Safety Evaluation of Signalized Restricted Crossing U-Turn Intersections

Chapter 5. Results

Sites

The project team identified 28 signalized RCUTs from 5 States (Alabama, Michigan, North Carolina, Ohio, and Texas). After some investigation, they eliminated two sites in Michigan because the RCUTs were installed in the mid-1990s, and crash data were no longer available from the before period. The project team obtained aerial photos of the remaining 26 sites to review for substantial changes in the roadways or surrounding land uses between the before period and after period. Such changes would likely confound the RCUT installation and make determination of the effects of the RCUT installation impossible. On this basis, the project team eliminated 15 sites. The most common confounding change resulting in site elimination was that the RCUT was installed at the same time the intersection traffic control changed from unsignalized to signalized. Table 3 identifies the final sites selected for evaluation, which included 11 sites in 4 States where signalized RCUTs were installed during years lending themselves to evaluation, and the RCUT installation was the only substantial change at the location.

Table 4 shows the major geometric characteristics of the treatment sites after RCUT construction. Two numbers in a cell means that two approaches to the intersection had those two different values for the parameter. The treatment sites have much in common besides having signalized RCUTs. They are all in suburban areas, on four-lane or six-lane divided arterials, and characterized by high-speed traffic and minimal crossing pedestrians; as a result, the CMFs emerging from this effort will apply to this type of site. Most of the signalized RCUT installations of which the project team is aware that are planned for the next few years share these characteristics.

N/A = not applicable; 1 = Alabama-Plum; 2 = Alabama-Retail; 3 = North Carolina; 4 = Ohio-Symmes; 5 = Ohio-Tylersville; 6 = Ohio-Hamilton-Mason; 7 = Texas-Evans; 8 = Texas-Stone Oak; 9 = Texas-New Guibeau; 10 = Texas-Shaenfield; 11 = Texas-71.

After identifying treatment sites and potential comparison sites, the project team contacted officials in each of the four States for the necessary data. State officials provided information on the RCUTs’ construction dates and crash data for the most recent 5 full years before the start of construction, through the construction year(s), up to the most recent available data. The project team decided 5 years of before data was desirable to boost the sample size and to provide the longest possible test of the quality of the comparison sites. The project team also reviewed the available aerial photos for each treatment site to ensure that no substantive changes had occurred during the 5-year before period.

States also provided data on all reported crashes that occurred within 1,500 ft of the main intersection along the major street and within 500 ft along the minor street for treatment sites. These distances ensured that all crashes related to the RCUT—especially those at the U-turn crossovers—were included in the dataset, while midblock crashes were excluded. The project team adjusted the data collection boundary in several instances due to geometric design differences. Two sites in Texas had longer distances to crossovers, so the boundary was extended. Two Alabama sites—US-231 northwest of Dothan at Plum Road and Retail Drive—had other intersections 1,200 ft and 1,000 ft from the main intersections. Thus, the team set the data collection boundary at 1,000 ft at Plum Road and 800 ft at Retail Drive.

Additionally, States provided traffic volume data at and around each treatment site and information on changes to the intersection geometry accompanying the RCUT construction. The States also provided the following crash reporting thresholds:

• Alabama: By law, there is no crash reporting threshold.
• North Carolina: $1,000 total crash damage or a reportable injury.
• Ohio: Fatality, injury, or $1,000 or more in property damage.
• Texas: Fatality, injury, or $1,000 or more in property damage to any one person’s property.

The thresholds did not change in any State during the study period. It is important to note the difference between Alabama’s threshold and the other three States’ threshold.

In the course of contacting officials in the 4 States for crash and other data for the 11 RCUT treatment sites and associated potential comparison sites, the project team also asked the officials why the RCUTs were installed. This information was important when selecting the analysis methodology. If treatments were installed as the result of a high-crash site identification process, regression to the mean is likely to result in a natural decrease in the number of crashes in the after period; methods to adjust for this threat should be used. On the other hand, if the treatments were installed for reasons other than to mitigate safety concerns, regression to the mean is not a potential bias to the study. Regression to the mean may occur from the before period to the after period, but it is just as likely to result in an increase as a decrease in the number of crashes. In this case, methods to adjust for regression to the mean are not needed and may be harmful if they are not executed appropriately. The responding officials in all four participating States indicated that all RCUTs were chosen for operational reasons, including to accommodate future growth in traffic, not due to a high number of crashes. All officials indicated safety was considered and that RCUTs were expected to perform well on the safety dimension but was not a case of selecting the design as a safety treatment for an identified high-crash site.

Table 5 summarizes the crash data provided for the treatment sites. There were some deviations from the desired 5 years of before-period data at some sites due to data availability, most prominently at the Texas sites where no crash data were available before 2009. Also, before-period data from 2003 to 2006 were collected at the Hamilton-Mason Road site in Ohio because crash data from 2007 to 2009 had reporting inconsistencies. The project team collected the oldest 5 years of after-period data available for the North Carolina site.

One way to confirm whether regression to the mean is a serious threat to the validity of an analysis is to look for a significant drop in crashes on average in the year before treatment compared to previous years. The decision to install the countermeasure is likely made without benefit of the crash data from that last year before construction, so if regression to the mean were to manifest in the time series, it would likely do so starting in that year. The project team was able to examine this trait at 8 of the 11 treatment sites, since data for Ohio-Hamilton-Mason were not available in the year before construction, and only 1 year of before-period data was available at the Texas-Evans and Texas-Stone Oak sites. Table 6 shows a summary of the year before construction data and reveals that only two of the eight sites—Alabama-Retail and Texas-Shaenfield—had a significant drop in crashes during the year before construction. At the other six sites, the number of crashes either matched the average from the previous years or rose. It appears that regression to the mean was indeed not a serious threat to the validity of this analysis.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.
*Construction years.
--No data.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.

Adjustment for Traffic Volume

This section reports the results from the analysis that adjusted for an estimated change in traffic volume between the before and after periods. Traffic volume may change due to development in the area, economic changes that affect travel, or changes in overall levels of motorization. Except for the Texas-71 site, the project team obtained traffic volume data from at least one official count or estimate by the responsible agency for at least one station near each site in the before period and the after period. If the responsible agency provided a count or estimate at more than one station and/or year in a before or after period at a particular site, the project team averaged all available counts. Table 7 summarizes the traffic volume data. The data for the Texas-Evans and Texas-Stone Oak sites are from counts on the major and minor streets, while count data for the other sites are from the major street only.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.
*Construction years.
--No data.

Table 8 shows the results from the analysis of all crashes after accounting for changes in traffic volumes from the before period to the after period. The safety effect is represented by a CMF. The crash reduction factor (CRF) is calculated as 1 minus the CMF value, converted from a proportion to a percentage. As discussed in the methodology section, a CMF equal to 1.0 implies that the treatment is not expected to change the number of crashes, while a CMF below 1 implies that the treatment is expected to reduce crashes. There was no available traffic volume estimate for the Texas-71 site; therefore, the project team could not compute CMF with this method.

Table 8 also shows the estimated SD of each CMF, which can be used to produce an approximate confidence interval for the CMF and determine whether it is significantly different from 1.0 at a particular significance level. If a CMF is 1 SD from 1.0, that corresponds to a 68-percent confidence level that the actual value is different from 1.0. If a CMF is 1.96 SDs from 1.0, that corresponds to a 95-percent confidence level. Among other things, the SD of a CMF is a useful indicator of the adequacy of the sample size, with small SDs usually associated with large samples of crashes.

As shown in table 8, eight sites had a CMF less than 1.0, and two sites had CMF values greater than 1.0. The North Carolina site, both Alabama sites grouped together, and all three Ohio sites grouped together had CMF values well under 1.0, while the Texas sites as a group had a CMF close to 1.0. All 10 sites grouped together had a CMF of more than 1.0, due to the influence of the large sample size from the Texas-Stone Oak site sample size. SDs were reasonable, ranging from 0.09 to 0.25.

The results in table 8 are based on the sums of the traffic volumes from the individual sites. This gives rise to the possibility that the results were skewed by higher volume sites. To account for this possibility, the project team reanalyzed the last five rows in table 8 using volume indices rather than totals (i.e., normalizing the volume at each site in the before period to a value of 1.0). The reanalysis provided only minor changes in the CMFs and the SDs of the CMFs, suggesting that summing the volumes was an acceptable technique for this evaluation.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.

Comparison Groups

The primary analysis method in this evaluation used comparison sites to adjust for the possibility of changes during the study period, such as development patterns, significant weather events, driver behavior, vehicle fleets, and crash reporting tendencies. As discussed in the methodology section, the key to a successful comparison site analysis is to find comparison sites that match the crash patterns of the treatment sites year by year in the before period. It is believed that the comparison sites change year by year in the after period, reflecting what would have happened at the treatment sites if the treatment had not been installed.

The project team selected comparison sites using the following criteria:

• Geographically located in proximity to the treatment site to be affected by the same trends.
• Geographically located far enough from the treatment site so as not to be affected by the treatment itself.
• Suitable crash sample.
• No obvious changes (e.g., large development or construction projects) between the before and after periods that would affect the crash pattern at a comparison site.
• Sites similar to the treatment sites in terms of design, such as large signalized intersections.

The project team narrowed the list of potential comparison sites using the time series of aerial photos available via the Internet. This review revealed construction or development at many of the potential comparison sites during the time periods of interest, thus disqualifying those sites from further analysis. The project team consulted with local or State transportation officials responsible for the remaining potential sites to gather additional input on high-quality sites.

Table 9 provides a list of potential comparison sites that met the criteria to warrant the collection of crash data. The before and after periods of crash data collected at all comparison sites were equal to the periods for the corresponding treatment site. There were four potential comparison sites for each treatment site.

OH = Ohio; TX = Texas.

The project team employed the odds ratio test, as Hauer described, to determine the best comparison site(s) for each treatment site or group of treatment sites.⁽⁵⁾ The odds ratio test measured the change in crash frequency at a comparison site from one year to the next relative to the change in crash frequency at the treatment site. The project team employed the odds ratio test for each available pair of years in the before period at each site and group of sites. Thus, the project team calculated four odds ratios at most at each site, with 5 years of crash data at most in the before period. An odds ratio of 1.0 meant that the comparison and treatment datasets moved together in complete harmony. The project team examined the mean of the odds ratios calculated, as well as the SD of that mean, with desired values of 1.0 for the mean and 0 for the SD. For each individual treatment site, the project team conducted the odds ratio test on each of its potential comparison sites, on potential comparison sites identified for nearby treatment sites, and on each combination of potential comparison sites. The project team used a combination of the best comparison sites found for each individual treatment site for each group of treatment sites.

In this dataset, the project team conducted tests of comparison sites for 9 of the 11 sites where at least 2 years of before-period data were available. They did not test Texas-Evans Road and Texas-Stone Oak comparison sites because only 1 year of before data were available at those sites. Rather, the project team conducted an analysis using all four comparison sites.

Table 10 displays the test results for the best comparison site or sites with each of the nine other treatment sites. To illustrate the quality of the comparison sites, figure 9 and figure 10 show the correspondence between the numbers of crashes per year in the before period at treatment and chosen comparison sites for the Ohio-Symmes site and for the set of all treatment sites, respectively. Most of the nine treatment sites had good to excellent comparison site(s), with means of the odds ratios within 0.1 of 1.0, SDs of the means less than 0.4, and relatively high sample sizes. The sites with the strongest comparison sites included the following:

• Alabama-Plum: Retail 2 and 4.
• Alabama-Retail: 1.
• North Carolina: 1 and 3.
• Ohio-Symmes: 1 and 3.
• Ohio-Tylersville: 1.
• Ohio-Hamilton-Mason: 2.
• Texas-New Guibeau: 1, 2, 3, and 4.
• Texas-Shaenfield: 1, 2, 3, and 4.
• Texas-71: 1, 3, and 4.

Note that, at the Alabama-Plum Road site, the best comparison site was from the Retail Drive set of potential sites. This is due to the proximity of those two treatment sites.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.
--No data.

Comp. crashes = comparison crashes.

The most important results from the evaluation—the results for all crashes using comparison sites to adjust for potential simultaneous event biases—are shown in table 11. This includes a statistic called Var{w}, which is an intermediate step in the calculation of the CMF and the SD of the CMF.⁽⁵⁾ Anyone attempting to replicate the calculations will need Var{w}. For cases in which the Hauer formula suggested a negative value for Var{w}, the project team used a value of 0.0055, which is an intermediate value in the range suggested by Hauer.⁽⁵⁾ Table 11 shows that eight sites had CMF values less than 1.0, and three sites had CMF values greater than 1.0. For nine sites, the CMF estimate was more than 1 SD away from 1.0. The group results showed CMF values less than 1.0. The CMF for all sites was 0.85 with an SD of 0.16, so the CMF was less than 1 SD away from 1.

Injury Crashes

The results in table 11 are based on all reported crashes. This subsection provides the results of the analysis on fatal and injury crashes. There were not enough fatal crashes in either the before or after periods to analyze in any detail. In the before period, the following was the fatal crash history at the treatment sites:

• Ohio-Hamilton-Mason: 1.
• Texas-Evans: 1.
• Texas-71: 3.

There were no fatal crashes at any treatment site after converting it to an RCUT.

The sample of injury crashes was large enough to allow for the same analyses as for all crashes, namely, traffic volume adjustment and comparison site adjustment.

Table 12 contains the results for the analysis of injury crashes adjusting for changes in traffic volume. The analysis again did not use data from the Texas-71 site because no after-period volume data were available. The results are very similar to the results of the traffic volume adjustment analysis on all crash data, with seven sites having a CMF less than 1.0 and two sites having a CMF greater than 1.0. The overall CMF from this analysis was 1.07 with an SD of 0.17.

Table 13 shows the results from odds ratio tests on the best comparison site or sites for each treatment site or group of treatment sites for injury crashes. The odds ratios and SDs were not as strong for injury crashes as all crashes because the sample sizes were smaller. Nonetheless, the odds ratios were mostly at 0.85 or above, and most SDs were below 0.3.

Comp. crashes = comparison crashes.; AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.

AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.
--No data.
Comp. crashes = comparison crashes.

The strongest comparison sites included the following:

• Alabama-Plum: Retail 3 and 4.
• Alabama-Retail: 3.
• North Carolina: 3 and 4.
• Ohio-Symmes: 3.
• Ohio-Tylersville: 1 and 3.
• Ohio-Hamilton-Mason: 1, 2, 3, and 4.
• Texas-New Guibeau: 2, 3, and 4.
• Texas-Shaenfield: New Guibeau 1 and 2.
• Texas-71: 3 and 4.

Note that, at the Alabama-Plum and Texas-Shaenfield sites, the best comparison sites were from the Retail Drive and New Guibeau sets of potential sites, respectively. This is due to the proximity of those treatment sites to each other.

Table 14 shows the results from the analysis of injury crashes using comparison sites. Note that some of the crash counts for groups of comparison sites do not match the sums of the component individual sites because comparison site data were not double-counted. The CMF values were much like those in the previous table of results, with eight sites having CMF values less than 1.0 and three sites having CMF values greater than 1.0. The CMF for all sites, 0.78, was lower than that for all crashes. This result suggests that signalized RCUTs have a larger positive effect on injury crashes than on property-damage-only crashes. The CMF for injury crashes at all sites was greater than 1 SD from a value of 1.0.

Comp. crashes = comparison crashes.; AL = Alabama; NC = North Carolina; OH = Ohio; TX = Texas.

Spatial Patterns

The crash data obtained for this evaluation specified location well enough to allow for a spatial patterns analysis. Notable clusters of crashes included the following:

• Alabama-Plum: Clusters of rear-end crashes southbound on the major street.
• North Carolina: Large clusters of rear-end crashes in both directions of the major street.
• Ohio-Symmes: A cluster of rear-end crashes northbound on the major street and a cluster of rear-end crashes eastbound on the minor street.
• Ohio-Tylersville: Smaller clusters of rear-end and sideswipe crashes in both directions on the major street.
• Texas-Evans: Large clusters of rear-end crashes in both directions of the major street, a cluster of sideswipe crashes northbound on the major street, and a cluster of crashes involving northbound left-turn vehicles.
• Texas-Stone Oak: Large clusters of rear-end crashes on the major street southbound.
• Texas-New Guibeau: Clusters of rear-end crashes northbound on the major street on the side with the stem of the T-intersection.
• Texas-Shaenfield: Clusters of rear-end crashes in both directions on the major street and notable clusters of sideswipe and fixed object crashes southbound on the major street on the side with the stem of the T-intersection.

Thus, with a few exceptions, the crash data mostly showed clusters of rear-end crashes occurring at the RCUT sites. These patterns are somewhat different from those seen at typical conventional intersections, which would tend to feature more prominent clusters of turning and angle crashes.

Other Variables

The collected data allowed for examination of other variables besides severity and location. The appendix provides more information on the changes to crash type, light, weather, road condition, and occasionally other variables (when available) from the before period to the after period at each site. The Alabama-Retail, Ohio-Hamilton-Mason, and Texas-71 sites did not have enough crashes in the after period to detect any important changes. Despite the available data at the Texas-Stone Oak site, no significant changes in the variables occurred at this site. Several important changes occurred at the remaining sites between the before and after periods, including the following:

• Alabama-Plum: The crash type changed drastically, from 68.5 percent right rear angle crashes in the before period to 52.5 percent rear-end center in the after period.
• Alabama-Plum: The percentage of daylight crashes decreased from 74.4 to 60.
• Alabama-Plum: The percentage of crashes in the rain decreased from 20.2 to 7.5.
• North Carolina: The percentage of left-turn, same roadway crashes decreased from 29 to 10.7.
• Ohio-Symmes: The percentage of fixed object crashes increased from 0 to 10.
• Ohio-Symmes: The percentage of wet crashes increased from 32.7 to 45.7.
• Ohio-Tylersville: The percentage of angle crashes decreased from 15.4 to 4.2.
• Ohio-Tylersville: The percentage of wet crashes decreased from 26 to 12.5.
• Texas-Evans: The percentage of same direction, both going straight, rear-end crashes increased from 11.7 to 22.8.
• Texas-Evans: The percentage of dark, no light crashes decreased from 15.5 to 4.9.
• Texas-New Guibeau: The percentage of same direction, both going straight, sideswipe crashes increased from 2.9 to 16.1, while the percentage of same direction, one straight, one stopped crashes decreased from 57.3 to 38.7.
• Texas-Shaenfield: The percentage of same direction, both going straight, sideswipe crashes increased from 2.5 to 19.5, while the percentage of same direction, one straight, one stopped crashes decreased from 60.5 to 24.4.
• Texas-Shaenfield: The percentage of dark, lighted crashes increased from 13.6 to 29.3.

The most prominent changes with RCUT installation appear to be decreases in angle crashes and increases in sideswipe crashes.