Analysis of Naturalistic Driving Study Data: Roadway Departures on Rural Two-Lane Curves (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

44 C h a p t e r 8 This chapter discusses the third research question: What is the relationship between driver distraction; other driver, roadway, and environmental characteristics; and roadway departure risk? Research Question 3 investigates how driver behaviors in conjunction with roadway and environmental factors affect the likelihood of a roadway departure on rural two-lane curves. The team had examined a number of different models— including generalized linear models, Bayesian models, and regression tree analysis—during a similar study. That study used NDS data to evaluate roadway departures as part of SHRP 2 Project S01E, Evaluation of Data Needs, Crash Sur- rogates, and Analysis Methods to Address Lane Departure Research Questions Using Naturalistic Driving Study Data (Hallmark et al. 2011). Logistic regression was determined to be the best model. Logistic regression models the probability (odds) of a given type of roadway departure based on driver, roadway, and environmental characteristics. Odds ratios are the probability that an event happens in relation to the probability that it does not happen. Logistic regression evaluates the association between a binary response and explanatory variables. The natural logarithm of the odds is related to explanatory variables using a linear model. The value of using logistic regression is that the model output (odds ratio) can be easily understood by transportation agencies. As a result, multivariate logistic regression was used to model the probability (odds) of a roadway departure based on driver, roadway, and environ- mental characteristics. Data are aggregated to the event level for this analysis. The objective of Research Question 3 was to identify which roadway, environmental, and driver factors are related to road- way departure risk. Models were developed that assessed the probability of a right-side encroachment, probability of a left- side encroachment, probability that the driver will enter the curve 8 km/h (5 mph or more) over the advisory speed if pres- ent or posted speed limit if not present, and probability that the driver will enter the curve 16 km/h (10 mph or more) over the advisory speed if present or posted speed limit if not present. Data Sampling and Modeling approach for research Question 3 Data at the event level were used for this analysis. Data were aggregated by epoch: each epoch is one row of data and rep- resents one driver trip through a single curve. The amount of time (decoseconds of data) a driver was traversing the curve was used to normalize epochs with different durations. One row of data (one observation) included information for a section 200 m upstream of the curve and within the curve. Two hundred meters was identified as the curve area of influence in Research Question 1. A total of 583 observations were included in the analysis. The sample included 57 right-side roadway departures and 40 left-side roadway departures. The advisory speed when present or the posted speed limit when the advisory speed was not present was exceeded by more than 8 km/h (5 mph) for 245 observations and was exceeded by 16 km/h (10 mph) for 123 observations. Data were aggregated to the event level for this analysis. As mentioned above, the team had previously examined a number of different models—including generalized linear models, Bayesian models, and regression tree analysis—during its work on a similar study (Hallmark et al. 2011). The team considered the objective of Research Question 3 and the type of data being modeled. Additionally, model output was con- sidered because the output of some models is more easily understood than that of others. For instance, logistic regression provides the probability or odds of a certain event happening, which can easily be understood by the stakeholders who are expected to use the results of this research, including state and local transportation agencies. Given these considerations, logistic regression was determined to be the best model. Analysis for Research Question 3

45 Variables Used for research Question 3 Table 8.1 shows the reduced kinematic variables used in this analysis. The two response variables included in the analyses are (1) the probability of a right-side or left-side roadway departure and (2) the probability that the driver entered the curve at 8 km/h or 16 km/h (5 mph or 10 mph) over the curve advisory speed when present or the posted speed limit when not present. Driver, roadway, and environmental factors were included in the analysis as independent variables. Static roadway characteristics were extracted as described in Chapter 4. Envi- ronmental characteristics (e.g., night, raining) were consid- ered to be consistent across the event and were also reduced as described in Chapter 4. Some driver characteristics were also static (e.g., age, gender). While a driver’s age changed over the study period, the age of the driver at the time the study commenced was generally used. Kinematic driver characteristics (e.g., glance location and distraction) were reduced and reported at the same resolu- tion as the time series data from the DAS (10 Hz). Kinematic driver and vehicle characteristics were summarized for 200 m upstream of the point of curvature and through the curve. Table 8.1. Description of Reduced Kinematic Variables Variable Measure UpOncoming Fraction of time oncoming vehicles are present for 200 m upstream of the curve (e.g., number of 0.1-s intervals) Oncoming Fraction of time oncoming vehicles are present within the curve UpFollow Fraction of time oncoming subject vehicles following lead vehicle 200 m upstream of the curve Follow Fraction of time oncoming subject vehicles following lead vehicle Up_Spd Speed averaged over 200 m upstream of the curve (m/s) Up_Std Standard deviation of speed for 200 m upstream of the curve Speed Speed averaged over curve (m/s) Std Standard deviation of speed within the curve Ent_Spd Speed at which vehicle entered curve (m/s) UP_FR Fraction of time driver glance location is forward roadway for 200 m upstream of the curve FR Fraction of time driver glance location is forward roadway over the curve UP_SA Fraction of time driver glance location is to roadway-related tasks for 200 m upstream of the curve SA Fraction of time driver glance location is to roadway-related tasks over the curve UP_NR Fraction of time driver glance location is to non-roadway-related tasks for 200 m upstream of the curve NR Fraction of time driver glance location is to non-roadway-related tasks over the curve UP_PASS Fraction of time driver glance location is away from roadway on passenger-related tasks for 200 m upstream of the curve PASS Fraction of time driver glance location is away from roadway on passenger-related tasks over the curve UP_InVeh Fraction of time driver glance location is away from roadway on in-vehicle-control-related tasks for 200 m upstream of the curve InVeh Fraction of time driver glance location is away from roadway on in-vehicle-control-related tasks over the curve UP_CELL Fraction of time driver glance location is away from roadway on cell phone–related tasks for 200 m upstream of the curve CELL Fraction of time driver glance location is away from roadway on cell phone–related tasks over the curve UP_PerHY Fraction of time driver glance location is away from roadway on personal hygiene–related tasks for 200 m upstream of the curve PerHY Fraction of time driver glance location is away from roadway on personal hygiene–related tasks over the curve UP_EAT Fraction of time driver glance location is away from roadway on eating/drinking-related tasks for 200 m upstream of the curve EAT Fraction of time driver glance location is away from roadway on eating/drinking-related tasks over the curve Up_AVG_SR Average length of glance away from forward roadway to roadway-related locations (seconds) for 200 m upstream of curve AVG_SR Average length of glance away from forward roadway to roadway-related locations (seconds) through curve Up_AVG_NR Average length of glance away from forward roadway (seconds) to non-roadway-related locations for 200 m upstream of curve AVG_NR Average length of glance away from forward roadway (seconds) to non-roadway-related locations through curve

46 These characteristics were reduced to the event level (see Table 8.1). For instance, vehicle speed for each 0.1-s interval over the curve was averaged. Random effect variables were included to account for multiple samples for the same driver or same curve. Separate models were developed for right- and left-side roadway depar- tures because they are likely to be affected by different factors and are two distinct events. Description of analytical approach and results for research Question 3 Right-Side Encroachment Logistic regression was used to model the odds of a right-side encroachment (0 for no right-side encroachment, 1 for right- side encroachment) for each event indexed by i as a random variable, γi, which follows a Bernoulli distribution with prob- ability of departure, pi. ∼ ( )γ Bernoulli pi i For the logistic regression, the log odds of a right-side encroachment were modeled as follows: log 1 0, 0, 0 1 1 2 2 3 3 4 4 5 5 2 2 p p x x x x x Normal Normal i i i i i d i c ∼ ∼ ( ) ( ) −   = β + β + β + β + β + β + α + γ α σ γ σ where x1 = proportion of time the driver is glancing at the forward roadway 200 m upstream of the curve; x2 = indicator for curve direction (0 = outside; 1 = inside); x3 = radius of the curve (m); x4 = indicator for presence of a guardrail (0 = not present; 1 = present); x5 = indicator for presence of a curve warning sign, either a W1-6 sign or curve advisory sign (0 = not present; 1 = present); ai = random effect for subject; and γi = random effect for curve. This model was chosen as the best fit model for the data by model selection using PROC LOGISTIC in SAS and by com- paring the AIC/BIC values to various other models that were examined. All models were fit using the glmer() command in the lme4 package in R. The fitted model parameters, p-values, and 90% Wald confidence intervals are shown in Table 8.2. Wald intervals were calculated as follows: pˆ .95z si iβ ± where bˆi = the estimate of the parameter given above; z.95 = the 95th percentile of the standard normal distribution; and si = the standard error of the estimate (not given). The odds ratios for this model, which are equivalent to exp(bi) for i = 1, . . . , 6, are given in Table 8.3. For the dummy variables, the odds of a right-side encroachment change by a factor of exp(bi) when the object (traveling on the inside of the curve, etc.) is present relative to when it is not present. For the numeric variables, the odds of a right-side encroachment change by a factor of exp(bi) when the covariate xi increases by 1 unit. Table 8.3 shows the 90% Wald intervals for these estimates, which are calculated by exponentiating the bound- aries of the confidence intervals in Table 8.2. Because Up_FR is a proportion, a more appropriate estimate of the odds ratio might correspond to the change in odds of lane departure when forward glance time increases by 0.10, or Table 8.2. Parameter Estimates for Right-Side Encroachments Parameter Estimate p-value 5% 95% b0 -1.4039 0.1998 -3.2049 0.3972 b1 -1.9968 0.0388 -3.5866 -0.407 b2 1.9116 2.57E-06 1.2194 2.6039 b3 -4.00E-04 0.0399 -8.00E-04 -1.00E-04 b4 -1.121 0.1072 -2.2656 0.0237 b5 0.2337 0.6144 -0.5293 0.9967 s2d 1.1754 NA NA NA s2c 0.4243 NA NA NA Table 8.3. Confidence Intervals for Right-Side Encroachments Variable Odds Ratio Estimate 5% 95% Up_FR 0.1358 0.0277 0.6656 Direction (inside versus outside) 6.7642 3.385 13.5167 Radius 0.9996 0.9992 0.9999 Guardrail (present versus not present) 0.326 0.1038 1.024 Curve warning (present versus not present) 1.2632 0.589 2.7093

47 10% of total time upstream of the curve. This value is given by the following: exp 0.10 0.81901( )β = Thus, the odds of lane departure increase by a factor of 1/0.8190 = 1.221 when the proportion of forward glance time upstream of the curve decreases by 0.10. The results also indicate that a right-side lane departure is 6.8 times more likely on the inside of a curve compared with the outside of the curve. Lane departures are slightly more likely (1.3 times) for curves with any type of curve advisory sign (including W1-6). It is unlikely that the presence of the warning sign leads to increased probability of a right-side encroachment. Rather, it is likely that advisory signs are more likely to be present on curves of a cer- tain type (i.e., those with sight distance issues, sharper curves), and encroachments are also more likely for those road types. Additionally, the results suggest that the simple presence of curve warning signs does not mitigate roadway departures. A statistically significant but small correlation exists between radius of curve and probability of a right-side encroachment. Drivers were 0.33 times less likely to have a right-side encroachment on roadways where a guardrail is present. A guardrail is used to decrease the severity of a crash when a vehicle leaves the roadway. It is not a countermeasure to prevent roadway departures. The presence of a guardrail may suggest to the driver that roadway conditions are less safe, resulting in better driver attention. Additionally, few delineation countermeasures (e.g., chevrons) were present in the curves included in the analysis. As a result, a guardrail may provide some delineation of the curve, which provides feedback to the driver about the sharpness of the curve. Left-Side Encroachment The log odds of left-side encroachment were modeled as follows: log 1 0, 0, 0 1 1 2 2 3 3 2 2 p p x x x Normal Normal i i i i i d i c ∼ ∼ ( ) ( ) −   = β + β + β + β + α + γ α σ γ σ where x1 = dummy variable for driver gender (0 = female; 1 = male); x2 = dummy variable for the direction of the curve (0 = outside; 1 = inside); x3 = radius of the curve; ai = random effect for drivers; and γi = random effect for curve. Parameter estimates, p-values, and 90% Wald confidence intervals are shown in Table 8.4. Odds ratios and 90% Wald confidence intervals are shown in Table 8.5. As noted, males are more than four times more likely to have a left-side encroachment, and drivers traveling on the inside of the curve are 0.1 times less likely to have a left- side encroachment than drivers traveling on the outside of the curve. The impact of radius was statistically significant but minor. Left-side encroachments are likely to be drivers who “cut the curve.” Although driver intent is difficult to determine, in several cases the driving manner as evidenced in the forward videos suggested that the driver was intentionally crossing the centerline. Probability of Exceeding Posted or Advisory Speed by 8 km/h (5 mph) The log odds of a vehicle traveling 8 km/h (5 mph) or more over the posted or advisory speed limit was also modeled using logistic regression as follows: log 1 0, 0, 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 2 2 p p x x x x x x x x Normal Normal i i i i i d i c ∼ ∼ ( ) ( ) −   = β + β + β + β + β + β + β + β + β + α + γ α σ γ σ Table 8.4. Parameter Estimates for Left-Side Encroachments Parameter Estimate p-value 5% 95% b0 -3.2435 2.00E-04 -4.6796 -1.8074 b1 1.4888 0.0753 0.112 2.8656 b2 -2.2653 7.00E-04 -3.3661 -1.1645 b3 -7.00E-04 0.06 -0.0012 -1.00E-04 s2d 1.879 NA NA NA s2c 2.383 NA NA NA Table 8.5. Confidence Intervals for Left-Side Encroachments Variable Odds Ratio Estimate 5% 95% Gender (male versus female) 4.4318 1.1186 17.5587 Direction (inside versus outside) 0.1038 0.0345 0.3121 Radius 0.9993 0.9988 0.9999

48 where x1 = driver age (years); x2 = fraction of time following another vehicle 200 m upstream of the curve; x3 = average speed 200 m upstream of the curve (m/s); x4 = indicator for roadway markings (0 = visible markings; 1 = obscured markings/not present); x5 = indicator for visibility (0 = clear; 1 = any reduced visibility); x6 = the radius of the curve (m); x7 = indicator for presence of a paved shoulder (0 = not present; 1 = present); x8 = indicator for raised pavement markings (0 = not present; 1 = present); ai = random effect for drivers; and γi = random effect for curve. Again, the categorical variables can be thought of as having more than one coefficient, with the interpretations of estimates and odds ratios the same as in the previous section. Parameter estimates, p-values, and 90% Wald confidence intervals are shown in Table 8.6. Odds ratios and 90% Wald confidence intervals are shown in Table 8.7. Results show that drivers who are following other vehicles or driving under reduced visibility conditions are less likely to enter the curve at 8 km/h (5 mph) or more over the posted or advisory speed. Drivers traveling at higher speeds upstream are much more likely to enter the curve at 5 mph over the speed limit, as expected. Additionally, when pavement markings are obscured or not present, drivers are significantly more likely to enter the curve more than 8.0 (5 mph) over the posted/advisory speed. Lane line markings may provide curve delineation, which aids drivers in gauging the sharpness of the curve so that they are better able to select curve entry speeds. The odds ratio estimates for the paved shoulder and raised paved markings indicator variables are extremely small, though they are still significantly different from zero in the model. This suggests that drivers are less likely to speed when paved shoulders or raised pavement markings are present. However, these variables are best used within the whole model, instead of being considered separately, as their estimates are almost nonsensical. However, their extreme values do indicate that their presence significantly decreases the odds of speeding more than 5 mph over the posted/advisory speed. Probability of Exceeding Posted or Advisory Speed by 16 km/h (10 mph) The log odds of a driver exceeding the posted or advisory speed by 16 km/h (10 mph) or more were modeled using the following equation: log 1 0, 0, 0 1 1 2 2 3 3 4 4 5 5 2 2 p p x x x x x Normal Normal i i i i i d i c ∼ ∼ ( ) ( ) −   = β + β + β + β + β + β + α + γ α σ γ σ where x1 = average speed 200 m upstream of the curve (m/s); x2 = average amount of time driver glance is away from the road and engaged in driving tasks; x3 = radius of the curve (m); x4 = indicator for paved shoulder (0 = unpaved; 1 = paved); x5 = indicator for raised pavement markings (0 = not present; 1 = present); ai = random effect for drivers; and γi = random effect for curve. Table 8.6. Parameter Estimates for 5 mph Over the Speed Limit Parameter Estimate 5% 95% p-value b0 -26.0071 -34.617 -17.3973 6.75E-07 b1 -0.0683 -0.1219 -0.0146 0.0365 b2 -2.0832 -3.2748 -0.8916 0.004 b3 1.8159 1.4302 2.2016 9.64E-15 b4 7.2735 3.8652 10.6818 4.00E-04 b5 -4.94 -7.4111 -2.4688 0.001 b6 -0.0015 -0.0022 -8.00E-04 3.00E-04 b7 -10.3751 -15.0862 -5.664 3.00E-04 b8 -7.3919 -10.3264 -4.4574 3.42E-05 s2d 41.215 NA NA NA s2c 4.831 NA NA NA Table 8.7. Confidence Intervals for 8 km/h (5 mph) Over the Speed Limit Parameter Estimate 5% 95% Age 0.934 0.8852 0.9855 UpFollow 0.1245 0.0378 0.41 UpSpd 6.1465 4.1795 9.0393 Markings 1441.5916 47.7128 43556.112 Visibility 0.0072 6.00E-04 0.0847 Radius 0.9985 0.9978 0.9992 PvdShd 3.11998E-05 2.80638E-07 0.0035 RPM 6.00E-04 3.27568E-05 0.0116

49 Parameter estimates, p-values, and 90% Wald confidence intervals are provided in Table 8.8. The odds ratios and 90% Wald confidence intervals are shown in Table 8.9. As noted in Table 8.9, the average speed upstream signifi- cantly increases the likelihood that a driver will also exceed the curve advisory/posted speed limit by 16 km/h (10 mph) or more. The length of the glance away from the forward roadway to roadway-related tasks (e.g., glances at rearview mirror) decreases the likelihood that drivers will exceed the posted/ advisory speed. This indicates that drivers may tend to slow down when they engage in longer glances away from the for- ward roadway. The probability of exceeding the posted/advisory speed is correlated to radius. As radius increases, drivers are less likely to exceed the posted/advisory speed. Curves with smaller radii are more likely to have an advisory speed, and presumably drivers are more likely to exceed lower speed limits. The odds ratio estimates for the paved shoulder and raised paved markings indicator variables are extremely small, though they are still significantly different from zero in the model. This suggests that drivers are less likely to speed when paved shoulders or raised pavement markings are present. However, these variables are best used within the whole model, instead of being considered separately, as their estimates are almost nonsensical. However, their extreme values do indicate that their presence significantly decreases the odds of speeding over 16 km/h (10 mph). Summary of Crash/ Near-Crash events Only one crash and three near crashes were present in the data set. Therefore, a model could not be developed for this research. All three near crashes appeared to be near rear-end collisions, with the roadway departure caused by the driver swerving to avoid the potential rear-end crash. A review of the data for the crash indicated that speeding was likely a major factor because the driver was 15 km/h (9 mph) over the posted speed limit of 72 km/h (45 mph) and 55 km/h (34 mph) over the curve advisory speed of 32 km/h (20 mph). The driver was not engaged in any distracting tasks and only glanced away from the forward view once (a glance to the steer- ing wheel, which lasted 0.6 s). The crash occurred between midnight and 3:00 a.m. There was some evidence of drowsi- ness because the driver was resting his head on his right arm or hand. The radius of curve was 50 m, and no shoulders, chevrons, or other countermeasures were present. Summary and Discussion The objective of Research Question 3 was to assess the relation- ship between driver, roadway, and environmental factors and risk of a roadway departure. The crash surrogates used for this research question were probability of a right-side encroach- ment, probability of a left-side encroachment, probability that the driver exceeded the posted or advisory speed by 5 mph or more, and probability that the driver exceeded the posted or advisory speed by 10 mph or more. Logistic regression was used to model observations at the event level. Key Findings Four different models were developed. The model for right-side encroachments indicated that the probability of a right-side encroachment increases as drivers spend less time glancing at the forward roadway. The results also indicate that a right-side lane departure is 6.8 times more likely on the inside of a curve compared with the outside of the curve. Lane departures are slightly more likely (1.3 times) for curves with any type of curve advisory sign (including W1-6). A statistically significant but small correlation exists between radius of curve and prob- ability of a right-side encroachment. Drivers were 0.33 times Table 8.8. Parameter Estimates for 16 km/h (10 mph) Over the Speed Limit Parameter Estimate p-value 5% 95% b0 -57.7719 6.77E-07 -76.9007 -38.6432 b1 3.3555 1.70E-10 2.4912 4.2198 b2 -1.8275 0.0485 -3.3513 -0.3038 b3 -0.0028 0.0429 -0.005 -5.00E-04 b4 -24.6413 0.0048 -39.002 -10.2807 b5 -16.3668 0.0259 -28.4486 -4.285 s2d 339.603 NA NA NA s2c 2.904 NA NA NA Table 8.9. Confidence Intervals for 16 km/h (10 mph) Over the Speed Limit Variable Odds Ratio Estimate 5% 95% UpSpd 12.0763 12.0763 68.0172 Avg_SA 0.035 0.035 0.738 Radius 0.995 0.995 0.9995 PvdShd 1.15252E-17 1.15252E-17 3.42885E-05 RPM 4.41499E-13 4.41499E-13 0.0138

50 less likely to have a right-side encroachment on roadways with a guardrail. The model for left-side encroachments indicated that males are more than four times more likely to have a left-side lane departure, and drivers traveling on the inside of the curve are 0.1 times less likely to have a left-side encroachment than drivers traveling on the outside of the curve. The impact of radius was statistically significant but minor. The probability that a driver will be 8 km/h (5 mph) or more over the posted/advisory speed is higher for younger drivers, when drivers have a higher average speed upstream, and when edge line markings are obscured or not present. The amount of time a driver spends following another vehicle, presence of lower visibility conditions, and presence of paved shoulders and RPMs decreases the probability that a driver will enter the curve 8 km/h (5 mph) or more over the posted/advisory speed. The probability that a driver will be 16 km/h (10 mph) or more over the posted/advisory speed is higher when drivers have a higher average speed upstream. The probability is lower when the average glance at roadway-related tasks is longer and when paved shoulders and RPMs are present. Implications for Countermeasures The presence of warning signs increased the likelihood of a right-side encroachment. It is unlikely that the presence of a warning sign itself increases the probability. Rather, it is likely that advisory signs are more likely to be present on curves of a certain type (i.e., those with sight distance issues, sharper curves), and encroachments are also more likely for those road types. Additionally, the results suggest that the simple presence of curve warning signs does not mitigate roadway departures. The presence of a guardrail decreased the probability of a right-side encroachment. The purpose of a guardrail is to mitigate the consequences of a driver leaving the roadway rather than to keep the driver from leaving the roadway. Consequently, a guardrail in and of itself does not mitigate roadway departures. The presence of a guardrail may suggest to the driver that roadway conditions are less safe, resulting in better driver attention. Additionally, few delineation counter- measures (e.g., chevrons) were present in the curves included in the analysis. As a result, a guardrail may provide some delin- eation of the curve, which provides feedback to the driver about the sharpness of the curve. The probability that a driver would exceed the posted/ advisory speed by 5 mph or more was higher for curves with obscured/missing edge lines. Presence of RPMs decreased the probability of exceeding the posted/advisory speed by 8 km/h and 16 km/h (5 mph and 10 mph). Taken together, these results indicate that better curve delineation may allow drivers to better gauge upcoming changes in roadway geometry, resulting in better speed selection and decreased risk of a roadway depar- ture, and it may help decrease speed. Delineation counter- measures include chevrons, the addition of reflective panels to existing chevron posts, reflective barrier delineation, RPMs, post-mounted delineators, edge lines, and wider edge lines. The speed models suggest that driver age and upstream speed have a significant impact on drivers’ speed within a curve. As a result, speed management countermeasures that affect tangent speed will also decrease curve speeds. The results also indicate that speed management is appropriate to get drivers’ attention before entering a curve. Countermeasures specifically targeted to reduce speed on curves include dynamic speed feedback signs, on-pavement curve warning signs, and flashing beacons. Limitations The most significant limitations are sample size and represen- tation of different curve and driver characteristics. More than 700 potential curves were initially identified. This represented a wide range of roadway characteristics and countermeasures. However, some countermeasures, such chevrons and rumble strips, were not widely available in the study areas, and some countermeasures, such as post-mounted delineators, were not available at all. Additionally, only one-third of the full NDS data set was available for query at the time the data request was made, and data were only found for 148 curves, which reduced the number of roadway characteristics that could be included. A total of 583 observations were included in the analysis. However, only 57 right-side roadway departures and 40 left-side roadway departures were present. Another limitation is that crashes/near crashes were not available, so the relationship between encroachments or speed and roadway departure crash risk could not be established. Additionally, although a model was developed for left-side encroachments, this model is likely to include drivers who cut the curve. Although it is difficult to determine driver intent, in several cases the driving manner as evidenced in the forward videos suggested that the driver was intentionally crossing the centerline.