Estimating Effectiveness of Safety Treatments in the Absence of Crash Data (2023)

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.

CHAPTER 5 Case Studies This chapter describes two case studies that focus on demonstrating how surrogate measures of safety from different data sources can be used for safety evaluations of two countermeasures that fit the high-priority countermeasure categories in the previous chapter. The first case study evaluates the implementation of leading pedestrian intervals (LPIs) at signalized intersections using conflict data obtained from video cameras. The second case study evaluates traffic signal coordination using high- resolution traffic signal data to produce measures that identify conflicts in a traffic stream as well as more macroscopic, traffic-level measures. In both case studies, the main purpose of the analysis to demonstrate how surrogate measures may be used to assess safety performance, with the goal of providing steps and guidelines that can be repeated for other datasets or safety treatments. Given the contextual framework for the two case studies, the estimated safety effects may not represent a comprehensive or robust evaluation of the treatments of interest as these might differ based on the surrogate methodology being used or other factors. The quantitative findings presented here are meant primarily for demonstration and may provide some indication of the magnitude and direction of the safety impacts expected. In the case of the LPI evaluation, the goal is not to replace or improve on existing crash-based evaluation results, though there is an opportunity to examine whether the results are consistent between the crash-based analysis and this surrogate safety measure evaluation. Case Study 1: Evaluation of LPI This section describes the research activities and results of the first case study, an effort to demonstrate the application of surrogate measures for evaluating LPIs. Chapter 4 identified treatments related to pedestrian and bicyclist facilities and control and treatments related to intersection design and control as two priority areas for applying surrogate safety measures. LPIs fall under both priority areas. The following sub-sections provide background, describe the data collection effort, and present the analysis results. Background An LPI is a traffic signal timing-based countermeasure that provides pedestrians a 3- to 7-second “head start” to cross the intersection before motor vehicles are given a green signal indication (FHWA, n.d.a). This allows the crossing pedestrians to enter the crosswalk and establish their presence at the intersection before vehicles begin moving. The goal of implementing an LPI at a crosswalk is to increase pedestrian visibility to drivers and improve driver yielding, thus improving pedestrian crossing safety. Goughnour et al. (2018) performed a crash-based safety evaluation of LPIs in 2018. The research examined 105 sites where LPIs had been implemented, as well as 451 comparison sites, across 3 U.S. 82 

cities: Chicago, IL; New York, NY; and Charlotte, NC. The study resulted in a set of CMFs for different crash types, crash severities, and LPI characteristics (namely, whether LPIs were implemented at all crossings at the intersection or only for crossings across the minor road, parallel to the major road). The CMFs are included in the CMF Clearinghouse and are highly rated (FHWA, n.d.b). The study computed a CMF of 0.87 for vehicle/pedestrian crashes of all severities at either type of implementation arrangement, corresponding to an estimated 13-percent reduction in crash frequency when implementing LPIs. The study also estimated a CMF of 0.87 for crashes of all types and severities. Aside from the Goughnour et al. research, there is little evidence on the safety effects of LPI, although one recent study did use microsimulation for 15 intersections in Toronto, Canada and found that LPI implementation can result in a reduction of vehicle-pedestrian conflicts of approximately 50 percent (Hassanpour and Persaud, 2022). The safety performance of pedestrian treatments remains a significant gap in crash-based safety analysis, as evidenced by the pedestrian- and bicyclist-related treatments category having the highest number of prioritized treatments in the Task 4 prioritization effort, shown in Table 11. This gap is due in part to lack of exposure data, the relative infrequency of pedestrian and bicyclist crashes, and the inability of more aggregate crash-based evaluations to capture nuanced effects on safety, such as those related to the numbers of lanes that must be crossed, the presence and type of refuge, and vehicle movement speeds. Furthermore, pedestrian facility improvements often induce increased demand, which can be difficult to assess using traditional before- after analysis techniques. The induced demand effect of LPIs is not well understood but is a factor that should be considered. A surrogate safety measure analysis may be beneficial in further exploring the safety performance effects of LPIs and may provide a basis for further surrogate analysis of pedestrian and bicyclist safety treatments. The following are the primary questions that the surrogate safety analysis sought to answer. 1. What challenges exist for conducting a surrogate safety analysis using video-derived conflict data? 2. Do LPIs decrease the frequency of vehicle/pedestrian conflicts? 3. Do LPIs decrease the frequency of vehicle/vehicle conflicts? 4. Can crash-conflict models be developed and used to make inferences on changes in crash frequency from changes in conflicts? 5. Are the results consistent with the results of the crash-based evaluation performed by Goughnour et al. (2018)? The surrogate safety analysis set out to answer these questions by using video-derived conflict data to estimate crash reductions after LPI implementation. The following section details that data and the data collection process, describes the analysis, and presents the results. Conflict-Based Surrogate Safety Analysis of LPIs Data Collection The City of Bellevue, Washington, provided video-derived conflict data which had been collected as part of a pilot analysis of LPIs in the City. The data set corresponds to 20 intersection sites in Bellevue that had LPIs implemented on at least 1 crossing as part of the pilot analysis. The 20 intersection sites contain a mix of treatment crossings, comparison crossings, and excluded crossings: 83 

 Treatment crossings: crossings where an LPI was added in between the before and after data collection periods.  Comparison crossings: crossings without an LPI.  Excluded crossings: crossings not visible in the camera view or crossings where an LPI was already implemented prior to data collection. The 20 intersection sites include 40 treatment crossings, 26 comparison crossings, and 14 excluded crossings. Of the 26 comparison crossings, 20 do not have LPIs and 6 had another traffic signal timing countermeasure implemented: “Rest in WALK Flashing Yellow Arrow (FYA).” Rest in WALK FYA is a signal timing-based countermeasure where the pedestrian phase has an automatic recall, remains in the WALK indication for the entire vehicle phase, and runs concurrently with a permissive left-turn phase that has an FYA display. Of the 14 excluded crossings, 5 added LPIs but are out of view of the camera, 4 do not have LPIs, 4 had existing LPIs at the time of data collection, and 1 does not have a marked crosswalk. LPIs were implemented at the treatment crossings on October 25th, 2020. The original video data collection occurred over two 6-hour periods before LPI implementation and two 6-hours periods after implementation. The before periods were Saturday, October 9th and Thursday, October 14th, 2020 from 12:00 to 18:00. The after periods were Thursday, October 28th and Saturday, October 30th, 2020 from 12:00 to 18:00. In addition to this core data set, the City performed more extensive video data collection at three of the sites that were noted to have higher pedestrian volumes. The three sites were Bellevue Way NE at NE 4th St, 106th Ave NE at NE 4th St, and 108th Ave NE at NE 4th St. These three sites include five treatment crossings, four comparison crossings, and three excluded crossings (one had an LPI added but was not in the camera view and two did not have LPIs). Before period data at these three sites were collected from Tuesday, October 6th to Thursday, October 15th, 2020 during the 8:00 AM to 10:00 PM period and after period data were collected from Sunday, October 26th to Sunday November 8th, 2020 during the 7:00 AM to 9:00 PM period. The additional data collected at the three sites did not undergo the same data cleaning that the original data set had. Because of this, it may present some errors in average speed estimates and certain vehicle trajectories (mostly commercial vehicles). This is due to camera coverage limitations. Table 12 summarizes the conflict video data collection information at the 20 City of Bellevue intersections. 84 

Table 12. Bellevue conflict video data collection intersection sites. Treatment Comparison Excluded ID N/S Street E/W Street Before Hours After Hours Crossings Crossings Crossings 1 100th Avenue Main Street 2 2 0 6 6 2 102nd Avenue Main Street 2 0 2 6 6 3 Bellevue Way NE NE 4th Street 2 1 1 140 196 4 Bellevue Way Main Street 2 1 1 6 6 5 106th Avenue NE NE 4th Street 2 1 1 140 196 6 130th Avenue NE NE 20th Street 3 0 1 6 6 7 108th Avenue NE NE 4th Street 1 2 1 140 196 8 156th Avenue NE Northup Way 4 0 0 6 6 9 112th Avenue Main Street 4 0 0 6 6 10 112th Avenue NE NE 4th Street 3 0 1 6 6 11 100th Avenue NE NE 4th Street 2 2 0 6 6 12 112th Avenue NE NE 2nd Street 1 2 1 6 6 13 100th Avenue NE NE 5th Street 1 2 1 6 6 14 140th Avenue NE NE 20th Street 2 1 1 6 6 15 110th Avenue NE NE 4th Street 2 2 0 6 6 16 156th Avenue NE NE 8th Street 2 2 0 6 6 17 156th Avenue NE NE 15th Street 2 2 0 6 6 18 156th Avenue NE NE 10th Street 0 2 2 6 6 19 156th Avenue NE NE 13th Street 2 2 0 6 6 20 158th Avenue NE NE 8th Street 1 2 1 6 6 Total 40 26 14 522 690 85 

The City of Bellevue contracted with the Advanced Mobility Analytics Group (AMAG) to process these video data and generate the conflict data set. The AMAG platform detects and tracks road user trajectories in the video data to identify conflicts between road users based on surrogate safety metrics. It tracks different road user types including passenger cars, pedestrians, bicyclists, buses, commercial vehicles, motorcycles, trucks, and vans. The data set compiled for this analysis does not include any conflicts involving trucks or vans. The conflict types identified by the platform include:  Adjacent approach conflicts.  Head-on conflicts.  Opposing approach conflicts.  Rear-end conflicts.  Sideswipe conflicts.  Parallel lanes turning conflicts.  U-turn conflicts.  Parked car conflicts.  Pedestrian conflicts.  Bicyclist conflicts. The data set compiled for this analysis does not include any head-on, U-turn, or parked car conflicts. The AMAG platform defines any two road users that appear in the camera view at the same time as “having spatial and temporal proximity,” and computes the potential conflict metrics for the two users, chiefly the TTC and PET values. When two road users experience a “conflict that is indicative of crash risk,” the conflict is defined as a critical conflict. AMAG bases this determination of critical conflicts on established research in crash risk and crash severity risk. AMAG uses a TTC and PET threshold of 1.5 seconds to define a critical conflict. Only a small percentage of road user pairs “having spatial and temporal proximity” register as critical conflicts when the metrics are computed. These critical conflicts are what make up the dataset used in this analysis. The data set includes a variety of information for each critical conflict: the conflict type, road user types involved, value of the chosen surrogate safety measure (depending on conflict type and surrogate measure values), road user movements, latitude and longitude, conflict angle, difference in velocity (Δ𝑉), and timestamp. The research team used the road user movement field to determine which crosswalk each conflict-involved pedestrian was using, and thus whether that conflict belonged to the treatment sample (i.e., crosswalk with observations both before and after LPI implementation) or comparison sample (i.e., crosswalk where no LPI implementation occurred between the before and after periods). The research team performed a similar treatment/comparison assignment exercise for the rear-end conflicts. Of the different vehicle/vehicle conflict types, rear-end conflicts are of particular interest in the context of LPIs. This is because the LPI allows pedestrians to establish their presence in the crosswalk prior to vehicle movements receiving a green signal indication. One goal of this is to improve pedestrian visibility to drivers, and thus reduce situations where motorists turning onto an adjacent intersection leg do not see a pedestrian walk into the crosswalk. As a result, a motorist must engage in a severe deceleration to avoid a collision, potentially leading to a rear-end conflict with a following vehicle. The research team established the convention that a given rear-end conflict would be assigned to the treatment group if one or more of the vehicles involved in the conflict were turning onto an intersection leg with a treated crossing. The example of Conflict A shown in Figure 9 illustrates this. Because 86 

Conflict A involves a vehicle turning onto an intersection leg with a treated crossing, it would be assigned to the treatment group. On the other hand, Conflicts B and C would be assigned to the comparison group because they do not involve any vehicle movements turning onto an intersection leg with a treatment crossing. Additionally, if a given conflict involved two through-moving vehicles originating from an intersection approach with shared through and turning lanes (shared through/left, shared through/right, or shared through/left/right) and one of the adjacent intersection legs had a treatment crossing, the conflict would be assigned to the treatment group. This was a conservative assumption taken by the research team to account for situations where driver indecision or uncertainty in making a turning movement may have led to a rear-end conflict. This situation is illustrated by Conflict D in Figure 9. While neither of the through movements involved in the conflict would actually turn onto an intersection leg with a treatment crossing, the shared through/right-turn lane makes this a possibility. Figure 9. Example of rear-end conflict assignment to treatment or comparison group. 87 

The conflict totals for vehicle/pedestrian and rear-end conflict types are shown in Table 13. The research team excluded any conflicts with a reported TTC or PET value of zero, as recommended by AMAG. One challenge that the research team encountered was that 26 of the 214 total vehicle/pedestrian conflicts (12 percent) had null values for the road user movements, and thus could not be assigned to either the treatment or comparison groups. Similarly, 1,354 of the 15,412 rear-end conflicts (9 percent) had null values for the road user movements. Therefore, the sum of the treatment and comparison group total conflicts for each of the before and after periods is not equal to the total number of vehicle/pedestrian and rear-end conflicts for the period. Table 13. Conflict counts. Vehicle/Pedestrian Rear-End Conflicts Conflicts Before After Before After Treatment 15 18 809 772 Comparison 69 86 5,256 7,221 Unassigned (Percent) 13 (13%) 13 (11%) 833 (12%) 521 (6%) Total 97 117 6,898 8,514 Analysis and Results Based on AMAG-Identified Critical Conflicts Table 14 shows the results for the two conflict types. These are based on the data in Table 13 and the methodology documented in Gross (2017), using the accompanying spreadsheet (available from https://safety.fhwa.dot.gov/hsip/evaluating.cfm) for performing the calculations. Table 14. Conflict reduction results using the before-after comparison group (C-G) method. Conflict Type Conflict Std. Error Reduction (𝑹) Vehicle/pedestrian 12% 34% Rear-end 31% 4% The result for vehicle/pedestrian conflicts using the comparison group method indicates an estimated 12-percent decrease in vehicle/pedestrian conflicts following the implementation of LPIs when considering the change that occurred at the comparison crossings, but this result is not statistically significant. For rear-end conflicts, a reduction of 31 percent was estimated with a standard error of 4 percent that indicates a highly significant effect. As indicated in the guide, changes in conflicts may be used to infer changes in crashes. This process does, however, require an SPF relating crashes to those conflicts. A search of the literature did not reveal any directly relevant SPFs that could be used for this study. However, two SPFs were found that could at least be used, with due caution, for the demonstration purposes of this study. One (Saleem et al., 2014) did pertain to rear-end conflicts at signalized intersections, but the conflicts were derived from microsimulation. The other (El-Basouyny and Sayed, 2013) used data from video observations, but these pertained to all vehicle conflicts, not specifically to vehicle/pedestrian or rear-end conflicts. Both studies identified conflicts based on TTC thresholds of 1.5 seconds, the same as for the AMAG data. The pertinent SPFs are shown in Equations 5 and 6 below. 88 

Saleem et al.: . . 𝐶𝑟𝑎𝑠ℎ𝑒𝑠⁄𝑦𝑒𝑎𝑟 0.523 𝑃𝑒𝑎𝑘 ℎ𝑜𝑢𝑟 𝑟𝑒𝑎𝑟 𝑒𝑛𝑑 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠 ∗ 𝑃𝑒𝑎𝑘 ℎ𝑜𝑢𝑟 𝑣𝑜𝑙𝑢𝑚𝑒⁄𝐴𝐴𝐷𝑇 5 El-Basouyny and Sayed: . 𝐶𝑟𝑎𝑠ℎ𝑒𝑠⁄𝑦𝑒𝑎𝑟 0.036 𝐴𝑣𝑒𝑟𝑎𝑔𝑒 ℎ𝑜𝑢𝑟𝑙𝑦 𝑑𝑎𝑦𝑡𝑖𝑚𝑒 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠 6 From the Saleem et al. SPF, it follows that the 31-percent reduction in vehicle conflicts indicated in Table 14 would correspond to a CMF of [(100-31)/100]0.307 = 0.89. Similarly, a CMF of 0.76 would be estimated from the El-Basouyny et al. SPF, which is based on all conflicts, not only rear-end ones. It is worth noting that the CMF of 0.89, which is based on a model for all vehicle crashes based on rear- end conflicts, and is therefore more relevant, is closest to the crash-based CMF of 0.87 estimated in the Goughnour et al. (2018) empirical Bayes before-after study. Supplementary Analysis and Results Based on Extreme Rear-End Conflicts This analysis involved developing subsets of “extremely severe conflicts” identified using a machine learning technique to help select a threshold of what is considered extreme from the set of recorded conflicts. Anomaly detection machine learning approaches identify an anomaly as an extreme value data point that does not conform with the rest of the data by learning the inherent data characteristics. The use of extremely severe conflicts logically provides a more realistic surrogate for crashes compared to the universe of conflicts based on some arbitrary threshold (Coles et al., 2001). The approach was applied in this research only for rear-end vehicle conflicts as there were too few vehicle- pedestrian conflicts to achieve meaningful results. The approach allows the consideration of frequency (TTC) indicators as well as severity indicators that are based on conflicting vehicle speeds. A commonly used severity indicator for conflicts is Delta- v or ∆𝑣, which can be defined as the change in vehicle velocity in the event of a collision. As described by Laureshyn et al. (2017), Delta-v can be determined based on a rear-end collision in which the lead and trailing vehicles are perfectly inelastic. The interaction severity in a conflict between two vehicles can be determined by the larger of two Delta-v values. These values were provided in the AMAG dataset. Details of the general approach can be found in Hussain et al. (2022) and Hassanpour et al. (2022). The frequency of conflicts is plotted based on an “anomaly score” that represents severity, which is calculated from the machine learning algorithm. Figure 10 plots this distribution for the AMAG rear- end conflict data. This histogram helps in identifying an anomaly score decision threshold, which is akin to determining a TTC threshold for identifying the universe of conflicts that may be regarded as surrogates for crashes. A sharp decrease in the distribution suggests that the anomaly score decision threshold should be close to that value. The conflicts that exceed the threshold are labeled as extremely severe ones. Accordingly, in Figure 10, the blue line represents the anomaly score, and the red rectangle suggests the potential threshold area. 89 

Figure 10: Histogram of conflict frequency based on anomaly scores. Based on the data for the histogram in Figure 10, the research team extracted extreme conflicts for three anomaly score thresholds (i.e., anomaly score greater than 1.00, 1.25, and 1.50). The higher the anomaly score threshold, the more extreme the selected conflicts are. It should be noted that, while the magnitude of these anomaly score thresholds is numerically close to the TTC/PET threshold of 1.5 seconds used in the AMAG analysis, these are different metrics that cannot be directly related to each other. Conflict reductions were estimated for each anomaly score threshold, again using the methodology documented in Gross (2017) and the accompanying spreadsheet tool. The results are shown in Table 15. This table also shows the results presented earlier for all rear-end conflicts with TTC values < 1.5 seconds for the purpose of comparison. As seen in the table, the estimated conflict reduction gets smaller as the conflict severity increases and as those conflicts logically become more representative of actual crashes. Of note is that the rear-end extreme conflict reduction results with anomaly scores (AS) greater than 1.25 and greater than 1.50 bracket the CMF of 0.87 estimated by Goughnour at al. (2018) for all vehicle crashes. This result, though interesting, is not conclusive since Goughnour et al. did not compute a CMF specific to rear-end crashes, and a direct comparison was not possible. Table 15. Results of supplementary extreme rear-end conflict analysis. Treatment Control Estimated Conflict Anomaly Score Reduction (std. Threshold Before After Before After error) All conflicts 809 772 5,256 7,221 31% (4%) Extreme conflicts with anomaly 172 128 950 841 17% (11%) score > 1.00 Extreme conflicts with anomaly 127 93 847 710 14% (13%) score >1.25 Extreme conflicts with anomaly 87 66 731 598 9% (16%) score > 1.50 90 

Exploration of a Surrogate-Crash Linkage This section builds on the earlier analysis of effects of LPI on surrogate measures by describing an exploratory effort to investigate a surrogate-crash linkage for potentially translating those effects into CMFs. The research team used the AMAG Bellevue data for this exploratory research, as crash data were available for the same intersections used for the analysis based on surrogates. Of necessity, the research team confined this investigation to rear-end crashes (and to corresponding conflicts) and to crashes of all severities due to sample size limitations. Data Ten candidate intersections where rear-end conflicts were recorded in all three AS categories identified in Table 15 for at least two approaches were selected from the list shown in Table 12. A total of 32 approaches at the 10 candidate intersections met these criteria. The research team developed a regression model to relate the number of crashes to conflicts on those approaches after establishing that these two variables are correlated. The 10 intersection sites contained a mix of treatment crossings and comparison crossings. The conflict data used for the analysis pertained to the period before LPI implementation on October 25, 2020. The crash data pertained to the 5-year period before LPI implementation. Table 16 summarizes the conflict and crash data for the six intersections. Table 16. Summary of Bellevue data used for exploring the surrogate-crash linkage. Extreme Rear-End Total Number of All Conflicts/hour Rear-End Intersection ID approaches Rear-End Crashes/ year observed Conflicts/hour AS>1 AS>1.25 AS>1.5 for the approaches Bellevue Way NE at 4 28.70 5.44 4.74 4.06 2 NE 4th St 106th Ave NE at NE 4 3.36 0.99 0.89 0.74 0.2 4th St 108th Ave NE at NE 4 3.99 0.10 0.08 0.04 1.6 4th St 156th Ave NE at NE 4 33.33 20.17 18.33 16.67 3 8th St 112th Ave NE at NE 4 39.00 4.00 3.00 2.33 0.8 4th St 156th Ave NE at 2 26.83 3.50 1.83 1.17 1.6 Northup Way 112th Ave at Main St 2 2.50 0.50 0.50 0.33 0.6 112th Ave NE at NE 3 21.33 0.67 0.50 0.17 0.4 2nd St 156th Ave NE at NE 2 1.33 0.50 0.50 0.33 0 13th St 110th Ave NE at NE 3 37.33 4.83 4.17 2.67 1.8 4th St AS= Anomaly Score. 91 

Analysis and Results Table 17 presents the results of the crash-conflict correlation analysis. As is evident, the correlation is relatively strong, suggesting a) that it may well be feasible to establish a formal statistical relationship for estimating a CMF from such a model in the absence of crash data, and b) that in the absence of such a relationship, inferences may be made on safety effects of LPI on the basis of conflicts alone. Table 17. Correlation coefficients – rear-end crashes/year and conflicts/hr. Regular Extreme Rear-End Conflicts Rear-End Conflicts AS>1 AS>1.25 AS>1.5 0.59 0.80 0.78 0.76 The research team used the data in Table 16 to explore the development of a statistical relationship between crashes and conflicts. Various model forms were explored, including a generalized linear form, as was the possibility of using data for each intersection approach separately. The research team assessed models based mainly on the statistical significance of the coefficient for the conflict variable. In the end, the most promising models were developed using the combined data for all approaches at an intersection and a linear regression model form with an intercept. Table 18 displays the results. Table 18. Linear regression model results for total rear-end crashes. Intercept Slope Conflict Type R-squared (Std. Error) (Std. Error) All rear-end 0.478 0.036 0.351 conflicts (0.432) (0.017) Extreme conflicts 0.686 0.126 0.634 with AS>1 (0.237) (0.034) Extreme conflicts 0.732 0.136 0.609 with AS>1.25 (0.239) (0.038) Extreme conflicts 0.788 0.145 0.583 with AS>1.5 (0.240) (0.043) Discussion The results of this exploratory analysis, though promising, may not be regarded as robust since they are only based on observations at 10 intersections. For this reason, the models depicted in Table 18 are not recommended for use in estimating CMFs. Nevertheless, the analysis provided some useful insights. First, the slope for extreme conflicts is positive, increasing for higher AS thresholds, and highly significant; this suggests that it seems feasible to assemble additional data to develop more robust models that can be used for estimating CMFs in the absence of crash data. In particular, such models can be used for quickly evaluating implemented treatments, especially new and innovative ones that are being piloted. Second, the use of the subset of extreme conflicts, as opposed to all conflicts, materially improves model performance as evidenced by the substantially higher R-squared values, and the increased size and statistical significance of the slope term. 92 

Conclusions The video-derived conflict data safety performance evaluation of LPIs described in this section resulted in conflict reduction estimates for vehicle-pedestrian conflicts and rear-end conflicts, both with surrogate measure values (TTC and PET) less than 1.5 seconds. The conflict analysis included rear-end conflicts because the operational characteristics of LPIs affect driver braking behavior on turning movements, which may affect the frequency of rear-end conflicts. The frequency of other vehicle-vehicle conflict types may be influenced by the introduction of LPIs, but for the purpose of this analysis, only rear-end conflicts could be reliably assigned to either the treatment or comparison group thus enabling the use of the comparison group before-after evaluation method. That method estimated a reduction in vehicle/pedestrian conflicts following implementation of LPIs, though the result was not statistically significant, suggesting that a larger sample size would be required for teasing out a robust estimate. The analysis result for rear-end conflicts did benefit from a larger sample size and resulted in a statistically significant conflict reduction of 31 percent. Based on available crash-conflict safety performance functions, this reduction was seen to be indicative of a CMF that is reasonably consistent with those estimated in an empirical Bayes crash-based before-after study. A supplementary analysis was conducted based on extreme rear-end conflicts in which an autoencoder machine learning approach detects an anomaly as an extreme value data point that does not conform with the rest of the data. This analysis considered conflict severity by using vehicle speeds, in addition to using TTC to assess conflict frequency. The results indicate that reductions in these extreme conflicts, which are logically more representative of actual crashes, are reasonably consistent with that estimated recently in the empirical Bayes crash-based before-after study. The research team built off the previous analyses to explore a potential surrogate-crash linkage using crash data available for the same intersections used in the analysis based on surrogates. The analysis included development of linear regression models for all rear-end conflicts and extreme conflicts. The results showed that extreme conflicts increase for higher AS thresholds and use of a subset of extreme conflicts improved models, as opposed to using all conflicts. While the results are promising, they should not be used to develop CMFs, as the data included only ten intersections. The two sets of analyses did not directly consider exposure data, and thus any changes in demand (including any pedestrian demand induced by the implementation of LPIs) could not be assessed. Although the comparison group method that was applied could mitigate this issue, it should be noted that the treatment and comparison sites are somewhat interconnected since “sites” were crosswalks, and, in some cases, a given intersection featured both treatment and comparison sites. Thus, it is possible that there may be some other unknown effects at the intersection. Furthermore, six of the comparison crossings featured a different traffic signal timing-based pedestrian crossing countermeasure, “Rest in WALK Flashing Yellow Arrow (FYA),” as mentioned previously. However, the purpose of this case study was to demonstrate the method, not to develop a defensible evaluation of LPIs. There are other intersection characteristics which may be important to consider when conducting a full and defensible evaluation of LPI implementation. These include:  Intersection geometry, including pedestrian crossing distance, lane arrangements, the presence of pedestrian refuge islands, curb radii, intersection sight distance, and other characteristics. 93 

 Left-turn traffic signal phasing (i.e., protected, protected/permitted, permitted and/or leading vs. lagging left turns).  LPI duration and pedestrian WALK phase duration.  Vehicle movement speeds. This investigation demonstrated an approach to using video-derived conflict data to assess the safety performance of a countermeasure. Datasets such as the one used in this analysis represent an emerging approach to widespread application of surrogate safety measures in determining the safety performance of a treatment. Case Study 2: Evaluation of Traffic Signal Coordination Signal coordination ranked at the top of the treatment prioritization efforts summarized in Chapter 4. This highlighted an opportunity to explore the use of field-collected data to derive surrogate measures and apply them to estimate potential safety effects of signal coordination. As a result, the team identified high-resolution data from the Automated Traffic Signal Performance Measures (ATSPM) system as a favorable target for a case study. ATSPM systems support data collection and analysis to produce performance measures aimed at enhanced safety, mobility, and efficiency of traffic signal operations, maintenance, management, and design (FHWA, 2022). Intersections that are part of an ATSPM system are typically equipped with advance and stop bar detectors to monitor vehicle volume and presence, as well as data processing units to record timestamped vehicle detection activations and deactivations from each sensor, along with the state of traffic signal indications. Among several locations with ATSPM implementations in the U.S., the Utah DOT (UDOT) has incorporated almost the entirety of its traffic signal network into the system, with well over 1,000 signals to date. In addition, UDOT maintains user interfaces that allow for data extraction, including processed outputs, as well as high-resolution raw datasets. The objectives of the case study focused on exploring potential safety effects of signal coordination using surrogate measures, mainly using two types of movements: 1) through movements, where vehicles are traveling in the same direction (e.g., two northbound vehicles), and 2) left-turn movements, where permitted left-turning maneuvers may experience opposing through traffic as a conflicting movement (e.g., a northbound vehicle turning left and a southbound vehicle traveling through). It is expected these two movements will be associated with the frequency of crashes on the same roadway, mostly rear-end crashes, and left-turning crashes, respectively. The project team defined surrogate measures to leverage the high-resolution data from ATSPM and identify instances that could result in vehicle conflicts for the two movements selected above. Individual sensor activations and deactivations were used for surrogates at the microscopic level to identify closely spaced vehicles, and variations of traffic flow and speed were used for a higher-level surrogate at a more macroscopic scale. Ultimately, the objective of the case study is to illustrate the exploration of surrogate measures to assess safety performance in the absence of crash data, and key elements to consider when defining surrogate measures, evaluating the relationships, and practical issues that could be faced when conducting a surrogate analysis. The following sub-sections describe the data collection process and identification of sites, the definition of surrogate measures, the analysis to answer our case study questions, and a set of recommendations drawn from the analysis. 94 

The proposed case study integrates detailed signal phasing and timing, vehicle sensor activations and deactivations, historical crash records, and geometric characteristics of selected intersections along corridors. A brief introduction of the ATSPM system is provided next, followed by a description of the site selection, data extraction, and the analysis of surrogate measures, and concludes with a summary and recommendations from the analysis. The ATSPM System Standard ATSPM interfaces are used in multiple states to produce metrics to monitor traffic operations, including traffic volumes, signal progression, traffic speed, etc. by recording activations and deactivations of sensors placed on the approach to an intersection and at the stop bar of the intersection. These interfaces use concepts from the Purdue Coordination Diagram (Day et al., 2010), to visualize the temporal relationship between the coordinated phase indications and vehicle arrivals on a cycle-by-cycle basis. Individual changes in phases and vehicle detectors are collected following the Indiana Traffic Signal Hi Resolution Data Logger Enumerations (Sturdevant et al., 2012), produced from a joint transportation research program that involved Indiana DOT, Purdue University, Econolite, PEEK, and Siemens. Utah’s ATSPM interfaces also allow for the retrieval of raw datasets from traffic controllers. These datasets are presented following the Data Logger Enumerations and include timestamped signal phasing and timing, and individual traffic data from sensors at or near an intersection. Raw datasets can typically contain millions of rows in a span of only a few days, so special handling is needed to process the data and extract surrogate measures. UDOT collects ATSPM data through a centralized system based at the State’s Traffic Operations Center (TOC) in the Salt Lake Valley. ATSPM repositories contain data from the majority of traffic controllers in the state, including most of the state’s signalized intersections in the Salt Lake Valley, as shown in Figure 11 below, where the blue circles indicate locations of controllers with active data collection from ATSPM. 95 

Figure 11. Illustration of intersection controllers part of the ATSPM system in the Salt Lake Valley. (Sources: Esri, DeLorme, HERE, USGS, Intermap, iPC, NRCAN, Esri Japan, METI, Esri China [Hong Kong], Esri [Thailand], MapmyIndia, OpenStreetMap contributors, and the GIS user community.) One disadvantage of the ATSPM interfaces is the difficulty in automating the extraction of some of the metrics, particularly when the outputs are only provided in the form of an image. For example, when querying the percent of vehicles arriving in red for a given intersection and time frame, results are displayed for each signal phase using charts and chart foot notes that cannot be captured automatically with a standard script to read text. Even if the images were saved automatically, processing them will likely require a large implementation effort for text recognition, or ultimately manual processing to extract the image processed outputs. Therefore, some metrics that are useful for managing the ATSPM are not easily extracted for research and the project team had to consider the accessibility of data in selecting the performance metrics. Given this practical data access limitation, the team decided to write custom scripts to process the raw high-resolution datasets and compute metrics of interest, including surrogate measures. The project team used open-source code in R Studio, generating not only output files but also visual representations of the signal timings, vehicle arrivals, and potential conflicts. A partial screenshot of the custom interface is shown for a sample approach in Figure 12, where the top portion of the figure shows the northbound through signal indication within the cycle time (green, yellow, and red lines) and the dots represent an individual vehicle activation, while the lower portion shows 30-minute volumes and the percent of vehicles arriving in green. The automated tool creates these figures for both directions of travel along the selected corridor and exports the processed data and summaries for additional analysis. 96 

Figure 12. Sample custom visualization of signal timing and vehicle activations (top) and traffic volume and percent of arrivals in green (bottom) for one intersection and one direction of travel. Site Selection and Data Extraction The team explored potential intersections for use in this study along arterials part of ATSPM in the Salt Lake Valley, where the largest population of the state is located. Data from ATSPM are typically available for multiple years and can be retrieved in near time, for example, allowing for extraction of same-day datasets with only a small delay after being recorded by the controller. In terms of coverage of the transportation network, notable corridors within ATSPM include several state routes that carry a wide range of traffic volumes spanning AADTs of about 20,000 to over 70,000 vehicles per day. Exploration of the data included verification of data availability, as well as a minimum array of sensors along the corridor to allow for detailed analysis of same direction and left-turning movements. More specifically, the team verified that potential sites had the following minimum detection capabilities:  At the stop bar: Lane-by-lane and count detection zones for through and left-turn movements. 97 

 At advance locations: Approach-level presence or count detection zones. The requirement for stop bar presence detection was set to enable leader-follower analysis and to identify potential for conflicts of vehicles in the same direction and in the same lane. Left-turn detection was needed to estimate both counts and vehicle presence, allowing to estimate not only the movement demands but also to identify vehicles departing the zone and starting the turning movement. Similarly, detection at an advance location was used to verify vehicle volumes, and more importantly to provide an indication of arrivals during green when the sensor data were integrated with traffic signal phasing and timing. The final sites selected included five corridors, three of them with four subject intersections each, and the remaining two corridors with five intersections each, for a total sample of 22 intersections. Although the lane configurations varied slightly along all intersections within a given corridor, the selected locations provided adequate detector outputs and enough raw data to perform the analysis. Extracted data were focused on weekdays, with the objective of capturing changes in signal coordination designed to address fluctuations in traffic due to commuting patterns. In addition, hours were limited to daytime hours, avoiding low traffic during nighttime when signal coordination is not expected to be active. This resulted in data analysis periods covering between 7 AM and 6:30 PM, for a total of 11.5 hours per day. Data were further divided into three main daytime periods, with the goal of capturing expected changes in traffic patterns and also with approximate cutoff times for signal timing plans. The three main time periods included a morning peak from 7 AM to 9:30 AM, an off-peak period from 9:30 AM to 4 PM, and an afternoon peak from 4 PM to 6:30 PM. Data selection was limited to days without adverse weather such as rain or snow or any other evidence of crashes or non-recurring events. The extraction was conducted in days between April and June of 2022. Once the datasets were extracted, the team mapped specific traffic controller outputs to the corresponding signal phase and travel direction, and each vehicle detector output channel to a type of zone (stop bar or advance), the type of detection (presence or count), and the associated lane or lanes. Intersection mappings are also part of the ATSPM interfaces made available by UDOT, and this effort was conducted by the team using manual extraction. Table 19 shows the selected corridors, intersections, as well as the hourly vehicular volumes in the morning peak periods, off-peak times, and afternoon peaks for each direction of traffic. Hourly volumes in bold font correspond to the highest hourly volume of the three periods, with each corridor showing specific patters that are repeated for all intersections. For example, all intersections for Corridor 1 along 700E Street have their highest hourly volumes between 7:00 AM and 9:30 AM in the northbound direction, and between 4:00 PM and 6:30 PM in the southbound direction. Similar patterns were observed along corridors on 900S Street and 1400S Street. The remaining two corridors, 4500S and State Street, had the highest volumes for the afternoon in both directions of traffic. The project team found significant discrepancies between counts from advance and stop bar detection zones in the process of extracting hourly volumes, prompting the team to review both sources of information to identify the most appropriate estimates considering nearby locations and daily traffic patterns. That is, the project team expected that the two detection zones would yield similar vehicle counts. The ATSPM documentation suggests using either advance or stop bar sensor outputs, not specifying which is the preferred for volume data. This count discrepancy issue was not encountered systematically across all intersections; rather, it seemed to be isolated to specific instances and the source of the issue was not found. It highlights the importance of redundant data whenever possible in 98 

a case study. A specific example was observed at the intersection of 700E and 500S in Corridor 1, where discrepancies on a single specific day were in the order of more than 50 percent when volumes were measured at advance or stop bar locations. Volumes in Table 19 reflect values after necessary corrections were made due to discrepancies caused by detection errors. Table 19. Lane configurations and volumes at selected intersections and corridors. Number of Lanes Average Hourly Volume Direction Through Morning Afternoon Corridor Intersection Right- Left- Off-Peak of Travel and Peak Peak Turn Turn (9:30am- Shared- (7:00am- (4:00pm- Only Only 4:00pm) Right 9:30am) 6:30pm) North 3 1 1 1,758 1,312 1,396 500s Street South 3 1 1 918 1,365 1,852 North 3 1 1 1,063 903 1,011 600s Street 1 – 700E South 3 1 1 697 1,063 1,380 Street North 3 1 1 900 756 845 800s Street South 3 1 1 811 1,187 1,550 North 3 1 1 1,197 974 1,045 900s Street South 3 1 1 738 1,076 1,412 East 1 1 1 565 683 781 Highland Dr. West 1 1 1 471 504 561 East 1 1 1 607 689 820 2– 1300e Street West 2 1 1 473 544 623 4500S Street East 2 1 1 567 616 757 900e Street West 2 1 1 610 685 797 East 2 1 2 983 985 1,183 700e Street West 2 1 2 793 866 1,024 North 3 1 1 756 933 1,008 5300s Street South 3 1 1 572 911 1,264 North 3 0 1 742 847 916 5900s Street 3 – State South 3 0 1 443 597 678 Street North 3 0 1 941 1050 1,162 6100s Street South 3 0 1 511 1036 1,715 North 3 1 1 844 974 1,059 6400s Street South 3 1 1 415 983 1,286 East 3 1 2 1,277 1,016 1,042 Redwood Rd West 3 1 2 975 1,305 1,758 East 2 1 1 1,135 1,010 989 2200w Street West 2 1 1 705 921 1,286 4– East 2 1 1 1,026 897 938 9000S 2700w Street Street West 2 1 1 772 932 1,085 East 2 1 1 1,189 1,008 1,235 3200w Street West 2 1 1 838 929 1,171 East 2 1 1 1,174 987 1,138 3450w Street West 2 1 1 707 725 965 1300w Street East 3 1 2 1,209 998 1,067 99 

Number of Lanes Average Hourly Volume Direction Through Morning Afternoon Corridor Intersection Right- Left- Off-Peak of Travel and Peak Peak Turn Turn (9:30am- Shared- (7:00am- (4:00pm- Only Only 4:00pm) Right 9:30am) 6:30pm) West 3 1 2 684 953 1,219 East 3 0 1 1,394 1,066 1,175 1600w Street West 3 0 1 677 1,053 1,552 4– East 2 1 2 1,141 811 892 Redwood Rd 10400S West 2 1 2 671 1,060 1,593 Street East 2 1 1 1,146 836 928 2200w Street West 2 1 1 744 975 1,444 East 2 1 1 998 747 845 2700w Street West 2 1 1 472 742 1,087 Surrogate Measures Individual vehicle detection activation and deactivations from ATSPM hold significant promise to develop surrogate measures of safety at microscopic and macroscopic scales. In our context, surrogate measures should serve as reliable indicators of observed safety performance, so general trends in terms of crash frequencies are expected to have an association (direction and some relationship in magnitude) to those obtained from surrogates. In addition, and in line with the project objectives, the surrogate measures and crash frequencies are also analyzed with respect to different degrees of signal coordination, assessing if measurable effects can be detected and if surrogates and crashes point to trends in the same direction. This is an exploratory work with the objective to illustrate potential surrogates, but further in-depth study is recommended to confirm or refine results presented in this chapter. Microscopic Surrogates In our context, a microscopic surrogate could be defined as a measure of the outcomes of disaggregated events from or between individual elements part of the traffic system. In this case study, the elements for the microscopic surrogates are individual vehicles. At intersections equipped with ATSPM, it is possible to track the arrival and departure of individual vehicles using vehicle activation and deactivation times at presence zones, and thus, they can be used to produce estimates of volume, occupancy, and speed for use in microscopic surrogate evaluation. These activations can be extended to assess vehicle interactions when evaluating data from a vehicle following another in the same lane. This following behavior can be considered an example of a leader-follower vehicle pair, which could be used to explore potential same-direction conflict based on their spacing, speed differential, and/or combinations of the two metrics. Similar ideas can be extended to left-turning vehicles when the departure of a vehicle from a left-turn lane detector can indicate a potential conflict if a vehicle in the opposing through movement also departs its detection zone. The analysis is centered on the exploration of different sets of measures that could show potential as valid surrogates. The team used two commonly applied surrogate measures to derive custom definitions for our particular detection configurations: TTC and PET. A standard definition of TTC between a follower (F) and a leading (L) vehicle is shown below in Equation 7: 100 

𝑋 𝑋 𝑙 𝑇𝑇𝐶 , 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑉 𝑉 0 7 𝑉 𝑉 Where: Xi = Position of vehicle i, i={L,F} Vi = Speed of vehicle i, i={L,F} LL = Length of the leading vehicle From the TTC definition, the spacing between vehicles was measured as the time difference between a sensor deactivation and the subsequent activation by a follower, and speed differentials were computed based on individual speed assessments using the duration of the detector call. When the speed of the following vehicle is higher than the speed of a leader, TTC can be calculated based on their spacing and speed differential. This points to shorter spacings and larger speed differentials to be more closely associated to same-direction conflicts, in this case, rear-end conflicts. These definitions pose some challenges. First, in the case study the computation of metrics is limited to a small portion of the approach where the stop bar sensors are located, preventing analysis of vehicle behavior along the approach; second, while the data resolution is adequate, and in the order of 0.1 seconds, the speed differential calculations assume the leader and follower vehicles are of similar length. This limitation could result in inaccuracies but given the urban setting and low large vehicle composition for the 22 intersections in this case study, the expectation is for the metric to generally point to trends in the correct direction. Regarding metrics targeting left-turning movements, the team defined potential conflicts using a concept similar to a PET. Typically, PET can be calculated as the time difference between the encroaching vehicle departure from the conflict area and the arrival of the vehicle with the right of way. Given that vehicle sensors do not provide complete vehicle trajectories, this time was assessed as the time difference between the departure of a vehicle in the left-turn lane and the departure of a through vehicle in the opposing through lane. A number of refinements to these metrics could also be considered. For example, the addition of a small-time extension to the departure of the left-turning vehicle could help account for the length traversed between the stop bar and the point where the turning vehicle starts the maneuver. Also, the time gaps between vehicles traveling in the opposing through lanes could help identify more precisely when potential gaps are available for the turning vehicle to start the maneuver. Calculations to produce the metrics described above were completed using custom scripts applied to the raw ATSPM datasets for each of the time periods and each intersection in the two directions of traffic. It’s important to highlight the tradeoff between the time and effort developing new scripts and manual data processing using already existing interfaces. The two alternatives need to be given careful consideration when making the decision to develop custom solutions or maximize the use of existing tools. Macroscopic Surrogates Additional sets of potential surrogates using aggregate measures were also explored as an alternative to the microscopic measures described above. Instead of focusing on individual vehicle interactions, aggregated measures described patterns observed from a collection of vehicles and could be viewed as macroscopic descriptors of traffic flow characteristics. Objectives behind exploring macroscopic metrics are similar to those for microscopic measures and involve determining if they display 101 

significant associations to safety performance. However, collecting data for macroscopic measures may not require the level of detail and disaggregation that microscopic measures entail, easing data collection requirements and making them more accessible in general. As described above, at least two metrics are readily available from detectors: 1) measurements of detector call durations, serving as a proxy for speed, and 2) measurements of the time spacing between consecutive vehicles. Aggregate measures from these metrics could use central tendency values to describe their expectation and variation. For example, using their mean and standard deviation during different time periods (e.g., peak or off-peak times) and levels of signal coordination. It is important to highlight that potential surrogates should relate to a specific characteristic (or a set of characteristics) of the traffic behavior being analyzed. Let’s first consider macroscopic metrics related to detector call durations, where larger mean values may indicate lower speeds and slower- than-expected vehicle discharge rates. Also, larger standard deviations may point to larger fluctuations in speeds and thus an increased need for drivers to accelerate and decelerate more often, potentially impacting safety. Now let’s consider the second metric (measurements of the time spacing between consecutive vehicles), where short mean vehicle spacing may be indicative of tighter groups of vehicles with less space to maneuver, and smaller variations may confirm that shorter spacing is prevalent among most vehicles in the subject approach. While some of these assumptions may be reasonable, it is important to recall that even if analysis results hold to our expectation, detector call durations and vehicle spacing are only measured at the stop bar. Therefore, they only provide a limited snapshot of the traffic state and their effectiveness to describe behavior along the approach is significantly constrained. The team calculated aggregated metrics for each available scenario, where a scenario can be thought of a unique combination of given time period, direction of traffic, and intersection. For example, to calculate average vehicle spacing for intersection i in the direction of traffic j (e.g., northbound), and for time interval k (e.g., the morning peak period), all spacing values from stop bar detectors in lane l and consecutive vehicle pairs (m, m+1) were included in the mean. Recall that the spacing is measured as the time difference between the deactivation or the zone from a leading vehicle and the subsequent activation from the following vehicle. The same concept was applied to calculate the standard deviations. The expressions below illustrate the aggregations for the spacing metrics and are analogous to those used to estimate the mean and standard deviation of the detector call durations. 1 𝑀𝑒𝑎𝑛𝑆𝑝𝑎𝑐𝑖𝑛𝑔 𝐷𝑒𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛𝑇𝑖𝑚𝑒 ,, , , 𝐴𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛𝑇𝑖𝑚𝑒 ,, , , 8 𝑁 , , , 𝐷𝑒𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛𝑇𝑖𝑚𝑒 ,, , , 𝐴𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛𝑇𝑖𝑚𝑒 ,, , , 𝑀𝑒𝑎𝑛𝑆𝑝𝑎𝑐𝑖𝑛𝑔 𝑆𝑡𝑑 𝐷𝑒𝑣. 𝑜𝑓 𝑆𝑝𝑎𝑐𝑖𝑛𝑔 9 𝑁 , , , Where N is the total number of vehicle pairs found in the dataset. 102 

Crash Datasets In addition to the ATSPM datasets, the team accessed crash data covering the analysis area, including crash narratives and crash diagrams. Data availability included multiple years with both crash-level and vehicle-level details so that crashes could be classified not only based on their geographic location using coordinates, but also given the crash type, and the travel directions and maneuvers of the vehicles involved. For example, a specific crash could involve vehicles traveling northbound in the same direction and resulting in a rear-end crash, or a vehicle involved in a left-turn crash when traveling southbound and colliding with northbound through vehicle. These elements were important in the analysis as the evaluation of safety performance and surrogates considers specific approaches and directions of travel, and constitute a higher level of detail than is typically needed for intersection-level analysis. Crashes were filtered based on the route along the selected corridors, excluding crashes associated to subject intersections but occurring on the crossing streets, and using a standard 250-feet distance from the center of the intersection. In addition, filters were applied to exclude crashes in adverse weather conditions as indicated in the fields describing the roadway surface condition and a weather indicator variable encompassing all types of significant weather in the crash report. A total of four years of data were selected, including data from 2016 through 2019. Finally, filters were also applied to specifically identify crash events that occurred only on weekdays and within the time frames for the specific time intervals defined (i.e., two peak periods and one off- peak period). With the crash data subdivided in such specific “bins” by day of the week, time of day, and specific directional approach, the expectation is to be able to assess safety performance under similar sets of traffic demands and with consistent traffic operations from a signal coordination standpoint. Then, modeling of crashes and conflicts, as well as their relationship will help determining the types of metrics with added potential as surrogates. A significant limitation of such detailed approach is the need for a full-scale research study, where crash data, traffic demands, and signal operations are known from enough locations that crash frequencies can be accurately assessed for each individual bin. Nonetheless, this case study provides an initial exploration and provides insights on the evaluation of surrogates in the context of signal coordination effects on traffic safety. Data Analysis After the initial identification of corridors and intersections, and the subsequent extraction of the ATSPM and crash data for the analysis period, the team began the analysis by exploring traffic demand and signal coordination patterns, providing an empirical basis for the data disaggregation by direction of travel and time periods (i.e., off-peaks and morning/afternoon peaks). Then, potential associations of same-direction and left-turning crash frequencies are explored in relation to the proportion of vehicles arriving at the intersection during the green indication. These associations serve as a reference to explore adequacy of potential surrogates, and finally, a direct relation between crashes and potential surrogates are analyzed. 103 

Traffic and Signal Coordination Patterns Hourly volumes and signal coordination summaries for each intersection have been shown in Table 19 above, but additional details are presented in this section to determine if these average trends were obtained from observations with large variations. Figure 13 shows the daily variation of 30-minute traffic volumes (top) and proportion of vehicles arriving in green (bottom) for the three time periods analyzed along three intersections from the same corridor on 700E Street. In the figure, time period 1 corresponds to the morning peak, time period 2 corresponds to the afternoon peak, and time period 3 to the off-peak hours. As noted from the small range of each box plot, 30-minute volumes (top of Figure 13) were consistent for each time period and each approach. For example, the first approach to the left of the figure, northbound at 600 S Street, showed a very small range of volumes for each of the three time periods. Likewise, the percent of vehicles arriving in green (bottom of Figure 13) displayed consistency for all approaches and time periods, although with slightly larger variations than volume. This data exploration was extended to all intersections along the five selected corridors, and supported the idea of summarizing intersection safety performance and extraction of potential surrogate measures for each direction of traffic and each time interval. 104 

Figure 13. Sample daily variation of traffic volumes (top) and proportion of vehicles arriving in green (bottom) for intersections on 700e corridor. Effects of Signal Coordination on Same-Direction and Left-Turn Crash Frequencies Using detailed crash records, the team first explored the relationships between crash frequencies, traffic volumes, and signal coordination. An NB model was developed with the crash frequency as a dependent variable, and the average hourly volume, the percent of vehicle arrivals during green, and if the time period corresponded to peak or off-peak hours as independent variables. Results of the models 105 

for same-direction and left-turn crashes are shown in Table 20, where the percent of vehicles arriving in green are shaded in grey. Table 20. Crash frequency models including signal coordination metric. Crash Independent Variables Estimate Std. Z Pr(>|z|) Theta Std. Type Error Value Error (Intercept) -2.983 1.923 -1.551 0.121 Same- Log (Hourly volume) 0.769 0.280 2.748 0.006 ** 2.328 0.814 Direction Percent green arrivals -2.680 0.448 -5.990 >0.001 ** Time period -0.334 0.183 -1.824 0.068 * (Intercept) -33.42 11.20 -2.984 0.003 ** Log (Hourly LT Volume) 3.10 1.048 2.958 0.003 ** Left- Log (Hourly Opposing Through) 3.178 1.104 2.877 0.004 ** 5.77 6.28 Turn (Hourly volume*Opposing -2.86e-5 1.03e-5 -2.769 0.006** Through) Percent green arrivals 0.323 0.874 0.370 0.7115 * Significant at 90% confidence level; ** Significant at 95% confidence level From the same-direction crash model results, a significant effect was identified for the average hourly volume, the average percent of vehicles arriving during green, and the time period. The positive sign of the coefficients indicates that crash frequencies increase as volume increases, as expected. For the percent green arrivals, the negative sign of the coefficient indicates that as signal coordination improves, increasing the vehicles arriving in green, the crash frequency was observed to be lower. Results from the left-turn model also show that as left-turning volumes and opposing through volumes increase, crashes increase. However, no significant association was found between the percent of vehicles arriving in green and left-turn crash frequencies. After establishing crash-based outcomes using the models, similar interactions between covariates and conflict frequencies are now expected to be obtained from successful surrogates. The next sections explore the microscopic and macroscopic metrics in light of the crash-based modeling results. Modeling Microscopic Surrogate Measures, Traffic Volume, and Signal Coordination A similar modeling exercise was conducted to investigate associations between same-direction conflicts from potential surrogates and traffic demand, as well as the percent of vehicles arriving in green. For left-turning movements, the opposing through demand was also incorporated into the model exploration. The same-direction conflicts are explored first, followed by the left-turn conflicts. Same-Direction and Left-Turn Surrogate Measures As described above, metrics related to TTC were used to identify conflicts between vehicles traveling in the same direction. The team explored a number of variations with potential to serve as surrogate measures, including different thresholds for the following:  Spacing between vehicles.  Speed differential between vehicles.  TTC measured as the ratio of spacing and speed differential. 106 

In addition, conflicts could also be defined based on the signal indication status, whether they occurred during the green phase or during the yellow and all-red transition. This gives an additional dimension to explore when identifying potential surrogates. Metrics showed inconsistent trends in terms of the effects of traffic volumes and also the effect of vehicles arriving in green, pointing to a potential revaluation of the type of conflicts being captured at the stop bar zones. Out of the surrogates described above, using combinations of spacing, speed, and TTC, a surrogate based on short vehicle spacing events during the yellow and red transitions showed the most similar associations to those observed from the crash-based models. However, models did not display the same level of significance for all associations between potential surrogates and covariates, indicating only a partial success of the metrics. Table 21. Conflict models for same-direction metrics including signal coordination. Std. Std. Metric Independent Variables Estimate Z Value Pr(>|z|) Theta Error Error Same- (Intercept) -5.654 2.6189 -2.159 0.031* Direction Log (Hourly LT Volume) 0.926 0.382 2.424 0.015* Conflicts in Percent green arrivals -0.2646 0.6482 -0.408 0.683 0.991 0.232 Yellow or Red – Time period -0.5088 0.255 -1.995 0.046* Gap<0.5s (Intercept) 2.35 4.30 0.547 0.58457 Log (Hourly LT Volume) 1.16 0.383 3.014 0.00258* Left-turn Log (Hourly Opposing Conflicts -0.499 0.443 -1.126 0.26019 Through) 2.12 0.34 - Gap<2.5s (Hourly volume*Opposing 2.71E-07 3.09E-06 0.088 0.93004 Through) Percent green arrivals -0.375 0.596 -0.63 0.52893 * Statistically significant at a 95 percent confidence level. Regarding the same-direction conflicts, Table 21 indicates that hourly volumes and the percent of vehicle arrivals in green point to the same direction as those observed for the crash model. A significant effect of volume was observed in the positive direction, with higher volumes producing higher conflicts, but the negative relation between the percent of vehicle arrivals in green and conflicts showed low significance. This points to potential limitations of the surrogate to reproduce the same relations displayed for crashes, perhaps due to issues associated to limited spatial data collection. Similarly, the associations between left-turn surrogates and volume and vehicles arriving in green were not strong to show significant effects, but similar outcomes indicating no significance of signal coordination on left-turn safety were also observed from crash-based models. Macroscopic Surrogates Given the significance of vehicles arriving in green on the models for same-direction crashes, the exploration to evaluate macroscopic measures was focused on this crash type. It is important to recall that macroscopic measures do not represent frequencies of individual events, but aggregations that could point to potential trends with significant effects on safety. As described above, the team defined metrics based on the mean and standard deviation of two microscopic measures: 1) vehicle spacing and 2) detector call durations (as a proxy for speed). 107 

First, the correlation of these potential surrogates with other variables of interest were explored, as shown in the correlation matrix in Figure 14. The figure includes crash frequencies, the percent of vehicles arriving in green, and hourly traffic volumes in addition to the potential surrogates. Numeric values in Figure 14 show the correlation coefficients between variables, and the shades of blue or red applied to each cell indicate their proximity to a perfect positive or negative correlation (i.e., a value of 1 or -1), respectively. The data indicate that the mean and standard deviation of vehicle spacing have a negative association with crash frequency, with shorter spacing and smaller variations in spacing being indicative of higher crash frequencies. On the other hand, positive correlations with crash frequencies were observed with the mean and standard deviation of the detector call durations. So, longer call durations, and therefore lower speeds, and higher speed fluctuations were correlated to an increase in crash frequencies. In addition, it is not surprising to observe a negative correlation between call durations and vehicle spacing. The shorter the call durations, the higher the vehicle speeds, resulting in higher vehicle spacing. Conversely, longer average call durations are expected to lead to shorter vehicle spacing. Figure 14. Correlation of dependent and independent variables including potential surrogate measures. In addition, exploration of the potential surrogates was conducted using an NB model, with the objective of establishing the magnitude and significance of the effect of the covariates on crash frequency. Table 22 shows the model outcomes for each of the potential surrogates, when modeled together with hourly volumes and the percent of vehicles arriving in green. As expected, the direction and significance of the effects of hourly volumes and the percent of vehicles arriving in green remained unchanged across all models. In terms of the potential surrogates, the highest significance out of the four models was observed for the standard deviation of the call duration (or the standard deviation of speed, shaded in grey in the table). However, none of the other potential surrogates was significant at a 90-percent confidence level or higher. 108 

Table 22. Model Outcomes for Potential Surrogates Std. Std. Metric Independent Variables Estimate Z Value Pr(>|z|) Theta Error Error (Intercept) -4.5097 2.6411 -1.708 0.088* Average Detector Log(Hourly LT Volume) 0.7790 0.4193 1.858 0.063* 2.69 1.04 Call Percent green arrivals -2.9687 0.5792 -5.125 >0.001** Duration Average Call Duration 0.6088 0.6213 0.980 0.3272 Standard (Intercept) -4.3051 2.6237 -1.641 0.100* Deviation Log(Hourly LT Volume) 0.6199 0.4122 1.504 0.133 Detector Percent green arrivals -2.5953 0.6325 -4.103 >0.001* 2.87 1.15 Call Standard Deviation Call Duration 2.3590 1.3488 1.749 0.080* Duration (Intercept) -3.4308 3.5297 -0.972 0.3311 Average Log(Hourly LT Volume) 0.8843 0.4446 1.989 0.047** Vehicle 2.64 1.01 Spacing Percent green arrivals -3.077 0.6228 -4.940 >0.001** Average Spacing -0.170 0.5704 -0.298 0.7655 (Intercept) -2.173 3.8929 -0.518 0.6043 Standard Deviation Log(Hourly LT Volume) 0.7766 0.4449 1.746 0.0809 2.69 1.04 Vehicle Percent green arrivals -2.9908 0.5959 -5.019 >0.001** Spacing Standard Deviation Spacing -0.779 1.084 -0.719 0.4723 * Significant at 90% confidence level; ** Significant at 95% confidence level Modeling results point to the standard deviation of the call durations (i.e., vehicle speed) as a potential measure for surrogate safety analysis, where a positive correlation between the metric and crash frequency was identified. Summary and Discussion The exploration presented in this section describes the data collection, identification, and evaluation of potential metrics as surrogate measures of safety at microscopic and macroscopic levels, with a particular focus on potential effects of traffic signal coordination. High resolution detector and signal phasing data were collected from the ATSPM in Utah, including 5 corridors and a total of 22 intersections. Crash-based models were developed for same-direction crashes (i.e., rear-end crashes) and left-turning crashes between turning and opposing through vehicle. The models identified significant associations of signal coordination with same-direction crashes, where the signal coordination was assessed in terms of the percent of vehicles arriving during the green phase. However, no significant effect was found for left-turn crashes. Then, a number of measures at microscopic and macroscopic metrics were defined and evaluated as potential surrogate measures. At the microscopic level, vehicle conflicts were defined using concepts similar to those from a traditional TTC definition. The team evaluated metrics with different thresholds and occurring at different phases within the traffic signal cycle, and each metric was used to produce a conflict prediction model. For same-direction conflicts, none of models produced the same strong associations observed in the crash models, where the percent of vehicles arriving during green was 109 

identified as a significant contributing factor. The metric with the highest potential as a surrogate for same-direction conflicts was based on short gaps between vehicles (<0.5 seconds) when a follower was reducing the spacing to the leader and this interaction occurred during the yellow and red phase transitions. Similarly, models using metrics for left-turning conflicts showed no significant associations with percent of vehicles arriving during green, but in this case such outcomes resembled the same findings from crash-based models. These results led to a macroscopic level analysis pursuing further surrogates only for same-direction movements. Macroscopic candidates for surrogates were defined as aggregates of microscopic metrics in terms of mean and average durations of detector calls (as a proxy for speed) and vehicle spacing. Crash- based prediction models evaluated the macroscopic metrics as covariates along with traffic volumes and the percent of vehicles arriving in green. Models including the potential surrogates as covariates for crash models pointed to the standard deviation of the call durations as a significant predictor of crash frequencies. From the models, higher standard deviation of vehicle call durations was associated with higher crash frequencies. In addition, a relatively high correlation in the negative direction was identified between the percent of vehicles arriving during green and the standard deviation of the call duration, providing additional support for this metric as potential surrogate. Additional validation of the results provided in this exploration is needed to verify the robustness of the two potential surrogates derived from ATSPM. At the microscopic level, the frequency of conflicts during the yellow and red indication between vehicles traveling at shorter spacing (i.e., lower than 0.5 seconds) and with a measurable TTC (i.e., when the speed of the follower is higher than the speed of the leader); and at the macroscopic, level the variation in the detector call duration (in this case measured as the standard deviation). Overall, improved signal coordination was associated with potential safety benefits in relation to same-direction crashes from both crash-based and surrogate-based models, but no significant association in a particular direction was obtained for left-turn crashes. This exploration also highlighted challenges to identify surrogates from field data, including identification of potential data quality issues, and limitations from the ATSPM sensor data (i.e., data are collected only at the stop bar). However, the case study identified one microscopic and one macroscopic metric to have potential as surrogates, and are worthy of future exploration to identify their robustness as a safety indicator, and as metrics to assess the effects of signal coordination along arterials. 110