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.

51 C h a p t e r 6 This chapter conducts a series of analyses targeting the research question, What are the most dangerous glances away from the road, and what are safer glances? It also examines the research question, Can risk from distracting activities (secondary tasks) be explained by glance behavior? 6.1 Distribution of Glances Over time The plot of Eyes off Path over time surrounding the precipi- tating event (Figure 5.2) clearly coincided with the odds ratios in Chapter 5 and may be a useful tool to understand more precisely which mechanisms are behind the risk from Eyes off Path. To examine further the more specific relation- ships of Eyes-off-Path glance behavior in the next sections, it is useful to examine the distribution of glances and the distri- bution of glance locations over time. Glance Location Timelines Figures 6.1 through 6.4 show timelines that indicate the pro- portion of various glances in the 12 seconds preceding the crash, the 12 seconds preceding the minimum time to colli- sion for the near crashes, and a random point for the baselines. These figures show glance locations over time for the crash, near-crash, matched baseline, and random baseline events. The glance locations include the transition time toward the locations according to ISO 15007. The crash data are aligned with the crash point at zero seconds, and the near-crash data are aligned with the minimum time to collision at zero sec- onds. Matched and random baselines are aligned to a ran- domly chosen reference point at zero seconds. It should be noted that the values for each glance location, at each point in time, are mean values for the all observations. For example, Figure 6.1 cannot be interpreted to mean 30% of the drivers looked at the forward path the full time. In fact, for the crashes, only four drivers (9%) looked at the forward path for the full 12 seconds preceding the crash or minTTC, and for the matched baselines, only 32 (12%) did. Figure 6.1 indicates that the glance locations in the crash events are predominantly toward the cell phone and interior objects, followed by left and right windows and mirrors. Noticeably, there is a reduction of forward path location view- ing up until about 1.5 seconds before the crash. The eyes return quickly to the forward path location after the 1.5-second mark. Figure 6.2 shows a similar, but less pronounced pattern for near crashes. These patterns contrast sharply with those for the matched and random baseline events, shown in Figures 6.3 and 6.4; a much larger proportion of eyes is directed to the forward path—dark blue dominates Figures 6.3 and 6.4 in a way that it does not in Figures 6.1 and 6.2. In the matched baseline events (Figure 6.3), there is proportionally more forward path viewing; glances to the cell phone and interior objects are also predominant, but to a lesser degree. In the random baselines (Figure 6.4), glance locations are comparable to the matched baselines (Figure 6.3); however, there seems to be proportionally more forward path viewing, and this seems to be because there is less cell phone viewing. These differences have important implications for generalizing the odds ratios calculated from the matched baselines to more general driv- ing situations. It may be more due to the fact that there are differences in driver’s willingness to engage in secondary tasks than differences in the traffic situation. The drivers included in the matched baselines were those who eventually crashed or nearly crashed. These drivers thus seemed to have a higher prevalence of phone interaction than those in the random baselines. To shed further light on potential mechanisms behind the protective effect of Talking/Listening on a Cell Phone (OR 0.1; see Figure 4.2), we compared the glance locations for matched baseline and near-crash events in which at least one of Distrac- tion 1, 2, or 3 was coded as Talking/Listening on Cell Phone with glance locations for all remaining matched baselines. Risk from Eyes off Path Before Crash or Minimum Time to Collision

52 Figure 6.1. Percentage of glance locations over time in crash events for the 12 seconds before and 1 second after the crash point (at 0 seconds). Figure 6.2. Percentage of glance locations over time in near-crash events for the 12 seconds before and 1 second after the minimum time to collision (at 0 seconds).

53 Figure 6.3. Percentage of glance locations over time in matched baseline events for the 12 seconds before and 1 second after a randomly chosen reference point (at 0 seconds). Figure 6.4. Percentage of glance locations over time in random baseline events for the 12 seconds before and 1 second after a randomly chosen reference point (at 0 seconds).

54 (Recall from Chapter 4 that Talking/Listening on Cell Phone was not present in any crash.) In this calculation, the mean values rather than time series were used. The results are shown in Figure 6.5. The comparison of main interest is between matched baselines with and without phone conversation. The proportion of off-path glances (i.e., not on forward path) is very similar between talking/listening events and the remaining events, with values just above 20% proportion Eyes off Path. However, the distribution of the various glance locations seems to differ. In particular, the pro- portion of windshield and mirror glances is substantially higher for talking/listening cases while glances related to other secondary tasks than the phone substantially fell, especially glances to interior objects. Somewhat surprisingly, the matched baseline events with Talking/Listening on Cell Phone still contain some glances to toward the phone. To understand this result, it is important to understand exactly how the beginning and end of the Talking/ Listening on Cell Phone distraction type was annotated. Recall from Section 4.2 that Distractions 1, 2, and 3 were always coded during the 6-second time window including 5 seconds before and 1 second after the precipitating event (or a random point for the baselines). In many cases, the talking/listening was already ongoing at the beginning of the time window and continued after the end of the time window. However, in some events, the talking/listening began or ended during the time window. In such cases, the start and end points of the talking/ listening were defined as follows (SHRP 2 dictionary v2.1): • Event start point. “Phone is at the driver’s ear. If using an earpiece, it begins when the driver has pushed the last but- ton on his or her phone.” • Event end point. “Phone is away from the ear and the driver has let go of the phone, OR (if driver does not release phone) the phone is no longer moving (i.e., driver puts the phone down in their lap but doesn’t let go of the phone). Once they put the phone in their lap and it is still (even if still holding it), this should be recorded as ‘Cell phone, other.’ If they are using an earpiece, it is when they push a button on their phone to end the call.” From this definition, it is clear that talking/listening may also involve glances related to the visual-manual interaction required for hanging up, which is the reason some glances toward the phone are coded as belonging to the talking/listening time seg- ment. If the end point of talking/listening had been defined so Figure 6.5. Percentage of glance locations for matched baselines in which talking/listening was coded as distraction compared with remaining matched baselines, and glance locations for the four near crashes that occurred while Talking/Listening on Cell Phone. Glance locations only represent the time periods when the talking/listening distracting activity was coded (i.e., for 5 seconds before and 1 second after the reference point in the baselines and for 5 seconds before and 1 second after the precipitating event in near crashes).

55 that glances associated with the hanging up activity were excluded, the total proportion of off-path glances would have been somewhat lower for the talking/listening cases and thus indicative of a general gaze-concentration effect. Moreover, almost 100% of the glances in the talking/listening events would have been driving-related (left or right windshield and mirrors). In the four near-crash events involving Talking/Listening on Cell Phone, glances away from the road hardly occurred at all. However, due to the limited number of events it is difficult to draw any firm conclusions based on this result. Eyes-off-Path Timelines Figure 6.6 provides a concise summary of the preceding fig- ures, showing only the percentage of Eyes off Path for crashes, near crashes, matched baselines, and random baselines. Fig- ure 6.6 plots Eyes off Path over the 12-second period preced- ing the crash point (in crash events), the minimum time to collision (in near-crash events), and the reference points for the baselines. Similar to Figures 6.1–6.4, there seem to be gen- erally more Eyes off Path in crashes than in the other events, and Eyes off Path are increasingly off road up until 1.5 sec- onds before the crash. In near crashes a similar, but less pronounced, effect is shown. The Figure 6.6 also shows the location of the precipitating events relative to the crash point, with most precipitating events occurring between 1 second and 5 seconds before the crash point. Glance Histograms for Eyes on Path and Eyes off Path Figure 6.7 shows the frequency distribution and the cumula- tive distribution of Eyes off Path. A Kolgomorov-Smirnov test for the equality of the Eyes-off-Path distributions shows that all event-type distributions are significantly different from each other (at the p < 0.05 level), except the crash and near-crash glance distributions, which are not significantly different from each other (p = 0.115). These results generally confirm the impression conveyed by Figures 6.1 through 6.4. Drivers in the matching base- lines drivers look away from road more than those in the random baselines. The difference between the random and matched baselines might be due to the same drivers doing more secondary tasks or it might reflect the other matching variables, such as road type or being part of the same trip. Figure 6.7 suggests a difference between crashes and near crashes, with crashes having more long glances, but the KS Figure 6.6. Percentage of glances off path (for each event type at each time point) in relation to minimum time to collision or crash point (at 0 seconds), and a histogram of the time of the precipitating events associated with each crash and near crash. The precipitating events correspond to the lead-vehicle brake light onsets.

56 test shows no significant difference. This likely reflects a lack of statistical power associated with the small sample of crashes. 6.2 eyes-off-path risk in time Segments preceding Crash or Minimum time to Collision Similar to the analysis of the proportion off-path glances asso- ciated with time windows preceding the precipitating event, an analysis was conducted for time windows preceding the crash point or minimum time-to-collision point. Odds ratios were calculated for six windows at six points relative to the crash point. One window encompasses the 1 second preceding and 1 second following the reference point (Off1to1after). Five other windows capture the influence of glances that occur earlier. Off3to1 is a window of from 3 seconds to 1 second before the reference point (crash or minTTC), and Off5to3 is from 5 seconds to 3 seconds before. The other windows step sample the timeline even earlier, with the last one considering the proportion of eyes off the forward path from 11 seconds to 9 seconds before the crash point. For example, within the time from 3 seconds to 1 second preceding the reference point, we have calculated the proportion of time eyes are off the road out of these 2 seconds. As with the analysis of the 2-second windows over the 6 seconds surrounding the precipitating event (in Figure 5.4), these variables are coded as the propor- tion of Eyes off Path for the stated window and so have a maxi- mum value of 1.0. The odds ratio is relative to zero for each variable. Total Eyes-off-Path Risk Preceding Crash or Minimum Time to Collision Figure 6.8 shows that the proportion of glances off path in the window immediately preceding the crash point or minTTC is most strongly associated with crashes and near crashes, but the proportion of glances off path that overlap the crash point is not. The odds ratio for the window that overlaps the crash point is small, which corresponds to the tendency of drivers to return their eyes to the road just before or just after a crash, associated with the time window immediately pre- ceding the crash point. The odds ratio declines as the window used to summarize the off-path glances moves further from the crash point. Figure 6.7. Distributions of Eyes-off-Path Glance Lengths (percentage and cumulative percentage) for the 12-second period before the crash points (crash events) and minimum time to collision (near-crash events), and a random point for random and matched baselines.

57 Cumulative Risk for Time Segments Preceding Crash or Minimum Time to Collision The preceding analysis of the proportion of time eyes were off the forward path shows that eyes off the forward path dur- ing windows of time near the precipitating event or the mini- mum time to collision are particularly risky but that earlier windows may also be influential. That is, the odds ratios are highest for the window from 3 seconds to 1 second before the crash point, but the odds ratios in the two preceding windows also achieve statistical significance. Extended periods of eyes off the road might be expected to add to the risk associated with the eyes being off the road in the period immediately preceding the crash. This analysis considers the degree to which the proportion of Eyes off Path in the earlier windows contributes to risk after the most risky periods are considered. This is done by fitting a series of models starting with the most risky window— the window at 3 seconds to 1 second preceding the crash or minTTC. The models were assessed using the Akaike infor- mation criterion (AIC); lower values of AIC indicate better models, and differences of less than 2 suggest the models do not differ substantially (Burnham et al. 2010). The models for each time window were created so that the estimated odds ratios of crashes and near crashes are the same as those in Figure 6.8. The best of these models, Off3to1, was extended by adding the other time windows. That is, the other time windows were added as predictors in addition to Off3to1. Of these models, the one including only Off3to1 emerged as most likely; the other most likely model included both the time window 2 seconds before the minimum time to collision and the time window 10 seconds before the mini- mum time to collision. Interestingly, none of these models had a lower AIC value than the model that used only the 2-second time window from 3 seconds to 1 second preceding the crash or minimum time to collision (Off3to1). This Off3to1 model has an odds ratio of 5.9 (see Figure 6.8) for cases in which the driver looks away from the forward path compared with a driver who looks to the forward path. Table 6.1 summarizes these models, with the first column indicating the predictors used in the model. The model description indicates the linear combination used in the logistic regression, which is based on the proportion of time Figure 6.8. Odds ratios and confidence intervals for various windows in the 12 seconds before and 1 second after the crash point.

58 the eyes are off the forward path during a 2-second window. All these are additive, except for the case in which the inter- action was tested (i.e., Off3to1+ Off11to9+ Off3to1X Off11to9). The second column shows the AIC, the third shows the difference from the model with the lowest AIC, and the final shows the model likelihood. A lower AIC indicates the model fits the data better, but a difference of less than 2 is typically required to justify including another variable based on statistical significance. Based on the AIC criterion, the Off3to1 model is best, and adding the proportion of time off road in the time from 11 to 9 seconds before the crash does not lower the AIC enough to justify including this variable. The model likelihood, indicated in the last column, shows the simplest model is substantially more likely than the others. This analysis shows there is no substantial cumulative effect from the proportion of Eyes off Path in the windows preced- ing the Off3to1 variable. Further examination of the cumulative effect of propor- tion of Eyes off Path can be made by comparing these results (Table 6.1) with the glances leading to and overlapping the precipitating event. Table 6.2 indicates the combined influ- ence of Eyes-off-Path metrics during windows leading to and overlapping the precipitating event. Similar to the odds ratios in Figure 5.4, the AIC values show that the best model is one that includes the window overlapping the precipitating event and the window that includes the entire 6-second period. The AIC of other models that include the overlapping window differ by less than 2, suggesting that they also fit the data well and that the period overlapping the precipitating event is an important determinant of risk. Models with only the earlier time windows of 5 to 3 seconds before and 3 to 1 second before the precipitating event perform relatively poorly. However, the Delta AIC between the best glance model in Table 6.2 (pe.Off1to1after_pe.Off5to1after, AIC 337.86) and Table 6.1. Contribution of Cumulative Effect of Proportion of Eyes off Path Model AICc Delta_AICc Model Likelihood Off3to1 320.39 0.00 1.00 Off3to1+Off11to9 320.66 0.28 0.87 Off3to1+Off9to7 321.47 1.09 0.58 Off3to1+Off7to5 322.40 2.01 0.37 Off3to1+Off5to3 322.40 2.02 0.37 Off3to1+ Off11to9+ Off3to1X Off11to9 322.66 2.27 0.32 Off5to3 349.30 28.92 0.00 Off7to5 352.89 32.51 0.00 Off11to9 353.03 32.65 0.00 Off9to7 357.74 37.35 0.00 Note: The models are assessed using Akaike information criterion (AIC). Lower values of AIC indicate better models, and differences of AIC less than 2 suggest the models do not differ substantially (Burnham et al. 2010). Table 6.2. Contribution of the Cumulative Effect of the Proportion of Eyes off Path in the 5 Seconds Before and 1 Second After the Precipitating Event Model AICc Delta_AICc Model Likelihood pe.Off1to1after_pe.Off5to1after 337.86 0 1 pe.Off1to1after_pe.Off5to3 338.88 1.02 0.6 pe.Off1to1after_pe.Off3to1 339.16 1.3 0.52 pe.Off5to1after 339.3 1.44 0.49 pe.Off1to1after 339.43 1.57 0.46 pe.Off3to1 347.54 9.68 0.01 pe.Off5to3 351.96 14.1 0

59 the best model in Table 6.1 (Off3to1, AIC 320.39) is 17.47. Delta AICs from the best models indicate how much inferior they are. A Delta AIC difference of 2 is significant; a difference of 10 or more means there is little support for the competing model (p < 0.00001). Thus, the risk estimation from the period overlapping the precipitating event is substantially poorer than the risk estimation 3 to 1 second before the crash (Table 6.1); this further confirms the observation that there is no substantial cumulative effect from the proportion of Eyes off Path in the windows preceding the Off3to1 variable. 6.3 risk from Glance Sequences preceding Crash or Minimum time to Collision The preceding analysis focused on the proportion of time the driver’s eyes were off the forward path, without consideration of individual glance characteristics. The next step is to assess the risk contributions of different glance characteristics, where glances are considered the unit of analysis rather than propor- tion of eyes off the forward path during a window of time. In the following analysis of glance characteristics, the cumulative effect of multiple glances is considered using an estimate of the hypothetical uncertainty that builds when the eyes are off the forward path and then diminishes when the eyes return to the forward path (Senders et al. 1967; Zwahlen et al. 1988). Figure 6.9 summarizes the association between the glance parameters and crashes and near crashes. Some of these parameters are closely related to the window-based mea- sures of proportion of Eyes off Path. For example, overlap. off is the duration of the glance that overlaps the point 2 seconds before the crash point, and pre.overlap.on is the duration of the glance preceding the overlapping glance that is directed to the forward path. Not surprisingly, a glance that overlaps this point is associated with an increase in crashes and near crashes. Importantly, the units of these metrics is seconds rather than proportion, so the odds ratios Figure 6.9. Odds ratios and confidence intervals for glance characteristics.

60 shown in Figure 6.9 represent the incremental increase in odds associated with one unit of change in the glance metric. Glances can last 13 seconds, whereas time proportion of the time win- dows in the previous section ranged between zero and 1. Although the duration of off-path overlapping glances has a strong association with crashes, the risk contribution of the on-path glance that precedes this over lapping glance is small. Three related glance parameters had a strong influence on crashes and near crashes: the number of off-path glances longer than 2 seconds (twosec.glances), the maximum duration of off- path glances (max.off), and the mean duration of off-path glances (mean.off). In contrast, the number of glances (glances) and the minimum duration of on-path glances had little effect (min.on). The complexity of the glance pattern (complexity), shown at the bottom of the figure, shows an odds ratio larger than any other metric, but the confidence intervals are also very wide. Further investigation of this parameter seems warranted because it might provide a holistic assessment of how the com- bination of glances over the 12 seconds preceding a crash might contribute to risk. One such holistic glance metric is uncertainty, which is examined in the next section. The remaining glance characteristics in the figure are Total Glance Time off path (tg.off), duration of a single glance on path immediately pre- ceding the off-path glance that overlaps with the 2-second point before the crash or minTTC (pre.overlap.on), and duration of off-path glance that overlaps (intersects) with the 2-second point before the crash or minTTC (overlap.off). Figure 6.10 shows the level of uncertainty as defined by an exponential accumulation and dissipation model, in which uncertainty increases when the driver’s eyes are off path and decreases when they are on path (Senders et al. 1967). The timelines in Figure 6.10 are for a selection of five crash events; they show that with extended off-path glances, uncertainty can increase to 1, and with extended on-path glances, uncer- tainty can decline to zero. Uncertainty arises from long glances away from the road, as well as from short glances to the road. It can also accumulate over time if drivers fail to look toward the road long enough to recover from previous off-path glances. The parameters for the uncertainly model were not fit to the data but were fixed to a level that produces reasonable behavior, such as no uncertainty with a glance to the forward path of several seconds. Optimizing these parameters to maxi- mize the odds ratio associated with differences in uncertainty is likely to produce a substantially better account of the risk associated with a sequence of off-path glances. Figure 6.11 shows several summary metrics of uncertainty and their corresponding odds ratios. “Uncertainty.3to1” is the mean uncertainty over the window from 3 seconds to 1 second before the crash or point of minimum time to collision. Mini- mum uncertainty and maximum uncertainty over the 12 sec- onds are labeled “min.uncertainty” and “max.uncertainty,” respectively. Finally, mean uncertainty in the 12 seconds is labeled “m.uncertainty.” All the uncertainty metrics have both high odds ratios and large confidence intervals for crashes, with reductions in both for near crashes. Minimum uncertainty and maximum uncertainty do not reach significance as predictors. However, both mean uncertainty metrics (either 3 seconds to 1 second or over the 12 seconds) have odds ratios significantly higher than 1. Near-crash and CNC odds ratios for both are similar, but crash odds ratios are quite high for both (OR 11.5 Figure 6.10. Timelines of glances and corresponding uncertainty for five crashes. 12 8 4 0 Time from TTCmin at 0 seconds Un ce rta in ty a nd e ye s of f r oa d

61 Figure 6.11. Odds ratios and confidence intervals for various uncertainty characteristics. for the 3-seconds-to-1-second window and OR 22.2. for the full window). These confidence intervals overlap each other. Combinations of Glance Metrics Within the 12 Seconds Before Crash or minTTC Three classes of glance metrics were considered for combina- tion in a model: metrics based on the proportion of glances off- path during various time windows, metrics based on individual glances, and metrics based on uncertainty associated with off- path glances. An important application of these metrics is to assess the incremental risk associated with distracting activities (secondary tasks) beyond what can be expected from the glance distribution (see Section 6.4). For this, we need to identify the most sensitive glance metric combinations. The most promis- ing metric from each class was selected: Off3to1 from the window-based approach (Off3to1), mean off-path glance dura- tion (mean.off), and mean uncertainty from the uncertainty function (m.uncertainty). Models were fit with each of these metrics and their interactions. The models were assessed using Akaike information criterion (AIC); lower values of AIC indi- cate better models, and differences of AIC less than 2 suggest the models do not differ substantially (Burnham et al. 2010). As shown in Table 6.3, a model based on a linear combina- tion of all three metrics (Off3to1_mean.off_m.uncertainty) emerged as most predictive of crashes and near crashes as it had the lowest AIC value (AIC 317.07). This three-metric model was substantially better than the second most predictive Table 6.3. Glance Metric Combinations Most Closely Associated With Crashes and Near Crashes in the 12 Seconds Before the Crash or minTTC Model AICc Delta_AICc Model Likelihood Off3to1_max.off_m.uncertainty 317.07 0.00 1.00 Off3to1 320.39 3.31 0.19 Off3to1Xmax.off 321.44 4.37 0.11 Off3to1Xmax.offXm.uncertainty 321.59 4.51 0.10 Off3to1Xm.uncertainty 322.01 4.94 0.08 max.off 337.58 20.50 0.00 m.uncertainty 345.86 28.79 0.00

62 model based only on the proportion of glances off-path (AIC 320.39), at a Delta AIC difference of 3.31. As an indi- vidual metric, the Off3to1 metric was a very good model in comparison with max.off and m.uncertainty. 6.4 risk Contributions from Distracting activities Over and above What Can Be explained by Glance Behavior Recall the research question, Can risk from distracting activ- ities (secondary tasks) be explained by glance behavior? The risk of distracting activities (secondary tasks) might be explained largely by the off-path glances they induce; alter- natively, the risk associated with secondary tasks might be associated with other properties of the distracting activities. Risk from additional properties such as manual distraction or cognitive distraction may add to the risk explained by Eyes off Path. To assess the incremental risk (of crash and near-crash events) associated with distracting activities beyond that expected from the Eyes off Path, we had originally intended to compare the glance behavior within the full 12-second period leading up to the crash/minTTC and the full periods within baselines, with the presence of distracting activities within those periods of time. Unfortunately, due to a misunderstanding in interpretation of the data requirements document, the distract- ing activities were not coded for the full periods. Distracting activities were only coded for the 5 seconds preceding and 1 sec- ond after the precipitating event in crashes and near crashes or the reference point in baselines (as discussed in Chapters 4 and 5). Thus, to assess the potential additive risk associated with distracting activities over and above the Eyes-off-Path glance metrics, we need to use the most predictive glance model for the 6 seconds surrounding the precipitating event (the 5 seconds preceding and 1 second after the precipitating event). The combined effects of glance metrics for the timeline anchored by the precipitating event showed a different pattern of glance metrics (Table 6.4) than in Table 6.3. The best indi- cator included the proportion of eyes off the road in the 2 sec- onds overlapping the precipitating event (pe.Off1to1after, AIC 339.43). For the relatively short timeline of 5 seconds preceding and 1 second after the precipitating event, the other glance metrics failed to improve the prediction substantially. Note that the three-metric glance model of the longer glance history preceding the crash/minTTC (Off3to1_max.off_m. uncertainty, AIC 317.07, in Table 6.3) was a far better predic- tor of risk of a crash or near crash than the 2 seconds overlap- ping the precipitating event (pe.Off1to1after, AIC 339.43), with a Delta AIC of 22.36 between these two models. As already discussed, the distracting activities were only coded for the 5 seconds preceding and 1 second after the pre- cipitating event. For this reason, the proportion of eyes off path in the 2 seconds overlapping the precipitating event (pe. Off1to1after) was used to assess the contribution of distract- ing activities because the comparison has to use the glance behavior from the same period of time. Table 6.5 shows that the proportion of Eyes off Path in the 2 seconds overlapping the precipitating event (pe.Off1to1after, AIC 339.43) is substantially more predictive than the models based on distracting activities alone (e.g., Texting), as the dif- ference compared with the pe.Off1to1after model is greater than 2—a significant difference. Because the model based on the proportion of eyes off path in the 2 seconds overlapping the precipitating event is so superior to the models based on distracting activities, it might be expected to fully account for the effect of distracting Table 6.4. Glance Metric Combinations Most Closely Associated With Crashes and Near Crashes in the 5 Seconds Before and 1 Second After the Precipitating Event Model AICc Delta_AICc Model Likelihood pe.Off1to1after 339.43 0 1 pe.Off1to1afterXm.uncertainty 339.86 0.43 0.81 pe.Off1to1after_max.off_m.uncertainty 340.21 0.77 0.68 pe.Off1to1afterXmax.off 340.67 1.23 0.54 pe.Off1to1afterXmax.offXm.uncertainty 342.36 2.93 0.23 pe.max.off 343.81 4.38 0.11 pe.m.uncertainty 345.69 6.26 0.04 pe.complexity 351.37 11.94 0

63 activities (secondary tasks). To test this assertion, a series of models was created in which the effect of each class of dis- tracting activities was added to the effect of the glances met- rics. If the proportion of Eyes-off-Path metric fully accounts for distraction risk, then adding the distracting activities to the model will result in no reduction in the AIC value of 339.43 for the pe.Off1to1after model. Table 6.6 explicitly shows the contribution of the dis- tracting activity to that of the proportion of Eyes-off-Path metric (pe.Off1to1after). When Talking/Listening on Cell Phone is combined with the proportion of Eyes-off-Path metric (pe.Off1to1after.TalkingListening, AIC 333.03), the model is substantially better than the pe.Off1to1after metric alone (a Delta AIC reduction of 6.41). Similarly, when Text- ing is combined with the proportion of Eyes-off-Path met- ric (pe.Off1to1after.Texting, AIC 334.23), the model is substantially better than the pe.Off1to1after metric alone (a Delta AIC reduction of 5.2). In both cases the activity does contribute to improving the risk estimation over and above the Eyes-off-Path metric, but the activity has more affect for talking than for texting. In both cases, not considering Table 6.5. Risk Predicted by the Proportion of Eyes off Path in the 2 Seconds Overlapping the Precipitating Event and Distracting Activities in the 5 Seconds Before and 1 Second After the Precipitating Event Model AICc Delta_AICc Model Likelihood pe.Off1to1after 339.43 0 1 Texting 341.66 2.23 0.33 TalkingListening 348.53 9.1 0.01 VisualManual 349.59 10.16 0.01 Table 6.6. Contribution to Risk Estimation from Distracting Activities in the 5 Seconds Before and 1 Second After the Precipitating Event and Proportion of Eyes off Path in the 2 Seconds Overlapping the Precipitating Event Model AICc Delta_AICc Model Likelihood pe.Off1to1after.TalkingListening 333.03 0 1 pe.Off1to1after.Texting 334.23 1.21 0.55 pe.Off1to1after 339.43 6.41 0.04 Texting 341.66 8.63 0 pe.Off1to1after.VisualManual 347.92 14.90 0 Talking 348.53 15.51 0 glances provides a very poor indicator of risk (as the indi- vidual AICs for Texting and Talking/Listening are substan- tially lower than pe.Off1to1after). Thus, in the case of Talking/ Listening on Cell Phone, the combined model provides a bet- ter estimation of the risk reduction, and in the case of Text- ing, the combined model provides a better estimation of the risk increase. When the Portable Electronics Visual-Manual activity is combined with the proportion of Eyes-off-Path metric (pe. Off1to1after.VisualManual, AIC 347.92), the model is substan- tially poorer than the pe.Off1to1after metric alone (a Delta AIC increase of 8.49 between these two models). Thus, for more general visual-manual interactions, the risk estimation is not improved by adding the distracting activity. Rather, the risk increase is better explained by the proportion of Eyes off Path. The interpretation of this pattern is that the effect of Texting and Talking/Listening on Cell Phone is not fully accounted for by the particular Eyes-off-Path model, but the effect of the gen- eral category of Portable Electronics Visual-Manual is fully accounted for. The risk-increasing (Texting) and risk-decreasing (Talking/Listening on Cell Phone) influences may go beyond glance patterns. This hypothesis should be explored in future work, but it is intriguing to consider the possibility that task demands (e.g., cognitive distraction) or other characteristics (e.g., drivers may keep longer headways when talking on the phone) for these tasks may provide additional disbenefit or benefit in terms of risk in rear-end striking crashes, beyond their influence on glances. Note that this analysis of the risk contributions from dis- tracting activities over and above what can be explained by glance behavior was limited by the fact that the distracting activities were not coded for the period of time up until the crash or near-crash minTTC. Future work should compare the much more powerful model based on a linear combina- tion of all three metrics (Off3to1_mean.off_m.uncertainty, AIC 317.07) with distracting activities occurring in a

64 comparable window of time closer to the crash or near- crash minTTC. It is possible that the influences of distracting activities can be explained by this more powerful three-metric glance model. 6.5 Conclusions This chapter considered the contribution to risk of three classes of glance metrics: proportion of eyes off the forward path during the time window preceding crashes and near- crash minTTC, summary of metrics of glance sequences, and uncertainty that reflects the sequence of glances. Indi- vidually, each class of metric indicated risk more strongly and precisely than the classes of distracting activities. The risk indicated by each class of glance metric is not redundant with the others because a model that includes the interactions between metrics provides the most precise indicator of crashes and near crashes. Interestingly, the protective effect of cell phone conversation (Talking/Listening on Cell Phone) is not fully accounted for by this combination of glance metrics and neither is the risk asso- ciated with activities such as texting, but the risk associated with the more general class of Visual-Manual Distractions associated with visual-manual interactions is accounted for by glance metrics. An important limit of the glances model, including 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). However, risk likely stems from both drivers’ attention to the road and the demands of the road. The following chapter con- siders how the changes in the lead-vehicle kinematics affect the risk of a crash or near crash.