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.

30 C h a p t e r 3 This analysis primarily focuses on providing the context for which subsequent analyses can be interpreted. It provides an orientation regarding the context, conditions, and driver demographics from which the data were collected. It also aims to confirm the matching criteria for the matched baselines. For each descriptive variable, differences were inspected among the four event types: random baseline, matched base- line, near crash, and crash. In some cases, such as demo- graphic variables (e.g., age), it was not appropriate to include a matched baseline because the variable is guaranteed to match the combined crash/near-crash population. However, random baseline, near-crash, and crash events can vary on demographics. Event-type constraints are indicated in each section. Logistic regression using either a binomial (for comparing two event types) or multinomial (for comparing more than two event types) distribution function was used to test whether descriptors differ among event types. Event type was treated as the dependent measure, and each descriptor was tested individually. 3.1 Overview of Differences Crashes differed significantly from near crashes in the follow- ing ways. There were more younger drivers and more older drivers in crashes than near crashes, and the crash-involved drivers had driven less mileage in the previous year. There were more “lead vehicle stopped” (precrash Scenario 26) in crashes and more “lead vehicle decelerating” (precrash Sce- nario 25) in near crashes. Similarly, regarding the precipitat- ing event types, there were more “vehicle slowed and stopped for less than 2 seconds” in crashes and more “vehicle deceler- ating” in near crashes. Crashes involved a greater proportion of “no driver reaction” (17%) and “braking with lockup” (15%); near crashes involved more “braking without lockup” and more “braking and steering.” Crashes were also more likely to include a visual obstruction than near crashes, and more likely to occur in rainy or clear weather than near crashes, which were more likely to occur in cloudy condi- tions. Crashes and near crashes were found with different triggers. Crashes and near crashes did not significantly differ with regard to preincident maneuver, relation to junction, traffic flow, locality, traffic density, number of travel lanes, or mean speed. When comparing all four event types, we found significant differences in relation to junction, traffic flow, locality, traffic density, and travel lanes, with a marginally significant differ- ence for weather. Among demographic variables, gender and driving experience did not differ by event type. With regard to the 11 matching criteria for matched baselines and their corresponding CNC events, the degree of matching was high: 1. Same driver—required and met. 2. Must not overlap in time with the crash or near-crash event—required and met. 3. Must have a maximum amount of 1.5 seconds of stand- still (<0.1 km/h)—required and met. 4. Traffic flow should be matched to divided or undivided— met. 5. The relation to intersection should be matched (not including interchange)—met for crashes; near crashes were more likely to be at intersections or intersection- related than their matched baselines, which were more likely to not be at a junction. 6. The event should be taken from the same trip—78% of the crash–matched baselines and 79% of the near-crash– matched baselines were taken from the same trips. 7. Speed should not vary more than ±15 km/h in compari- son with the crash or near-crash event—met. 8. The event should be matched according to adverse weather (if present in crash or near crash)—met. 9. Locality should be matched to limited access or not—met. Differences Between Event Types in Descriptive Variables

31 10. Traffic density should be within one category—met. 11. Daylight should be matched to day or night—met. 3.2 Differences in Descriptive Variables Age The distribution of age groups is shown in Figure 3.1. The fig- ure indicates that both older and younger drivers are more rep- resented in crashes than in other event types. Logistic regression using age groups as indicated in the figure showed a signifi- cant effect of age [Wald X2 (14) = 29.9, p = 0.0079]. The model indicates that, as shown in the figure, older and younger drivers are significantly overrepresented in crashes, relative to ran- dom baseline and near crashes. Random baseline events also include more older drivers than near crashes do. Gender Driver gender did not significantly differ between event types. The male (M)/female (F) distribution was as follows: ran- dom baselines (53% M, 47% F), matched baselines (52% M, 47% F), near crashes (52% M, 47% F), and crashes (52% M, 48% F). Mileage In the driving history questionnaire, drivers responded to the question “Approximately how many miles did you drive last year?” Mileage was treated as a continuous variable in the logistic regression, and it was marginally significantly associ- ated with event type [Wald X2 (2) = 5.72, p = 0.0570]. Mileage for crashes (x– = 9,000) was lower than for near crashes (x– = 12,840) and random baselines (x– = 12,082). Precrash Scenario Types The precrash scenario types, as defined by Najm and Smith (2007), are shown for crashes and near crashes in Figure 3.2. All but one crash and six near crashes fell into the two catego- ries of Lead Vehicle Decelerating and Lead Vehicle Stopped. Lead Vehicle Stopped is a common source of confusion, as it is often interpreted as the subject vehicle approaching a lead vehicle that was stopped from the beginning. Recall that the team decided to use the reaction-time point as the deciding time (if stopped when the driver started to react, then Sce- nario 26; if decelerating but not yet stopped or never stops, then Scenario 25). Logistic regression indicates that crashes are significantly more likely to fall into the Lead-Vehicle- Stopped Scenario, while near crashes are more likely to fall into Figure 3.1. Percentage of random baselines, near crashes, and crashes by driver age.

32 The Lead-Vehicle-Decelerating Scenario [Wald X2 (1) = 8.44, p = 0.0037]. Precipitating Events The distribution of precipitating events for crashes and near crashes is shown in Figure 3.3. Many categories were repre- sented by only one or two events, so for analysis purposes, these were recoded into three categories: (1) Other or subject vehicle decelerating, (2) Other or subject vehicle slowed and stopped for less than 2 seconds, and (3) Other or subject vehi- cle stopped in roadway for more than 2 seconds. Other cate- gories in Figure 3.3 had too few observations to include in analysis, so they were dropped. The distribution of these cat- egories differed significantly for crashes and near crashes [Wald X2 (2) = 8.67, p = 0.0131] such that crashes were more likely to involve a vehicle slowed and stopped for less than 2 seconds, while near crashes were more likely to involve a vehicle decelerating. Trigger Type Trigger type for finding crashes and near crashes was not ana- lyzed statistically but is presented here in Figure 3.4. The trigger represents the filter component by which a given event was included in the set of cases to be evaluated further through video review. These triggers can introduce certain qualities in the events that are likely to drive the overall differences between crashes and near crashes. The near crashes were almost entirely found using a high deceleration threshold. Crashes were found primarily through site reports and Auto- matic Crash Notification (ACN). ACN is a VTTI-proprietary real-time crash detection algorithm with unknown trigger values. Given the differences in how crashes and near crashes were found, it is unlikely that the events are homogeneous in origin. Wu and Jovanis (2012) discuss the implications of using dissimilar filters when crash and near-crash events are analyzed jointly. Driver Reaction Driver reaction was significantly different for crashes and near crashes [Wald X2 (7) = 25.29, p = 0.0007]. The pattern of reaction is shown in Figure 3.5. Crashes involve a greater proportion of no reaction (17%) and braking with lockup (15%) than near crashes. In near crashes, braking without lockup made up 75% of cases, and braking and steering made up another 18% of near crashes. Figure 3.2. Percentage of precrash scenarios in near-crash and crash events, according to precrash Scenarios 22–26 from Najm and Smith (2007).

33 Visual Obstructions Visual obstructions were originally categorized into a variety of elements as shown in Figure 3.6. However, since there were few observations in each unique category, we recategorized the data into “Obstruction Present” and “No Obstruction.” Based on this simple categorization, crashes were significantly more likely to include an obstruction compared with near crashes [Wald X2 (1) = 4.54, p = 0.0331]. Nonetheless, 70% of crashes and 83% of near crashes involved no visual obstruction. Speed Mean speed did not differ for crashes and near crashes, but it did differ between crash/near-crash events and random base- line events [Wald X2 (3) = 28.61, p < 0.0001]. Mean speed was 57.1 km/h during random baselines, 52.2 km/h during matched baselines, 46.6 km/h during near crashes, and 39.8 km/h during crashes. The lower speeds during crash and near-crash events may be due to slowing in response to the precipitating event. Maximum speed also varied as a function of event type across the four events [Wald X2 (3) = 17.32, p = 0.0006], but the difference between near crashes and crashes was only marginally significant [Wald X2 (1) = 3.18, p = 0.0746]. Maxi- mum speed was highest in random baselines (62.2 km/h on average), followed by matched baselines (59.0 km/h), near crashes (55.7 km/h), and crashes (48.1 km/h). See Figure 3.7. Speed Similarity Between Matched Baselines and Crashes/Near Crashes A matching criterion was requested such that the matched baseline events should be within ±15 km/h of their corre- sponding crash or near-crash event. Mean speed was signifi- cantly lower for near crashes compared with their matched baselines [Wald X2 (1) = 6.98, p = 0.0082] but within the ±15 km/h criteria. Crashes were not different from their matched baselines. Similarly, maximum speed was margin- ally lower for near crashes compared with their matched baselines [Wald X2 (1) = 3.33, p = 0.0679] but not different for crashes. Means across events are shown in Table 3.1. Traffic Flow Traffic flow is defined as “roadway design (including the pres- ence or lack of a median) at the start of the precipitating Figure 3.3. Percentage of precipitating events in near-crash and crash events.

34 event. If the event occurs at an intersection, the traffic flow conditions just before the intersection should be recorded.” Crashes and near crashes were not different from their matched baseline cases or from each other on traffic flow. However, crashes, near crashes, and all matched baselines were different from random baselines on this variable [Wald X2 (12) = 36.50, p = 0.0003]. See Figure 3.8. In particular, random baseline events were more likely than the other three event types to occur on undivided roads. Relation to Junction Relation to junction is defined as “the relation of the involved driver or drivers to a junction (point where two roads meet) at the time of the start of the precipitating event. If the inci- dent occurs off of the roadway, the relation to junction is determined by the point of departure. Note that this is differ- ent than GES in that this database records relation to junction at the beginning of the precipitating event, while the GES manual codes this variable at the beginning of the first harm- ful event.” Crashes did not differ from their matched baselines on relation to junction, but near crashes did [Wald X2 (7) = 33.60, p < 0.0001]. Near crashes are more likely to be at intersections or to be intersection-related, while their matched baselines are more likely not to be at a junction. Across all event types, this pattern was significant, with crashes and near crashes more likely to be at or related to an intersection and both baseline types not at a junction [Wald X2 (24) = 85.41, p < 0.0001]. This pattern is shown in Figure 3.9. Locality Locality is defined as “the best description of the surround- ings at the time of the start of the precipitating event. If there are any commercial buildings, indicate as business/industrial Figure 3.4. Percentage of trigger types that found near-crash and crash events.

35 Figure 3.5. Percentage of driver reactions for near-crash and crash events. Figure 3.6. Percentage of visual obstructions in near-crash and crash events.

36 Table 3.1. Comparison of Speed Similarity Between Matched Baselines and Crashes/Near Crashes Event Type Mean Speed (km/h) Maximum Speed (km/h) Crash 39.8 48.1 Matched baseline for crash 46.2 53.3 Near crash 46.6 55.7 Matched baseline for near crash 53.5 60.1 area (this category takes precedence over others). Indicate school, church, or playground if the driver passes one of these areas at the same time as the beginning of the event (these categories take precedence over any other categories except business/industrial).” Locality did not differ between crashes and their matched baselines or between near crashes and their matched base- lines. However, random baseline events were significantly more likely to occur in residential areas and less likely to occur on interstate highways than the other three event types [Wald X2 (30) = 57.47, p = 0.0018]. See Figure 3.10. Traffic Density Traffic density is defined as “the level of traffic density at the time of the start of the precipitating event, based entirely on number of vehicles and the ability of the driver to select the driving speed.” The results for traffic density mirror those of locality. Crashes and near crashes were not differ- ent from each other or their matched baselines. However, the random baseline was significantly more likely to occur in free-flowing traffic than any of the other event types [Wald X2 (18) = 165.98, p < 0.0001]. The pattern is shown in Figure 3.11. Weather Weather is defined as “the weather condition at the time of the start of the precipitating event.” Relative to their matched baselines, crashes were not different with respect to weather. Crashes are more likely than near crashes to occur in rainy or clear weather [Wald X2 (1) = 10.32, p = 0.0354]. See Fig- ure 3.12. Near crashes were significantly more likely to occur in cloudy conditions and less likely to occur in clear con- ditions than their matched baselines [Wald X2 (4) = 10.87, p = 0.0281]. Figure 3.7. Percentage of maximum speed by event type.

37 Figure 3.8. Percentage of types of traffic flow by event type. Figure 3.9. Percentage of relation to junction by event type.

38 Figure 3.10. Percentage of locality type by event type. Figure 3.11. Percentage of traffic density by event type.

39 Figure 3.12. Percentage of weather type by event type. Figure 3.13. Percentage of lighting type by event type.

40 Surface Condition Surface condition is defined as “the type of roadway surface condition that would affect the vehicle’s coefficient of fric- tion at the start of the precipitating event.” Surface condition did not differ among any of the event types, including com- parisons between crashes, near crashes, and their respective matched baselines. Lighting Lighting is defined as “the lighting condition at the time of the start of the precipitating event.” Lighting did not vary between near crashes and crashes, or between each of these and their matched baselines. However, random baselines were signifi- cantly more likely to occur in dark-but-lighted conditions than the other groups [Wald X2 (12) = 29.41, p = 0.0034]. See Figure 3.13. Same Trip Of the 46 crashes, 36 (78%) matched baselines came from the same trip (preceding the crash). This was also true for 167 (79%) of the 211 near crashes. The remaining matched baselines were selected from a different trip by the same driver.