one.