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.
25 The purpose of this step is to identify treatable crash risk factors that can be used to identify locations across the network for potential systemic treatments. Using the database developed in Step 2, the next step is to analyze the data to identify factors associated with the focus pedes- trian crash types. A sound risk analysis is also needed to help agencies to reliably estimate future crashes, select appropriate countermeasures, and perform economic analyses on treatment options. This will support a data-driven process that will maximize the expected safety benefits from investment in a systemic program. As this chapter will show, there will be an opportunity to compare risk identification approaches that take into account activities in later steps in the systemic process such as loca- tion prioritization. Readers are encouraged to keep in mind basic risk concepts as they work through the various stages of the systemic process. The following are roadway conditions that are generally perceived (and established in the literature) to increase risk of a pedestrian crash: â¢ High volumes of vehicles, but infrequent interaction with pedestrians, which may lead to lower driver expectancy (such as a rare pedestrian crossing on a high-volume road in a sub- urban area) â¢ High volumes of pedestrians â¢ Long time or a long distance in which pedestrians are exposed to oncoming traffic (such as when crossing multiple lanes) â¢ Conflict points in roadway design and operations (such as when a vehicle crosses a sidewalk at a driveway or crosswalk at an intersection to make a turn) â¢ Lack of separation (in space and/or time) between pedestrian and motor vehicle paths (such as when a pedestrian walk signal is concurrent with motor vehicle permissive left turn signal) â¢ Higher speed traffic (particularly roads with speed limits posted 30 mph or higher) on roads with significant pedestrian activity (such as near bus stops) â¢ Dark or sparsely lit roads or inconspicuous crossing locations â¢ Long distances (e.g., block lengths) or wait times (due to signal timing) between roadway crossing opportunities (which may lead pedestrians to misuse or ignore available facilities) Select Approach to Determine Risk Factors Prior crash histories, unadjusted for the normal expected crashes for similar facilities, have not always proved to be reliable indicators of what will happen in the future, particularly for pedestrian crashes or other crash types that tend to be widely dispersed. This is a key reason a systemic approach is necessary, that is, to be more proactive about identifying where crashes may occur without waiting for them to happen in high numbers. If crash histories are not very reliable for predicting the future, how does an agency prioritize locations for treatment that have not necessarily experienced any prior crashes? The HSM established procedures to help address C H A P T E R 4 Step 3: Determine Risk Factors
26 Systemic Pedestrian Safety Analysis these challenges and to produce reliable estimates of crash potential that could be used to pri- oritize locations for treatment. In a nutshell, the HSM recommends the development of model equations known as safety performance functions (SPFs; a definition is provided in the Glossary and in Step 2). These mathematical models estimate relationships of risk-related variables (or crash predictors) to crash frequencies and can be used to produce estimates of potential crashes based on the modeled relationships. These estimates have been found to be more reliable than observed prior crash frequencies alone (because of the tendency of crashes to move around) by accounting for traffic volume and other potential measures of crash exposure that vary across the network. The crash estimation values that can be derived from predictive models will come into play further in Steps 4 and 6 but are important to consider for a systemic pedestrian safety process, or any safety program, that aims to cost-effectively apply resources where they may do the most good. Without good estimates of the potential contribution of each risk factor and the overall crash potential posed by a variety of risks (including those that are associated with crashes but cannot necessarily be treated), it can be challenging to determine which sites are more important to treat and to select cost-effective treatment scenarios for a systemic program. Ideally, the two objectives of identifying treatable risk factors and prioritizing locations for systemic treatment can be accomplished by adapting the HSM procedures and developing safety performance functions, also known as crash count prediction models. If SPFs cannot be devel- oped for the network, there are lists of high risk factors shown later in the guidebook (Tables 7 and 8) that may be considered for use in performing other types of analyses, or for screening the network for locations of potential higher risk. Many states are familiar with developing and using SPFs for road safety decision making, especially regarding motor vehicleâonly crash types. Traffic volume is essential for SPF develop- ment. SPFs have been less widely developed and used for pedestrian safety. Traffic and pedes- trian volumes are highly desirable for developing pedestrian SPFs. Until recently, pedestrian volume data has tended to be less available. Other roadway and site characteristics that are use- ful for identifying pedestrian crash factors and locations can be included in the analysis, ideally allowing for identification of important treatable risk factors while also aiding the prioritization and evaluation processes. This guidebook recommends adapting the HSM method of estimating SPFs for use in a sys- temic pedestrian safety process. Important data types for potential inclusion in the analysis data- base have already been described in Step 2. The following section provides more information on the benefits and challenges of other major methods for determining risks. In reality, the method agencies use may be some combination of these different basic approaches. Definition: Risk Risk is the probability of a crash between a pedestrian and a motor vehicle at a specific location within a defined period. While true risks are rarely known, the traffic engineering field creates estimates of risk by identifying attributes of locations on a roadway network that are associated with crash frequencies or severities. See the definitions for crash predictors, their relationship to risk factors, and how crash predictors are used in SPFs in Step 2.
Step 3: Determine Risk Factors 27 Risk Determination from Crash Count Models of Jurisdictional Data: SPF Method Developing SPFs by use of negative binomial regression modeling has been widely used and tested and is a defensible method to model the relationship of variables to crash frequencies. Other analysis approaches are also available to perform predictive modeling of crash frequen- cies. Some analysts have, for example, performed regression tree modeling (using methods such as random forest or conditional random forest) to identify risk factors. These analysis methods provide an alternate to negative binomial modeling for systemic process as they can be used to identify factors that have an independent risk relationship to crashes, but there may still be a need for baseline SPFs or other weighting methods to use in prioritization. Regardless of the specific analysis approach selected, it is necessary to have traffic and pedes- trian volume data, or valid surrogate measures, to perform these types of analyses. Exposure measures, surrogate measures, and caveats were described in Step 2. The roadway descriptors of interest to pedestrian safety are also neededâin particular, factors that may play a role in crash occurrence and potential treatment for the focus location and crash types. Finally, land use, transit measures (if transit is present in the jurisdiction), and other measures of the built and social environment (typically available in census, land use, and business type databases) have been found to be important predictors of pedestrian crashes in several prior studies, even though they may in part be correlated with pedestrian and traffic volumes. These measures, as mentioned in Step 2, should also be considered for the analysis whether or not volume data are available. Troubleshooting Another option to developing SPFs is also available. Agencies may con- sider calibrating a model developed with data from other jurisdictions for use in assessing risks, such as those included in the first edition of the HSM (for signalized intersections only) (AASHTO 2010). Calibration of models with a limited set of predictive factors could be used primar- ily for crash estimation purposes for use during prioritization, similar to an approach used by Oregon DOT (Case Example 2). If such models are used for estimating crash potential, then one of the methods in this step will still be needed to identify risk factors. Consult the addi- tional resources at the end of the chapter for more information on SPF calibration. Appendix B of the technical report references additional pedestrian studies that produced SPFs for different locations and crash types. A model from a similar jurisdiction could also be considered for calibration, but it would be necessary to develop the same data elements. Table 6 compares three different basic approaches to determining pedestrian crash risk fac- tors, which can then be used to identify sites for potential safety improvement needs. The first approach is the one just described, that is, to develop SPFs by modeling crash counts using networkwide data and a meaningful set of traffic, roadway, land use, and other characteristics
Table 6. Comparison of methods for determining risks to use in a systemic pedestrian safety process. Strengths Limitations Count Models (SPFs) â¢ Uses network data. â¢ Provides estimates that can be used to determine high- potential crash locations (as well as higher risk locations) specific to the jurisdiction. â¢ Identifies risks while controlling for other important factors such as traffic and pedestrian volume. â¢ Data determine risks based on crash prediction. â¢ Provides âweightsâ of variable importance within model. â¢ Provides ability to estimate crashes for prioritization, economic analysis, and treatment evaluation. â¢ Requires effort during Step 2 to compile or estimate pedestrian volume data from different sources (roadway, crash, and other). Otherwise, data needs are similar to other methods. â¢ Requires more modeling expertise than other methods. â¢ May provide misleading identification of risk factors or a biased list of sites if important variables are missing from the data and modeling. â¢ Does not require local crash data matched to locations. â¢ Uses local roadway characteristics for screening. â¢ May be simple to perform initially. â¢ Does not require initial use of pedestrian volume data. â¢ Smaller jurisdictions could assess risks through road safety audits. â¢ Assumes risk factors are similar to those from other studies or jurisdictions. â¢ Requires local knowledge and expertise to determine risk factors. â¢ Still requires compiling relevant data types to screen the network for risks. â¢ May require more effort at later steps to compile additional data (to account for pedestrian demand/exposure) to prioritize zero-frequency crash locations (Step 6), if these measures are not included in the initial risk screening. â¢ May require judgment to apply weighting factors for prioritization. â¢ Does not produce crash estimates for project evaluation or economic analysis. â¢ Does not produce SPFs that can be used to evaluate treatments. Frequency-Based Method â¢ Uses network data. â¢ May seem more intuitive to apply. â¢ May make a priori determinations of crash types and roadway factors that are treatable for use in identifying systemic issues. â¢ Expert judgment needed to make determinations of conditions relevant for countermeasures application (e.g., traffic volume and speed). â¢ Is not built on analysis of risk factors that may contribute to crashes across the network while controlling for other factors such as traffic volume. â¢ May not account for regression-to-the mean/random effects. â¢ Disaggregation may obscure risks for pedestrians, especially if based on vehicle concerns. â¢ May identify sites having features correlated with high traffic and high pedestrian volumes but potentially miss other locations with elevated risk. â¢ May require more effort at later steps to compile additional data (to account for pedestrian demand/exposure) to prioritize zero-frequency crash locations (Step 6), if these measures are not included in the initial risk screening. â¢ Does not produce crash estimates to evaluate projects (economic analysis) or treatments. Research/Local Judgment
Step 3: Determine Risk Factors 29 to determine risks. The other two methods are determining risk factors from a combination of prior research and local knowledge and using systemwide crash data to identify locations in the network where target crash types have occurred and the prevalent characteristics of those loca- tions. These methods are described in depth. Risk Determination Based on Prior Research and Expert Knowledge It is most desirable to model crash relationships using roadway, land use, and population data for the network under consideration since risk relationships and crash factors may vary by jurisdiction. Many land use and roadway factors, while intercorrelated, also vary considerably across different urban forms, geographies, and roadway networks and thus it can be difficult to isolate factors that contribute to increased risk. If modeling network crashes is not an option, an agency may consider determining risk fac- tors from a combination of prior research and local knowledge. An important caveat with this approach is to exercise local judgment and expertise to be sure to consider other factors that have been previously identified or appear to be associated with pedestrian crash types within the network. A potential challenge in using this approach is to determine how to weight different risk characteristics, either in site identification or during prioritization steps. See Table 6 for these and other limitations. Looking Ahead Additional work may be needed for other steps in the process if SPFs are not developed to identify risk factors. There may be a need to consider additional data on the built environment, populations, or land uses, to help determine which locations may warrant treatment. Subjective weighting factors may also need to be developed to help with prioritization. See Case Example 3 for an example of how Arizona dealt with these issues. Oregon DOT used a blended approach in their initial systemic effort. They identified crash risks from a mix of expert judgment and analysis and then developed SPFs based on crash and pedestrian volumes to help prioritize locations based on SPF-predicted crashes. See Case Example 2 for this approach. Risks from Prior Research for Consideration If despite the challenges using a set of predetermined risk factors is the best option, Table 7 provides a summary of factors that have been found (at the time of this publication) to have consistent relationships in the expected direction to crashes. These factors might be considered, among other locally determined factors, for risk-based screening. Agencies should ensure they are considering relevant characteristics for their network and focus crash types. For example, if crashes involving left-turning vehicles are an issue, it is important to identify locations that lack turn restrictions or leading pedestrian intervals (LPIs), since these measures have been associated with improved safety. Conversely, the lack of restrictions or LPIs would be a factor associated with increased crash potential.
30 Systemic Pedestrian Safety Analysis Table 8 summarizes conditions associated with increasing pedestrian injury severity. In general, the evidence for some of these measures associations with pedestrian or crash injury severity are quite strong, as there have been many crash-based studies analyzing relative severity outcomes. As discussed in Step 2, some of these risk factors may be captured to some extent through road- way and built environment data. For example, speed limits and proportion of truck/bus traffic are included in the list of potential roadway data needs in Tables 3 and 4. Others could potentially be measured to some extent through land use and population-based measures, as shown in Table 5. Risk Factor Estimation Based on Cross-Tabulations or Frequency-Based Methods The most basic approach to identify potential risk factors is to simply use historical crash data for the entire system to identify types of locations across the network where target Variable/Risk Factors Intersections Segments Traffic volume Positive (generally positive but not linear) Positive (generally positive but not linear) High-turning volumes Unknown threshold Unknown at present Functional classesâarterials and collectors compared with local streets Positive Positive Proportion of truck/bus traffic in traffic stream Positive (crash severity) Positive (crash severity) Proportion of local streets at intersection (potential surrogate for AADT) Negative Unknown at present Pedestrian volume Positive (but not linear) Positive (but not linear) Number of legs > 3 (may also be partial traffic surrogate) Positive Unknown at present Total lanes on largest leg (5+) Positive Unknown at present No median/median island Positive (less certain than for segments) Positive Presence/number of transit stops Positive Positive Presence of on-street parking Positive Positive Presence/number of driveways Positive Unknown (theoretically yes) Presence of signal Positive with crash frequencies Negative with crash severity Unknown at present Lack of separate turning movements from walk phase (all red walk phase, or walk and restricted turn phase) (signalized intersections) Positive Unknown at present Lack of leading pedestrian interval (signalized intersections) Positive Negative Presence of four or more through lanes Higher numbers of total lanes Theoretically yes Positive Presence of TWLTL Unknown at present Positive Speed limit > 25 mph Unknown at present Positive with crash severity; positive with frequency in a few studies Vehicle speed Positive with severity Positive with severity Note: Positive and negative denote correlations with crashes. Table 7. Potential roadway risk factors identified from prior research and relationship to pedestrian crashes.
Step 3: Determine Risk Factors 31 crash types have occurred and then to identify prevalent characteristics of those locations. This method is basically an extension of the crash tree or matrix methods used initially to identify focus crash types in Step 1. Mapping and spatial analysis techniques may also be used. FHWAâs Systemic Safety Project Selection Tool (Preston et al. 2013) describes these types of approaches, which again depend on an ability to link crashes to location characteristics. The assumption is that roadway characteristics most prevalent for high frequency crash types represent elevated risk. However, recall that crash frequencies, unadjusted for volumes of users and crash trends for similar types of facilities, may give misleading results about risk factors. In addition, this method still leaves questions for how to prioritize zero crash loca- tions. This method may potentially lead to identification of location types that are predomi- nantly high motor vehicle and pedestrian traffic areas, especially if data for only high-crash locations are used. The crash type frequency method may be appealing based on its apparent simplicity and offers agencies a choice for a method for risk factor identification that does not depend on modeling. However, this method may not be fully risk-based if it does not properly account for the influence of traffic and pedestrian volumes. Furthermore, the data needs can be similar to the data needs for modeling. Many of the same data types, including crash types, traffic volume, and roadway location descriptors are important to have on hand to identify risk relationships and to identify specific locations with the risk characteristics. Land use, census, and other spatial data types may also be used. This method may also still require a significant amount of exper- tise to determine the crash and location type subsets or combinations that are most likely to represent treatable risk patterns. See the summary of strengths and limitations of this approach in Table 6. Light conditions Dark, with and without street lighting or unspecified Positive Strong Crash data Speed limit Higher speed limits (> 25 mph) Positive Strong Roadway data Traffic control type Other than signal (stop sign) or no control Positive Moderate Roadway data Vehicle type Variedâlarger compact to smaller, especially trucks or buses Positive Strong Crash data; traffic data (% heavy vehicles) Pedestrian age ~65 years and higher Positive Strong Crash data or census data (area population %) Pedestrian impairment Pedestrian under influence; alcohol use suspected or detected Positive Strong Crash data; locations of alcohol vendorsâmay be available in GIS as a potential population- level surrogate Pedestrian action Pedestrian crossing roadway (with/without signal or at midblock) Positive Moderate Crash or crash type data Note: Strong = six or more studies with consistent direction of effect; moderate = five to six studies with consistent direction of effect. Variable Category (if relevant) Relationship Evidence Potential Data Source Table 8. Roadway, crash, and person factors associated with increasing injury severity in pedestrian crashes.
32 Systemic Pedestrian Safety Analysis Perform Analyses and Identify Risk Factors If an agency is analyzing network data (developing SPFs or performing another type of modeling or analysis), this step describes a few considerations for performing those analyses. If risks are being identified using prior research and other means, then there may be useful information in this section on risk factors to potentially use in screening the network to identify sites for treatment. Perform Analysis Depending on methods used and numbers of variables available for analysis, the analysts may wish to perform initial data mining analysis prior to SPF development. For example, random forest or conditional random forest methods have been used to narrow the list of potentially important crash predictors to test in regression models. Such methods could also be used to identify potentially important crash types and location characteristics and to confirm or revisit decisions made in Step 1. The primary goal of the analysis is to identify treatable risk factors, but it is also important to generate models that are reliably predictive. SPFs that have too many factors, even if they are statistically significant, can reduce crash prediction efficiency by including more random noise in the model. The resources referenced at the end of this step provide more information on modeling statistics, which can help to reduce the chances of over-specifying the model while including important crash predictors (defined in Chapter 3). Determine Risk Factors for Use in Subsequent Steps In examining the model results, agencies will want to particularly consider the variables found to have strong positive associations with crash frequencies (i.e., crash predictors and/or risk fac- tors). If any of the variable associations do not conform to expected relationships, additional steps may be needed to revisit the model-building steps and discuss how those variables should be interpreted or applied in subsequent steps in the systemic process. At this point, agencies will need to determine which of the model variables associated with higher crash frequencies are to be considered treatable risk factors (i.e., there is an associated countermeasure that could be applied systemically; see Table 15 for some options). For example, if a model identifies the presence of midblock crosswalks as a crash predictor, sites with this feature can then be identified and treated systemically. The subset of relevant variables selected by the model will be applied in Step 4 to screen the network to identify candidate sites for treat- ment; several examples are provided. Variables that are not necessarily treatable still have value in a systemic process in that they help to improve the predictive ability of the model to better estimate where crashes are more Noteworthy Practice Case Example 3 provides an example of the process for determining risk factors to use in identifying treatable sites from prior research and expert knowledge and additional steps carried out for prioritization. Case Example 4 describes an example of the application of a frequency- based method to identify potential systemic risk patterns.
Step 3: Determine Risk Factors 33 likely to occur. These crash estimates raise the priority of those sites for further treatment con- sideration. Variables such as urban density/development type and traffic volume, regardless of whether they are included in the model, also provide important context for selecting appropriate countermeasures. Example Table 9 summarizes variables found to be associated with two types of pedestrian crashes at roadway segments (motor vehicle traveling straight and pedestrian crashes under dark conditions) based on analyses described in Chapter 3 of the technical report using Seattle, Washington, data. Based on the model results shown in Table 9, an agency might select the following variables to be used in identifying potential treatment locations (all were significant for one or both models) and treatments. â¢ Presence of a midblock crosswalk â¢ Presence of a two-way left-turn lane â¢ Presence of four, five, or more through lanes â¢ Presence of on-street, striped parking â¢ Speed limits above 25 mph Table 9. Variables predicting pedestrianâmotor vehicle crashes on roadway segments in Seattle, Washington. Predictive Factor Variable and/or Category Straight at Segment Pedestrian Crashes Under Dark Conditions at Segment Traffic volume Log transformation of pred_ada predicted ADT Positive â Pred_rfr predicted ADT/10000 â Positive Pedestrian volume Logarithmic term of AADP_MB Negative Negative Midblock pedestrian volume Positive Positive Built environment/ surrogate exposure measures Number of buses stopping nearby Positive Positive Commercial property density Positive Positive Mean income area residents/10,000 Negative Negative Light poles per 100 feet on segment Positivea Positivea âUrban villageâ development intensity category with increasing intensity Positive Positive Roadway factors Midblock crosswalks Positive Positive Two-way left-turn lane presence Positive Positive Four or more (5+) lanes compared with one lane Positive None detected Striped parking lanes (1 or 2+) Positive None detected Speed limit category (30 or 35 compared with 25; positive trend for higher speeds) None detected Positive One-way traffic flow None detected Negative Presence of right-turn-only lanes at one adjacent intersection Positive None detected Note: Dashes indicate that for each of these two traffic volume measures, the other measure was a better predictor in one or the other models. a Light poles are a potential surrogate for other traffic, design, or built environment features or an inadequate measure of lighting quality. Motor Vehicle Traveling
34 Systemic Pedestrian Safety Analysis â¢ One-way traffic flow â¢ Presence of right-turn only lanes at an adjacent intersection Steps 4, 5, and 6 will build on these results to demonstrate how these variables could be used to identify potential treatment sites and countermeasures and to prioritize systemic safety projects using economic analyses. Additional Resources Case Example 1 in this guidebook and Chapters 3 and 4 in the technical report provide more information and a detailed example analysis using the SPF development method. Tables 11 and 12 and Appendix B in the technical report provide more details on specific variables that have been analyzed. Below are additional resources on SPF development and systemic safety practices. Crash Modification Factor Clearinghouse website http://www.cmfclearinghouse.org/resources_spf.cfm FHWAâs Safety Performance Function Decision Guide: SPF Calibration versus SPF Development FHWAâs Safety Performance Function Development Guide: Developing Jurisdiction-Specific SPFs https://safety.fhwa.dot.gov/rsdp/downloads/ spf_development_guide_final.pdf NCHRPâs Userâs Guide to Develop Highway Safety Manual Safety Performance Function Calibration Factors FHWAâs Systemic Safety Project Selection Tool https://safety.fhwa.dot.gov/systemic/fhwasa13019/ FHWAâs Reliability of Safety Management Methods: Systemic Safety Programs FHWAâs Evaluation of Four Network Screening and Performance Measures https://safety.fhwa.dot.gov/rsdp/downloads/ fhwasa16103.pdf Resource Link https://safety.fhwa.dot.gov/rsdp/downloads/ spf_decision_guide_final.pdf http://onlinepubs.trb.org/onlinepubs/nchrp/docs/ NCHRP20-07(332)_FinalGuide.pdf https://safety.fhwa.dot.gov/rsdp/downloads/ fhwasa16041.pdf