Influence of Infrastructure Design on Distracted Driving (2022)

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.

96 Chapter 9. Safety Framework to Assess the Impact of Overhead Dynamic Message Signs on Distraction Using the SHRP2 NDS Data 9.1 Introduction 9.1.1 Background Similar to the Safety Framework described in Chapter 8, this Safety Framework utilized the SHRP2 NDS data. This data set is described in more detail in Sections 4.1 and 8.2.1. The main advantage of using NDS data in general to evaluate IRD is that driver behavior can be linked to actual events (normal or safety critical) and roadway characteristics. While the Safety Framework described in the previous chapter used the SHRP2 NDS data to evaluate railroad crossings, the Safety Framework described in this chapter used the SHRP2 NDS data to evaluate the relationship between distraction and an overhead DMS. DMS are utilized on roadways to provide drivers with real-time information. Common uses of DMS include providing drivers real-time travel time information, information about congestion, information about nonrecurring events (e.g., crashes, special events), safety messaging (number of crashes to date, safety slogans, etc.), and alerting drivers to road weather. Figure 22 shows a typical overhead DMS. Source: rawf8, Shutterstock. Figure 22. Typical overhead DMS. DMS should only be active when relaying pertinent information to drivers. Past research has found high driver compliance with the messaging provided by a DMS (Erke et al. 2007), which suggests that drivers are paying attention to them. However, the sign may be distracting if drivers focus overly long on reading and comprehending the message or engaging in other distracting activities such as route planning or cell phone usage based on information provided by the DMS (e.g., when the message is “congestion ahead”).

97 Prior studies have examined effects of DMS on driver behavior. One study using eye tracking software found that drivers fixated on DMS more than on standard road signs (Anttila et al. 2000). A study by Erke et al. (2007) looked at changes in a driver’s route, speed, and other behaviors such as braking when they encountered two DMS alerting them to a closed roadway ahead. The study found that speeds decreased while speed variation increased. This reduction in speed as well as an increase in braking near the DMS was attributed to attention overload or driver distraction. A simulator study by Spell et al. (2014) found that drivers slowed down when encountering a DMS with quantitative information displayed on a 65 mph roadway. No statistically significant changes were seen, however, on 55 mph roadways. In a recent study by Ermagun et al. (2021), the researchers looked at distractions caused by a compromised DMS. Through stated preference surveys, they found that drivers were cognitively, visually, and manually distracted by a compromised DMS. While these studies found potential links between the presence of DMS and driver distraction, they all focused on small samples of one to two locations and included a variety of data sources, including on-road, simulation, and stated preference surveys. The use of SHRP2 NDS data to evaluate this question provides the advantage of including multiple drivers on a variety of roadway sites. The forward video typically included with NDS data allows other potentially confounding factors (e.g., congestion) to be coded and accounted for. In-vehicle video can be used to determine information on glance location and duration as well as any distraction the driver was involved in during events of interest. 9.1.2 Scope of Framework The objective of the analysis presented in this chapter was to develop a Safety Framework to evaluate the relationship between distraction and the presence of an active overhead DMS using the SHRP2 NDS data and to determine whether these data are feasible for this purpose. A general list of research questions relevant to assessing driver distraction due to overhead DMS was formulated, and the following three questions were selected as being the most feasible given the available data and the project objectives: • What is the relationship between an active overhead DMS and driver glance performance? • What is the relationship between an active overhead DMS and driver lane keeping? • What is the relationship between an active overhead DMS and driving speed? Since data relevant to all of these driver behaviors were available in the curated data sets utilized for this Safety Framework, an attempt was made to evaluate the potential effects of DMS on distraction by addressing all three questions. 9.2 Data Sets Utilized The study utilized several SHRP2 NDS data sets to develop a Safety Framework, as described in this section. Data acquisition and data reduction are discussed in the following section.

98 9.2.1 Curated SHRP2 NDS Data Sets This Safety Framework utilized SHRP2 NDS data. More information about this data set is provided in Section 4.1 and 8.2.1. VTTI provides access to SHRP2 NDS data sets that were curated for other researchers. These can be obtained at low or no cost through a data request. Each record in the data sets typically provides a time series trace (speed, acceleration, position, and other kinematic vehicle characteristics at 0.1-second intervals), forward roadway view video, and rear roadway view video. In some cases, reduced data such as glance location or distraction are available. Several research projects were identified that had the potential of including an overhead DMS. Each was reviewed, and multiple data sets were excluded because they did not include forward video, which was needed to confirm DMS presence and status (on/off). From the review of the curated SHRP2 NDS data sets, it was determined that the following data sets were the most feasible for answering the identified research questions: • Driver Behavior and Performance in the Vicinity of Closely Spaced Interchange Ramps on Urban Freeways – Proof of Concept (doi:10.15787/VTT1/RCXYJV). The data had been provided to the University of Utah. The data included time series traces, forward video, and variables about cell phone use for traces through roadway segments located in Washington and North Carolina that included an entrance ramp followed by a downstream exit ramp. • Driver Behavior in the Vicinity of Closely Spaced Interchange Ramps on Urban Freeways – Phase 2 (doi:10.15787/VTT1/IMC8YV). A total of 18,000 traces through 60 freeway segments on closely spaced interchange ramps located in Washington and North Carolina had been provided to the University of Utah. The data included time series traces, video, and information on distraction and cell phone interaction and were grouped by entering, exiting, and through drivers. 9.2.2 Safety-Critical Events Another source of existing SHRP2 NDS data is a data set of approximately 8,790 crashes, near- crashes, or conflicts (SCEs) that VTTI identified and coded. Time series data and a forward roadway video clip are available for each SCE and are housed on the SHRP2 NDS website, InSight (https://insight.shrp2nds.us/). Other information can be extracted for SCEs if needed. For instance, the presence of relevant infrastructure items can be coded using forward video or the SHRP2 RID. Additionally, data for specific features of interest can be requested. For instance, if a series of billboards is identified, driving traces in the vicinity can be provided. The research team reviewed 3,207 crashes and near-crashes in the InSight database coded as involving a secondary task of “Other External Distraction” (i.e., looking at an object external to the vehicle), and only seven were found in the vicinity of an overhead DMS, the majority of which were variable speed limit signs. After looking at the narratives associated with each of these traces, none was found to involve a DMS. The narratives were also reviewed to see whether they contained the words “DMS,” “overhead,” or “dynamic,” but none of these words were found. As a result, no additional data were available from this source.

99 9.3 Data Request and Data Reduction IRB approval was obtained through Iowa State University for the use of the SHRP2 NDS data, after which a data request was made to VTTI. However, it was found through the request to VTTI that neither the second data set nor the cell phone data for the first data set were currently available. Available data included the following: • 240 traversals (80 entering, 80 exiting, 80 through) at eight locations (1,920 traversals total): o Time series data. o Forward video. The data included an area extending 10 seconds upstream of an entrance ramp to 10 seconds downstream of an exit ramp. Of the eight locations in the data request, only two from North Carolina had a traditional DMS and were therefore the only locations utilized for this Safety Framework. In addition to the existing data, the team had VTTI code glance location and visual distraction for 20 traces from the two North Carolina sites, including 10 traces where the DMS was on and 10 where the DMS was off. (Only 20 traces were reduced due to available project resources.) The coded data resulted in only four visual distractions. Therefore, glance location was utilized as the sole measure of visual distraction for this Safety Framework. This Safety Framework also aimed to answer questions related to surrogate measures of distraction, including driver speed and lane keeping ability, so only traces that had accurate speed and lateral position data were included. The speed data were considered accurate if speed readings were available for a minimum of 50% of the trace, which correlates to at least one speed reading every 0.2 seconds. A confidence measure provided as part of the SHRP2 NDS time series data was used to determine accuracy for lateral position. The following two variables were utilized for reliability: vtti.right_marker_probability and vtti.left_marker_probability. These variables indicate the probability that the vehicle-based machine vision lane marking evaluation is providing correct data for the right and left side lane markings, respectively. The value for each reading is a number from 0 to 1,024, and VTTI considers anything above 512 to be accurate. Traces were further filtered to only include those for which at least one of the two marking probabilities was above 512 for at least 75% of the readings. This filtering for traces with accurate speed and lateral position data resulted in 225 potential traces that were then manually coded. For each trace, the presence and status (on/off) of the DMS in the traversals utilized were confirmed. The following variables were reduced for each trace: • Location of DMS: Timestamp where the driver encountered the DMS, reduced from forward traversal video. • Time of day (e.g., day, night): Reduced from forward traversal video. • Ambient conditions (e.g., dry, rain): Reduced from forward traversal video. • Status of the DMS (e.g., on, off): Reduced from forward traversal video. • Congestion status (e.g., congested, slightly congested, free flow): Reduced from forward traversal video.

100 • Location of lane change: Timestamp where driver finished changing lanes, reduced from forward traversal video. • Driver ID: Driver information was not provided; the forward video was utilized to determine if the same car and therefore same driver was used in multiple traces; a unique identifier was given to each driver. During the manual coding, traces were removed for a variety of reasons, including the driver exiting before encountering a DMS, the presence of congestion or construction, the driver making multiple lane changes, or other scenarios that did not represent normal driving conditions and could skew the results. These removals resulted in 174 remaining traces that were utilized for the analyses of speed and lane position. Using timestamps, features of interest were manually identified and the location noted in the time series data (e.g., the timestamp associated with the location of the overhead DMS). Using speed data and the known time of each row of data (with each row representing 0.1 seconds), the distance of the subject vehicle from each infrastructure element could be calculated at any point (e.g., to determine glance location at 800 ft upstream of the DMS). 9.4 Analysis 9.4.1 Variables Included The use of several dependent variables was investigated. Analyses were ultimately conducted using average speed, standard deviation of speed, standard deviation of lateral position, and glance location as dependent variables. Glance location was reduced for 20 of the traces within the data set, as noted above. Glances were coded as follows: • Forward (glances to the center, left, or right that involve little or no head movement and appear to be mostly directed to the left or right portions of the windshield). • Center console (e.g., glances toward climate controls or radio). • Steering wheel (glances toward speedometer, fuel gauge, or cruise control). • Down (glances toward something in the driver’s lap or on the floor). • Up (glances toward the upper portion of the windshield, often associated with visor or sunroof). • Left (glances to the left of the A-pillar). • Right (glances toward the right that involve both eye and head movement). • Rear view mirror. • Over the shoulder (glances over the driver’s left or right shoulder and requiring the driver’s eyes to pass the B-pillar). • Other (blinks, squints, or closed eyes that last more than 10 frames). • Missing (poor video, eyes obscured, or glance cannot be determined). The timestamp, glance location, and any visual distraction associated with a glance away from the roadway was coded for each video frame at 15 Hz. This information was connected to the time series data collected at 10 Hz by matching the timestamp in the video to the closest

101 timestamp within the time series data. As a result, the number of times a driver glanced at a particular location and the length of each glance could be determined. VTTI had coded visual distractions for only four of the traces, as noted above, and none of these distractions appeared to be directly related to the DMS. Therefore, glance location was used as a surrogate for distraction. Drivers distracted by an overhead DMS may be expected to focus their glance on the DMS for an extended period of time or move their focus from other locations differently than if not distracted. As a result, one of the distraction metrics included the proportion of the time drivers spent glancing at the forward roadway within the DMS influence area when the DMS was on compared to the time drivers spent glancing at forward roadway within the DMS influence area when the DMS was off. The proportion of time was used rather than actual time because the time each driver was within the DMS influence area differed. A distance of 800 ft upstream of the DMS was chosen as the DMS influence area because it is the recommended legibility distance of changeable message signs per the Manual on Uniform Traffic Control Devices (MUTCD) (FHWA 2009). It was assumed that a driver would likely be more distracted by a DMS when the message on the sign could be read rather than when the sign was simply visible. In addition to glance location, the team also summarized various metrics that could be used as surrogates for distraction. These included average speed, which can be impacted by distraction because drivers often compensate for risk when distracted by lowering their average speed (Oviedo-Trespalacios et al. 2019); standard deviation of speed, because drivers who are distracted tend to have higher deviations in speed; and standard deviation of lateral position, because drivers who are distracted tend to move around more in their lanes (McDonald et al. 2020). Steering wheel position has also been used as a surrogate for distraction, but this metric was not available for the majority of traces and therefore was not able to be utilized in these analyses. The metrics utilized as dependent variables included the following: • Proportion of the time spent glancing at the forward roadway within the DMS influence area when the sign was on compared to the time spent glancing at the forward roadway in the DMS influence area when the sign was off. • Average speed within the DMS influence area when the sign was on compared to when the sign was off. • Standard deviation of speed within the DMS influence area when the sign was on compared to when the sign was off. • Standard deviation of lateral position within the DMS influence area when the sign was on compared to when the sign was off. The time series data for each trace were utilized to calculate the above metrics within the area 800 ft upstream of the DMS. Since a driver changing lanes is expected to have a larger deviation in lateral position, any traces where the driver changed lanes within the area 800 ft upstream of the DMS were removed. This left 103 traces for modeling the standard deviation of lateral position.

102 Independent variables included the following: • Time of day (day, night). • Weather (dry, rain). • Road ID (used to account for repeated samples). • Driver ID (used to account for repeated samples). • Sign status (on, off). • Movement (entering, exiting, through). • Lane change (yes, no). • Congestion (free flow, slightly congested). 9.4.2 Modeling Approach and Results for Glance Location An exploratory analysis of the data was conducted for each dependent variable. This section describes the analysis of glance location. Figure 23 shows the percentage of time drivers spent looking at the forward windshield within the DMS influence area when the sign was on (right-hand box) compared to when it was off (left-hand box). Ten observations were available for when the sign was on, and 10 observations were available for when the sign was off. As the figure shows, drivers are glancing forward approximately 91% of the time when the DMS was on compared to around 82% when the sign was off. Figure 23. Percentage of time drivers spent glancing forward when sign was on (right) or off (left).

103 The expected outcome is a difference in glance location when the sign is on compared to when it is off. As a result, the statistical model utilized needed to detect those differences as well as the impact of other covariates. Additionally, the exploratory analysis indicated that the data did not follow a normal distribution. Therefore, a gamma-distributed generalized linear model was utilized. After testing for potentially random effects due to repeated samples of Road ID and driver ID, it was found that these effects did not significantly influence the results of the model and therefore a mixed effects model was not needed. In a full analysis, the use of a mixed effects model should be evaluated due to the large sample of drivers and roadways that would be expected to be included. The best fit model that was developed for glance location included the status of the sign (on/off) and whether the driver changed lanes within the 800 ft influence area. The results of the model are presented in Table 10. Table 10. Percentage of time looking forward within DMS influence area. Variable Estimate Standard Error p-value Intercept 0.0118369 0.0006198 6.35e-13 Changed lanes within the influence area (base=no) 0.0040137 0.0011483 0.00277 Sign Status (base=off) -0.0021147 0.0008568 0.02449 Because the gamma distribution utilizes an inverse link function, interpreting the results is not straightforward. The impact of DMS status on the percentage of time drivers spent looking forward is given by the following: 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑇𝑖𝑚𝑒𝐹𝑜𝑟𝑤𝑎𝑟𝑑 = . = 84.48%, 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑇𝑖𝑚𝑒𝐹𝑜𝑟𝑤𝑎𝑟𝑑 = . . = 102.85%. Therefore, a driver was 1.22 (102.85/84.48) times more likely to be looking forward when the sign was on compared to when it was off. If a driver changed lanes within the area 800 ft upstream of a DMS, the driver’s glance location was 1.15 times more likely to be forward when the sign was on compared to when it was off. This suggests that drivers change the focus of their attention in the presence of a DMS and may suggest that drivers over-focus on DMS, which could have negative consequences. However, this focus may be positive if it increases driver attention. 9.4.3 Modeling Approach and Results for Speed and Lateral Position Models were also developed for three other surrogate measures: average speed, standard deviation of speed, and lateral position. Two different models were used: linear regression and gamma regression. The former is used when the response (dependent) variable either takes real values or the data are bell-shaped; linear regression was therefore utilized to model average speed (mph). The latter is used when the response (dependent) variable is nonnegative and the data are skewed; gamma regression was therefore used to model standard deviation of speed (mph) and lateral position.

104 Mixed-effects models are usually used to consider dependency introduced by countable groupings of variables, such as driver ID and Road ID. Regular (fixed effects) linear regression and gamma regression were used instead of a mixed-effects model because (1) the number of repeated driver ID values was not statistically significant and (2) the data come from only two different roads, which suggests that Road ID should instead be included as a fixed effect. The data thus favor a simpler model. The linear regression model has the following form: 𝑦 = 𝑥 𝛽 + 𝜖 where 𝑦 is the dependent variable observations, 𝑥 is the vector of dependent variables associated with the 𝑖-th trace, β is the vector of coefficients, and 𝜖 is the error. The errors 𝜖 are independent and identically distributed as N(0,σ2). In contrast, a gamma regression is used for a nonnegative response, and the covariates (or independent variables) are associated through a link function. That is, 𝑦 is distributed as a gamma with mean 𝜇 and variance 𝜇 , and the link function used is the inverse link, as follows: 𝑥 𝛽 = 1𝜇 Since 𝜇 = (𝑥 𝛽) μi = (xTiβ) − 1μi = (xiTβ) − 1, this means that in order to calculate the expected value of a given 𝑥, it is necessary to first compute 𝑥 𝛽, where 𝛽 is the vector of estimated effects, and then take its inverse. The models, either Gaussian or gamma, were adjusted with the function “gam” from the “mgcv” package in R. The final models were chosen using forward selection based on the Akaike information criterion (AIC), and the models were assessed with residual plots and generalized residual plots. The models for average speed and standard deviation of lateral position were unable to determine any correlation between sign status (on/off) and these surrogate measures. The best fit models instead included Road ID, congestion (free flow, slightly congested), and movement (entering, exiting, through) for the average speed and movement (entering, exiting, through) for the standard deviation of lateral position. Entering, exiting, or traveling through the roadway are expected to be correlated with both speed and variation in lane position. Drivers entering a highway slowly speed up, while drivers exiting a highway begin to slow down. Additionally, both entering and exiting a highway involves the potential for changing lanes and merging/diverging, which is associated with high variation in lateral positioning. Due to the limits of the available data and the small sample size, it was not possible to focus on only one movement or free flow conditions. The best fit model for standard deviation of speed was found to include these same variables along with some interactions. This was expected for the same reasons mentioned above. However, a lesser fit model was developed that did demonstrate a correlation between sign status and speed variation. This model included only an intercept and a binomial variable for sign

105 status. The estimates of the model are shown in Table 11 and demonstrate that standard deviation of speed is expected to be 1.1 times higher when the sign is active than when it is inactive or blank. Table 11. Standard deviation of speed within DMS influence area. Variable Estimate Standard Error p-value Intercept 0.456 0.0156 0 Sign Status (base = off) −0.044 0.0243 0.0713 9.5 Outcomes and Discussion 9.5.1 Summary A preliminary analysis was conducted to evaluate the impact of overhead DMS on distraction. A curated set of SHRP2 NDS data that included traversals on freeway sections featuring overhead DMS that were at times active and at other times inactive was utilized for the analyses. Twenty traversals (10 for each sign status [on/off]) coded for glance location and visual distractions 800 ft upstream of the DMS were reduced, and glance location was used as a surrogate for distraction. Modeling options were limited due to the small sample size. However, a gamma regression model was used to model the percentage of time a driver glanced at the forward roadway within 800 ft of the DMS. The model indicated that the proportion of time drivers spent glancing at the forward windshield was higher and statistically different when the sign was on compared to when it was off. The difference suggests that drivers focus more on the forward roadway when an overhead DMS is active. This behavior could be positive (attention to traffic control devices) or negative (fixation on the overhead sign rather than the road ahead). Models of average speed and lateral position were also created but were unable to capture the influence of DMS status. This may have been due to the small sample size as well as confounding factors that could not be controlled for due to the data set utilized, such as congestion and movement (entering, exiting, through). The presence of the sign itself may also have an effect on driver speed and lane positioning, but due to the data set utilized, upstream driving outside of the DMS influence area was unable to be utilized as control data. While not identified by the best fit model, it was found that the presence of an active DMS was correlated with an increase in driver speed variation. Although not conclusive, the results suggest a difference between glance behavior and a potential increase in speed variation when an overhead DMS is active compared to when it is inactive. 9.5.2 Performance Metrics Performance metrics suitable for measuring the success of the research activities and outputs were identified and are described below.

106 Feasibility of the Data Set Utilized This analysis used a reduced data set from the SHRP2 NDS. Naturalistic driving studies are well suited for this type of analysis because driver behavior can be compared for similar roadway environments or, in some cases, across multiple traces from the same driver on the same roadway near and away from the infrastructure element of interest (and, as in the case of a DMS, when the element of interest is active or inactive). Additionally, driver behavior can be captured when face position video data are reduced. The SHRP2 NDS data were uniquely suited to this analysis because the data set represents a large number of drivers over multiple years on a variety of roadway conditions. The presence of forward video also allows the status of dynamic infrastructure features (e.g., DMS, speed feedback signs, lowered railroad crossbucks) to be identified. In this case, DMS status (on/off) could be determined. In some cases, it was possible to determine the message that was displayed. The time series data collected as part of the SHRP2 NDS also include several variables that are potential surrogate measures of distraction, including lateral position, speed, and steering wheel position. Information on steering wheel position was not available in a sufficient number of traces in the data set utilized for this study, but it was present in a portion of traces. Encroachments into other lanes or the roadside can also be determined through lateral position, lane width, and vehicle track width data and then confirmed through the forward video. Additional discussion of the feasibility of using the SHRP2 NDS or another NDS to evaluate the relationship between distraction and infrastructure elements is provided in Section 8.5.4. Overall, the SHRP2 NDS data set is more than adequate for the purposes of evaluating the impacts of common roadway infrastructure elements such as DMS. The SHRP2 RID, the data for which were collected in conjunction with the SHRP2 NDS, includes a mix of rural and urban roads and provides information on the roadways used for a large percentage of the trips captured in the SHRP2 NDS. A mobile data collection van was used to collect data for about 12,500 centerline miles in the six SHRP2 NDS states. Other sources of existing data were collected and incorporated into the RID. The RID includes sign data and, more specifically, the presence of DMS. Approximately 1,000 DMS were found within the RID. These data, along with other data in the RID (e.g., speed limit, number of lanes, lane widths) can be used to select sites with specific characteristics. Data requests can then be made for traces near an identified DMS and can include specific requirements such as through movements only; present and accurate speed, lateral position, and steering wheel data for the area of interest; a mix of drivers and ages and potentially multiple samples from the same driver; specific times of day; a minimum number of samples from each site; and so on. Additionally, many states publish the locations of their DMS online. For instance, Indiana, one of the states in the SHRP2 NDS, provides the locations of its DMS in a map through its 511 website. These online sources can be utilized to identify potential roadway segments for data requests as well as the type of information posted on the signs (e.g., travel times only versus travel times or congestion messaging).

107 Sample Size As noted in Section 9.3, only 20 traversals were available for the glance location analysis and approximately 174 traces were available to model the other surrogates (i.e., average speed, standard deviation of speed, and standard deviation of lateral position). These 174 traces included a variety of drivers making different movements (i.e., entering, exiting, and through) and roadways with different levels of congestion. As a result, it was difficult to determine the statistical significance of the independent variables. However, it is presumed that the following independent variables are relevant and should be included in similar modeling efforts: • Driver age. • Type of vehicle (e.g., passenger vehicle, large truck). • Type of DMS (e.g., overhead, side mounted). • DMS messaging type (travel times, congestion messaging, safety messaging, etc.). • Lane widths. • Speed limit. • Traffic density (e.g., free flow, slightly congested, congested). • Time of day (e.g., night versus day). • Lighting. • Lane changes. • Driver ID. • Roadway ID. • Weather. It is suggested, when feasible, that analyses be constrained to free flow conditions. It is difficult to separate the impact of surrounding traffic on a driver’s behavior from the driver’s reaction to the infrastructure element. Moreover, it would likely be advantageous to restrict observations to only those captured during free flow conditions, those involving through movements, and those not involving lane changes because these variables can be expected to significantly impact the dependent variables. The above variables represent at least 20 different factors. Using an estimate of 50 samples per factor, a sample size of an approximately 1,000 traversals is needed. This number can be reduced if observations are restricted to those mentioned above or if only certain situations are examined (e.g., overhead DMS only or DMS that only display travel times). The curated data set utilized for this analysis, which was obtained through a data request to VTTI, included 240 traversals at two locations where a DMS was present for all movements and 160 traces at two locations where a DMS was present. The data set thus included a total of 800 traversals. It was not expected that these data on their own would be sufficient for analysis. As noted in Section 9.3, one of the data sets requested from VTTI would have expanded on the proof-of-concept data set that was ultimately used for the present analysis and may have provided traces for many more sites. However, the larger data set was unavailable at the time the research team’s data request was made. Moreover, as noted above, approximately 1,000 overhead DMS are coded within the RID, and additional driving traces through those locations could be requested.

108 As noted in Section 9.3, lane position and steering wheel position variables were only available for a subset of the data used in this analysis and are only reliably available for a small subset of the SHRP2 NDS data as a whole. As a result, if lane position, standard deviation of lane position, or deviation in steering wheel position are used as surrogates for distraction, it would be necessary to request only those traces for which these variables are present and accurate. Additionally, glance location would need to be reduced for at least 1,000 traces, which could be selected to represent a range of conditions. Location of the DMS within the time series data would need to be manually reduced for each traversal in order to determine vehicle position relative to the DMS (unless the position of the DMS is noted in the RID). Overall, it is expected that sufficient samples could easily be obtained to conduct a more robust analysis similar to the one described in this chapter. Moreover, the framework described for analysis using the SHRP2 NDS data can be applied to other NDS data sets. Performance of Statistical Models Because the analyses were used as proofs of concept to guide the development of the Safety Framework, sample sizes were limited to existing data that could be obtained within project resources. Due to sample size limitations, only simple metrics such as the p-value and AIC could be used to assess model performance. Additionally, due to the small sample size, only the impact of covariates could be determined; interactions between covariates could not be examined. However, a gamma generalized linear model did indicate that a difference in the percentage of time a driver glances forward could be derived when the DMS sign was active compared to when it was inactive. Variations in driver speed could also be derived when the sign was active compared to when it was inactive. Due to the limited sample size, a mixed-effects model was not necessary because the number of repeated driver ID values was not found to be statistically significant when included as a random effect and, with only two sites, Road ID was better suited to be included as a potential independent variable. With additional data, a linear mixed-effects model or a generalized linear model would be more appropriate because it can account for random effects due to the repeated samples from different roadway sections and drivers. The dependent variable in the glance location model was the proportion of time drivers spent looking at the forward roadway within the influence area of the DMS. This model included an intercept and binary variables for whether the driver changed lanes and whether the sign was on or off. The fit of the model was assessed using a likelihood ratio test to determine the final model. For the other distraction surrogates (i.e., average speed, standard deviation of speed, and standard deviation of lateral position), the final models were chosen using forward selection based on the AIC, and the models were assessed using residual plots and generalized residual plots. None of the best fit models included sign status (on/off) as a significant variable. (However, a lesser fit model of the standard deviation of speed did find sign status to be a significant variable.) The best fit models included congestion- and movement-related variables. As suggested earlier, further analysis may need to be conducted only for through movement vehicles in free flow conditions due to the strong correlation between the surrogate dependent variables and congestion and movement.

109 9.5.4 Discussion The objective of this Safety Framework was to assess the efficacy of using SHPR2 NDS data to evaluate distraction due to an overhead DMS. Although the sample size used for this analysis was small, the results indicate that drivers do glance differently and tend to vary their speed more when a DMS is active compared to when it is inactive. The evaluation also indicated that locations with a DMS can be obtained from several sources, thus allowing a variety of overhead DMS applications to be identified, including the display of travel times and messages about congestion or crashes ahead or the presentation of campaigns such as “Message Monday,” where an overhead DMS is used as a mechanism to alert the public about the importance of traffic safety through catchy phrases and slogans. A variety of DMS placements can also be determined (e.g., overhead, side mounted, over a single lane). As a result, further analysis of distraction related to DMS is feasible. This Safety Framework can also be used to explore the relationship between distraction and other roadway features using SHRP2 NDS data. A comprehensive summary of the advantages and disadvantages of using NDS data to assess distraction due to infrastructure elements is provided in Section 8.5.4. However, two additional recommendations resulted from the development of this Safety Framework. First, in urban areas, driving behavior can be impacted significantly by surrounding vehicles. As a result, the impact of distraction is most likely to be detected under free flow rather than congested conditions. Second, urban areas are characterized by higher traffic volumes and a large number of infrastructure elements, including traffic control, outdoor advertising, and urban landscaping. As a result, the impact of infrastructure on distraction may be best evaluated when particular features can be isolated from other elements that could also prove distracting.