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.
57 Background and Motivation To help realize its Vision Zero goal, Seattle DOT developed a series of pedestrian SPFs to help establish a more comprehensive approach for reducing pedestrian crashes throughout the city. Prior to the development of these SPFs, Seattle DOT employed a series of traditional crash frequency-based approaches to identify hot spot locations experiencing a high number of pedes- trian crashes. The efficacy of these approaches is, however, often limited by the failure to account for âregression to the mean,â in which crash frequency at high crash locations would be expected to decrease toward the average frequency for similar sites. These traditional crash frequency- based approaches also are not based on factors contributing to pedestrian crashes across the network; as a result, they may fail to identify the underlying factors that contributed to increases in pedestrian crashes. Therefore, Seattle DOT set out to create a series of SPFs that included parameters that could help identify locations with potential for systemic pedestrian and bicycle safety improvements. Step 1: Define Study Scope Intersection crashes were selected as the initial focus for the first round of systemic analyses. Intersections of all types accounted for 70% of total crashes (Figure 6) and 59% of fatal and serious injury pedestrian crashes in the city (severity data not shown in the figure). Crashes in which the motor vehicle was going straight through the intersection (i.e., no turns) had a higher severity percentage, but crashes involving turning vehicles occurred more often (City of Seattle Bicycle and Pedestrian Safety Analysis 2016). However, because this was a new effort, city staff C H A P T E R 9 Case Example 1: Seattle Department of Transportation Key Takeaways â¢ Used networkwide data for all intersections to develop pedestrian SPFs for different crash types. â¢ Applied crash predictions from model estimation to identify high risk sites. â¢ Conducted additional field investigations to identify and supplement missing data types. â¢ Planned LPIs at several intersections to address risks identified; crash predictions were used to help prioritize sites without prior crashes.
58 Systemic Pedestrian Safety Analysis and analysts decided to analyze all types of pedestrian crashes at intersections in lieu of left- turning-only crashes, as well as the more severe subset of crashes involving motorists traveling straight through. Step 2: Compile Data To develop its SPFs, Seattle DOT relied on a comprehensive relational crash and roadway database that incorporates spatial linkages between crashes and roadway segments, or inter- sections. Much of the crash and roadway data (e.g., lane types, lane widths, or traffic control type) were data Seattle DOT and many DOTs already collect and maintain. Building on the spatially referenced roadway inventory data, an intersection database for the entire network was compiled that joined crash counts per intersection with spatial data from transit agencies, land use, and census sources that could be useful in accounting for pedestrian activity-based expo- sure and other pedestrian crash risks. Table 19 provides a summary of the data sources used to develop the variables to use in analyzing the crash relationships through modeling. Figure 6. Pedestrian crash distributions by location type and crash type, based on data from 2008â2014 (Adapted from Figure 12 in City of Seattle Bicycle and Pedestrian Safety Analysis 2016). Data Source Comprehensive crash database City of Seattle Roadway network geodatabase Generalized land use Building footprints University locations (volume estimate model only) Schools Short-term, quarterly, and continuous user count data used in pedestrian volume estimation Census blocks and demographic/employment data Census Bureau National elevation data set U.S. Geological Survey Transit stop location and schedule data Sound Transit: Google Transit Feed Specification Table 19. Data and corresponding sources for variables tested in pedestrian crash risk SPFs.
Case Example 1: Seattle Department of Transportation 59 Step 3: Determine Risk Factors Limited pedestrian count data were available prior to the analysis but short-term counts were available for about 50 intersections across the city. Ballpark pedestrian and bicycle volume esti- mates were developed by modeling the count data and associated location characteristics. The mode was then used to generate estimates for all intersections, which were then parsed to seg- ments as well. The estimates were annualized using factors derived from a prior study in a city with similar amounts of walking and developed infrastructure (San Francisco). These proce- dures are explained in Sanders et al. 2017. The model equations, based on short-term counts at 50 locations in Seattle, used variables such as nearby population density, employment density, numbers of households, and others. The variables that were included in the best-fitting pedes- trian volume model included number of households within 0.25 mi of an intersection/10,000; number of commercial properties within 0.25 mi of an intersection; and whether the inter- section was within 0.25 mi of a university. The pseudo-R2 (a measure of the amount of variation in the count data accounted for by the model) was 0.76 (Sanders et al. 2017). Because the models were developed based on a limited number of short-term counts primarily from arterial loca- tions and there was a need to make other assumptions, it was important to also consider other measures of pedestrian activity to help account for potential biases in these estimates. Negative binomial regression modeling was used to test the relationships of variables to crashes for two intersection crash types: all pedestrian crashes at intersections and crashes involving crossing pedestrians struck by a through motor vehicle at an intersection. Variables that were significant in the final model, or safety performance function, were used as the basis for a GIS-based tool that would allow Seattle DOT to conduct screenings using various crash type predictions and ranking methods to infer future crash potential. The tool also allows filtering by various location characteristics that were correlated with crashes to help prioritize locations for further assessment. Variables that were positively associated with increasing pedestrian crashes at intersections included pedestrian volume (although a threshold volume was estimated when the relationship changed direction). Roadway-related variables included the presence of traffic signals, higher arterial classes, the numbers of entering legs, numbers of lanes at the intersection, and the pres- ence of parking. Some of these variables were likely correlated with traffic volume, which was not available for much of the network and so was not used in the analysis. Prior analyses had sug- gested that arterial class served as a fair proxy for traffic volume, so arterial class was included in the analyses. A recent modeling effort to estimate traffic volumes at counted locations across the network found that arterial class was in fact highly correlated with traffic volumes. At the time this model was interpreted, it was thought that the number of legs and numbers of lanes as well as traffic signals likely also correlated with volume-related risk, so the relationship of these crash predictors was interpreted cautiously (Thomas et al. 2017). In addition, increasing transit/buses stopping nearby were positively associated with crashes, as were commercial land uses. Mean income of area residents had a negative correlation with pedestrian crashes. Step 4: Identify Potential Treatment Sites Several SPF and empirical Bayes ranking metrics generated from the model results were avail- able to help identify high priority locations for potential treatment. While the SPF-predicted crashes come directly from the model equation and coefficients, the empirical Bayes estimate comes from a weighted blend of the SPF prediction and prior observed crashes. Given the uncer- tainties around the estimates of pedestrian volume and the lack of traffic volume data for this first analysis, the City is putting more weight on the empirical Bayes estimates when prioritizing
60 Systemic Pedestrian Safety Analysis sites, since this estimate helps to account for missing variables. These estimates are described in Step 6 in the main text. The data and ranking metrics were turned into a map-based tool that also allows Seattle DOT to visualize highly ranked locations and compare the outputs from various ranking methods, including prior crash frequency ranking. See the map in Figure 7, which is for illustration purposes only. Safety needs for specific locations have not been verified. These predictive outputs, along with all the variables, can also be exported to a spreadsheet tool to be used in screening and ranking sites for potential improvement. Some of the significant roadway factors from the model could also be used for initial screen- ing. The top 20 intersections, for example, are overwhelmingly signalized, 4 leg or 5+ leg inter- sections, and involve major arterials spanning at least four lanes across the largest leg. But because of limitations of the data, further field diagnoses were carried out at highly ranked sites to diag- nose the potential conflicts and risks. Traffic volume and unrestricted turning movements were identified as risk factors for many of the signalized intersections. In addition, traffic signals are also typically placed to control movements where conflicts and volume are high, so these con- siderations suggest that signals are not necessarily the source of increased risk. Nevertheless, the factor is useful to identify locations that may be at increased risk of pedestrian crashes. Other Steps and Lessons Learned to Date The model-derived estimates (SPF and empirical Bayes) for potential for crashes have also been used to help justify multi-modal treatments. For example, by considering estimated bicycle crashes (from similar bicycle models) at one intersection where a bicycle path entered, the City Figure 7. Illustration of highly ranked intersections from SPF-based calculations compared with prior observed high frequency crash sites (PTot = total pedestrian crashes at intersections) (Thomas et al. 2017).
Case Example 1: Seattle Department of Transportation 61 was able to justify implementing turning restrictions at an intersection that did not quite meet the Cityâs warrants for this improvement based on auto crashes alone. Seattle is working on various data improvements, including improvements to both traffic and pedestrian volume data, and hopes to improve other potential risk measures, including signal operations. These improvements should allow for improved identification of treatable risk fac- tors in future analyses. Although current data limitations created some challenges in interpreting how some of the roadway features contributed to risk, the supplemental census and land use data elements contributed to the estimation of activity-based risk. Results were also generally intuitive and useful for diagnosing potential solutions, and the model results have been used to rank and identify sites that may have potential for future crashes, even if those sites have not experienced prior crashes. In the interim, engineers visited highly ranked sites to identify miss- ing data types and found a common risk factor at signalized intersections that involved turn- ing traffic conflicts with pedestrians. Thus, several similar locations were identified for leading pedestrian intervals. The City is pursuing a mix of local and state highway safety improvement program funds to help implement LPIs at around 150 intersections, about 50 per year over the next few years. The City has used observed crashes to prioritize about 35 locations and, with the stateâs approval, is using the crash prediction estimates (empirical Bayes) derived from the SPFs to justify about 115 others. In this case, the City used empirical Bayes estimates as the primary ranking metric, which relative to the SPF predictions helps to better account for missing variables (such as the missing traffic volume data). Nevertheless, many locations with no prior crashes in the past 8 years were identified on the basis of empirical Bayes estimates. Based on casual observations, staff have observed crashes occurring at predicted crash loca- tions since the period covered by the analysis. City staff also mentioned that while the initial data compilation was time-consuming, the cost was not that high, and the resulting knowledge and tools were well worth the investment. Interns, for example, coded the initial 8 years of crash locations. The City plans to update the SPF predictions with new data using the current models within 3 years and to develop new models in approximately 5 years. As data improvements are made, including networkwide traffic estimates, the relationships identified from future models may be better able to isolate treatable factors that have been associated with prior crashes. In addition, the SPF and related empirical Bayesâbased ranking methods are useful in prioritizing the locations that may be most likely to experience future crashes, as well as to weed out those with low potential for pedestrian crashes. For more information, see the 2016 City of Seattle Bicycle and Pedestrian Safety Analysis, available at https://www.seattle.gov/Documents/Departments/SeattleBicycleAdvisoryBoard/ presentations/BPSA_Draft_Public_093016.pdf and Thomas et al. (2017).