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.

41 C h a p t e r 4 One central topic in naturalistic driving study research has been the calculation of risk associated with (human-identified) classifications of distracting activity types (also called second- ary tasks), such as talking, dialing, eating, and texting (Fitch et al. 2013; Klauer et al. 2006, 2010, 2014; Olson et al. 2009). This task risk approach has had particular influence on policy and rulemaking decisions regarding tasks that should or should not be done while driving. Thus, it is of particular relevance for the present research to be able to relate the team’s inattention per- formance approach (quantifying Eyes off Path and lead-vehicle closing in later chapters of this report) to the task risk approach. We began by examining task risk in distracting activities. 4.1 Methods For Chapters 4–6, we estimated the odds ratio of having a criti- cal event (crash, near crash, or both, depending on the analysis) as a function of various predictors. Chapter 4 covers activity types as predictors, Chapter 5 focuses on the precipitating event and behavior surrounding it, and Chapter 6 focuses on glance behavior preceding minimum time to collision. The primary method used to estimate odds ratios in this study is conditional logistic regression. However, it is useful to understand unconditional logistic regression first and then contrast the conditional approach. The purpose of logistic regression is to develop models of crash or near-crash risk as a function of various predictors associated with driver behavior and environment. Logistic regression is used in a wide variety of applications when the response variable has a small number of possible outcomes. The binary outcome case is most com- mon and is described here. Suppose the response variable, y, for a driving event is assigned a value of 1 if it resulted in a crash (or near crash) and is assigned the value of 0 if it did not result in a crash. The data set also has a set of r predictors, x1, x2 . . . xr, each of which describes a characteristic of the situation, driver, or behavior (e.g., glance pattern, distracting activity) in each case. The logistic regression model uses these data to predict the risk of crashing, given the predictors, according to the formula given in Equation 4.1. = + ∑( )− β + β = ˆ 1 1 (4.1) 0 1 p e xi ii r where pˆ is the predicted probability of crashing, b0 is the intercept, the bi are coefficients of the predictors, and the xi are the predictor values. The model-fitting process results in estimates of the bs and an additional error component that measures model uncertainty using any lack of fit of the model to the data. Logistic regression is a type of general linear model in that a simple transformation of the predicted outcome is related to a linear function of predictors (though individual predictors can also be transformed). Here, the odds of crashing are defined as p/(1–p). Logistic regression models the natural logarithm (ln) odds of crashing on the left side of the model equation as a linear function of predictor variables on the right side of the model equation. For N events, the model equation is shown in Equation 4.2. p p x x i Ni i i r irln 1 . . . 1, . . . (4.2)0 1 1 −     = β + β + + β = where b0 is the intercept parameter, bj, j = 1, . . . , r are regression coefficients or slope parameters, and xij are known predictor variables such as distracting activity. Note that this formulation is equivalent to Equation 4.1 but focuses on the linear portion of the equation. Like many models, logistic regression models are fit using the method of maximum likelihood. Since the left side of the model equation represents the ln odds of a crash, the slope parameters have interpretations as ln odds ratios. For a continuous predictor x such as inverse Tau, the slope parameter attached to it represents the ln odds of a crash for a unit increase in inverse Tau at fixed values of Risk from Distracting Activity Types

42 all other predictors. For a binary predictor such as cell phone use that is coded 0 if the occupant did not use a cell phone and 1 if the occupant did use a cell phone, the slope parameter represents the ln odds of a crash for occupants who used a cell phone, compared with those who did not. Because multiple predictors are included in the model, an odds ratio for a variable is interpreted under the assumption that it has been adjusted for the other predictors included in the model. One challenge with logistic regression is interpreting odds ratios. Most readers find risk ratios to be more intuitive, and there is a tendency to want to interpret odds ratios as risk ratios. However, risk has a significant analytical disadvantage in that it is bounded between 0 and 1. If the risk of an event is, say, 50% under Condition A, then the risk ratio cannot be more than 2 for Condition B compared with Condition A. However, odds can range from zero to infinity, and their ratio is not constrained by the size of the denominator. However, in circumstances when the outcome of interest is rare (<5% probability), an odds ratio can be interpreted as relative risk. In that case the odds ratio approximates a relative risk, and it is appropriate use logistic regression as an exposure-based risk model (Greenland and Thomas 1982). In this study, we used an adaptation of logistic regression that is often used in epidemiology to increase efficiency of sampling (Rothman 2012). Instead of selecting a large num- ber of events and identifying crashes among them, we selected crashes and near crashes and then found one matching event in the baseline for each crash or near crash. This is called a matched case–control study, and in this case, because matched events are from the same driver, it falls into the class of studies known as case-crossover. Because this method controls the percentage of analyzed cases that are crashes (50%), the intercept in logistic regres- sion is meaningless. Instead, conditional logistic regression estimates the probability that each specific event is the event in its group that resulted in a crash. Matching characteristics cannot be used as predictors because they do not distinguish between crashes and baseline within each group. However, other characteristics (e.g., gaze patterns or distractions) as well as interactions between matching variables and other characteristics can be used. Although the intercept, or b0, is not estimated with condi- tional logistic regression, the remaining coefficients are still interpretable in the same way as for logistic regression. Indeed, if there were only one group, conditional and unconditional logistic regression would produce the same estimates (other than the intercept). In addition, with unconditional logistic regression, coefficients of predictors are unbiased, even when the underlying sample is biased, as it is here (Breslow 1996). Although both forms of logistic regression produce com- parable coefficients, unconditional logistic regression is not appropriate for matched data sets. The matching itself is a feature of the sampling, not the underlying process, and therefore must be taken into account using conditional logis- tic regression. Although we control the base risk by using a 1–1 match, the odds ratios for predictors can be interpreted in this context as risk ratios. Because the odds ratios are unbiased, they are expected to be the same, even if sampled differently. In driving, crashes and even near crashes are extremely rare events that would fall under the “rare disease assumption” level. The next sections present odds ratios for crashes, near crashes, or both as a function of a variety of predictors. 4.2 Distracting activities Distracting activities were manually coded by reviewing video by VTTI according to the SHRP 2 data dictionary (Variables Distraction 1, Distraction 2, and Distraction 3 in the SHRP 2 NDS Event Definitions and Variables v 2_1 from December 2010). In accordance with the SHRP 2 data dictionary, dis- tracting activities were coded if present within the 6-second time window including 5 seconds before and 1 second after the precipitating event (or a random point for the baselines). There are more than 50 categories of distracting activities in the data dictionary and many subtle nuances in classification of the various categories. Therefore, caution should be taken when interpreting categories, and the data dictionary should be used as support for interpretation. For example, the Talking/ Listening on Cell Phone category excludes locating, reaching for, and answering a cell phone but actually does include hanging up (this is discussed in greater detail in Section 6.1). The rationale within the data dictionary is unclear for includ- ing an explicitly visual-manual interaction (hanging up) in an activity that seems to be clearly designated as nonvisual (talking/listening). This seems problematic as it creates a category that is not easily mapped onto clearly defined dimen- sions of distraction and makes it more difficult to draw con- clusions regarding nonvisual/nonmanual interaction. Classes of distracting activities (Distraction 1, 2, 3 event variables) present in the 5 seconds before the precipitating event and 1 second after were developed, because relatively few observations were present in the individual categories (see below). Note that the individual categories Lost in Thought (three cases), Looked but Did Not See (three cases), and Cogni- tive, Other (zero cases) were not included in the classes as they were believed to be questionable categories and difficult to group together with other nonvisual activities. Classes of distracting activities are as follows: 1. Portable Electronics Talking/Listening: Talking/Listening on Cell Phone. 2. Portable Electronics Visual-Manual: Dialing Hands-Free Cell Phone Using Voice-Activated Software; Texting on

43 Cell Phone; Dialing Handheld Cell Phone; Dialing Hand- held Cell Phone Using Quick Keys; Locating/Reaching/ Answering Cell Phone; Cell Phone, Other; Locating/ Reaching PDA; Operating PDA; Viewing PDA; PDA, Other. 3. Original Equipment: Reaching for object that is a manu- facturer-installed device; Adjusting/Monitoring Cli- mate Control; Adjusting/Monitoring Radio; Inserting/ Retrieving Cassette; Inserting/Retrieving CD; Adjusting/ Monitoring Other Devices Integral to Vehicle. 4. Nonelectronics, Nonvisual: Talking/Singing Lost in thought; Looked but did not see; Cognitive, other (deleted, as described above). 5. Nonelectronics Other: Dancing; Reading; Writing; Mov- ing Object in Vehicle; Insect in Vehicle; Pet in Vehicle; Object Dropped by Driver; Reaching for Object, Other; Object in Vehicle, Other; Reaching for Food-Related or Drink-Related Item; Eating with Utensils; Eating with- out Utensils; Drinking with Lid and Straw; Drinking with Lid, No Straw; Drinking with Straw, No Lid; Drink- ing from Open Container; Reaching for Cigar/Cigarette; Lighting Cigar/Cigarette; Smoking Cigar/Cigarette; Extinguishing Cigar/Cigarette; Reaching for Personal Body-Related Item; Combing/Brushing/Fixing Hair; Applying Makeup; Shaving; Brushing/Flossing Teeth; Biting Nails/Cuticles; Removing/Adjusting Jewelry; Removing/Inserting/Adjusting Contact Lenses or Glasses; Other Personal Hygiene. 6. Nonelectronics, Passenger-Related: Passenger in Adjacent Seat—Interaction; Passenger in Adjacent Seat—No Interaction or Cannot Tell; Passenger in Rear Seat— Interaction; Passenger in Rear Seat—No Interaction or Cannot Tell; Child in Adjacent Seat—Interaction; Child in Adjacent Seat—No Interaction or Cannot Tell; Child in Rear Seat—Interaction; Child in Rear Seat— No Interaction or Cannot Tell. 7. Vehicle-External Distraction: Looking at Previous Crash or Incident; Looking at Pedestrian; Looking at Animal; Looking at an Object External to the Vehicle; Distracted by Construction; Other External Distraction. 8. Inattention to the Forward Roadway: Inattention to the Forward Roadway—Left Window; Inattention to the Forward Roadway—Left Mirror; Inattention to the For- ward Roadway—Center Mirror; Inattention to the For- ward Roadway—Right Mirror; Inattention to the Forward Roadway—Right Window; Inattention to the Forward Roadway—Back Window. 9. Other: Other nonspecific Eyeglance; Inattention, Other; Unknown Type (Distraction Present); Unknown. 10. Not Distracted: Not Distracted. Figure 4.1 shows the odds ratio for several variables. The precise OR is shown in the center of each dot, and the lines surrounding the dots indicate the 95th percentile confidence interval. Odds ratios are significant when the confidence interval does not cross 1 on the x-axis. For each activity three ORs are shown: the near-crash (NC) situations, the crash and near-crash (CNC) situations combined, and the crash (C) situations. In some cases, there are insufficient data to calcu- late precise confidence intervals, such as the OR of crashes associated with phone conversations. There were no crashes when drivers were engaged in Portable Electronics Talking/ Listening and so the OR and its confidence interval (CI) are undefined, represented by the CI that extends across the full width of the graph. Note that Portable_electronics_talking_ listening in Figure 4.1 corresponds to the Talking_listening_ on_cell_phone category in Figure 4.2. Talking_listening_on_cell_phone in Figure 4.2 is part of the Portable_electronics_talking_listening aggregate category in Figure 4.1. Not surprisingly, Texting has a high OR of 5.6 (CI 2.2–14.5) for crash and near-crash situations, but Talking_ listening_on_cell_phone has an OR of only 0.1—representing a large reduction in risk. The magnitude of the risk reduction (a protective effect) can be directly compared with the magni- tude of the risk by reversing the sign of the coefficient before converting it to an odds ratio. In this case, the odds ratio would be approximately 10—greater than the risk of texting. 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. Individual categories, such as Locating/ Reaching for/Answering a Cell Phone or Adjusting/Monitor- ing the Radio, or other aggregate categories, such as Original Equipment or Vehicle External Distraction, were not signifi- cantly risky. 4.3 Driver Impairments The Driver Impairments event data variable was examined (see Figure 4.3). Distraction occurred frequently in the events collected for this study, and the odds ratios associated with these distractions suggest they contribute to crashes. There is an association of distraction and crashes [OR = 3.0, CI (1.1, 8.3)], near crashes [OR = 1.6, CI (1.0, 2.4)], and crashes and near crashes together [OR = 1.8, CI (1.2, 2.6)]. Impairments associ- ated with drowsiness, drugs, and alcohol were much less fre- quent. Figure 4.3 shows the distribution of these impairments relative to distraction. Of these impairments, only drowsiness occurred frequently enough to merit investigation. A condi- tional logistic regression model found an odds ratio of 2.7, but the relatively few cases leads to large confidence intervals and a failure to achieve statistical significance [Wald X2(1) = 2.1, p = 0.147]. Distraction is dominant, and the other behaviors or impairments are rare in this sample of driving.

44 Figure 4.1. Odds ratios (numbers inside circles) and confidence intervals (horizontal lines) for classes of distracting activities in crashes, near crashes, and crashes and near crashes combined. Odds ratios are significant only when confidence intervals are fully above or below 1 (do not cross vertical line at 1). Presence of a distracting activity was coded between 5 seconds before and 1 second after the precipitating event. Figure 4.2. Odds ratios (numbers inside circles) and confidence intervals (horizontal lines) for specific distracting activities. Odds ratios are significant only when confidence intervals are fully above or below 1 (do not cross vertical line at 1). Presence of distracting activity was coded between 5 seconds before and 1 second after the precipitating event.

45 4.4 Conclusions The effect of these distracting activities on the likelihood of crashes and near crashes varies tremendously. Some activities, such as Talking/Listening on Cell Phone, have an odds ratio substantially less than one, suggesting a protective effect; others, such as Texting, have an odds ratio much greater than one, suggesting a substantial risk. Texting (OR 5.6, CI 2.2–14.5) and the aggregate category of Portable Electronics Visual- Manual (OR 2.7, 1.4–5.2) had the highest odds ratios, suggest- ing a substantial risk. Talking/Listening on Cell Phone was found to decrease 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). In line with predictions regarding sample-size limitations, sig- nificant differences for distracting activities were not found for crash events alone as there were too few observations in this category. However, notably, there were no crashes when drivers were engaged in Talking/Listening on a Cell Phone. The limitation in sample size also influenced the ability to detect significant odds ratios in other distracting activities (such as dialing a cell phone or adjusting the radio). Note that the limitation in sample size affects the odds ratio analysis of distracting activities because distracting activities are not found in all events. The odds ratio analysis of glance data is not affected in the same manner because off-path glances occur in almost all events, providing more data to work with. Other behaviors and impairments (drowsiness, drugs, and alcohol) are rare. Of these, only drowsiness occurred fre- quently enough to merit investigation and showed an odds ratio of 2.7, but was not statistically significant. In sum, distracting activities occur frequently, much more frequently than impairments such as drowsiness. Some activ- ities, such as phone conversations, have an odds ratio sub- stantially less than one, suggesting a protective effect, and others, such as texting, have an odds ratio much greater than one, suggesting a substantial risk. Figure 4.3. Distribution of driver impairments relative to distracting activities across all event types together (crashes, near crashes, matched baselines, and random baselines).