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.
46 C h a p t e r 5 Chapters 5 and 6 focus on the research question, What are the most dangerous glances away from the road, and what are safer glances? To begin the analysis of which glance character- istics are associated with the most risk, the logical place to start is to use the SHRP 2 NDS data to perform a replication of or comparison with the 100-car analyses (Klauer et al. 2006, 2010; Liang et al. 2012; Victor and Dozza 2011). The 100-car glance analyses focused on comparing the relative influence of various glance characteristics on risk (expressed as ORs). Conducting a replication analysis makes it possible to set the SHRP 2 OR results into context with previous research. Previous 100-car reanalysis research analyzed 6 seconds of data aligned to the precipitating event; this is the reason for the setup of alignment and the 6-second window in this chapter. The analyzed events here comprise a 6-second interval, 5 seconds before the precipitating event (PE) and 1 second after, for each crash, near-crash, and baseline event. 5.1 risk from total eyes off path time at the precipitating event Previous analyses have focused on analyzing crash/near-crash risk from Total Eyes off Roadway Time (TEORT) at the pre- cipitating event (Klauer et al. 2006, 2010, 2014; Fitch et al. 2013; Olson et al. 2009). The present analysis replicates these earlier analyses so that we can directly compare OR estimates between the present SHRP 2 data and previous data. The TEOR variable is directly comparable to this projectâs variable, Total Eyes off Path Time (TEOPT). The only difference is that we believe it is more correct to refer to the variable as Eyes off Path than Eyes off Road, even though it is coded in the same manner. The odds ratios were calculated in two ways. First, the crude odds ratios were calculated to allow direct compari- sons with similarly calculated odds ratios in previous studies (e.g., Klauer et al. 2006). Crude odds ratios are based directly on the ratios of event occurrence (e.g., crashes) and non occurrence (e.g., random baseline) for situations that contain the potential risk factor (e.g., cell phone conversation) and those that do not. Random baselines have no matching variables and so are ana lyzed using standard logistic regression, which produces odds ratios that are exactly equivalent to crude odds ratios when no other variables are present. Second, conditional logistic regression was used for com- parison with previous case-crossover analyses (e.g., Klauer et al. 2010). Here the team used matched baselines instead of random baselines and estimated ORs for the same set of predictors. Using conditional logistic regression is critical with data that include matched baselines because the crude odds ratios provide biased estimates of the true odds ratios for this data typeâtypically underestimating the odds ratios. Crude odds ratios also neglect the dependence between events and the baselines and so can underestimate the confidence interval surrounding the estimates. Crude odds ratios are provided here to facilitate comparison with previous analyses and are shown as small solid dots in the figures. TEOPT Risk in 5 Seconds Before and 1 Second After the Precipitating Event Eyes off Path time was calculated for a window of 5 seconds before and 1 second after the precipitating event. Here, Eyes off Path time is the amount of time (in seconds) the eyes were off path longer than a certain amount within this 6-second period. Bins of 0.1â0.5 second, 0.5â1.0 second, 1.0â1.5 sec- onds, 1.5â2.0 seconds, and 2.0â2.6 seconds were used. Note that each bin is compared with the case in which the driverâs eyes were on the path for the entire 6 seconds (Off0). Figure 5.1 shows the odds ratios and 95th percentile con- fidence intervals calculated with matched baselines using conditional logistic regression (OR), with matched baselines as controls. The figure also shows the odds ratios for the crude odds ratios as small dots. Risk at the Precipitating Event
47 Only the TEOPTs greater than 2 seconds (Off2.0-6.0) pro- duce an odds ratio with a confidence interval that does not include 1.0. The small filled dots show the crude odds ratios; for the near crashes (NC) and the combination of crashes and near crashes (CNC), the crude odds ratios are lower than the odds ratios from the conditional logistic regression, indicated by the large open circles and the numeral at the center of those circles. In analyses of risk in the 6 seconds around the time of the precipitating event (5 seconds leading up to the precipitat- ing event and 1 second after), TEOPTs above 2 seconds were shown to be significantly risky for crashes (OR 2.1), near crashes (OR 1.9), and crashes and near crashes combined (OR 2.1). Lower TEOPTs were not significantly risky. This result is directly comparable to the very similar 100-car study findings (Klauer et al. 2010) concerning Total Eyes off Roadway Times (TEORTs). Klauer et al. showed an odds ratio of 1.6 (CI 1.3, 2.0) compared with this projectâs odds ratio of 2.1 (CI 1.2, 3.6) for the case-crossover (matched baselines) comparison. In the case-control (random baselines) com- parison, Klauer et al. showed an odds ratio of 2.1 (CI 1.7, 2.8), while this project found an odds ratio of 2.0 (CI 1.2, 3.2). Eyes-off-Path Timeline Figure 5.2 shows a potential explanation for the OR results: the percentage of glances directed off the road increases in the 5 seconds before the event onset and 1 second after for crashes and near crashes. Each point in the graph represents the aver- age proportion of time eyes are off road at every tenth of a second. For reference, the SHRP 2 data are plotted with the 100-car data (Victor and Dozza 2011). A key observation in Figure 5.2 is that the Eyes-off-Path pat- terns for crashes and near crashes are markedly different. For near crashes, the peak is lower. For crashes, there is actually a dip just before the PE, and the highest peak is about a second after. The SHRP 2 data in Figure 5.2 show an increase in eyes off the forward path as the precipitating event approaches in time, similar to that found in the 100-car data. However, in the SHRP 2 data, the increase continues after the PE for the crashes. Figure 5.2 also seems to indicate that there is generally more Eyes off Path time in crashes and near crashes, that there is generally more Eyes off Path time in matched baselines than random baselines, and that these baselines are comparable to the 100-car baselines. Figure 5.2 also shows variation in when the crash point (in crashes) and minimum TTC (in near crashes) occurs relative to the precipitating event. Another general observation that should be noted is that proportions of crashes to near crashes matter in the combined crash/near-crash values. When combining crashes with near crashes, the fact that there are many more near crashes influ- ences the combined crash/near-crash values toward near crashes, a simple weighting effect. Thus, we need to be mind- ful of this weighting effect in which proportions of event types change the data in a combined data set. This is particularly Figure 5.1. Odds ratios and confidence intervals for total off-path glance durations in the 5 seconds before and 1 second after the precipitating event.
48 important because incidents and near crashes are often much more numerous than crashes. Reconsidering the 100-car Eyes- off-Path data and other naturalistic driving study data with this in mind may prove important, as the patterns that emerge from crashes may be very different from near crashes and inci- dents, if separated. Timing Difference Between Precipitating Event and Brake Light Onset To understand in more detail what the precipitating event was, a comparison of whether the precipitating event repre- sented the brake light onset was calculated. Figure 5.3 clearly shows that the precipitating events coincided with a brake light onset. Thus, in our data set, the precipitating event (PE) can be largely seen as brake light (BL) onset. That is, it was largely the brake light onset that was annotated as âthe state of environment or action that began the sequence under analysis.â 5.2 eyes-off-path risk in time Segments at the precipitating event An important feature of Figure 5.2 is that the proportion of Eyes off Path is not uniformâthere is a pronounced increase around the precipitating event. Consequently, analysis of windows of time preceding the precipitating event might Figure 5.2. Percentage of glances off path in relation to the precipitating event, crash points, and minimum time to collision (TTC) (for near crashes). The zero point on the x-axis indicates the onset of the precipitating event.
49 reveal how the timing of glances away from the road influence crash and near-crash risk. Odds ratios were calculated for four windows of time before the precipitating event. One window encompasses the 1 second preceding and the 1 second following the precipitat- ing event (Off1to1afterPE). Two other windows capture the influence of glances that occur earlier. Off3to1PE is a window from 3 seconds to 1 second before the precipitating event, and Off5to3PE is a window from 5 seconds to 3 seconds before the precipitating event. The final period, Off5to1afterPE, cal- culates the overall proportion the eyes were off the forward path during the entire 6-second period. These variables are coded as proportion of Eyes off Path for the stated window and so have a maximum value of 1.0. The odds ratio is relative to zero for each variable. See Figure 5.4. Figure 5.4 has several notable features. First, the odds ratios shown in this figure are notably greater than those in Figure 5.1. The scale in Figure 5.1 ranges from 0 to 5, and the scale in this figure extends from 0 to 15. The proportion of eyes off the road over various windows can be a very sensitive measure of distraction-related risk. Second, the highest odds ratios occur with the time window that overlaps the precipitating event, as shown by the points at the bottom of the figure. The proportion of time the eyes are off the forward path during the time of the precipitating event is a particularly good pre- dictor of crash and near-crash involvement. Third, the crude odds ratios are systematically lower than those estimated through conditional logistic regression. Most interesting, the odds ratios associated with crashes are substantially greater than those associated with near crashes; however, the relatively few crashes lead to large confidence intervals and preclude any definitive interpretation of these results. 5.3 Conclusions In contrast to the analysis (in Figure 5.1) that calculated Eyes off Path time as the amount of time (in seconds) the eyes were off path between 2 and 6 seconds within the 6-second period (i.e., the Off2.0-6.0 variable), the Off5to1afterPE vari- able calculated the proportion of time the eyes were off path in the 5 seconds before until 1 second after the precipitating event. The key difference between these calculations is that in Off5to1afterPE, all Eyes off Path data are included, whereas the Off2.0-6.0 only counts instances with values above 2 seconds. This is an important difference because the Off5to1afterPE variable showed significantly elevated odds ratios for crashes (OR 13.2), near crashes (OR 2.8), and crashes and near crashes combined (OR 3.6). Figure 5.3. Comparison between when the precipitating event (PE) occurred (at 0 seconds) and the brake light (BL) onset that was closest in time for crashes and near crashes.
50 The analysis that divided the 6-second window at the pre- cipitating event into three 2-second segments revealed signifi- cantly high odds ratios. The 2-second segment surrounding the precipitating event (Off1to1afterPE) was riskiest, showing significantly high odds ratios for crashes (OR 9.3), near crashes (OR 3.7), and crashes and near crashes combined (OR 4.3). Similarly, for the 2-second segment 3 seconds to 1 second before the precipitating event, significance was shown for crashes (OR 8.2) and for crashes and near crashes combined (OR 2.1). The only other significant effect using the 2-second segments was found for 5 seconds to 3 seconds before the pre- cipitating event in near crashes (OR 1.9). This analysis indi- cates that the risk is highest closer in time to the precipitating event. The results show that higher risk for larger proportions of Eyes off Path closer to the precipitating event is strongest in crash events. This effect is clearly visible in Figure 5.2, which shows how the percentage of Eyes off Path is greatest in crashes and greatest around the precipitating event for both crashes and near crashes in the SHRP 2 data and in the 100-car data. Figure 5.2 also seems to indicate that there is generally more Eyes off Path time in crashes than near crashes, that there is generally more Eyes off Path time in matched baselines than random baselines, and that these baselines are comparable to the 100-car baselines. In sum, one objective of the analysis presented in this chapter was to replicate previous findings. The analysis shows generally similar results that are consistent with previous findings and are within the margin of error of the studies. Interestingly, these consistent findings were achieved with far fewer baselines as a comparison [e.g., Klauer et al. (2006) used five baselines per crash or near crash, while this analysis used only one]. Analysis of eyes off the road in the time windows preceding and overlapping the critical event shows that the timing of glances mattersâglances away from the road during the precipitating event are particularly strongly associated with crashes and near crashes. Figure 5.4. Odds ratios and confidence intervals for various windows in the 5 seconds before and 1 second after the precipitating event.
