Introduction

Transportation research has repeatedly demonstrated that crashes are caused by multiple factors. These factors may include driver behavior, vehicle malfunction, characteristics of the roadway, and/or surrounding infrastructure. Improving our understanding of these different risk factors can help transportation researchers direct their efforts at reducing the highest risk factors to better improve driver safety. This paper addresses the risks associated with roadway infrastructure.

Traditional methods to understand crash risk factors have included the use of epidemiological data (crash databases and incident rate data) and simulation data, as well as in-depth investigation of actual crash sites. For example, Kar & Datta (2007) followed a method previously used with Michigan state data, to calculate risk using collision rates and roadway throughput data (AADT). These risk rates were calculated for various state and county jurisdictions in the development of an algorithm to identify the riskiest segments of roadway. For these algorithms, risk was calculated using crash rates for various segments of roadway and compared to the AADT for the same roadway.

Similarly, Shankar, Mannering, and Barfield (1994) assessed a segment of I-90 to determine how specific roadway geometries (grade, curvature, or combination) and weather contribute to crashes. These authors quantified the impacts of geometric characteristics on motor vehicle accidents by using existing crash data that specified the roadway geometry and weather conditions on a specific road segment. Using the Poisson distribution, the risks associated with these factors were calculated to determine the riskiest portions of the roadway. The results suggested optimal speeds around horizontal curves and also suggested that designers make every effort to avoid steep road grades coupled with horizontal curves.

Likewise, traffic simulation studies are very beneficial for identifying or assessing the potential safety benefits for specific roadway geometries or selected roadway infrastructure. These simulation studies run thousands of potential traffic scenarios to assess traffic patterns and estimate crash rates given certain traffic volumes. For example, Saccomano, Cunto, Guido, & Vitale (2008) conducted a traffic simulation to assess the number of crashes and near-crashes that would occur at a pre-specified location if a typical intersection were present versus a roundabout.

Although traditional methods of studying crash causation have provided valuable information regarding the risk associated with specific segments of roadway or for specific geometric design, these methods do have some limitations. Simulation studies can generate thousands of possible traffic patterns very quickly, with those patterns only reflecting the parameters that were set specifically per simulation evaluation. In other words, random characteristics that are present in real-world traffic can not be simulated.

Crash database records have limitations in that they are restricted to police-reported crashes and the data that an investigating police officer records on them. Dingus, Klauer, Neale, Petersen, Lee, et al. (2006) found that there are approximately 5 non-police-reported crashes for every police-reported crash, thus there are a significant number of crashes that are not included in these reports. Also, police accident report forms vary from state to state with some states not reporting important driver behavior information such as cell phone use, and so forth. Seat belt use and cell phone use, for that matter, may not be accurately reported by the accident victims or bystanders because they forget or do not want to admit any level of fault.

Therefore, crash risk calculated using these methods provides only a gross assessment of risk whereby aspects of the driver, driver behavior, weather conditions, lighting, and so forth are not considered. For example, if a specific road segment/intersection happened to be located near a high school, the calculated risk of this segment may be higher simply because of the high percentage of teen drivers present, not necessarily because there is a dangerous design flaw to the intersection.

To address these issues, recent advances in technology have been used to create a new method to study crash contributing factors: naturalistic driving. The naturalistic driving method is a technique of data collection whereby sophisticated sensors and data acquisition equipment is installed in a vehicle that is used by a driver for daily use. The instrumentation of vehicles and naturalistic data collection on a large scale provides a unique opportunity to examine a variety of crash factors and determine the factors that lead to a statistically higher risk of involvement in a crash or near-crash.

For example, Klauer, Sudweeks, Hickman, & Neale (2006) used naturalistic driving data to calculate the risk of unsafe driving behaviors. With the premise that drivers who engage in one risky driving behavior typically also engage in other behaviors, risk calculations for unsafe driving behaviors were assessed while controlling for the correlations that exist among other unsafe driving behaviors. The result was a more accurate assessment of the crash risk of driving inattention, following too closely, traveling faster than surrounding traffic, etc.

The purpose of this paper is to present crash, near-crash, and incident frequency data from the 100-Car Naturalistic Driving Study (Dingus et al., 2005) to identify roadway infrastructure characteristics that are frequent contributing factors in the Northern Virginia/Washington DC geographic area. In addition, crash/near-crash risk was calculated for a variety of roadway characteristics to assist roadway designers in identifying problematic roadway characteristics. Based on crash rate information, the following Virginia Department of Transportation state-maintained roadways were of interest:

I-66 (Custus Memorial Parkway)
US-50 (Lee Jackson Memorial Highway, John Moby Highway, Arlington Blvd.)
US-29 (Lee Highway, Arlington Blvd.)
SR-267 (Dulles Greenway, Dulles airport access road)
SR-28 (Sully Rd.)

Method

The method used for the 100-Car Naturalistic Driving Study is explained in detail in Klauer, Dingus, Neale, Sudweeks, and Ramsey (2006). Those methods relevant to the analyses contained herein are described.

Instrumentation

The 100-Car instrumentation package was designed and developed by staff at VTTI. The system consisted of a Pentium-based computer that received and stored data from a network of sensors distributed around the vehicle. Data storage was achieved via the system’s hard drive. Sensors included a vehicle network box that interacted with the vehicle network, an accelerometer box that obtained longitudinal and lateral kinematic information, a headway detection system to provide information on leading or following vehicles, side obstacle detection to detect lateral conflicts, an incident box to allow drivers to flag incidents for the research team, a video-based lane tracking system to measure lane keeping behavior, and video to validate any sensor-based findings. The video subsystem was particularly important as it provided a continuous window into the happenings in and around the vehicle. This subsystem included five camera views monitoring the driver’s face and driver side of the vehicle, the forward view, the rear view, the passenger side of the vehicle, and an over-the-shoulder view for the driver’s hands and surrounding areas. A frame of compressed 100-Car video data is shown in Figure 1. Other relevant subsystems were cellular communications and a global positioning subsystem (GPS) that collected information on vehicle position. The GPS and the cellular communications were often used in concert to allow for vehicle localization and tracking.

video1

Figure 1. A compressed video image from the 100-Car data. The driver’s face (upper left quadrant) is distorted to protect the driver’s identity. The lower right quadrant is split with the left-side (top) and the rear (bottom) views.

Subjects

One-hundred drivers who commuted into or out of the Northern Virginia/Washington, DC metropolitan area were recruited as primary drivers to have their vehicles instrumented or to receive a leased vehicle for this study. Drivers who had their private vehicles instrumented (78) received $125.00 per month and a bonus at the end of the study for completing necessary paperwork. Drivers who received a leased vehicle (22) received free use of the vehicle, including standard maintenance, and the same bonus at the end of the study for completing necessary paperwork.

As some drivers had to be replaced for various reasons (for example, a move from the study area), 109 primary drivers were included in the study. Since other family members and friends would occasionally drive the instrumented vehicles, data were collected on 132 additional drivers.

The age by gender distribution of the primary drivers is shown in Table 1. The distribution of miles driven by the subjects during the study appears as Table 2. As presented, the data are somewhat biased compared to the national averages in each case, based on TransStats, 2001. Nevertheless, the distribution was generally representative of national averages when viewed across the distribution of mileages within the TransStats data.

Table 1. Driver age and gender distributions.

	Gender
Age		Female	Male	Grand Total
18-20		9	7	16
		8.3%	6.4%	14.7%
21-24		11	10	21
		10.1%	9.2%	19.3%
25-34		7	12	19
		6.4%	11.0%	17.4%
35-44		4	16	20
		3.7%	14.7%	18.3%
45-54		7	13	20
		6.4%	11.9%	18.3%
55+		5	8	13
		4.6%	7.3%	11.9%
Total N		43	66	109
Total Percent		39.4%	60.6%	100.0%

Table 2. Actual miles driven during the study.

Actual miles driven	Number of Drivers	Percent of Drivers
0-9,000	29	26.6%
9,001-12,000	22	20.2%
12,001-15,000	26	23.9%
15,001-18,000	11	10.1%
18,001-21,000	8	7.3%
More than 21,000	13	11.9%

Vehicles

Six vehicle models were selected based upon their prevalence in the Northern Virginia area. These included five sedan models (Chevrolet Malibu and Cavalier, Toyota Camry and Corolla, and Ford Taurus) and one SUV model (Ford Explorer). The model years were limited to those with common body types and accessible vehicle networks (generally 1995 to 2003). The distribution of these vehicle types was:

· Toyota Camry – 17%

· Toyota Corolla – 18%

· Chevy Cavalier – 17%

· Chevy Malibu – 21%

· Ford Taurus – 12%

· Ford Explorer – 15%

Data Reduction

Data reduction refers to a process of finding driving events of interest in the data stream, then reviewing the video to determine the details of the event. This process began with establishing post-hoc “triggers” to identify segments of data that may potentially contain a safety-related event. A post-hoc trigger uses either a single signature (e.g., any lateral acceleration value greater than ±0.6 g) or multiple signatures (e.g., forward TTC value > 3 s plus a longitudinal deceleration value > -0.5 g) in the driving performance data stream to identify those points in time when it was likely that a driver was involved in an incident, near-crash, or crash. The list of dependent variables ultimately used as triggers used to identify crashes, near-crashes, and incidents is presented in Table 3.

Table 3. Dependent variables used as event triggers.

Trigger Type	Description
1. Lateral acceleration	· Lateral motion equal to or greater than 0.7 g.
2. Longitudinal acceleration	· Acceleration or deceleration equal to or greater than 0.6 g. · Acceleration or deceleration equal to or greater than 0.5 g coupled with a forward time-to-collision (TTC) of 4 s or less. · All longitudinal decelerations between 0.4 g and 0.5 g coupled with a forward TTC value of ≤ 4 s and that the corresponding forward range value at the minimum TTC is not greater than 100 ft.
3. Event button	· Activated by the driver by pressing a button located on the dashboard when an event occurred that he/she deemed critical.
4. Forward time-to-collision	· Acceleration or deceleration equal to or greater than 0.5 g coupled with a forward TTC of 4 s or less. · All longitudinal decelerations between 0.4 g and 0.5 g coupled with a forward TTC value of ≤ 4 s and that the corresponding forward range value at the minimum TTC is not greater than 100 ft.
5. Rear time-to-collision	· Any rear TTC trigger value of 2 s or less that also has a corresponding rear range distance of ≤ 50 ft and any rear TTC trigger value in which the absolute acceleration of the following vehicle is greater than 0.3 g.
6. Yaw rate	· Any value greater than or equal to a plus and minus 4 degree change in heading (i.e., vehicle must return to the same general direction of travel) within a 3-second window of time.

After the events of interest were found, events were examined in detail by data reductionists. The method for recruiting, training, and performing quality checks of reduced data by data reductionists is described in detail in Dingus et al. (2005). Data reductionists viewed 90-second epochs for each event (60 s prior to and 30 s after), then reduced and recorded information concerning, among other things, the nature of the event and the roadway infrastructure. Over 110,000 triggers were reviewed in order to validate 9,125 events. The distribution of the total number of reduced events by severity is shown in Table 4. For the data-reduction process on valid events, data reductionists determined the severity of the event and recorded all of the required variables (discussed below) for the event type. The operational definitions of the event severity levels are presented in Table 5.

Table 4. The total number of events reduced for each severity level.

Event Severity	Total Number
Crash	69 (plus 13 without complete data)
Near-crash	761
Incidents (Crash-relevant Conflicts and Proximity Conflicts)	8,295

Table 5. Operational definitions of event severity.

Severity Level	Operational Definition
Crash	Any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably transferred or dissipated. Includes other vehicles, roadside barriers, objects on or off of the roadway, pedestrians, cyclists, animals, etc.
Near-crash	Any circumstance that requires a rapid, evasive maneuver by the subject vehicle, or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. A rapid, evasive maneuver is defined as a steering, braking, accelerating, or any combination of control inputs that approaches the limits of the vehicle capabilities.
Crash-Relevant Conflict	Any circumstance that requires a crash-avoidance response on the part of the subject vehicle, any other vehicle, pedestrian, cyclist, or animal that is less severe than a rapid, evasive maneuver (as defined above), but greater in severity than a “normal maneuver” to avoid a crash. A crash avoidance response can include braking, steering, accelerating, or any combination of control inputs. A “normal maneuver” for the subject vehicle is defined as a control input that falls outside of the 95 percent confidence limit for control input as measured for the same subject.
Proximity Conflict	Any circumstance resulting in extraordinarily close proximity of the subject vehicle to any other vehicle, pedestrian, cyclist, animal, or fixed object where, due to apparent unawareness on the part of the driver, pedestrians, cyclists, or animals, there is no avoidance maneuver or response. Extraordinarily close proximity is defined as a clear case where the absence of an avoidance maneuver or response is inappropriate for the driving circumstances (including speed, sight distance, etc.).

Of interest to the present analysis, the data reductionists recorded environmental variables and roadway characteristics that may have contributed to the presence of the crash, near-crash, or incident. The categories of environmental and roadway variables evaluated were: weather, light, surface condition, traffic density, relation to junction, traffic flow, traffic control, alignment, and infrastructure. Appendix A, Table 1 presents each environmental or roadway characteristic variable with the corresponding data reduction options for that variable.

In addition to the trigger-related event database, a baseline database was created which was comprised of approximately 20,000 6-second segments (referred to as an epoch). Kinematic triggers on driving performance data were not used to select these baseline epochs. The epochs were selected at random throughout the 12- to 13-month data collection period per vehicle. A 6-second segment of time was used 5 s prior to and 1 s after the onset of the conflict. The number of baseline epochs selected per vehicle was stratified as a proportional sample based upon vehicle involvement in crashes, near-crashes, and incidents. This stratification, based on frequency of crash, near-crash, and incident involvement was conducted to create a case-control dataset in which multiple baseline epochs are present to compare to each crash and near-crash. Case-control designs are optimal for calculating odds ratios (also referred to as relative near-crash/crash risk) due to the increased power that a case-control data set possesses. Greenberg, Daniels, Flanders, Eley, & Boring (2001) argue that using a case-control design allows for an efficient means to study rare events, such as automobile crashes, even though smaller sample sizes are used. Given that relative near-crash/crash risk calculations were an objective of the following analyses, the creation of a case-control data set was deemed important.

Considering that the number of baseline epochs was dependent upon the number of crashes, near-crashes, and incidents of vehicle involvement, not driver involvement, an analysis was conducted to determine the percentage of baseline epochs that were attributable to the primary driver and secondary drivers. The results indicate that 88.2 percent of the baseline epochs randomly selected were from the 109 primary drivers and 11.8 percent were from the 132 secondary drivers. Therefore, the vast majority of the baseline epochs were primary drivers. Four vehicles did not have any crashes, near-crashes, or incidents and were therefore eliminated from the baseline database.

Figure 2 shows the number of events in which each vehicle was involved (y-axis) and the corresponding number of baseline epochs that were identified for that vehicle (x-axis). Note that the vehicles that were involved in multiple crashes, near-crashes, and incidents also had a larger number of baseline epochs. Note that two drivers (data points on the far right side) were over-represented in their crash, near-crash and incident involvement.

Figure 2. The frequency of each vehicle’s involvement in crash, near-crash, and incident events versus the number of baseline epochs selected for each vehicle.

Training procedures were implemented to improve both inter- and intra-rater reliability given that data reductionists were asked to perform subjective judgments on the video and driving data. Reliability testing was then conducted to measure the resulting inter- and intra-rater reliability. The average inter-rater reliability compared to expert reductionists across all tests is 88.4 percent. The Kappa statistic was also used to calculate inter-rater reliability, which indicated that the association among raters is significant (K = 0.65, p <0.0001). Furthermore, a tetrachoric correlation coefficient is a statistical calculation of inter-rater reliability based on the assumption that the latent trait underlying the rating scale is continuous and normally distributed. The average of the pair-wise correlation coefficients for the inter-rater analysis is 0.86. The coefficients for the intra-rater analysis were extremely high with nine raters achieving a correlation of 1.0 among the three reliability tests and five raters achieving a correlation of 0.99. Inter-rater reliability tests were also conducted for the baseline events. Reductionists’ responses were compared to an expert data reductionist’s responses. Given these three methods of calculating inter-rater reliability, it appears that the data reduction training coupled with spot-checking and weekly meetings proved to be an effective method for achieving high inter- and intra-rater reliability.

RESULTS AND DISCUSSION

Descriptive Statistical Analysis

Data reduction was completed on 69 crashes, 761 near-crashes, and 8295 incidents resulting in a total of 9125 events. Of these 9125 events, 1159 events were identified as having infrastructure-related contributing factors. These events comprised 12.38% of the total events. Each infrastructure-related crash, near-crash and incident was plotted using Microsoft MapPoint. Appendix B, Figures B-1, B-2, and B-3 show these respective plots.

Table 6 shows the relative frequency of crashes, near-crashes, and incidents for each combination of infrastructure-related variables. Note that some events contained multiple types of infrastructure-related variables (i.e. visual obstruction in a construction zone). Of the 69 total number of crashes, 761 near-crashes and 8295 incidents, infrastructure-related events accounted for 23.19% of crashes, 14.59% of near-crashes, and 12.09% of all incidents.

Table6. Number of crashes, near-crashes, and incidents for each infrastructure event type.

Infrastructure Type	Crash	Near-crash	Incident	Total
Roadway delineation	3	2	132	137
Visual obstruction	3	71	552	626
Roadway alignment	2	12	119	133
Weather visibility	2	2	14	18
Roadway alignment w/ visual obstruction	3	5	7	14
Traffic control device	1	1	10	12
Roadway delineation w/ visual obstruction	1	3	6	10
Weather visibility w/ visual obstruction	1	0	14	15
Roadway delineation in construction zone	1	2	19	22
Roadway sight distance	0	2	18	20
Construction Zone	0	9	110	119
Roadway sight distance w/ visual obstruction	0	2	9	11
Traffic control device w/ visual obstruction	0	0	2	2
Roadway alignment in construction zone	0	0	7	7
Traffic control device in construction zone	0	0	4	4
Visual Obstruction in construction zone	0	0	7	7
Roadway alignment in construction zone w/ visual obstruction	0	0	1	1
Roadway delineation in construction zone w/ visual obstruction	0	0	1	1
Total	17	111	1032	1159

Figure 3 shows the frequency of the top ten infrastructure-related crashes and near-crashes. Visual obstruction accounted for the majority of infrastructure-related near-crashes (64%); followed by roadway alignment (11%), and construction zones (8%). Visual obstruction and roadway delineation both accounted for 19% of infrastructure-related crashes.

Figure 3. Frequency of top ten infrastructure-related crashes and near-crashes.

Figure

Figure 4 shows the frequency of the top 10 infrastructure-related incidents. Visual obstruction again accounts for the majority of infrastructure-related incidents (53%); followed by roadway delineation (12%), roadway alignment (11%) and construction zone (10.7%).

Figure 4. Frequency of the top ten infrastructure-related incidents.

High-Incident Roadway Analysis

Specified roadways in or near the Washington, DC area were identified a priori for this analysis based upon high-incident rates. The high-incident roadways are as follows:

I-66 (Custus Memorial Parkway)
US-50 (Lee Jackson Memorial Highway, John Moby Highway, Arlington Blvd.)
US-29 (Lee Highway, Arlington Blvd.)
SR-267 (Dulles Greenway, Dulles airport access road)
SR-28 (Sully Rd.)

Interestingly, of the 1159 infrastructure-related events, only 25 of these events occurred on the high-incident roadways. Table 7 shows the total frequency (main road and exit/entrance ramps) of each event type by the high-incident roadway. Appendix C, Figure C-1 shows these events plotted on a map of Northern Virginia.

Table 7. Frequency of events by severity.

Roadway	Severity
Roadway	Crash	Near-crash	Incident
I-66	0	0	2
US-50	0	1	12
US-29	0	0	5
SR-267	1	1	3

Each specified roadway event that occurred on exit/entrance ramps for each roadway of interest was also identified; these events are also represented as part of the total in Table 8. Table 8 shows the frequency of exit/entrance ramp events by severity.

Table 8. Frequency of events on exit/entrance ramps.

Roadway	Severity
Roadway	Crash	Near-crash	Incident
I-66	0	0	1
US-50	0	1	2
US-29	0	0	0
SR-267	1	0	3

The type of infrastructure that was a contributing factor for each event on these specified roadways was recorded. Table 9 shows a frequency count for events by roadway and infrastructure type. Note that roadway alignment and roadway delineation were the most common contributing factor, although the majority of the events where roadway alignment was a factor occurred on US-50.

Table 9. Frequency of events by specified roadway and infrastructure type.

Roadway	Infrastructure Type
Roadway	Roadway Alignment	Roadway Delineation	Roadway Sight Distance	Traffic Control Device	Weather Visibility	Visual Obstruction
I-66	0	2	0	0	0	0
US-50	7	3	2	0	1	0
US-29	0	1	1	3	0	0
SR-267	1	1	0	1	2	0

Out of all the specified roadways, only one crash was identified; this crash occurred on SR-267 and was labeled as a Category 2 crash. During this event the subject drove through a toll booth at night traveling at approximately 50 mph, lost control and collided with a barricade. Appendix D, Figure D-1 provides a dynamic description of the crash.

Data Modeling and Odds Ratio Calculations

Assessment of Risk

In this section we quantitatively evaluate the associations between crashes and near-crashes and the presence of environmental and roadway characteristics, as observed via video. For this analysis, all 69 crashes and 761 near-crashes as well as 20,000 baseline epochs were used in the assessment of risk. Note that only crashes and near-crashes were used in the assessment of risk based on an analysis conducted by Dingus, Klauer, Neale et al. (2006) that demonstrated that crashes and near-crashes had similar kinematic properties. The kinematic properties of the incidents were much more variable and determined to not be as good of a crash surrogate as the near-crashes. Thus, incidents were left out of the crash/near-crash risk analysis.

Note that this analysis was conducted using the presence of various environmental and roadway characteristics but not the infrastructure-related contributing factor variable. This is because this variable could not be recorded for baseline epochs since there was no dangerous situation in which the infrastructure could contribute to the cause or occurrence of the event. As described below, risk is calculated using both variables reduced for crashes and near-crashes as well as baseline epochs.

Definition of an Odds Ratio Calculation. A commonly used measure of the likelihood of event occurrence is termed as the odds. The odds measure the frequency of event occurrence (i.e., presence of inattention type) to the frequency of event non-occurrence (i.e., absence of inattention type). That is, the odds of event occurrence are defined as the probability of event occurrence divided by the probability of non-occurrence. The 2x2 contingency table in Table 10 will be used to illustrate this and related measures.

Table 10. An example of a 2x2 contingency table used to calculate the moderate traffic density odds ratio.

	Inclement Weather Present	Clear Weather Present
Reduced Event	n₁₁	n₁₂	n_1.
Baseline Event	n₂₁	n₂₂	n_2.
	n_.1	n_.2	n_..

If the probability of success (inclement weather present) for the first row of the table is denoted by π_1
=n₁₁/n_1. and the probability of failure (clear weather present) is defined as (1 –π₁) = n₁₂/n_1., then the odds of success is defined as π₁/(1-π₁) = n₁₁/n₁₂. The odds of success for the second row are defined similarly with the corresponding success probability, π₂.

The ratio of the odds is a commonly employed measure of association between the presence of cases (crash and near-crash events) and the controls (baseline driving epochs). Odds ratios are used as an approximation of relative near-crash/crash risk in case-control designs. This approximation is necessary due to the separate sampling employed for the events and baselines and is valid for evaluations of rare events. (Greenberg, Daniels, Flanders, Eley, & Boring, 2001). Referring to Table 10, the odds ratio would be defined as:

Equation 1

and is a comparison of the odds of success in row 1 versus the odds of success in row 2 of table 10.

Algebraically, this equation can be rewritten as shown below. Basic odds ratios are calculated as shown in Equation 2.

Odds Ratio = (A x D)/(B x C) Equation 2

Where:

A = the number of crash/near-crash events where <inclement weather> was present

B = the number of crash/near-crash events where clear weather was present

C = the number of baseline epochs where <inclement weather> was present

D = the number of baseline epochs where clear weather was present

To interpret odds ratios, a value of 1.0 indicates no significant danger above normal, baseline driving. An odds ratio less than 1.0 indicates that this activity is safer than normal, baseline driving or creates a protective effect. An odds ratio greater than 1.0 indicates that this activity increases one’s relative risk of a crash or near-crash by the value of the odds ratio. For example, if inclement weather obtained an odds ratio of 3.0, then this indicates that a driver is three times more likely to be involved in a crash or near-crash while traveling in inclement weather than if he or she was driving in clear weather.

Model Description

In the context of this analysis, the environment/highway infrastructure variables have multiple levels, as described in the Method section. For example, weather conditions were categorized into cloudy, rainy, inclement weather (combination of snow, sleet, and fog), and clear weather conditions. The question of interest is whether inferior weather conditions pose a higher risk to drivers than a “normal” or “safe” condition, namely a reference level. In the case of weather, clear weather conditions served as the reference level. Quantitatively, a condition is considered a risk factor if the probability of a crash or near-crash occurring is significantly higher than that of the reference level. As introduced above, the odds ratio is a convenient comparison index.

Therefore, a model-based approach was adopted to estimate odds ratios. This approach allows multiple factors to be evaluated in a single model and the estimated odds ratio is adjusted for the other factors in the model.

Model Setup

A logistic regression approach was used to model the crash/near-crash event and baseline data. In this approach, an observation comes from a binomial distribution with two possible outcomes: event (crash/near-crash) or non-event (baseline). We are interested in the probability of the occurrence of these events and non-events.

To develop our notations, let be the observation for driver and denote the number of drivers as and the number of observations for driver as . Here an index for the drivers was used. We define

where is index for drivers, and is the index for observations for a driver.

The probability of crash/near-crash for outcome , , is considered to be related to the corresponding environmental and infrastructure variables. They are connected through a regression in which the logit of , (the logarithm of odds) is set to be a linear function of explanatory variables, i.e.,

where ’s are the regression coefficients that are common to all observations and is the value of the explanatory variables for driver observation. For multi-level categorical explanatory variable cases, the corresponding is a vector with each element corresponding to a specific level.

The primary research objective is to assess whether there is evidence of elevated risk for a given condition compared to the appropriate reference level; e.g., if the probability of having an accident on a rainy day is higher than a clear day. In other words, is the odds ratio of accident for rainy day versus clear day significantly higher (lower) than 1.0? Mathematically it can be shown that this odds ratio is the exponent of the difference between the element of which corresponds to the rainy and clear day variables, i.e.,

where and are the elements of corresponding to rainy and clear days, respectively.

When several factors are considered in a sample model, the above estimation is the odds ratio controlled or adjusted for other factors. That is, the odds ratio of interest is assessed with the assumption that all other factors remain the same. Using a model to calculate odds ratios reduces the bias caused by potentially confounding effects among potential risk factors.

One source of bias in this type of data arises from the individual differences associated with each driver. Each individual driver has his/her own characteristics, and it is commonly believed that observations from the same driver are similar to each other. This violates the assumption of the ordinary logistic regression model that observations are independent. Individual driver effects can be accounted for by using a Generalized Estimation Equation (GEE) model, instead. The GEE model assumes the observations from different subjects are independent of each other and that a given correlation structure exists for observations from the same individual. The interchangeable correlation structure can be adopted using the following form.

It has been well established that the GEE estimation is robust for different correlation structures and a sensitivity analysis using different correlation structures confirms this.

Model Fitting, Results, and Discussion

Eight variables were considered for this model: weather, roadway surface condition, relationship to junction, road alignment, lighting conditions, traffic control, lane types, and traffic density. The GEE estimation for the correlation coefficient,, is 0.013, which indicates a rather weak correlation among observations for the same driver. The odds ratio and the corresponding confidence intervals are estimated using the contrast method. The results are listed in Table 11followed by a discussion of the model output.

Table 11. GEE logistic regression odds ratio estimation results

Factors	Odds Ratio	Standard Error	Confidence Limits		P-value*
Factors	Odds Ratio	Standard Error	Confidence Limits		P-value*
Weather: Cloudy vs. clear	3.75	0.60	2.74	5.13	<.0001
Weather: Fog/Mist/Sleet/Snow vs. clear	1.26	0.62	0.48	3.28	0.639
Weather: Raining vs. clear	0.27	0.20	0.07	1.15	0.076
Surface condition: Icy/Muddy/Snowy vs. Dry	1.28	0.47	0.62	2.63	0.504
Surface condition: Wet vs. Dry	1.37	0.26	0.95	1.98	0. 093
Junction: Driveway vs. Non-Junction	7.46	2.75	3.63	15.35	<.0001
Junction: Entrance/exit ramp vs. Non-Junction	4.20	1.05	2.57	6.85	<.0001
Junction: Interchange Area vs. Non-Junction	9.26	2.81	5.11	16.80	<.0001
Junction: Intersection vs. Non-Junction	5.06	0.87	3.61	7.07	<.0001
Junction: Parking lot vs. Non-Junction	3.85	1.20	2.09	7.09	<.0001
Alignment: Curve Grade/Hillcrest vs. Strt Lvl	1.30	0.53	0.58	2.91	0.520
Alignment: Curve Level vs. Strt Lvl	1.46	0.17	1.17	1.83	0.001
Alignment: Strt Grd/Hillcrest vs. Strt Lvl	1.00	0.27	0.59	1.70	0.999
Lighting: Darkness lighted vs. daylight	1.13	0.14	0.88	1.45	0.357
Lighting: Darkness no lighted vs. daylight	1.01	0.17	0.73	1.39	0.965
Lighting: Dawn vs. daylight	5.49	1.79	2.90	10.38	<.0001
Lighting: Dusk vs. daylight	1.28	0.22	0.92	1.78	0.151
Control: Stop sign vs. No control	0.85	0.21	0.52	1.39	0.512
Control: Yield sign vs. No control	2.90	1.24	1.25	6.70	0.013
Control: Traffic signal vs. No control	0.87	0.14	0.64	1.18	0.359
Control: Lanes Marked vs. No control	0.60	0.16	0.35	1.00	0.052
Lanes: No lanes vs. divided	0.20	0.08	0.09	0.42	<.0001
Lanes: Not divided vs. divided	1.47	0.09	1.31	1.65	<.0001
Lanes: One-way vs. divided	0.72	0.17	0.46	1.14	0.160
Density: Forced/unstable vs. Free	3.28	0.77	2.08	5.20	<.0001
Density: Unstable temp vs. Free	6.82	1.18	4.86	9.57	<.0001
Density: With restriction vs. Free	1.18	0.10	0.99	1.40	0.067
Density: Stable vs. Free	4.01	0.46	3.21	5.02	<.0001

*Significant p values are in bold font.

Weather

Three levels for weather were used due to low frequency counts for snow, fog, mist, and sleet conditions. Thus the three levels included clear, cloudy, and inclement weather (combined fog, mist, sleet, and snow conditions). Recall that weather was reduced based upon a trained data reductionist reviewing the video surrounding the event. Reductionists recorded an event as ‘rainy’ if the windshield wipers were active, ‘cloudy’ when the windshield wipers were not activated but the roadway was wet, and ‘inclement’ if the wipers were active but the precipitation was something other than rain.

The odds ratio estimation indicates that there is a three-fold increase in crash/near-crash involvement during cloudy weather compared to clear weather. A potential explanation for this result is that drivers did not drive as cautiously on wet pavement when it was not actually raining as they did when it was raining. This result is also interesting since it may indicate that drivers are slowing down in rainy conditions more so due to reduced visibility than fear of losing traction or control of their vehicle. Thus, drivers do not account for a reduced ability to slow/stop their vehicle when the pavement is wet; but they will slow down in poor visibility conditions. These results support the development of innovative pavements with a higher coefficient of friction under wet conditions. The results from this analysis indicate that such improvements in pavement may very well result in saved lives.

Surface condition

There are three types of surface conditions assessed: wet, dry, and the combined icy, muddy, snowy surface condition. The last category is aggregated due to sparse data within these categories. The results indicate there is no increased crash/near-crash risk due to wet surfaces or icy/muddy/snowy surfaces. This result may at first appear to contradict the result for weather condition. However, these two results should be interpreted in relation to each other in that the wet pavement condition for this variable is essentially the aggregate of ‘rainy’ and ‘cloudy’ weather conditions. Thus, when these two conditions are aggregated, the risk level drops most likely because the cautious behavior during rainy, wet pavement conditions counteracts the risk levels associated with what was previously labeled cloudy, wet pavement conditions. This is most likely due to the higher frequency of occurrence of rainy plus wet pavement vs. cloudy plus wet pavement.

Relationship to junction

There are six categories for the relationship to junction factors:

a. driveway

b. entrance/exit ramp

c. interchange area

d. intersection

e. parking lot

f. non-junction section

The non-junction section was used as the reference level. The model output indicates that there are statistically significant increases in crash/near-crash probabilities for all categories. Results show that the interchange area is the most dangerous, with the odds of crash/near-crash involvement being over nine times that of driving in the non-junction sections. The odds for driveways, intersection, entrance/exit ramps, and parking lot all show elevated risks of 7.5, 5.0, 4.2, and 3.6, respectively. While it is well documented in transportation research that any roadway junction has a much higher rate of crashes than any straight road segment, the relative risk levels of each of these roadway junctions contains some useful information. While interstates and limited access highways are generally deemed to be safer than arterials and surface streets, the interchange and entrance/exit ramp sections have higher or comparable levels of risk to junctions on other road types. This may be an artifact of the high traffic volumes in the Washington, DC area. Regardless, this is a finding that road designers in high traffic density areas should take into account in new interchange and ramp design.

Road alignment

There are four road alignment types: curve grade/ hillcrest, curve level, straight curve/hillcrest, and straight level. The straight, level roadway was used as the reference level. The results indicated that only curve level was significantly different from the straight, level roadway with an odds ratio estimation of 1.46. This result indicates a slight increase in crash/near-crash risk for travel on curved, level roadways. Typically, roadways with reduced sight distance (i.e., curved roadway with a grade, straight roadway with grade, and some curved roadways) all possess a higher crash risk. These results or lack of heightened crash risk for graded roadways, are most likely an artifact of the geographic location where data were collected. The topography of the Washington, DC area is relatively flat and level with very few graded roadways. If data were collected in a particularly hilly, mountainous region (such as San Francisco) the odds ratios for curve grade/hillcrest and straight grade, hillcrest would mostly like be significant as well.

Lighting conditions

There were five ambient lighting conditions recorded by reductionists: darkness lighted, darkness not lighted, dawn, dusk, and daylight. Naturally, daylight is considered as the reference level. The results indicated that dawn is significantly more risky than daylight. However, it should be mentioned that the confidence interval is quite wide, implying high variability potentially caused by small sample size. No other lighting conditions were found to be significantly different from daylight.

Traffic control

There are five traffic control options that were recorded by trained reductionists: stop sign, yield sign, traffic signals, lanes painted with traffic control direction, and no traffic control device present. Traffic volume (AADT) and road geometry will have non-negligible effects for traffic control devices and will be very similar to the results from the relation to junction variable. Therefore, the inclusion of traffic control device for this analysis is mainly to adjust for these factors when assessing the risks associated with intersections, merge ramps, etc.

Lane type

There were four operationally defined levels of lane type: no lanes marked (i.e. parking lot), not divided, one-way road, and divided highway. The divided highway was used as the reference level. The resulting odds ratios indicated that no-lanes are significantly safer than a divided highway. Non-divided highways have a significantly higher risk than divided highways. Caution is urged in the interpretation of lane type as it could be biased without consideration for traffic demand and highway geometries. However, it makes sense that non-divided highways do possess a higher risk of crash/near-crash occurrence than divided roadways, given the risk of a middle lane line crossing. It also makes sense that parking lots are safer than divided highways due to the dramatically lower speeds.

Traffic density

Based on highway level of service (LOS) the following traffic density levels were recorded and defined as follows:

· LOS A: Free flow,

· LOS B: Flow with some restrictions,

· LOS C: Stable flow maneuverability and speed are more restricted,

· LOS D: Unstable flow- temporary restrictions substantially slow driver,

· LOS E: Flow is unstable – vehicles are unable to pass, temporary stoppages, etc.

· LOS F: Forced traffic flow condition with low speeds and traffic volumes that are below capacity.

Due to the relatively small sample size, LOS E and F were aggregated into a single category. The LOS A, free flow, was adopted as reference level and the analysis confirmed the lowest associated risk. The results indicated a relatively low risk for crashes/near-crashes for either very low or very high traffic densities (LOS A, B, and E/F) with a higher crash/near-crash risk for medium traffic flows (LOS C and D). LOS B did not significantly differ from LOS A. LOS D and E both demonstrated a significantly higher risk with odds of four and seven times greater compared to free flow. LOS F demonstrated a three-fold increase in the odds of a crash/near-crash. These results may indicate that it is relative safe when the interactions among vehicles are minimal, i.e., LOS A and B. When the traffic becomes very congested, the chance of risk is higher when compared to free flow traffic. As was found in Klauer, Sudweeks, Hickman, & Neale (2006), this may be due to drivers being more alert with regard to the surrounding environment when traffic patterns are erratic coupled with slower vehicle speeds. Interestingly, the moderately congested roadways (LOS C and D) had the highest risk levels. This may be due to the higher relative speeds associated with these traffic conditions. Driver expectation may also be a component as well. Under conditions of heavy traffic congestion (LOS E and F), drivers are expecting that traffic is going to stop in front of them. However, in moderate levels of traffic (LOS C and D), drivers may or may not expect traffic to suddenly slow or stop in front of them. Coupled with higher rates of speed, these moderate traffic densities become far riskier than either low or extremely high traffic densities.

conclusions

Naturalistic driving studies provide a unique and novel perspective for highway designers and engineers. While crash/incident rate data are valuable, there are inherent limitations that naturalistic studies can overcome. The collection of police-accident reports over multiple years can provide insight into those roadways and/or intersections that may be dangerous. Naturalistic driving studies can provide far more detail about driver behavior as well as rates of non-police-reported crashes, near-crashes, and lower severity safety-related incidents for the same geographic locations. The frequency rates of police-reported crashes to non-police-reported crashes to near-crashes to incidents can be extrapolated from a naturalistic study to a particular roadway by comparing the police-reported crash rate from the naturalistic study data and police-reported crash rate from traditional sources (i.e., crash database). Thus, the naturalistic study can provide a better understanding of the rate of all types of incidents in a specific area than can a crash rate count using only traditional methods.

Secondly, while the actual frequency counts of crashes are much smaller, more detail can be gleaned by looking at the types of infrastructure-related contributing factors as well as driver behavior in these situations. For example, the crashes that occurred in a toll lane highlighted the combined problem of high speed, narrow E-ZPass lanes, and cell phone use.

Third, naturalistic driving studies can also provide a more precise calculation of crash risk. Previous studies assessing the crash risk for various roadway characteristics and infrastructure are calculated based upon police-accident reports and AADT for particular roadways. While this is a commonly accepted method of analysis, it can only provide a gross estimate of crash risk. It does not include non-police-reported crashes (which some estimate to be 5:1 ratio of non-police-reported to police-reported crashes). In addition, these gross estimates do not necessarily account for details such as the typical type of driver for a given segment/section of interest. For example, some intersections close to high schools will have a higher percentage of teen drivers crossing that intersection than any other age group. Conversely, an intersection that is located close to a retirement community may have a far higher percentage of elderly drivers crossing that intersection. These driver populations in combination with various types of roadway infrastructure may have an elevated crash rate compared to a similar intersection that is not located near a high school or retirement community.

The results from the infrastructure-related contributing factor analysis indicated that visual obstructions, roadway alignment, and construction zones accounted for the highest frequency of infrastructure-related crashes and near-crashes in the metropolitan Washington, DC area. The same was true for incidents with the addition of roadway delineation in the top four contributing factors. An analysis of a priori selected high-incident roadways did not yield as useful results due to the low frequency of crash/near-crash/incidents on any one roadway . The one crash that occurred on a high-incident roadway was a crash in the near vicinity of a toll booth with an E-ZPass lane. This driver was talking on a cell phone while traveling through the E-ZPass lane. He attempted to pass through the toll booth at 50 MPH and lost control, colliding with a sand-filled barricade. In the future, E-ZPass lanes may require a vehicle speed display to alert drivers that they are traveling too fast and need to slow down to safely pass through the toll booth.

A future analysis could take these high-incident roadways or other high-incident intersections, for example, and look at the occurrence of police-reported crashes, non-police-reported crashes, near-crashes, and incidents. Given the a priori known rate of police-reported crashes, an extrapolation using the relationship of police reported to non-police-reported crashes to near-crashes to incidents could be conducted to gain a better estimate/understanding of the real rate of crashes and near-crashes that are occurring at given segments of roadway. This may help roadway designers better gauge the dangerous locations and where to place resources that will reduce these dangers and make roads safer to travel.

The Generalized Estimation Model that was used to calculate crash/near-crash risk while controlling for driver as well as the other variables in the model provided unique and novel results. First, the risk of driving on wet roadways immediately after a rain shower or during snow melt is riskier than driving during rainy or otherwise inclement weather. This may be due to the fact that drivers engage in more cautious driving behavior when their visibility is limited (when it is raining or snowing). If visibility is not diminished, drivers engage in normal driving and do not take into account the reduced coefficient of friction and the associated lengthened stopping times/distances that are present when roadways are wet. This provides an argument for the development of pavement with higher coefficient of friction during wet conditions. Light-vehicle drivers typically do not understand the limitations of their vehicles during these conditions which is highlighted by this high risk of crash/near-crash result.

There is also greater risk of crash/near-crash involvement on all areas of roadway junctions, which in and of itself is not a surprising finding. Of interest is that limited access roadways, typically considered to be safer than surface streets, were found to have higher or similar risks of crash/near-crash occurrence at the interchanges and on merge ramps as do surface street intersections. This may be an artifact of the high traffic congestion that is typical of a larger U.S. metropolitan city; however, this information may or may not generalize to other metropolitan areas in the U.S.

The risks associated with various levels of traffic density suggested that the risk of crash/near-crash involvement was lower for very low and very high traffic congestion; whereas the risk of crash/near-crash involvement was much higher for moderate levels of traffic congestion. This result is most likely due to the higher speeds associated with moderate levels of traffic density as compared to high levels of traffic density, but also to driver expectation. Under high levels of traffic density, drivers are most likely more attentive because they expect that traffic will slow and stop. Under moderate levels of traffic density, drivers are less likely to anticipate or expect traffic to slow/stop suddenly and thus may also be more likely to engage in secondary activities that direct their attention away from the forward roadway. Thus, when traffic slows/stops, the drivers are more likely to not be aware and become involved in a crash or near-crash. This result has implications for variable traffic message signs, both external and internal to the vehicle. Providing real-time, up-to-date traffic information could be very helpful in reducing crashes by increasing the driver’s expectation that traffic could slow and stop suddenly as well as increase roadway throughput by informing drivers when traffic is less likely to slow or stop.

Other risks included driving during the ambient lighting conditions present during dawn. While this was significant, the confidence interval is very wide due to low power. When aggregated with dusk in the attempt to create a ‘limited ambient light condition’ with higher statistical power, the effect was no longer present. Given the wide confidence interval and the results of the aggregated condition, caution is urged in the interpretation of this result.

Curved-level roadways were also found to have a higher rate of crash/near-crash occurrence than straight, level roadways. Interestingly, roads with a grade were not found to have higher associated crash/near-crash risk. This finding may be an artifact of the topography of the Washington, DC area; thus, this may be useful information for roadway designers in the Washington, DC area but not necessarily useful information for roadway designers in other parts of the country with different topography.

Future analyses with these data could be conducted to calculate the risks associated with multiple levels of these roadway variables to more precisely assess the riskiest roadway conditions. For example, a comparison of undivided versus divided roadways at each level of traffic density could yield very interesting crash/near-crash risk results and interactions that could be used for both traffic control and roadway design.

While this study represents the largest instrumented vehicle study conducted to date (100 vehicles for 12 months of data collection), there are some limitations to the study that need to be considered in the interpretation of these results. The data were collected in one geographical location in the U.S.: the metropolitan Washington, DC area. While many of these results may be generalized to other metropolitan cities in the country, others may not. These data should not be generalized or cautiously generalized to rural drivers.

Naturalistic data should not be used to identify high-incident areas or segments of roadway due to the nature of the data collection. These data are limited in detecting a particular area with high incidents given that the data collection system was not designed with this intent. These data can be used to extrapolate from known incident rates to rates of non-police-reported and near-crash rates to better understand rates associated with a wide variety of incidents. The details of these crashes and near-crashes, including the driver behavior associated with these crashes and near-crashes, can give highway designers insight into methods to improve highway safety and reduce fatalities.

REFERENCES

Dingus, T. A., Klauer, S. G., Neale, V. L., Petersen, A., Lee, S.E., Sudweeks, J., Perez, M. A., Hankey, J., Ramsey, D., Gupta, S., Bucher, C., Doerzaph, Z.R., Jermeland, J., & Knipling, R. R. (2005). The 100-Car Naturalistic Driving Study: Phase II – Results of the 100-Car Field Experiment. (Interim Project Report for DTNH22-00-C-07007, Task Order 6; Report No. TBD). Washington, DC: National Highway Traffic Safety Administration.

Greenberg, R.S., Daniels, S. R., Flanders, W. D., Eley, J. W., & Boring, J. R. (2001). Medical Epidemiology, 3^rd Edition. New York City, New York: McGraw-Hill.

Kar, K & Datta, T. K. (2008). An Approach to Identify Areas with Traffic Safety Issues Due to Driver Behavioral Factors – Arizona Experience. Presented at the 87^th Annual Meeting of the Transportation Research Board. CD-ROM.

Klauer, S. G., Sudweeks, J. D., Hickman, J. S., & Neale, V. L. (2006). How risky is it? An assessment of the relative risk of engaging in potentially unsafe driving behaviors. (Contract No. 51090). Washington, DC: AAA- Foundation for Traffic Safety.

Klauer, S. G., Dingus, T. A., Neale, V. L., Sudweeks, J.D., & Ramsey, D. J. (2006). The Impact on Driver Inattention on Near Crash/Crash Risk: An Analysis Using the 100-Car Naturalistic Driving Study Data. (Report No. DOT HS 810 594). Washington, DC: National Highway Traffic Safety Administration.

Rosenthal, T.J., Christos, J.P., Aponso, B.L., & Allen, W.R., (2004) “A Driving Simulator for Testing the Visibility and Conspicuity of Highway Designs and Traffic Control Device Placement.” Paper presented at the 83^rd Annual Meeting of the Transportation Research Board. CD-ROM.

Saccomanno, F. F., Cunto, F., Guido, G., & Vitale A. (2008). “Comparing Safety at Signalized Intersections and Roundabouts Using Simulated Rear-End Conflicts.” Presented at the 87^th Annual Meeting of the Transportation Research Board. CD-ROM.

Shankar, V., Mannering, M., & Barfield, B., (1994). “Effect of Roadway Geometrics and Environmental Factors on Rural Freeway Accident Frequencies.” Accident Analysis and Prevention, V27, No 3, pp. 371-389.

Yan, X., & Radwan, E. (2004). “Geometric Models to Calculate Intersection Sight Distance for Unprotected Left-Turn Traffic.” Paper presented at the 83^rd Annual Meeting of the Transportation Research Board. CD-ROM.

APPENDICES

APPENDIX A: Environmental and Roadway Characteristic Variables

Table A-1. List of Environmental and Roadway Characteristic Variables

Reduced Variable		Levels of Variable
Weather		Clear
		Cloudy
		Fog
		Mist
		Raining
		Snowing
		Sleeting
		Smoke dust
		Other
		Unknown
Light		Dawn
		Daylight
		Dusk
		Darkness, lighted
		Darkness, not lighted
		Unknown
Surface Condition		Dry
		Wet
		Snowy
		Icy
		Muddy
		Oily
		Other
		Unknown
Traffic Density		LOS A - Free flow
		LOS B - Flow with some restrictions
		LOS C - Stable flow, maneuverability and speed are more restricted
		LOS D - Unstable flow, temporary restrictions substantially slow driver
		LOS E - Flow is unstable, vehicles are unable to pass, temporary stoppages, etc.
		LOS F - Forced traffic flow condition with low speeds and traffic volumes that are below capacity. Queues forming in particular locations.
		Unknown
Relation to Junction		Non-Junction
		Intersection
		Intersection-related
		Driveway, alley access, etc
		Entrance/exit ramp
		Rail grade crossing
		On a bridge
		Crossover-related
		Other, non-interchange area
		Unknown, non-interchange
		Parking lot
Trafficway Flow		Not Divided
		Divided (median strip or barrier)
		One-way traffic
		Unknown
Traffic Control		No traffic control
		Officer or watchman
		Traffic signal
		Stop sign
		Slow or warning sign
		Traffic lanes marked
		No passing signs
		Yield sign
		One-way road or street
		Railroad crossing with markings or signs
		Railroad crossing with signals
		Railroad crossing with gate and signals
		Other
		Unknown
Alignment		Straight level
		Curve level
		Grade straight
		Grade curve
		Hillcrest straight
		Hillcrest curve
		Dip straight
		Up curve
		Other
		Unknown
Infrastructure		None
		Roadway Alignment
		Road Sight Distance
		Traffic Control Device
		Roadway Delineation
		Weather, Visibility
		Construction

Appendix B: Location of Infrastructure-related Crashes, Near-crashes, and Incidents.

Figure B-1. Location of infrastructure-related crashes within the Northern Virginia area.

Figure B-2. Location of infrastructure-related near-crashes.

Figure B-3. Location of infrastructure-related incidents.

Appendix C: The 25 Infrastructure-related Events that Occurred on the

High-incident Roadways of Primary Interest.

Figure C-1. Location of the infrastructure-related events that occurred on the high- incident roadways of primary interest.

Appendix D: Infrastructure-Related Crash Description

Guidance through automated toll pass

Crash Number –

0020307310154009590

The SV is heading northbound on route 267, a four lane high-speed toll road. The male driver is engaged in an emotional cell phone conversation with his girlfriend. While approaching the toll station driver decreased speed slightly from 60 mph to 50 mph. While passing through the narrow toll-in-motion lane the driver failed to keep the vehicle centered. As the vehicle exits the toll section the left side first contacted a sand barrier or curb. The driver reacted with a steering input to the right. Overcorrection reoriented the vehicle too far, resulting in a second collision with a sand barrier along the right edge of the lane. A second overcorrection to the left resulted in further contact between the left side of the vehicle and two or three cones. Significant body damage as well as a broken front axle results from the multiple impacts.