Traffic Control Devices and Measures for Deterring Wrong-Way Movements (2018)

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.

5 This chapter provides information regarding the method- ologies used to develop a divided highway multistate crash dataset. This dataset was used to determine the characteris- tics of wrong-way crashes and wrong-way drivers. The data- set was also used to analyze the effects of geometric features and traffic control devices upon wrong-way crashes on high- speed rural and urban divided highways. Divided Highway Multistate Crash Dataset In order to assess the characteristics of wrong-way crashes on divided highways and determine the effects of multiple design features and traffic control devices upon wrong-way crashes on divided highways, the research team compiled a multistate crash dataset. A recent analysis of wrong-way fatal crashes in the United States (Baratian-Ghorghi et al. 2014) found that the states with the highest number of wrong-way fatal crashes and fatalities were Texas (14 percent), California (10 percent), and Florida (8 percent) (based on the Fatality Analysis Reporting System data from 2004 to 2011). These three states also provided geographic representation from the western, central, and eastern United States. The following sections detail the data obtained from each state and the methodology used to identify the wrong-way crashes, wrong-way crash corridors, and control corridors. Preliminary descriptive statistics revealed that wrong-way crashes rarely occur on low-speed, rural divided highways (1 percent). In addition, researchers conducted a prelimi- nary statistical analysis of the severity of the wrong-way crashes in Texas and Florida by location (urban and rural) and speed limit. In Texas, researchers found that the prob- ability of a severe outcome (i.e., KAB crash) increased with increasing speed limit. In addition, the probability of a severe outcome at lower speed limits (less than 50 mph) was less than 25 percent. Although speed limit was not found to be a statistically significant variable in the Florida model selection process, the trend was comparable to that found for Texas. Based on these findings and the project panel’s input, the research team limited the divided highway crash analysis to high-speed (≥ 50 mph) roadways. Wrong-Way Crash Corridors Texas The research team obtained crash data for 2012 to 2014 from CRIS. Only crashes that occurred on Texas DOT road- ways and resulted in injury, death, or at least $1000 in damage were included in the dataset. Researchers used the contribut- ing factor variable to identify wrong-way-related crashes. CRIS lists up to three contributing factors and up to two possible contributing factors for each crash (up to five overall). A con- tributing factor code of 69 represents “wrong side—approach or intersection,” 70 represents “wrong side—not passing,” and 71 represents “wrong-way—one-way road.” Researchers included all three codes in order to capture all possible wrong- way crashes. Initially, the research team identified 2751 wrong- way crashes. Researchers then used the speed limit, road type, and func- tional system variables to reduce the dataset. The speed limit variable was used to retain wrong-way crashes that occurred on high-speed roadways. The road type variable was used to remove two-lane, two-way roadways and undivided roadways. While the functional system code included categories for urban and rural interstates, as well as other urban freeways, it did not include a category for divided highways. Further- more, a cursory review of the wrong-way crashes coded as occurring on an interstate or other urban freeway revealed that the functional system for some of these crashes was miscoded. Therefore, the research team decided to plot all the crashes using the geolocation information (i.e., latitude and longitude) and manually determine the type of roadway for each crash. Using the verified location information, the C H A P T E R 2 Divided Highway Crash Analysis

6 research team found that 255 wrong-way crashes occurred on high-speed divided highways. Next, researchers reviewed the crash diagrams and asso- ciated narratives for the wrong-way driving crashes that occurred on high-speed divided highways. Based on this review, 54 crashes (20 percent) were removed because they were determined not to be wrong-way crashes of interest. Many of these crashes were the result of a right-way driver losing control of his or her vehicle and ending up on the wrong side of the road. Researchers expected this might occur because Texas uses multiple codes to flag wrong-way driving as a contributing factor. All of these codes were ini- tially included in the dataset to ensure that no crashes of interest were excluded. The crash diagrams and associated narratives only men- tioned specific entry points for 32 of the 201 (16 percent) remaining wrong-way crashes. Using these limited data, the research team determined that the average wrong-way entry occurred within 777 ft (0.15 mi) of the crash with a standard deviation of 972 ft. Given these characteristics, it can be easily shown that most of the wrong-way entries were within 2682 ft (0.51 mi) of the crash and 41 percent of these crashes originated at T-intersections where a crossover to the oppo- site travel direction was provided. Based on these findings, the research team decided to create study corridors 1 mi in length (0.5 mi in each direction from the crash). The research team developed a process by which staff identified the study corri- dors surrounding the wrong-way crashes in Google Earth. The most likely wrong-way entry points were determined using crash data elements (e.g., direction of travel, side of road), crash diagrams (only available for Texas data), and engineer- ing judgment. The corridor length was extended, as needed, to include the most likely wrong-way entry point or to accom- modate multiple crashes. Researchers then used Google Earth to obtain various characteristics of the most likely wrong-way entry point (e.g., various cross-section measurements, median type, inter section type, interior and exterior traffic control devices present, lighting presence, and speed limit). A review of the traffic control devices present at each probable wrong-way entry point revealed that 18 locations were signalized intersec- tions. The research team removed these sites from the dataset since the focus of this research was wrong-way movements at unsignalized inter sections. Overall, the final Texas dataset con- tained 183 wrong-way crashes on high-speed divided highways (117 rural and 66 urban) and 168 study corridors (111 rural and 57 urban). Thirteen corridors contained multiple crashes. Florida The research team obtained the 2010 to 2013 crash point shape files from Florida DOT. These files contained crashes that occurred on the state highway system and involved: • Death, personal injury, or indication of pain of persons involved in the crash; • Leaving the scene involving damage to attended vehicles or property; • Driving while under the influence; and • A vehicle inoperable to a degree that required a wrecker or involved a commercial vehicle. The research team used the two-digit code that gives driver actions at the time of the crash to determine which crashes for each year were attributed to a wrong-way driver (i.e., 21— wrong side or wrong way). Additionally, the research team used the road category variable to determine the following for wrong-way crashes: • Type of roadway—freeway, divided roadway, or other; • Area type—urban, suburban, or rural; and • Other descriptors (e.g., interstate, ramp). Similar to the Texas dataset, the Florida dataset did not include a variable that specifically separated divided high- ways from other types of divided roadways. Therefore, the research team again used geolocation information (i.e., lati- tude and longitude) to plot each crash and visually confirm the type of roadway. Using a similar process, researchers identified the study corridors and most likely wrong-way entry points for the Florida wrong-way crashes. In addition, the signalized inter- section wrong-way entry points and the associated crashes were removed from the dataset. The final high-speed divided highway dataset for Florida included 160 wrong-way crashes (59 rural and 101 urban) and 142 study corridors (48 rural and 94 urban). Thirteen corridors contained multiple crashes. California The research team reduced and analyzed the California dataset (2008–2011) received from the Highway Safety Infor- mation System (HSIS). This dataset contains crashes investi- gated by the California Highway Patrol that result in injury, death, or at least $500 in damage. The research team used the two-digit code that gives the collision factor category of the crash to determine which crashes for each year were attrib- uted to a wrong-way driver (i.e., 05—wrong side of road). The research team then used the divided highway variable to determine whether the roadway was divided or not. Next, the research team used the roadway classification variable in the roadway file to determine the type of roadway (i.e., lim- ited access or divided non-freeway) and whether the road- way was urban or rural. The research team then removed all roadways with fewer than four lanes using the total number of lanes variable. Finally, the research team used the median

7 type variable to remove divided non-freeway facilities with two-way left-turn lanes or other miscellaneous median lanes (e.g., bus lanes). Similar to the Texas and Florida datasets, the variables in the California dataset did not separate divided high- ways from other types of divided roadways. However, the California dataset also did not include geolocation informa- tion (i.e., latitude and longitude). Using other variables and California roadway information, the research team success- fully identified the location of the California divided highway wrong-way crashes to within 1/10 of a mile and then plotted the crash locations using ArcGIS. Researchers then identified the study corridors and most likely wrong-way entry points for the California wrong-way crash data on high-speed divided highways. Wrong-way entry points at signalized intersections and their associated crashes were removed from the dataset. Overall, the final California dataset contained 66 wrong- way crashes (38 rural and 28 urban) and 48 study corridors (27 rural and 21 urban) on high-speed divided highways. Fifteen corridors contained multiple crashes. Summary Table 1 provides a summary of the high-speed divided high- way wrong-way crash dataset. Overall, the research team veri- fied 409 wrong-way crashes on high-speed divided highways that were the result of wrong-way entries from unsignalized intersections. Fifty-two percent of these crashes occurred in rural areas, while the other 48 percent happened in urban areas. These crashes occurred in 358 corridors that were identified by researchers using previously described methods. Because multiple crashes occurred in some of the selected corridors, the number of crash corridors was less than the number of wrong-way crashes. Control Corridors In order to avoid biasing the results by just focusing on sites with a reported history of wrong-way crashes, the research team wanted to supplement the dataset with data from sites with no recorded wrong-way crashes. In order to achieve this goal, the research team developed a procedure to draw statistically representative site samples from the three states under study. The research team used the roadway inventory files from each of the three states under study as sampling frames in order to perform the probability sampling of control corri- dors. However, the research team needed to ensure that seg- ments with a wrong-way crash history were avoided and that no overlap of crash segments and control segments occurred. Therefore, the research team developed a 2-mi buffer around the segments with previously identified wrong-way crashes and removed those segments from the sampling frame. After excluding the segments in close proximity to each identified wrong-way crash, the research team applied filters to identify urban and rural sites with speed limits greater than or equal to 50 mph, and with at least four through lanes (two lanes per direction). Using the three reduced roadway inven- tories as sampling frames, the research team utilized an open source software environment and package (Fox and Weisberg 2011; R Development Core Team 2011) to draw a random subset of roadway sections intended to identify candidate control corridors. In the cases of Texas and Florida, the filters to identify high- speed multilane highways could be applied with relative ease because the road inventories for these two states had all the required variables. However, for California, this task was especially challenging because the roadway inventory was obtained from the HSIS, and that database does not offer any geodesic information. Initially, the research team attempted to use an unfiltered sample of California seg- ments to identify control sections. However, this sample required a large amount of effort to identify segments with the required characteristics, yielding very few actual con- trol sites. The complication with lack of geodesic information in the California data had emerged earlier, when the research team determined the segments with wrong-way crashes. To over- come this issue, the research team estimated the approximate geolocation of the crashes by matching the linear referencing fields from the HSIS crash table to the corresponding linear referencing fields in a shape file obtained from Caltrans. This TexasData Florida California Overall Number (%) of Wrong-Way Crashes Urban Rural Total 66 (36%) 117 (64%) 183 (45%) 101 (63%) 59 (37%) 160 (39%) 28 (42%) 38 (58%) 66 (16%) 195 (48%) 214 (52%) 409 (100%) Number of Wrong-Way Corridors Urban Rural Total 57 111 168 94 48 142 21 27 48 172 186 358 Table 1. Overview of high-speed divided highway wrong-way crash dataset.

8 shape file consists of geolocation markers every 1/10 of a mile along California highways. Similar to the instance of defining segments with wrong-way crashes, the research team per- formed an analysis to match the HSIS road inventory table to the referenced shape file. Additional sample road segments were obtained from the matched databases and were subse- quently used to define California control segments. Table 2 contains a summary of the identified control cor- ridors for each state and overall. The more labor-intensive selection process for California resulted in fewer control sites compared to Texas and Florida. However, the number of con- trol sites for each state was similar to the number of crash corridors for each state. For each control corridor, researchers randomly chose an unsignalized intersection with a median opening and documented its characteristics (e.g., various cross-section measurements, median type, intersection type, interior and exterior traffic control devices present, lighting presence, and speed limit). Wrong-Way Crash Characteristics Before performing the statistical analysis, the research team conducted an exploratory analysis of the wrong-way crash data to examine trends. Table 3 shows the percentage of wrong- way crashes by crash severity. The Texas and Florida dataset were very similar, with most of the wrong-way crashes on high-speed divided highways resulting in a serious injury. In contrast, almost 68 percent of the wrong-way crashes on high-speed divided highways in California resulted in possible injuries or property damage only. Overall, 54 percent of the wrong-way crashes on high-speed divided highways resulted in serious injuries. Figure 1 shows the trends for wrong-way crashes on high- speed divided highways by day of the week for each state and overall. The Texas and Florida data exhibited similar charac- teristics by day of the week. For California, the research team identified a larger amount of wrong-way crashes on Tuesday and Friday and fewer wrong-way crashes on Saturday. Even so, overall, most of the wrong-way crashes on high-speed divided highways occurred on Friday, Saturday, and Sunday (53 percent). Table 4 contains the percentage of wrong-way crashes on high-speed divided highways by lighting condition. Again, the Texas and Florida datasets were very similar, with about two-thirds of the wrong-way crashes occurring at night. In contrast, more than half of the California wrong-way crashes occurred during the day. Even so, overall, most of the crashes on high-speed divided highways occurred at night. Based on a review of the Texas crash diagrams, crash descrip- tors (e.g., vehicle direction and side of the roadway), and engineering judgment, the research team identified the most likely wrong-way entry point in each crash corridor in each state. Table 5 shows the types of intersections from which drivers presumably entered divided highways and went the wrong direction by state and overall. In some locations, the divided highway would intersect with another major road- way and entrance/exit ramps were used to connect the two major roadways, leading to 8 percent of the wrong-way entries. In other situations, the driver crossed through the median (less than 1 percent) or median opening (7 percent) and began to drive in the wrong direction on the other side of the road. Entries from private driveways (e.g., individual homes/subdivisions and businesses), three-leg inter sections, and four-leg intersections accounted for 7 percent, 49 percent, and 29 percent of the wrong-way entries, respectively. More than two-thirds of the wrong-way entries from inter secting roadways occurred where median openings were present. These wrong-way entries at median openings resulted in State Number of Rural Number of Urban Total Texas 65 57 122 Florida 44 93 137 California 42 7 49 Total 151 157 308 Table 2. Summary of control corridors. State KAB CO Unknown Texas (n = 183) 60% 39% 1% Florida (n = 160) 57% 42% 1% California (n = 66) 32% 68% 0% Total (n = 409) 54% 45% 1% Note: KAB = Killed, incapacitating injury, and non-incapacitating injury; CO = Possible injury and not injured (property damage only). Table 3. Percentage of wrong-way crashes by crash severity.

9 Pe rc en ta ge o f W ro ng -W ay C ra sh es Figure 1. Percentage of wrong-way crashes by day of the week. State Daylight Dawn/Dusk Night Texas (n = 183) 32% 3% 65% Florida (n = 160) 30% 2% 68% California (n = 66) 54% 8% 38% Total (n = 409) 35% 3% 62% Table 4. Percentage of wrong-way crashes by lighting condition. Intersection Type Texas (n = 183) Florida (n = 160) California (n = 66) Overall (n = 409) Ramp 22 (12%) 3 (2%) 9 (13%) 34 (8%) Crossed median 2 (1%) 0 (0%) 0 (0%) 2 (<1%) Median opening only 14 (8%) 11 (7%) 3 (5%) 28 (7%) Private driveway With median opening Without median opening Total 10 (5%) 11 (6%) 21 (11%) 0 (0%) 1 (<1%) 1 (<1%) 2 (3%) 3 (5%) 5 (8%) 12 (3%) 15 (4%) 27 (7%) Three-leg intersection With median opening Without median opening Total 64 (35%) 9 (5%) 73 (40%) 66 (41%) 38 (24%) 104 (65%) 16 (24%) 6 (9%) 22 (33%) 146 (36%) 53 (13%) 199 (49%) Four-leg intersection with median opening 51 (28%) 41 (26%) 27 (41%) 119 (29%) Table 5. Intersection type at most likely wrong-way entry points by state and overall.

10 75 percent of the wrong-way crashes that occurred on high- speed divided highways with four lanes (two in each direc- tion) and a grass median. Table 6 shows the intersection type from which the wrong- way movement most likely occurred by area type (i.e., rural and urban). The most notable differences were found for ramps, three-legged intersections, and four-legged inter sections. In rural areas, slightly more wrong-way entries initiated at three- legged intersections with median openings and four-legged intersections. In urban areas, more wrong-way entries were attributed to ramps and three-legged inter sections without median openings. Based on a subset of the Texas crash data for which the crash diagrams and associated narratives mentioned specific entry points (n = 32), researchers found that 63 percent entered via at-grade intersections. Ninety percent of the wrong-way entries from at-grade intersections occurred when a driver turned left from an intersecting roadway into the near main lanes. The remaining 10 percent of the wrong-way maneuvers occurred when a driver crossed the near main lanes but then turned right into the far main lanes. In one of these occur- rences, the wrong-way driver was trying to enter a business driveway that was located to the right of the median opening. In the other case, the geometry of the intersection did not allow drivers on the minor approach to enter the near main lanes (i.e., near main lanes went over minor approach with no ramp). The wrong-way driver crossed under the near main lanes and then turned right into the far main lanes (at-grade intersection). These data are similar to those found in Iowa (Athey Creek Consultants 2016). Figure 2 shows the median width at the assumed wrong- way entry points at intersections with a median opening by state. The median width depicted was measured from the edge of the traveled way to the edge of the traveled way and included shoulders but not turn lanes (i.e., MUTCD defini- tion). This figure shows that most of the median widths were less than 100 ft. According to the MUTCD, for median widths less than 30 ft, no control is needed on the interior approaches in the median. For median widths 30 ft or wider, the interior approaches in the median should be controlled with a STOP or YIELD sign. For wrong-way entries at intersections with median openings, researchers found that 18 percent of the sites had median widths less than 30 ft and 82 percent of the sites had median widths 30 ft or wider. Figure 3 shows the type of control in the median open- ing (i.e., none, STOP sign, or YIELD sign) at the presumed wrong-way entry points by median width. No control was found at 110 (40 percent) of the wrong-way entry points with median widths that ranged from 3 to 241 ft. Seventy- two percent of the locations without control had median widths greater than 30 ft (denoted by line). All but three of these locations had median widths from 30 ft to 80 ft wide. The research team did verify with Google Earth that the larger median width values did not have any control in the median opening. STOP or YIELD signs were found in the median opening at 47 (17 percent) and 117 (43 percent) of the wrong-way entry points, respectively. Twelve percent of these locations had median widths less than 30 ft. Wrong-Way Driver Characteristics The research team also explored the characteristics of the wrong-way drivers involved in crashes on high-speed divided highways. This analysis only included data from Texas and Florida. There was no way to determine the at-fault driver from the California occupant data obtained from HSIS. Figure 4 shows the gender of the wrong-way drivers by state and overall. In all cases, females represented about one-third Intersection Type Rural (n = 213) Urban (n = 196) Overall (n = 409) Ramp 10 (5%) 24 (12%) 34 (8%) Crossed median 2 (1%) 0 (0%) 2 (<1%) Median opening only 17 (8%) 11 (6%) 28 (7%) Private driveway With median opening Without median opening Total 11 (5%) 7 (3%) 18 (8%) 1 (1%) 8 (4%) 9 (5%) 12 (3%) 15 (4%) 27 (7%) Three-leg intersection With median opening Without median opening Total 85 (40%) 7 (3%) 92 (43%) 61 (31%) 46 (23%) 107 (54%) 146 (36%) 53 (13%) 199 (49%) Four-leg intersection with median opening 74 (35%) 45 (23%) 119 (29%) Table 6. Intersection type at most likely wrong-way entry points by area type.

Figure 2. MUTCD median width at wrong-way entry points with median openings by state (n = 277). Figure 3. MUTCD median width by type of control in the median for wrong-way entry points with median openings (n = 277).

12 of the wrong-way drivers, while males represented two-thirds of the wrong-way drivers. Figure 5 shows the age of the wrong-way drivers on high- speed divided highways by state and overall. In Florida, older drivers (≥ 60) represented 32 percent of the wrong-way drivers, and those ages 21 to 29 accounted for 29 percent. In contrast, in Texas, older drivers accounted for only 16 percent of the wrong-way drivers. The largest portion of Texas wrong- way drivers were ages 21 to 29. The three middle-age catego- ries were relatively consistent between the two states and overall. Table 7 shows that most of the wrong-way drivers, independent of age, made wrong-way maneuvers on high- speed divided highways that led to crashes at night. Figure 6 shows the cumulative distribution of the blood alcohol concentration (BAC) levels for the 67 wrong-way drivers from both states that tested positive for alcohol. For the remaining 248 wrong-way drivers, either the driver tested negative for alcohol (n = 126, or 40 percent) or the entry was blank (n = 122, or 39 percent). Key findings from Figure 6 include: • More than 90 percent had a BAC level equal to or greater than the legal limit (0.08 g/dL). • Almost 70 percent had a BAC level equal to or greater than twice the legal limit (0.16 g/dL). • Almost 25 percent had a BAC level equal to or greater than three times the legal limit (0.24 g/dL). Statistical Modeling In addition to the exploratory analysis, the research team worked to develop models that could be used to explore the relationships with potential explanatory variables as changes in the odds of a wrong-way crash occurring. Overview of Statistical Analyses Initially, the research team proposed using the propor- tion of wrong-way crashes among all crashes in a study site as the response variable to analyze. This approach is appli- cable in situations where the dataset is collected either by using a random process to assure representativeness or by selecting a range or combination of independent variables under evaluation. However, in this effort, the dataset was compiled from two subsets: one of sites with a known his- tory of wrong-way crashes and one of random sites with Figure 4. Gender of wrong-way drivers on high-speed divided highways by state and overall.

13 Figure 5. Age of wrong-way drivers on high-speed divided highways by state and overall. Age Group Florida (n = 157) Texas (n = 158) Overall (n = 315) Day (n = 47) Night (n = 110) Day (n = 52) Night (n = 106) Day (n = 99) Night (n = 216) < 21 11% 89% 50% 50% 32% 68% 21–29 24% 76% 23% 77% 24% 76% 30–39 18% 82% 20% 80% 19% 81% 40–49 25% 75% 32% 68% 28% 72% 50–59 37% 63% 40% 60% 39% 61% 60 38% 62% 44% 56% 40% 60% Unknown 75% 25% 40% 60% 47% 53% Total 30% 70% 33% 67% 31% 69% Note: Shading indicates majority of the crashes. Table 7. Wrong-way driver age versus time of day.

14 no history of wrong-way crashes. Because of the different approach to developing the dataset, the method of analysis was modified slightly. Instead of using the proportion of wrong-way crashes as the response variable, the final method of analysis defined the response variable as the probability of wrong-way crashes. The data allowed estimating such prob- ability via the indicator variable “crash,” which equaled 1 if a data point represented an instance of wrong-way crash and equaled 0 if a data point represented a site with no recorded wrong-way crash (i.e., control corridors). Therefore, for a given set of conditions (e.g., traffic volume, median width), the average of the variable “crash” represented an estimate of the probability of a wrong-way crash. A logistic regression approach allowed researchers to model the relationships with potential explanatory variables as changes in the odds of a wrong-way crash occurring. The following relation defines the regression model: 0Logit WW( )π = β + aX i Where pWW = Proportion of wrong-way crashes, Logit(•) = Log-odds function, defined as Logit(v) = log(v/1 − v), X = Row vector of explanatory variables, a = Vector of coefficients to be estimated, and β0 = Intercept term (to be estimated). The impact of a change in one explanatory variable, X1, is interpreted as a factor applied to the odds of wrong-way crashes before the change in X1. This factor is known as the odds ratio and represents a multiplicative change in the odds of wrong-way crashes associated with a change in variable X1. An odds ratio of 1.0 indicates no change in the odds of wrong- way crashes, while values smaller than 1.0 indicate a reduction in the odds of wrong-way crashes (e.g., 0.3 times as large). Similarly, values larger than 1.0 indicate an increase in the odds of wrong-way crashes (e.g., 1.2 times as large). When discussing odds ratios, it is common to express them in terms of “x times as big” independent of whether the impact is a decrease or increase. While the odds ratios could be equiva- lently expressed as a present decrease or increase, percentages are typically reserved for proportions and probabilities. Modeling Process Early in the modeling process, an examination of the range of key variables suggested that the dataset be further reduced for statistical analysis. First, the research team acknowledged Figure 6. Cumulative distribution of BAC levels for wrong-way drivers involved in a crash on high-speed divided highways.

15 that crashes attributed to wrong-way maneuvers at ramps, median openings only (no intersecting roadway), and cross- ing the median were relatively small (8 percent, 7 percent, and less than 1 percent, respectively). The research team thought that these unique situations should not be evaluated in con- junction with the other types of wrong-way entries from intersecting roadways. Therefore, the research team removed these crashes from the dataset. The research team also recognized that intersections with and without median openings are signed differently and thus should be analyzed separately. However, the dataset for inter- secting roadways without a median opening included only 80 sites (11 percent), which did not provide an adequate sample size for further analysis. Thus, the focus of the statistical analysis was for sites with intersecting roadways with median openings. In addition, researchers included only sites with two through lanes in each direction (four through lanes total) because 90 percent of the sites had this main lane configuration. Of the remaining sites, only 17 had medians wider than 120 ft, and most of these sites had STOP signs in the median opening. In contrast, for median widths 120 ft or less, all three types of control in the median opening were represented (i.e., no control, STOP sign, and YIELD sign). Researchers reviewed the impact of removing median widths larger than 120 ft on the modeling process and found no undue influence in the model parameter estimates. Therefore, researchers restricted the range for statistical analysis to sites with median widths less than or equal to 120 ft. The modeling effort had various stages. Originally, a two- tier stepwise model selection scheme was implemented. First, for major design or operation variables (e.g., annual aver- age daily traffic [AADT], median width, presence of turning lanes) and then at convergence, a new stepwise model selec- tion procedure was started for traffic control device-type variables (i.e., signing and pavement markings). Given the small sample size, it was expected that the models would have limited statistical power to pick up variables with relatively small effects, a feature that would limit potential findings about variables in the second stage of model selec- tion. For this reason, the research team decided to develop linear combinations of signage variables through principal component analysis (PCA). As mentioned earlier, an addi- tional round of model selection was performed using these combined variables where improved fits were obtained. The final stage of modeling consisted of testing all models for some variable interactions and a careful examination and potential substitution of some variables for their collinear competitors that may have a more straightforward interpre- tation. Three criteria were computed to assess each model fit: cumulative residual plots, Hosmer-Lemeshow goodness-of-fit test, and receiver operating characteristic curves. All the final models scored acceptable to exceptional in all these metrics. Traffic Control Device Variables Considered in the Analysis A significant challenge associated with this analysis was the number of traffic control device variables collected for each site. Table 8 contains a list of the 29 traffic control device variables documented for each site. Figure 7 explains the sign location nomenclature used in Table 8. Ideally, when the associations between these variables and odds of wrong-way crashes are of interest, a large enough sample size is required. However, the sample at hand was of a size that would only provide statistical evidence to detect associations between a reduced number of variables and the odds of wrong-way crashes. In addition, many of the traffic control device variables were strongly correlated. This was expected since many traffic control devices are installed and function as a system, not individual devices. This strong cor- relation tends to inflate the standard error of regression esti- mates; thus, increasing the sample size is necessary. Given these characteristics, the research team anticipated that the modeling process would be able to identify a rela- tively small number of signage variables as key terms in the regression models. An exploration of the traffic control vari- ables revealed that there was only one KEEP RIGHT sign in the dataset; thus, KEEP RIGHT signs were not included in the modeling process. In addition, researchers chose to focus on the signs on the near side of the median opening. Any signifi- cant finding could then be applied to the intersection of the minor approach with the opposing travel lanes. Researchers performed a PCA on the set of remaining traffic control device variables in order to identify combinations of these variables that would capture the most variability in the traffic control devices of the sites. The combinations of variables and their weights, as suggested by the resulting PCA, were then consid- ered in a round of model selection. Results The following sections contain the results of the model ing process for four different analyses of the final dataset (i.e., intersecting roadways with median openings, four through lanes, and median width less than or equal to 120 ft): • High-speed rural divided highways—all wrong-way crashes, • High-speed rural divided highways—nighttime wrong- way crashes, • High-speed urban divided highways—all wrong-way crashes, and • High-speed urban divided highways—nighttime wrong- way crashes.

16 Expanded Variable Name Short Variable Name Variable Description Type of Intersection Control IntersectionTC_TTI Type of control at intersection of minor and major roadways. 1 is none; 2 is STOP signs; 3 is YIELD signs; 4 is traffic signal. Divided Highway Sign Near Right NRDivided_Hwy_Sign Presence of divided highway sign on the near right of the minor road approach. 1 for present; 0 for not present. Divided Highway Sign Near Left NLDivided_Hwy_Sign Presence of divided highway sign on the near left of the minor road approach. 1 for present; 0 for not present. Total Number of Divided Highway Signs TotalNo_DH_Sign Sum of divided highway signs. ONE WAY Sign Near Right Outside NROutOneWay_Sign Presence of ONE WAY sign on the near right outside. 1 for present; 0 for not present. ONE WAY Sign Near Left Inside NLInsOneWay_Sign Presence of ONE WAY sign on the near left inside. 1 for present; 0 for not present. ONE WAY Sign Near Right Inside NRInsOneWay_Sign Presence of ONE WAY sign on the near right inside. 1 for present; 0 for not present. ONE WAY Sign Far Right Inside FRInOneWay_Sign Presence of ONE WAY sign on the far right inside. 1 for present; 0 for not present. ONE WAY Sign Far Left Outside FLOutOneWay_Sign Presence of ONE WAY sign on the far left outside. 1 for present; 0 for not present. ONE WAY Sign Far Right Outside FROutOneWay_Sign Presence of ONE WAY sign on the far right outside. 1 for present; 0 for not present. Total Number of ONE WAY Signs TotalNo_OneWay_Sign Sum of ONE WAY signs. DO NOT ENTER Sign Near Left Outside NLOutDNE_Sign Presence of DO NOT ENTER sign on the near left outside. 1 for present; 0 for not present. DO NOT ENTER Sign Near Left Inside NLInsDNE_Sign Presence of DO NOT ENTER sign on the near left inside. 1 for present; 0 for not present. DO NOT ENTER Sign Far Right Inside FRInsDNE_Sign Presence of DO NOT ENTER sign on the far right inside. 1 for present; 0 for not present. DO NOT ENTER Sign Far Right Outside FROutDNE_Sign Presence of DO NOT ENTER sign on the far right outside. 1 for present; 0 for not present. Total Number of DO NOT ENTER Signs TotalNo_DNE_Sign Sum of DO NOT ENTER signs. WRONG WAY Sign Near Left Outside NLOutWW_Sign Presence of WRONG WAY sign on the near left outside. 1 for present; 0 for not present. WRONG WAY Sign Near Left Inside NLInsWW_Sign Presence of WRONG WAY sign on the near left inside. 1 for present; 0 for not present. WRONG WAY Sign Far Right Inside FRInsWW_Sign Presence of WRONG WAY sign on the far right inside. 1 for present; 0 for not present. WRONG WAY Sign Far Right Outside FROutWW_Sign Presence of WRONG WAY sign on the far right outside. 1 for present; 0 for not present. Total Number of WRONG WAY Signs TotalNo_WW_Sign Sum of WRONG WAY signs. Type of Interior Control InteriorTC_TTI Type of control in the median opening at wrong-way entry point. 1 is none; 2 is STOP signs; 3 is YIELD signs; 4 is traffic signal. KEEP RIGHT Sign KeepRight_Sign Presence of KEEP RIGHT sign in the median opening. 1 for present; 0 for not present. Stop or Yield Line in Median Opening Stopbar_Crossover Presence of stop or yield line in median opening. 1 for present; 0 for not present. Stop or Yield Line on Minor Approach Stopbar_MinorApproach Presence of stop or yield line on minor approach. 1 for present; 0 for not present. Centerline in Median Opening Centerline_Crossover Presence of centerline in median opening. 1 for present; 0 for not present. Wrong-Way Arrow Markings Near Left Near_WW Arrows Present of wrong-way arrow markings on the near left side of the intersection. 1 for present; 0 for not present. Wrong-Way Arrow Markings Far Right Far_WW Arrows Present of wrong-way arrow markings on the far right side of the intersection. 1 for present; 0 for not present. Total Number of Traffic Control Devices in Field of View TotalNo_TCD Total number of traffic control devices present (sum of all signs and pavement markings). Table 8. Traffic control device variables.

17 The probability of wrong-way crashes was used as the response variable in all the analyses. The datasets were made up of roughly a 50/50 split between sites with a history of wrong-way crashes and randomly selected control sites. High-Speed Rural Divided Highways— All Wrong-Way Crashes The overall high-speed rural divided highway dataset con- tained 297 sites. Table 9 contains the coefficients for the high- speed rural divided highway wrong-way crash model. This model shows that there was a multiplicative increase in the risk of wrong-way crashes associated with increasing traffic volume (i.e., AADT). It was estimated that each time the traf- fic volume doubles, the odds of wrong-way crashes increase by a factor of 1.91 (exp(2*0.934) = 1.910). An increase in traf- fic volume means there are more vehicles on the roadway. Therefore, a driver making a wrong-way maneuver would be more likely to encounter another vehicle. The presence of a centerline in the median opening was related with a reduction in the risk of wrong-way crashes. The odds of wrong-way crashes at locations with centerlines in the median opening were 0.316 times as large as the odds of wrong-way crashes at locations without this pavement mark- ing (exp(−1.153) = 0.316). Researchers believe the centerline in the median opening provides a visual cue for drivers that may draw their attention to the median opening and oppos- ing traffic lanes. Similarly, the presence of wrong-way arrow markings in the near main lanes near the intersection was found to be cor- related with a reduction in wrong-way crash risk. The odds of wrong-way crashes at locations with wrong-way arrow markings were 0.258 times as large as the odds of wrong- way crashes at locations without this pavement marking (exp(−1.353) = 0.258). Researchers believe that the wrong- way arrow markings indicate to drivers that both of the near main lanes are going in the same direction. The presence of the DO NOT ENTER sign on the near left inside of the intersection was also found to be associ- ated with a safety benefit. The odds of wrong-way crashes at locations with this sign were 0.351 times as large as the odds of wrong-way crashes at locations without this treatment (exp(−1.047) = 0.351). The placement of the DO NOT ENTER sign in the median would provide a warning to drivers attempt- ing to turn left into the near main lanes. As discussed previ- ously, this maneuver was the cause of most of the wrong-way crashes in Texas and Iowa for which the wrong-way entry Fixed Effects Coefficients Estimate Standard Error z-value p-value Significance log(AADT) 0.934 0.329 2.842 0.0045 ** Centerline in Median Opening −1.153 0.545 −2.117 0.0342 * Wrong-Way Arrows in Near Left Lanes −1.353 0.616 −2.195 0.0282 * Near Left Inside DO NOT ENTER Sign −1.047 0.526 −1.990 0.0465 * MUTCD Median Width (with Traffic Control in Median Opening) 0.015 0.013 1.203 0.2291 MUTCD Median Width (without Traffic Control in Median Opening) 0.075 0.022 3.397 0.0007 *** Random Effects Coefficients Mean −10.065 2.980 −3.378 0.0007 *** Standard Deviation 1.769 * Statistically significant at the 5% significance level. ** Statistically significant at the 1% significance level. *** Statistically significant at the 0.1% significance level. Table 9. High-speed rural divided highway—wrong-way crash model. Source: Texas A&M Transportation Institute. Figure 7. Sign location nomenclature reference.

18 point was known. Although not included in the modeling process, researchers believe this result would also be true for a DO NOT ENTER sign placed in the median on the far right side. Finally, the correlation between wrong-way crashes and the median width was found to be dependent upon the pres- ence of control in the median opening. At sites with a STOP or YIELD sign, there was no change in the risk of wrong-way crashes based on the median width (i.e., statistically equiva- lent to zero). Conversely, at sites without a STOP or YIELD signs, it was found that the odds of wrong-way crashes tended to increase by a multiplicative factor of 2.117 for every addi- tional 10 ft of median width (exp(0.075*10) = 2.117). Figure 8 shows the type of control (i.e., none, STOP sign, or YIELD sign) in the median opening for the rural sites. MUTCD Figure 2B-16 does not show control in the median opening when the median width is less than 30 ft. In the rural dataset, this was the case for 11 percent of the sites. MUTCD Figure 2B-15 shows control (either a STOP or YIELD sign) in the median opening when the median width is greater than or equal to 30 ft. Most of the sites (55 percent) exhibited these characteristics. At 4 percent of the sites, a STOP or YIELD sign was located in the median opening even though the median width was less than 30 ft. However, 30 percent of the sites had a median width greater than or equal to 30 ft with no control in the median opening. The median widths in these instances were typically between 30 ft and 80 ft. High-Speed Rural Divided Highways— Nighttime Wrong-Way Crashes The nighttime high-speed rural divided highway dataset also contained 297 sites. However, in this model the indica- tor variable “crash” equaled 1 if a data point represented an instance of a nighttime wrong-way crash and 0 if a data point represented a site with no recorded nighttime wrong-way crash (i.e., control corridors and wrong-way crash corridors with only daytime wrong-way crashes). Table 10 shows the coefficients for the high-speed rural divided highway nighttime crash model. As with the over- all high-speed rural divided highway model, the presence of a centerline in the median opening and wrong-way arrow markings resulted in a reduction in the risk of nighttime wrong-way crashes (exp(−0.828) = 0.437 and exp(−2.284) = 0.102, respectively). The presence of the DO NOT ENTER sign on the near left inside of the intersection also made it into the model, but the significance of the correlation with nighttime wrong-way crashes was less (only statistically sig- nificant at a 10 percent level compared to 5 percent for the overall model). Figure 8. Type of control in the median opening—rural sites.

19 The presence of either the near right outside ONE WAY sign or near left inside ONE WAY sign was associated with a reduction in the risk of nighttime wrong-way crashes. The odds of nighttime wrong-way crashes reduced by a factor of 0.487 at locations with either of these signs compared to locations without these signs (exp(−0.719) = 0.487). Both of these ONE WAY signs are currently required in the MUTCD. The optional ONE WAY sign on the near right inside did not make it into the model. Only the ONE WAY signs on the near side were included in the modeling process. Even so, researchers believe this result would also be true for ONE WAY signs placed in the median on the far right inside or far left outside (i.e., where the median opening intersects with the opposing travel direction). The presence of either a stop or yield line in the median opening was also associated with a reduction in the risk of nighttime wrong-way crashes. The odds of nighttime wrong-way crashes at locations with a stop or yield line in the median opening were 0.437 times as large as the odds of nighttime wrong-way crashes at locations without this treatment (0.437 = exp(−0.828)). Researchers believe that stop and yield lines in the median opening may provide yet another visual cue for drivers indicating that they should cross the near main lanes to reach the opposing travel lanes. Similar to the findings for all wrong-way crashes at rural sites, the median width was found to be dependent upon the presence of control in the median opening. At sites with Fixed Effects Coefficients Estimate Standard Error z-value p-value Significance Near Right Outside ONE WAY Signa −0.719 0.298 −2.414 0.0158 * Near Left Inside ONE WAY Signa −0.719 0.298 −2.416 0.0157 * Wrong-Way Arrows in Near Left Lanes −2.284 0.812 −2.815 0.0049 ** Near Left Inside DO NOT ENTER Sign −1.041 0.585 −1.780 0.0751 ~ Stop or Yield Line in Median Openingb −0.828 0.364 −2.277 0.0228 * Centerline in Median Openingb −0.828 0.364 −2.275 0.0229 * Median Opening Width 0.012 0.006 1.823 0.0682 ~ MUTCD Median Width (with Traffic Control in Median Opening) 0.030 0.016 1.918 0.0551 ~ MUTCD Median Width (without Traffic Control in Median Opening) 0.083 0.026 3.176 0.0015 ** Random Effects Coefficients Mean −3.555 0.386 −9.199 0.0000 *** Standard Deviation 1.861 a Jointly estimated as an unweighted linear combination suggested by PCA. b Jointly estimated as an unweighted linear combination suggested by PCA. ~ Statistically significant at the 10% significance level. * Statistically significant at the 5% significance level. ** Statistically significant at the 1% significance level. *** Statistically significant at the 0.1% significance level. Table 10. High-speed rural divided highway—nighttime crash model. STOP or YIELD signs, researchers found a weak correlation (only statistically significant at a 10 percent level) between median width and nighttime wrong-way crashes. The odds of nighttime wrong-way crashes tended to increase by a multi- plicative factor of 1.350 for every additional 10 ft of median width (exp(0.030*10) = 1.350). In comparison, at sites with- out a STOP or YIELD sign, researchers found that the odds of nighttime wrong-way crashes tended to increase by a multi- plicative factor of 2.293 for every additional 10 ft of median width (exp(0.083*10) = 2.293) (statistically significant at a 1 percent level). Researchers kept the median opening width in the final model because its presence improved the overall fit of the model. However, the correlation between the median open- ing width and nighttime wrong-way crashes was weak com- pared to the other variables (exp(0.012) = 1.012 and only statistically significant at a 10 percent level). High-Speed Urban Divided Highways— Wrong-Way Crashes The overall high-speed urban divided highway dataset contained 210 sites. Table 11 contains the coefficients for the high-speed urban divided highway wrong-way crash model. This is the only model in which a difference among the states was found. There was an increase in the risk of wrong-way crashes associated with urban sites located in California.

20 Given the small set of model covariates and the small sample of urban sites from California (n = 19), researchers speculate that there was a confounding, unaccounted factor that pro- duced the increased risk. Similar to the findings for the overall rural model, the median width was found to be dependent upon the presence of control in the median opening. At sites with a STOP or YIELD sign, there was no change in the risk of wrong-way crashes based on the median width (i.e., statistically equiva- lent to zero). Conversely, at sites without a STOP or YIELD sign, it was found that the odds of wrong-way crashes tended to increase by a multiplicative factor of 2.721 for every addi- tional 10 ft of median width (exp(0.1001*10) = 2.721). Figure 9 shows the type of control (i.e., none, STOP sign, or YIELD sign) in the median opening for the urban sites. Thirty- five percent of the sites had a median width less than 30 ft and no control in the median opening. In addition, 37 percent of the sites had a median width greater than or equal to 30 ft with either a STOP or YIELD sign in the median opening. At 8 percent of the sites, a STOP or YIELD sign was located in the median opening even though the median width was less than 30 ft. However, 20 percent of the sites had a median width greater than or equal to 30 ft with no control in the median opening. The median widths in these instances were typically between 30 ft and 50 ft. The presence of a stop or yield line on the minor approach to the intersection and the presence of the near left divided highway sign were both loosely associated with a reduction in the risk of wrong-way crashes (statistically significant at only a 10 percent level). However, their inclusion improved the over- all fit of the model. It should also be noted that the near left divided highway sign (which is optional in the MUTCD) was used at only nine sites and was always paired with a divided highway sign on the near right (which is required in the MUTCD for median widths greater than or equal to 30 ft). The near right divided highway sign did not make it into the model. High-Speed Urban Divided Highways— Nighttime Wrong-Way Crashes The overall high-speed urban divided highway dataset also contained 210 sites. As with the rural nighttime model, in the urban nighttime model, the indicator variable “crash” equaled 1 if a data point represented an instance of a night- time wrong-way crash and 0 if a data point represented a site with no recorded nighttime wrong-way crash (i.e., control corridors and wrong-way crash corridors with only daytime wrong-way crashes). Table 12 shows the coefficients for the high-speed urban divided highway nighttime crash model. The presence of a near left inside WRONG WAY sign was associated with a reduction in the risk of nighttime crashes. This sign was located on the same side of the road as the DO NOT ENTER sign that made it into the overall rural model and rural nighttime model (although its impact was less sig- nificant in the rural nighttime model). Again, signs placed in the median versus the outside of the roadway provide a warning to drivers attempting to turn left into the near main lanes. The odds of nighttime wrong-way crashes at locations with a near left inside WRONG WAY sign were 0.495 times as large as the odds of nighttime wrong-way crashes at locations without this treatment (exp(−0.704) = 0.495). The presence of the near left divided highway sign (which is optional in the MUTCD) was associated with a more sig- nificant reduction in the risk of nighttime wrong-way crashes (statistically significant at a 5 percent level). However, the small sample size (n = 9) limits the practicality of this finding. Fixed Effects Coefficients Estimate Standard Error z-value p-value Significance State of California 2.675 1.004 2.665 0.0077 ** Stop or Yield Line on Minor Approacha −0.776 0.461 −1.683 0.0923 ~ Near Left Divided Highway Signa −1.553 0.922 −1.683 0.0923 ~ MUTCD Median Width (with Traffic Control in Median Opening) −0.029 0.019 −1.550 0.1212 MUTCD Median Width (without Traffic Control in Median Opening) 0.100 0.030 3.363 0.0008 *** Random Effects Coefficients Mean −1.245 0.457 −2.722 0.0065 ** Standard Deviation 1.828 a Jointly estimated as a weighted linear combination suggested by PCA. ~ Statistically significant at the 10% significance level. ** Statistically significant at the 1% significance level. *** Statistically significant at the 0.1% significance level. Table 11. High-speed urban divided highway—wrong-way crash model.

21 Figure 9. Type of control in the median opening—urban sites. Fixed Effects Coefficients Estimate Standard Error z-value p-value Significance Near Left Inside WRONG WAY Signa −0.704 0.352 −1.999 0.0456 * Near Left Divided Highway Signa −2.111 1.056 −1.999 0.0456 * Stop or Yield Line on Minor Approacha −0.704 0.352 −1.999 0.0456 * MUTCD Median Width (with Traffic Control in Median Opening) −0.022 0.014 −1.624 0.1045 MUTCD Median Width (without Traffic Control in Median Opening) 0.049 0.019 2.594 0.0095 ** Median Opening Width −0.011 0.006 −1.812 0.0700 ~ Near Left-Turn Lane −0.715 0.396 −1.805 0.0711 ~ Random Effects Coefficients Mean −1.190 0.006 −198.297 0.0000 *** Standard Deviation 0.7304 a Jointly estimated as a weighted linear combination suggested by PCA. ~ Statistically significant at the 10% significance level. * Statistically significant at the 5% significance level. ** Statistically significant at the 1% significance level. *** Statistically significant at the 0.1% significance level. Table 12. High-speed urban divided highway—nighttime crash model.

22 The presence of a stop or yield line on the minor approach was also associated with a more significant reduction in the risk of nighttime wrong-way crashes. The odds of nighttime wrong-way crashes at locations with a stop or yield line on the minor approach were 0.495 times as large as the odds of nighttime wrong-way crashes at locations without this pave- ment marking (exp(−0.704) = 0.460). Similar to the findings for the other three models, the median width was found to be dependent upon the presence of control in the median opening. At sites with a STOP or YIELD sign, there was no change in the risk of wrong-way crashes based on the median width (i.e., statistically equivalent to zero). In contrast, at sites without a STOP or YIELD sign, it was found that the odds of nighttime wrong-way crashes tended to increase by a multiplicative factor of 1.632 for every additional 10 ft of median width (exp(0.049*10) = 1.632). Researchers kept the median opening width and the pres- ence of the near left-turn lane in the final model because these variables improved the overall fit of the model. How- ever, the correlation between these two variables and night- time wrong-way crashes was weak compared to the other variables (statistically significant at only a 10 percent level). Median Width Threshold for Control in the Median Opening Because the effects of median widths at sites without control in the median opening were found to monotonically increase (i.e., without peaking), the research team decided to use the high-speed rural divided highway crash model that included all wrong-way crashes to investigate if a minimal threshold for recommending the use of median opening control (i.e., STOP or YIELD sign) could be determined. Researchers selected the rural model instead of the urban model because of the wider range of median widths without control in the median opening in the rural dataset. Researchers compared the magnitude of the effect of median width to the amount of unexplained variability of differences between sites, after accounting for the effects of other variables. The rationale was that although this analy- sis found an increasing effect for increasing median widths, such an effect may not be practically significant when con- sidering the random variability in the odds of wrong-way crashes from site to site. If the variability was large com- pared to the change in odds linked to median width, such an expected change should be virtually undetectable. However, it was also anticipated that for a wide enough median, the expected increase in odds of wrong-way crashes associated with median width should be clearly observable, even when the random variation in the odds of wrong-way crashes due to differences between sites was present. Figure 10 shows the log-odds line of the effect of median width (compared to median width equal to 0 ft) and the 90 percent and 95 percent confidence envelope of uncertainty associated with variability between sites (i.e., the model ran- dom effect variance). It can be seen that, depending on the Median (ft) C h an g e in lo g -o d d s o f W W C ra sh Figure 10. Log-odds median width effect plus random effect (without control in the median opening).

23 level of confidence desired, the increase in wrong-way crash log-odds associated with median width should be clearly identifiable for median widths of about 50 ft or wider. Summary This chapter documented the exploratory and statistical analyses of wrong-way crashes on high-speed divided high- ways. Researchers found that most of the wrong-way crashes on these roadways resulted in serious injury, occurred over the weekend (Friday through Sunday), and happened at night. More than two-thirds of the wrong-way crashes resulted from a wrong-way maneuver from an inter secting roadway where a median opening was present. Overall, females comprised about one-third of the wrong-way drivers, while males rep- resented two-thirds of the wrong-way drivers. Older drivers (≥ 60) and those ages 21 to 29 comprised more than 60 per- cent of the wrong-way drivers. The safety analysis found that there were numerous sites where the traffic control in the median opening did not fully comply with the MUTCD with respect to treating the location as one or two intersections. Primarily this was represented in narrow medians (less than 30 ft) with a STOP or YIELD sign in the median opening (6 percent) or in wide medians (greater than or equal to 30 ft) with no STOP or YIELD sign in the median opening (26 percent). This may be an indi- cation that practitioners are using engineering judgment to determine the most effective installation of interior right- of-way devices to address safety and operations at divided highway junctions. NCHRP Report 375: Median Intersection Design (Harwood et al. 1995) found a similar deviation in the interior right-of-way traffic control treatments for median widths between 30 ft and 85 ft (median width included left- turn lanes, if present). There was also evidence from the safety analysis that the criterion may be 50 ft instead of 30 ft. Additional specific findings from the safety analysis include the following: • Greater use of ONE WAY signs (above those that are required) does not appear to deter wrong-way movements. • There was limited evidence that use of the required divided highway sign on the crossroad exterior approaches deters wrong-way movement. • The placement of DO NOT ENTER and WRONG WAY signs on the inside turn of a wrong-way movement (side of divided highway nearer the right-of-way line) does not deter wrong-way movements. • Treatments that appear to deter wrong-way movements include: – DO NOT ENTER and WRONG WAY signs on the out- side of a wrong-way turn, – Wrong-way arrow markings for the through lanes on the divided highway, – Presence of a centerline in the median opening, and – Use of stop or yield lines when interior right-of-way treatments are provided.