Estimating Bicycling and Walking for Planning and Project Development: A Guidebook (2014)

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.

30 4.1 Identification of Planning Needs and Assessment of Available Tools The goal of NCHRP Project 08-78 has been to provide planners and analysts needing to estimate the demand for bicycle or pedestrian travel with (1) a better understanding of the key underlying relationships and (2) planning tools that put those relationships to work. Planning needs involv- ing bicycle and pedestrian travel are wide ranging, from rep- resenting non-motorized travel activity levels and impacts in regional plans to estimating demand for an individual facility. In general, it is helpful to organize these needs in relation to geographic scale: • Regional Planning Scale: Exemplary of the plans and analyses performed by MPOs, particularly in relation to long-range regional transportation plan (RTP) updates or supporting areawide policy or investment analyses. Non- motorized planning needs include – Projecting areawide bicycle and pedestrian activity levels – Accounting for bicycle and pedestrian access in estimating transit use – Effects of bicycle/pedestrian mode choice on the demand for auto travel and subsequent impact on congestion and VMT – Impacts of bicycle/pedestrian travel on effectiveness of compact mixed-use development (i.e., smart growth), and the converse – Use in regional visioning or scenario planning. • Corridor and Subarea Analysis Scale: To support analysis of travel in corridors, activity centers, neighborhoods, or transit-oriented development (TOD) plans where success of a modal investment, viability of a local land use plan, or the magnitude of traffic impacts are closely tied to the interaction between the corresponding transportation and land use plans. These analyses may be part of local compre- hensive or master planning and involve stakeholders from the local planning, zoning, transportation, development, and residential communities. Visualization and the ability to support interaction are important needs, as is the degree to which walking and biking support local trip-making and circulation and access to transit. • Facility Demand and Project Development Scale: For project (facility) planning, it is important to (1) gage the impact of improvements in accessibility provided by the respective networks on walking and bicycling activity levels, (2) evaluate priorities for the most effective improvements, and (3) account for the corroborative effects of the built environment. In addition to the geographic scaling that differentiates these categories, there is an alignment with the types of entities who would be performing the analysis, the types of questions being asked, accuracy needs, response time, and the tools and exper- tise available. Table 4-1 characterizes these different audiences and the tools they are using. NCHRP Project 08-78 evaluated numerous existing tools and methods developed in relation to bicycle and pedestrian travel. The goal was to identify those tools that reflected the best existing practice in addressing the three categories of application needs above. Table 4-2 provides an overview of the range of tools evaluated, along with noteworthy examples of each. In some cases, the methods are free-standing tools; in other cases, they are enhancements or supporting techniques for existing tools. Some of these examples fall into the cate- gory of “research models,” developed mainly to investigate and quantify key relationships, although the models themselves are generally not suitable as planning tools. In relation to Regional Planning, standard practice consists of the traditional four-step trip-based regional forecasting models, which rely on TAZs as their geospatial structure when estimating trip generation, destination, and modal choice. These methods have difficulty representing non-motorized C H A P T E R 4 Best-Practice Methods for Estimating Bicycle and Pedestrian Demand

31 travel demand, largely because of coarse scale of analysis attributable to the TAZ aggregation of land use. If these mod- els are used to account for non-motorized travel, it is typi- cally limited to the trip generation step; non-motorized trip productions and attractions are estimated, but they are then removed from the remainder of the analysis, which focuses on motor vehicle trips. Three types of efforts have been made to improve the sen- sitivity of these widely used transportation planning models to land use and non-motorized travel: • Enhancements: Various types of enhancements to the steps of the modeling, including sensitizing trip generation to land use factors, reducing the size of TAZs, and taking advantage of the smaller zones to try to carry non-motorized trips further into destination choice and mode split. (A similar approach developed under NCHRP Project 08-78 is presented as one of the recommended methods.) • Post Processors: Development of ancillary models that use GIS methods to reflect differences in land use at a much finer level of geography (parcels or grid cells); these models are then used to modify preliminary results from the trip-based model. • Microsimulation: A new class of activity- or tour-based models developed using parcel or point-level information instead of TAZs to more closely associate travel choices with the adjacent (as well as regional) transportation and land use characteristics. The finer scale makes it possible to directly incorporate walking and biking as modes. (One of the new methods developed by the NCHRP 08-78 project takes advantage of such a tour-based structure.) Most of the reviewed methods fall into the category of facility-demand estimation tools. Because the regional mod- eling tools are not easily accessed or understood by many practitioners, nor realistically represent non-motorized travel, practitioners needing answers for planning bicycle or pedes- trian systems have been obliged to develop their own tools. Tools in this category include • Factoring and sketch-planning methods that estimate demand by projecting from a similar project or situation, relying on mode-choice information from Census Journey- to-Work statistics, or using various rules of thumb to relate bike/pedestrian use levels to existing or new population or activity levels. • Direct demand models, which are among the newest and most widely used tools in this genre, developed using regres- sion models to explain demand levels as recorded in counts as a function of measured characteristics of the adjacent environment (e.g., population, employment by type, major generators, and facility proportions). Scope Regional Corridor/Subarea Project/Facility Geographic scale Region Local (county, large municipality) Mulmodal corridor; Transit line/node; Acvity center; Neighborhood Development site; Travel network link; Intersecon Agency MPO, County Planning City Planning MPO, County, Municipality, Transit Agency County/Municipality; Developer; Praconer type Transportaon planner Travel modeler; Bike/Ped planner Transportaon planner; Bike/ped planner Traffic engineer Bike/ped planner Traffic engineer Key Quesons Walk/bike travel levels; Access to transit; Mode choice, VMT; Land Use viability Access to transit Person/vehicle conflicts Network coverage & connecvity Network coverage & connuity; Safety; Link demand levels Resources Computer tools & experse; GIS tools/data; Travel survey & other specialized data Computer tools & experse; GIS tools/data; Travel survey & other specialized data Simple methods: Maps & Counts Advanced methods: GIS tools/ data; Travel survey & other specialized data Current tools Regional models (trip based or acvity based); Scenario planning tools w/ land use sensivity Regional models; Scenario planning tools Planning standards Direct demand models; Planning standards; Professional judgment; Factoring methods Table 4-1. Framework for relating planning needs to applications and user characteristics.

32 Also among the tools in this genre that provide valuable insights on factor relationships but lack the structure to serve as complete or practical planning tools are • Aggregate demand methods that attempt to explain regional (or similar large area) activity levels of walking or biking based on aggregate population, employment, density, facility mileage, and even climate factors. • Route choice models that focus on the factors that affect choice of route. These models have their greatest value in quantifying the degree to which particular features (e.g., type facility, hilliness, and so forth) affect the utility and selection of a link or path. Tools for the middle of the planning spectrum, Corridor and Subarea planning, were found to be the leanest of the Applicaon Category/Approach Examples Regional Planning Trip generaon: trip generaon augmented by special models that esmate non motorized producons based on density, land use mix, accessibility, and/or urban design Atlanta (ARC), Ausn (CAMPO), Portland (Metro), Durham, NC; Buffalo Auto ownership: context enhanced auto ownership as input to non motorized trip producon Atlanta (ARC), Ausn (CAMPO), Portland (Metro), Los Angeles (SCAG) Desnaon choice: separate models to forecast trip generaon for inter and intrazonal trips based on land use/accessibility context factors Buffalo, Durham Mode choice: Special context sensive models to esmate non motorized mode split for intrazonal trips Buffalo, Durham Acvity/Tour based models: projected replacement to trip based models, spaal resoluon reduced to parcel level and individual travelers – remove TAZ aggregaon bias in clarifying non motorized mode use; travel treated as simple versus complex tours which impact mode choice Edmonton Transport Analysis Model; San Francisco (SFCTA), Sacramento (SACOG), many under development Corridor, Subarea and TOD Planning Scenario Planning Tools: Esmaon of non motorized travel and VMT reducon in relaon to alternave land use and transportaon investment scenarios US EPA Index 4D method (2001); Frank & Co. I-PLACES (2008); Ewing, et al.—MXD model (2010); Kuzmyak, et al.—Local Sustainability Planning Model (2010) Walk Trip Models: Models that resemble four step regional approach, but employ “pedestrian” zones instead of TAZs; create trip tables and assign to facilies PedContext – Maryland State Highway Administra”on and Univ of MD Nat Center for Smart Growth (2004/08); Cli˜on—MoPeD Model (2008) Facility Planning Factoring and sketch planning methods: aˆempt to predict facility demand levels based on peer comparisons, applicaon of trip generaon rates to sociodemographic data, associaon with other related data/trends, proximity rules, etc. Lewis & Kirk (1997); Wigan, et al. (1998); Goldsmith (1997); Ercolano, et al. (1997); Clark (1997); Krizek, et al. (2006) Direct Demand: Project bicycle or pedestrian volumes based on counts related to various context and facility factors through regression models Ashley & Banister (1989); Parkin & Wardman (2008); U.C. Berkeley—Seamless Travel (2010); Schneider, et al.—Alameda (2009); Liu & Griswold (2008); Fehr & Peers—Santa Monica (2010) Aggregate demand: Seek to quanfy relaonship between overall demand (e.g., annual regional bike trips) and underlying factors, oŽen as a way of gauging importance of infrastructure types and extents Baltes (1996); Dill & Carr (2003); Buehler and Pucher (2011); Nelson & Allen (1997) Route or path choice: Methods that try to account for the characteriscs of a transportaon network or its users in determining route choice, and for idenfying network improvement priories Hunt & Abraham (2006); Krizek (2006); Menghini, et al. (2009); Dill & Gliebe (2008); Hood, et al. (2011); Space Syntax—Raford and Ragland, Oakland pedestrian master plan (2003); McCahill & Garrick—Cambridge MA bike network (2008) Table 4-2. Overview of existing tools and methods for non-motorized planning.

33 offerings in the existing body of methods or focus of research. NCHRP Project 08-78 found that planning at this level is either done with a focused application of the respective regional model (albeit lacking sensitivity to land use and non-motorized travel), or without analytic tools and relying instead on trip generation rates and traffic level of service standards. Two variations on the focused regional model approach are • Scenario Planning tools, such as Envision Plus, Urban Foot- print, I-PLACES, and EPA’s Smart Growth Index, rely heav- ily on GIS to depict alternative land use and transportation configurations and estimate their effect on travel behavior. These tools may be used independently for local planning, or in tandem with the respective regional model for larger area assessments. (These tools served as a basis for NCHRP Project 08-78’s design and testing of a GIS-based accessibility approach, which expands the capability of these existing tools in impor- tant ways, particularly in relation to non-motorized travel). • Walk Trip Models: Two models were found to have inter- esting capability and relevance for this subarea level of analysis: PedContext and its sequel, the Model of Pedestrian Demand, or MoPeD. These models estimate pedestrian travel (only) in relation to land use and transportation net- work features. Both methods are similar to the four-step process, but operate at a much finer level of detail—PAZs— which are roughly the scale of a city block. Both perform trip generation (for walk trips only), create trip tables, and assign the trips to the local walk network to produce link-level and intersection-level activity estimates. The principal difference between the two methods is the degree of detail and rigor applied, with MoPeD being the less detailed of the two. Table 4-2 provides referenced examples for each type of tool or procedure. These and similar examples are docu- mented in greater detail in Appendix 7 of the Contractor’s Final Report. 4.2 Addressing the Gaps The review and evaluation of the existing tools corroborated initial perceptions that the current methods fell short in being able to address the range of planning and decision-making needs. In general, an overall paradigm to explain bicycle and pedestrian travel decisions in relation to travel demand theory and in consideration of the mode-specific factors of impor- tance identified in the research is lacking. This paradigm should attempt to account for the following elements: • Sociodemographic characteristics of the traveler and the traveler’s household • Trip purpose • Access to purpose-specific activities by each mode, as afforded by the patterns of land use and the design of the transportation network providing connectivity with those opportunities A model that includes such a structure is said to be “choice- based,” meaning that each of the factors enabling the indi- vidual to choose from among his/her destination or modal options is part of predicting their behavior. This choice is generally determined as a probability that the traveler will pick alternative A over alternative B, C, or D based on their comparative advantages (utilities) and how those ele- ments are weighed in importance by the particular type of individual. Although the regional models are regarded as choice- based, they neither include all of the relevant modes in the set of choices nor provide the detail to properly calculate the utility for the non-motorized choices (i.e., land use attrac- tions and facilities relevant for walk or bicycle travel). At the other end of the spectrum are the facility-demand models, which are not choice based. Rather, a set of descriptive environmental context variables are used to explain varia- tions in usage levels (through activity counts) across a sam- ple of sites; however, the counts—and hence the explanatory models—do not reveal behavioral motivation, in terms of traveler characteristics, trip purpose, origin-destination, or available alternatives. This dichotomy creates a dilemma with regard to designing the best user tools. Ideally, a choice-based approach should be used for most bicycle-pedestrian planning assessments. As characterized in Figure 4-1, the choice-based process pro- gresses from trip generation to destination choice, then mode choice, then assignment of trips by mode to the respective network. From the network assignment step it is possible to ascertain facility volumes (link or intersection). Assuming the choices are captured correctly in the respective models, this approach allows for multiple forces to interact toward the final outcome, while providing multiple places for testing planning interventions or other assumptions. In contrast, the facility-based approach focuses directly on explaining link or intersection counts. Although this approach is much less cumbersome for the planner, it also is consider- ably less informative as to composition or behavioral motiva- tion underlying the observed volumes. To improve the overall caliber of bicycle/pedestrian plan- ning tools, project research has focused heavily on forging a satisfactory choice-based approach, both to provide needed illumination about the behavioral relationships in non- motorized travel and, by accounting for those relationships, enabling planners to control for those variables in an analy- sis. Therefore, most of the tools featured in the guidebook

34 will embody some semblance of this holistic, choice-based behavioral structure. This is not to say that the facility-based methods have no value in bicycle/pedestrian planning. For certain types of analyses (e.g., extrapolating demand from a pre-existing situa- tion in response to incremental changes in local development) facility-based methods may be helpful in supporting localized decisions related to pedestrian connections or intersection crossings, grid traffic management, or additional bike travel. However, because facility-based methods lack a behavioral structure, their use for land use planning, planning for changes to the bike or walk networks as regards connectivity would be limited because they do not incorporate those relationships in their structure. The following guiding principles have been applied in developing and suggesting methods for bicycle and pedestrian planning and demand estimation: • The recommended planning tools should stress a choice- based structure. • The modeling tools developed or brought forward by the project should bring the choice-based option within the ability and resource range of more practitioners. • The tools should be able to assess the relative importance of land use features versus facility improvements, toward the ideal combination of both. • To the degree possible, the tools should be applicable in circumstances ranging from full deployment in regional planning to strategic use of the relationships in scenario planning or facility design. • The choice-based tools and relationships should be able to assist in improving the structure and accuracy of facility- demand tools. 4.3 Introducing the Guidebook Planning Tools Rather than a single all-purpose model, the guidebook features some tools that may be of particular value to prac- titioners, depending on the scale of the analysis, the decision being supported, skill level of the user, and available resources. Recommended tools are listed in Table 4-3. The tools are listed in generally declining order of complex- ity, which also roughly corresponds to the geographic scale at which they most likely will be applied. The first three tools were all created through research performed under NCHRP Project 08-78, taking advantage of willing local partners, suitable environments for walking and bicycling, and above- average data to support the research. Two of the projects were performed using data from the Puget Sound Regional Council (PSRC) in the Seattle area, while the third focused on Arling- ton County, VA, using data from the Metropolitan Washington Council of Governments (MWCOG). Tour-Generation and Mode-Split Models: In conjunc- tion with the Puget Sound Council of Governments’ efforts to develop a new tour-based model structure for the Seattle region, research team members took advantage of various new data and tools to develop a set of pedestrian and bicycle models. The set includes a procedure for generating tours (as opposed to trips) by purpose, and a pair of modal-split mod- els that predict walk, bike, transit, and auto choice for five tour purposes. The variables included in these models pro- vide access to a broad spectrum of sociodemographic, land use, transportation network characteristics, and accessibility in estimating (separately) bicycle and pedestrian demand, as well as the effect on transit use of non-motorized accessibil- ity. Although immediately suited to working in an activity- or Trip Mode Choice Network Facility Volume Facility Based Generaon Desnaon Choice Assignment Choice Based Process Figure 4-1. Choice-based versus facility-based activity estimation approaches.

35 tour-based environment, the methods may also be used to enhance conventional trip-based models, and a spreadsheet version of the model can be used for simultaneous testing of any of the relationships in the models or for creating sketch- planning tools. GIS-Based Walk-Accessibility Model: Using data from the Metropolitan Washington (DC) Council of Governments (MWCOG) for Arlington County, VA, the research team devel- oped a method for estimating walk trip generation and mode split that relies exclusively on GIS tools and data. The method uses geospatial overlay and network path-building procedures readily available in GIS to calculate measures of accessibility to or from any point by any mode and by type of attraction. By comparing the modal accessibilities, it is possible to esti- mate mode split and create walk trip tables by purpose. The current model does not perform network assignment of the walk trips, although it is assumed that users can apply such features in their existing transportation planning software to do so. Given insufficient data, the current model does not forecast bicycle demand, although the structure will readily Modeling Approach Source Characteriscs Tour Generaon/ Mode Split NCHRP 8 78 (Seale/PSRC data) Simple/complex tour generaon for 8 trip purposes (sociodemographic characteriscs, land use, local & regional accessibility) Mode choice (walk, bike, transit, auto) for 5 trip purposes (sociodemographics, land use, local & regional accessibility, Fully detailed walk and bicycle networks, physical aributes affect impedance GIS Accessibility Model NCHRP 8 78 (Arlington, VA/MWCOG data) Uses GIS layering to create accessibility scores for walk, bike, transit, and auto. Links mode choice with accessibility scores at trip origin and desnaon Esmates mode share at block level for HBW, HBO, NHB and WBO purposes Builds walk trip table (but does not assign) Highly visual presentaon Trip Based Model Enhancements NCHRP 8 78 (Seale/PSRC data) Strategic changes to tradional four step TAZ model to improve sensivity to land use and non motorized travel Sensizes auto ownership and trip generaon to land use characteriscs Performs pre mode choice to disnguish inter versus intrazonal trips Performs mode choice separately for intra zone (drive alone, shared ride, walk) and inter zone (drive, shared ride, transit, walk, bike) travel Pedestrian Demand Models PedContext and MoPeD (Univ. of MD/ Maryland DOT) Modified four step approach focused on esmang walk trips Walk trip generaon for several purposes at PAZ level Creates walk trip tables, assigns trips to walk network Bicycle Route Choice Models San Francisco County Transp. Authority Portland State Univ. Models built from GPS data to predict choice of route for bicycle riders Quanfies importance of route characteriscs (type facility, gradient, directness, traffic exposure) Facility Demand Models Fehr & Peers (Santa Monica) Separate bicycle and pedestrian direct demand models Predict PM peak hour bicycle demand based on employment density, proximity to bike facilies, land use mix, and intersecons Predict PM peak hour walk demand based on employment density, proximity to shopping, PM bus frequency, and traffic speeds Table 4-3. Bicycle/pedestrian planning tools included in guidebook.

36 accommodate such an enhancement when adequate data are available. This approach offers a new and intuitive way of interpreting modal choice that is very responsive to changes in the built environment (land use) or the travel networks such as would occur in corridor or subarea planning, using generally available data and with relative independence from the respective regional travel model. Enhancements to Trip-Based Models: Research team members also worked with the PSRC data in Seattle to create a template for systematically enhancing a conventional TAZ/ trip-based regional model to improve its sensitivity to land use and non-motorized travel. Advanced statistical methods were used to create enhancements to the Auto Ownership, Trip Generation, Trip Distribution, and Mode-Choice steps in the existing PSRC regional model. Measures of auto and non- motorized accessibility play a major role in these enhance- ments. Although pedestrian and bicycle mode choice are still constrained by the TAZ structure, the methods improve on the current process by introducing a “pre-mode-split” step, which first divides trips into intra- versus interzonal groups and then performs a mode-split specific step to those groups. Although the enhanced regional model may not be as fluid as the tour-based or GIS-accessibility approaches in overcom- ing TAZ aggregation issues, the enhanced regional model takes advantage of the new smaller TAZs adopted by many MPOs and provides considerably more sensitivity in existing models. Walk Trip Generation and Flow Models: The PedContext and MoPeD models developed and tested in Maryland offer a set of methods for estimating walking trips and creating walk trip tables at a block level. Both methods follow a varia- tion of the four-step process, and both assign the walk trips to the local walk network to estimate link and intersection activity levels. The difference in the methods is the degree of detail each applies at each step, with the MoPeD model being the less detailed of the two. Another tool grouped in this set is a pedestrian model recently produced for Portland Metro that is similar to MoPeD, but which serves more as a support procedure for the regional travel model. Facility Demand: Two types of models are included in this category: bicycle route choice models (e.g., those devel- oped by the San Francisco County Transportation Author- ity and Portland State University) and direct demand models which predict walk or bike facility use and volumes based on observed counts and context-driven regression models. A third type of model reviewed falls into the cate- gory of “network simulation,” and is most exemplified by the Space Syntax model, which estimates network flows using network geometric relationships. This model is pre- sented in discussion which follows, but was not included in the list of recommended tools because (1) it is proprietary, and (2) it was difficult to acquire enough information on the inner workings to be able to fairly evaluate its perfor- mance and validity. The next section provides the user with an overview of each of the tools. The objective is to give enough information on how the tools were developed, their structure, and how they work to establish a basic understanding of what they are and what they do. Chapter 5 then integrates and synthesizes this information to help users distinguish among the tools and determine which to choose for their particular applica- tion needs. Users may want to refer to the profiles below in Section 4.4 as they become more involved in looking at the tools and their capabilities. Full documentation for all the models is also provided: for those tools developed directly by NCHRP Project 08-78, Appendices 1 (Seattle Tour-Based Model), 2 (Arlington Walk-Accessibility) and 3 (Trip-based Model Enhancements) of the Contractor’s Final Report contain full descriptions of each tool. Citations and website addresses are provided for the other tools. 4.4 Overview of Recommended Guidebook Tools Tour-Based Approach The researchers used data and resources from the PSRC in Seattle to create a new set of models that estimate the demand for walking and bicycle travel based on characteristics of the traveler, purpose of the trip, opportunities present in the pre- vailing land use, and the accessibility provided by the respec- tive travel networks. These models offer important insights into non-motorized travel behavior, in large part because the extremely high level of detail is much more effective in cap- turing factors that influence walk or bicycle mode choice. The models were developed using a tour-based model struc- ture. Tour- and activity-based models have gained increased attention from transportation planners and planning agencies for use in regional and even statewide planning. They are dif- ferent from conventional TAZ-level trip-based models in the following ways: • Parcels instead of TAZs: Analysis is performed at a much finer scale of geospatial resolution, generally working with land use parcels as opposed to TAZs. This allows for much sharper characterization of the travel environment and the factors that affect non-motorized travel. • Tours instead of Trips: Travel is portrayed in the form of complete “tours” rather than a series of individual “trips.” This is more reflective of how travel actually occurs (i.e., with one or more purposes accomplished before complet- ing a “round trip”) and has an important bearing on mode use. Travelers in more urban, mixed-use environments make more journeys as simple out-and-back tours, while journeys in lower density/separated land use settings more

37 commonly occur as multi-stop complex tours, for efficiency. The multi-stop tours are generally made by auto, while simple tours are more likely to be made by walking, biking, or transit. • Individuals instead of Households: Tour or activity-based models focus on the travel of individuals, rather than aggregate households. This allows inclusion of key socio- demographic factors such as age, gender, driver status, and employment/student status, along with household compo- sition (income, size, and vehicle ownership). These factors have been found to be fairly important in explaining non- motorized travel behavior tendencies. As with most of the guidebook tools, accessibility—the mea- sure of the opportunities that can be reached with the land use patterns and modal options at hand—is a central theme in this approach. Determining accessibility for non-motorized modes is more challenging than with auto, given the need to perform the assessment at a much finer geographic level and the importance of physical factors in gauging the performance of the travel networks. The Seattle approach used the follow- ing steps to measure accessibility: • Explicit Networks: Travel distance and time are particularly important factors in the non-motorized travel decision. All else being equal, people considering a walk or bicycle trip rank straight line distance as the number one factor in assessing their travel options, which is heavily deter- mined by the coverage and connectivity of the respective travel network. However, cyclists and pedestrians are also highly sensitive to personal safety, and so will prefer routes with less exposure to vehicle traffic, even if they involve longer distances; steep hills are a similar discouragement. To ensure that the tour models accurately reflected these sensitivities, considerable care was devoted to mapping and quantifying the various attributes of the bicycle and pedestrian networks. The result is a sharper depiction of the service characteristics provided by the respective net- work when calculating the statistical relationships. • Buffering Walk and Bike Opportunities: Accessibility to opportunities by walk and bicycle from each potential trip origin or destination was estimated using a buffering process, which sums the number of opportunities within a “reasonable” distance (1 mile for walk, 2 miles for bike), with each opportunity discounted by its respective over- the-network distance (not travel time because that infor- mation was not available for walk/bike). The effect of longer distances making far-away destinations less desirable is represented through a logistic distance-decay relationship, as illustrated in Figure 4-2. These curves were plotted using data from the regional household travel survey. The flat portions at the beginning of each curve imply that distance is not important for the first block or two, but then utility falls off rapidly with longer distance. • Competing Opportunities: Travelers who have autos or transit service available make tradeoffs in choosing between travel to a nearby destination by walking or biking or mak- ing a vehicle trip to a more remote location. This competi- tion is measured through a comparison of local and regional accessibility, with the latter accounting for all opportuni- ties, regardless of distance, by all modes. In the tour-based model, regional accessibility is represented through a log- sum measure, which is a summation of the accessibilities of auto driver, shared-ride, transit, and bicycle weighted by modal share as taken from the denominator of the mode/ destination choice model. More details on the composition of this measure are provided in the model documentation. With this “dual accessibility” structure, the models can be used when attempting to ascertain how walking or biking 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 m ile s 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 1 m ile 1. 1 1. 2 1. 3 1. 4 1. 5 1. 6 1. 7 1. 8 1. 9 2 m ile s Pe rc en to fT ot al U l ity Trip Distance Buffer 1 (Walk) Buffer 2 (Bicycle) Figure 4-2. Logistic travel decay rates by distance.

38 would benefit from improvements in local land use or net- work coverage that directly improve local accessibility, or from changes that might occur regionally (e.g., new highway or tran- sit line, congestion delay, or changes in fuel prices) that would affect the desirability of longer trips by driving or transit. Two types of predictive models were developed for bike/ pedestrian analysis: • Tour Generation: A model that predicts the number of daily tours that a given individual will generate, the travel purpose of the tour, and whether the tour will be simple or complex. • Mode Choice: A set of models predicting choice of mode (walking, bicycle, transit, or auto) for different trip pur- poses (home-based work, home-based school, home-based recreational, home-based other, and work-based other). Two different mode-choice model formats were developed: – Origin-Destination—incorporates information on land use, network, and accessibility at both origin and destination, as well as origin-destination travel time and/or cost; using this version of the model is appropri- ate when the location of the trip destination is known (e.g., a work or school trip). – Origin Only—includes information on land use, net- work, and accessibility at the origin end only; using this version is appropriate when the location of the destination is unknown (e.g., a shopping or personal business trip). A quick overview of the models, the variables they contain, and how they work are presented below. Guidelines on how to apply the models for planning are provided in Chapter 5, including introduction to a spreadsheet version of the com- bined model set which accompanies the guidebook. Detailed specifications for each model (e.g., coefficients, estimation statistics, and sample size) are presented in Table A-1 (Appen- dix A) of the guidebook. Full documentation describing development of the models and preparation of the data is provided as Appendix 1 to the Contractor’s Final Report. In a full application, the Seattle tour-based models follow a sequence in which the number and type of tours are first estimated for a given individual. The tours are then processed by one of the mode-choice models to estimate the probabil- ity of traveling by walking, bicycle, transit or auto for each tour purpose. A simplified portrayal of the tour generation/ complexity model is shown in Figure 4-3, illustrating the following steps: • Number of Tours: The first step estimates the likelihood (probability) that the person will make any tours at all on the given day, and then whether they will make a second, third, or fourth tour on the same day. • Tour Complexity: The next calculation is in whether the tour will be simple or complex; this is not a yes/no answer, but again a probability that separates the given tour into a simple and complex portion. • Tours by Purpose: The third step is to determine the pur- pose of the tour. The model predicts travel for work, school, escort, personal business, shopping, eating a meal, and social/recreational. If a home-based work tour is made, a separate tour generation model estimates the number and type of work-based tours that will be made (followed by mode split). The tour-generation model is applied to the population of potential travelers, typically represented through a synthetic population (a specially drawn demographically representa- tive sample meant to represent the overall population). The estimated tours are then assigned to travel modes, using either of the two mode-choice models (origin-only information or both origin and destination). A simplified presentation of the origin-only mode-choice model is shown in Table 4-4, and the origin-destination model is shown in Table 4-5. The tables show the estimated coefficients used to compute the modal shares for each tour purpose. The previous sepa- rate tour generation models for home-based shopping, eating a meal, and personal business have been collapsed to a single home-based other model for mode-choice purposes. Differ- ent variables come into play for any given mode, depending on the trip purpose, and walk and bike modes use different specifications for the buffer measures, with Buffer 1 reflecting the range for walk trips (including walk access to transit) and Buffer 2 addressing bicycle. Table 4-6 summarizes the variables contained in these models. Planners can use these relationships in various ways, from building their own models to enhancing existing mod- els, post-processing model results, or creating sketch-planning models for sensitivity testing or project comparisons. Sug- gestions for use, along with an inter active spreadsheet version of the models, are provided in Section 5.4 of Chap- ter 5. Elasticities for the mode-choice models are provided in Section 5.3. GIS-Based Walk-Accessibility Approach The other original research conducted by the NCHRP Project 08-78 team focused on developing a direct accessibility approach that would take maximum advantage of the capabil- ities offered by modern GIS tools and data. Although the tour- based approach developed in Seattle is an effort to expand the limits of what is currently possible in regional-scale models, the GIS-based accessibility seeks to create something much simpler and more intuitive in concept that might be accessible to many users and a range of applications.

39 Planners as well as non-planners are familiar with Walk Score, the internet application that attempts to quantify the level of walkability for any given place on a scale of 1 to 100. This statistic is widely used to assess the richness of access to local activities and is even employed by the real estate indus- try as an added-value attribute when marketing properties. Although the NCHRP project did not set out to replicate Walk Score, the research showed that similar types of mea- sures, if properly constructed and interpreted, could provide the basis for a practical and fairly accurate procedure for bicycle and pedestrian planning. Although early uses of GIS focused on its rich mapping capabilities, its true value is the ability to perform complex mathematical tasks using geospatial overlay methods to exchange information among multiple layers. From a trans- portation perspective, this makes it possible to overlay the cov- erage and service provided by transportation networks in one layer onto the characteristics of the corresponding land use environment, leading to very realistic measures of connection between the two. For reasons of data quality, tools, and a highly diverse transportation/land use setting, the accessibility analysis Figure 4-3. Tour generation/complexity calculations (shown values are coefficients, not elasticities).

Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Constant 7.31 3.61 3.78 3.82 7.69 2.94 2.92 4.84 5.78 3.03 6 8.5 3.49 3.49 0.986 Income < $25k 0.379 1.14 2.38 0.647 0.813 Income > $100k 0.546 0.42 0.256 1.81 Male 0.337 0.676 0.32 0.711 1.96 0.72 Age <35 1.38 0.412 1.25 0.26 Age > 50 0.833 0.991 2.17 0.486 0.338 Zero car HH 4.69 5 3.09 3.6 Adults > Cars 1.21 1.16 0.799 0.417 Buffer 1 aractions for purpose 0.403 0.423 0.262 0.36 Buffer 2 aractions for purpose 0.22 Mode/desnaon logsum with zero cars 0.245 0.289 0.0922 0.355 0.699 Mode/desnaon logsum with full car own 0.154 0.0944 0.04 Buffer 1 household density 0.00026 Buffer 1 net intersecon density 0.0043 0.00007 0.00048 0.0101 0.00014 Buffer 2 net intersecon density 0.0087 0.0127 Buffer 1 average fracon rise 29.2 35.5 Buffer 2 average fracon rise 62.6 31.4 92.5 Buffer 2 fracon Class 1 bike path 2.4 3.15 Buffer 1 percent no sidewalk 1.04 1.38 0.769 2.96 1.12 1.6 3.89 Buffer 1 transit stops 0.737 0.291 0.121 0.296 0.312 Buffer 1mixed use index 0.716 0.454 1.36 0.791 0.559 Walked to work Bike to work 2 Transit to work 0.574 Car to work 1.67 Tour Complexity 1.45 1.08 0.781 2.21 2.18 0.314 1.33 0.628 0.693 1.3 1.59 0.361 1.61 2 0.677 Work BasedHome BasedWork Home Based School Home Based Social/Rec Home Based Other Table 4-4. Tour mode choice models with origin-only information (shown values are estimated coefficients, not elasticities).

Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Walk Bike Transit Auto Constant 1.07 2.92 4.74 0.91 4.12 1.6 2.96 4.53 3.15 1.81 3.74 5.61 3.34 8.82 8.05 Income < $25k 0.863 0.961 0.36 3.02 0.0615 0.702 0.468 Income > $100k 0.412 0.447 0.669 0.0075 0.498 1.15 0.121 1.57 Male 0.534 0.859 0.186 0.578 1.71 0.0012 0.0356 2.05 0.325 0.119 0.842 0.215 1.16 Age <35 1.45 0.398 0.36 0.0084 0.285 0.458 2.25 Age > 50 0.863 1.26 0.518 Zero car HH 4.7 5 3.32 4.32 10 Adults > Cars 1.4 1.27 0.976 0.633 0 Route choice generalized distance 0.113 0.277 0.0874 0.276 0.331 Distance (over network) 0.942 1.45 1.6 1.87 1.88 Pct Class 1 path Pct Class 2 path Fracon wrong way Turns/mile Fracon rise Dest Buffer 1 Tot Emp 3.80E 05 2.70E 05 Dest Buffer 2 Tot Emp Dest Buffer 2 Emp Density 3.70E 07 Orig+Dest Buffer 1 Avg Intersecon Density 0.005 0.0111 Orig Buffer 1 Intersect Density 1.50E 04 Orig Buffer 2 Intersect Density 0.0061 Orig+Dest Buffer 2 avg Fracon Class 1 Path 4.97 3.01 Orig+Dest Buffer 1 Avg Fracon Rise 61.3 9.85 15.6 Orig Buffer 1 Fracon Rise 36.2 Orig Buffer 2 Avg Fracon Rise 77.8 Orig Buffer 1 Transit stops 0.539 0.334 0.608 0.214 Dest Buffer 1 Transit Stops 0.179 0.268 0.825 0.606 1.73 Orig Buffer 1 Pct. No Sidewalk 0.84 0.715 1.07 1.44 Dest Buffer 1 Pct. No Sidewalk 0.872 4.26 Walked to work 10 Bike to work 10 Transit to work 0.224 Car to work 2.3 Complex Multi stop Tour 1.24 0.782 0.501 2.55 2.25 0.785 2.14 1.61 5.00E 15 1.51 1.95 0.647 2.71 2 1.5 In vehicle me 0.01 0.02 0.01 0.02 0.01 0.02 0.01 0.02 0.01 0.02 Wait me 0.02 0.02 0.02 0.02 0.02 Fare 0.2 0.2 0.2 0.2 0.2 Dest Parking Cost 0.06 0.06 0.06 0.06 0.06 Work BasedHome BasedWork Home Based School Home Based Social/Rec Home Based Other Table 4-5. Tour mode-choice models with both origin and destination information (shown values are estimated coefficients, not elasticities).

42 focused on Arlington County, VA. Given that Arlington is part of the Washington, DC, region, its selection provided access to the resources of both the County and the MWCOG, including a recent (2008) regional household travel survey with excellent coverage in Arlington. The following data and tools were used to create the acces- sibility relationships that formed the basis for the eventual walk-accessibility model: • A regional employment database prepared by Dun & Bradstreet and accessed through MWCOG, providing information on the type (4-digit NAICs code), size, and point location of all employers; this information was used to represent trip attractions. • A complete-streets transportation network developed by NAVTEQ and accessed through MWCOG; the base net- work was enhanced to include any missing bicycle or pedes- trian links; GTFS data were used to represent the transit network. • Complete information on 9,100 trips from the regional travel survey having at least one trip end in Arlington County. By knowing the block-face location of each of the 9,100 trips (both origin and destination), it was possible to estimate accessibility for all modes (i.e., walk, bicycle, auto, and transit) using the respective travel networks in conjunction with the Dun & Bradstreet data. This was done by using the Network Analyst program within ArcGIS to ascertain the shortest time path between the respective trip end and each opportunity represented in Dun & Bradstreet, using the actual network for that mode. Individual opportunities were discounted by the amount of travel time required to reach them, applying a logarithmic time-decay relationship similar to the approach used in Seattle but with the values drawn from distributions of the Arlington trip data. The discounted opportunities were then summed into a total accessibility value for each mode. A strong relationship was identified between the calculated walk-accessibility score at either the trip origin or destination and the mode which was used for the trip as recorded in the travel survey data. These relationships are illustrated in the graphs of Figures 4-4 for home-based work travel and 4-5 for home-based non-work travel. The figures illustrate the percentage of trips made by auto, transit, and walking for dif- ferent levels of walk-accessibility, ranging from under 200 to over 1200, with both walk share and transit share increasing directly with higher values of walk-accessibility. There were too few bicycle trip observations in the survey data to enable inclusion of bicycle as one of the primary modes, although the accessibility approach appears suitable for bike travel. The black curves in the figures represent the plotted data, while the red curves are those fitted by to the data by Excel. The mathematical functions describing the fitted curves are Sociodemographic Land Use/ Accessibility Transportaon/ Network Characteriscs Tour Generaon & Complexity Gender Age Work/Student status Income Car ownership/ compeon Children in HH Land use mix (entropy) Purpose-specific buffer acvity Purpose-specific logsum Intersecon density Distance to transit stop Gradient Class I or II bike path Mode Choice (origin only) Gender Age Income Car ownership/ compeon Land use mix (entropy) Household density Purpose-specific buffer acvity Mode/desnaon logsum Intersecon density Transit stop density Gradient Percent Class 1 bike facilies Percent no sidewalks Mode Choice (origin-desnaon) Gender Age Income Car ownership/ compeon Land use mix (entropy) Employment density Purpose-specific buffer acvity Mode/desnaon logsum Intersecon density Transit stop density Trip Distance Gradient Percent Class I & II bike facilies Percent wrong way Turns per mile Percent no sidewalks Auto & transit travel me Auto & transit cost Table 4-6. Variables included in Seattle tour-based models.

43 Figure 4-4. Mode choice in relation to walk-accessibility score—home-based work travel. Walk Score 100 300 500 700 900 1100 1300 All 100 300 500 700 900 1100 1300 All Auto 685 68 24 16 25 8 30 856 237 97 61 55 43 37 326 856 Transit 260 57 23 16 25 14 30 425 54 16 7 19 18 10 307 431 Walk 12 7 3 4 12 3 6 47 6 5 3 6 7 4 16 47 957 132 50 36 62 25 66 1,328 297 118 71 80 68 51 649 1,334 Auto 72% 52% 48% 44% 40% 32% 45% 64% 80% 82% 86% 69% 63% 73% 50% 64% Transit 27% 43% 46% 44% 40% 56% 45% 32% 18% 14% 10% 24% 26% 20% 47% 32% Walk 1% 5% 6% 11% 19% 12% 9% 4% 2% 4% 4% 8% 10% 8% 2% 4% Home Based Work (HBW) Origin Home Based Work (HBW) Desnaon Mode Shares Mode Shares Walk Score Walk Score y = -0.126ln(x) + 1.2663 R² = 0.8399 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Auto Mode Share y = 0.0766ln(x) - 0.0497 R² = 0.6348 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Transit Mode Share y = 0.049ln(x) - 0.2166 R² = 0.5669 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Walk Mode Share y = 0.903e-3E-04x R² = 0.6556 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Auto Mode Share y = 0.1143e0.0008x R² = 0.4996 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Transit Mode Share y = 0.0181ln(x) - 0.0586 R² = 0.2751 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Walk Mode Share

44 also shown in each graph, illustrating both a logarithmic rela- tionship in each curve and a high R2 value reflecting goodness of fit. Table 4-7 shows that projected walk mode split for home- based work trips increases from about 1% at the lowest walk-accessibility level at the origin to 14% at the highest accessibility location, while transit share also increases from 30% to 50% and auto share declines from 65% to 35%; at the destination end, the increase in walk share is somewhat less (3% to 9%), while transit again increases by 20 per- centage points and auto declines by 25 percentage points. For non-work travel, walk shares are higher overall and the increase with improvements in walk-accessibility are greater, particularly at the origin end: Walk share increases by 22 percentage points and transit increases by 16 percent- age points, while auto declines by 40 percentage points; at the destination end the effect is not quite as dramatic, with auto share dropping by only 18 percentage points while walk increases by only 3 percentage points and transit increases by 12 percentage points. Figure 4-5. Mode choice in relation to walk-accessibility score—home-based non-work travel. Walk Score 100 300 500 700 900 1100 1300 All 100 300 500 700 900 1100 1300 All Auto 1,342 134 42 39 47 10 43 1,657 777 331 153 87 95 40 174 1,657 Transit 68 18 8 5 6 8 16 129 16 12 9 6 17 3 66 129 Walk 146 28 14 22 30 4 30 274 111 57 20 14 21 7 44 274 1,556 180 64 66 83 22 89 2,060 904 400 182 107 133 50 284 2,060 Auto 86% 74% 66% 59% 57% 45% 48% 80% 86% 83% 84% 81% 71% 80% 61% 80% Transit 4% 10% 13% 8% 7% 36% 18% 6% 2% 3% 5% 6% 13% 6% 23% 6% Walk 9% 16% 22% 33% 36% 18% 34% 13% 12% 14% 11% 13% 16% 14% 15% 13% Home Based Non-Work (HBW) Origin Home Based Non-Work (HBW) Des na on Walk klaWerocS Score Mode edoMserahS Shares y = -0.158ln(x) + 1.6157 R² = 0.9539 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Auto Mode Share y = 0.0502e0.0011x R² = 0.4914 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Transit Mode Share y = 0.0891ln(x) - 0.3206 R² = 0.5847 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Walk Mode Share y = 0.9074e-2E-04x R² = 0.6389 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Auto Mode Share y = 0.0169e0.0018x R² = 0.8183 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Transit Mode Share y = 0.1197e0.0002x R² = 0.3762 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 1400 Walk Mode Share

45 Perhaps as important as the effect of walk-accessibility on walk mode share is the effect that higher walk-accessibility has on transit share, particularly at the destination end. This may be due simply to destinations being more walk accessible to transit users, but may also provide evidence that travel- ers are more likely to use transit if they do not have to be dependent on personal vehicles once they reach their primary destinations. The relationship between the walk-accessibility score and the patterns of land use is shown in Figure 4-6, which uses color-shading to represent the level of walk-accessibility for each of the MWCOG survey trip ends. The map shows clear patterns between the level of walk-accessibility and location in Arlington County, particularly highlighting the areas of high walkability along the Orange Line (Rosslyn-Ballston) corridor, in Crystal/Pentagon City, and in Washington, DC. When applying the walk-accessibility model, the basis shifts from the survey trip ends which were used to calibrate the models, to census blocks. The user defines the “study area” of interest, as well as the surrounding walk shed of oppor- tunities that can be reached by walking from the study area. The census blocks for the walkshed are identified, and their centroids become the reference points for model application. Walk-accessibility scores are computed for each block by accumulating the opportunities present in each block in the walkshed as represented by their employment or population, discounted by the network travel time between the respective blocks. This approach—both calibration of the base model and its application to a block-specified study area—has been com- piled into a custom spreadsheet program provided with the guidebook. Step-by-step instructions on its structure and use are provided in Section 5.4, illustrating how the model may be used for trip generation, distribution, and mode-split analysis. As part of the presentation, the model is applied to an actual setting in the Shirlington area of south Arlington County, where changes are made to both existing land use and the networks, and the results run through the model to exhibit WALC Score HBW Origin HBW Des na on Auto Transit Walk Auto Transit Walk <200 65% 30% 1% 85% 10% 3% 200 55% 37% 5% 79% 17% 5% 400 50% 43% 8% 70% 21% 6% 600 43% 45% 10% 67% 24% 7% 800 40% 47% 11% 65% 27% 7% 1000 38% 48% 13% 62% 29% 8% >1200 35% 50% 14% 60% 30% 9% WALC Score HBO Origin HBO Des na on Auto Transit Walk Auto Transit Walk <200 88% 2% 10% 88% 1% 12% 200 75% 8% 17% 81% 3% 13% 400 65% 12% 23% 79% 8% 14% 600 59% 15% 26% 76% 10% 14% 800 54% 16% 29% 74% 11% 15% 1000 51% 18% 31% 72% 12% 15% >1200 48% 18% 32% 70% 13% 15% Table 4-7. Mode split for HBW and HBO trips in relation to walk-accessibility score at origin and destination.

46 changes in overall walking levels—by origin-destination block pair and by trip purpose. A complete documentation of the development of the walk-accessibility model is also provided as Appendix 2 of the Contractor’s Final Report. Strategic Enhancements to TAZ Trip-Based Models The two techniques presented in the preceding sections represent new approaches to the analysis of pedestrian and bicycle travel demand. They will offer assistance in not only non-motorized travel demand analysis, but how the principles of accessibility are used to understand bicycle and pedestrian demand and how the demand can be influenced by changes to land use and the transportation networks. Many planners—particularly those in metropolitan or local planning agencies—may also be seeking near-term options for improving the capability of their existing regional forecasting models to do a better job of accounting for non-motorized travel. For this reason, a third research approach was devel- oped by NCHRP Project 08-78 to identify how conven- tional trip-based models might be enhanced to improve their sensitivity to land use and non-motorized travel. This research also took advantage of the special data resources developed in the Seattle area and used by the tour-based modeling team. Enhancing trip-based models, particularly to improve their sensitivity to differences in land use and accessibility factors, is not a new concept. The background research identified and reported on some such efforts, some of which were referenced in Table 4-2 (e.g., Durham and Buffalo) and can be reviewed in greater detail in Appendix 7 of the Contractor’s Final Report. The approach developed by NCHRP Project 08-78 incorpo- rates some similar methods, particularly in trying to bring walking and biking further along in the modeling than trip generation. However, there was also a deliberate effort to Figure 4-6. Walk-accessibility scores—illustrative values in Arlington County.

47 sensitize as many of the steps in the process to important land use effects as possible. The primary limitation posed by most trip-based models when trying to analyze non-motorized travel is the aggrega- tion inherent in the use of TAZs. Although this may be an acceptable simplification of detail when analyzing vehicle travel, it eliminates the very detail necessary to understand non-motorized travel. This detail applies to (1) the level and mix of activity within reasonable travel distance by walking or biking and (2) the accessibility provided by the respective transportation networks. Recently, many MPOs have updated their models using a finer-grained system of TAZs. This shift has resulted in zones now more the size of a census block group than a Census Tract, increasing the number of zones overall by a factor of 3 or 4 to 1. Although smaller, the block-group-sized zones are still much larger than the parcels, blocks, or walkable buffers featured in the previous two methods. However, the downsiz- ing provides more resolution and also opens the opportunity for including non-motorized modes in the trip distribution and mode-choice steps of the model. (The large scale of pre- vious TAZs allowed the assumption that most non-motorized trips would remain within the TAZ in which they originated.) Figure 4-7 illustrates the standard four-step modeling pro- cess, depicting how non-motorized travel is generally accounted for and shows where the NCHRP Project 08-78 enhancements were targeted. The boxes in the Enhanced Approach highlight those steps where new relationships were developed, largely by drawing on the rich database of land use characteristics developed by PSRC using parcel buffering methods. The database of available land use (also presented as built environment, or BE) measures included the following • Number of persons and households (¼ mile buffer); • Employment (# jobs) by type (¼ mile buffer); • Parking supply: daily and hourly paid spaces, free off-street spaces (¼ mile buffer); • Parking cost: average daily or hourly cost (¼ mile buffer); • Street grid: # of dead ends, 3-way and 4+ way intersections (½ mile buffer); • Distance to transit: nearest express bus stop, local bus stop (miles); • Bus stop density: number of express, local stops (¼ mile); and • General home location indicator: urban, suburban or rural. In addition, two key measures of accessibility were developed (for each TAZ) and had important roles in the new models: • Single-occupant vehicle accessibility index (SOV AI): created from the logsum (denominator) of the destination choice model, based on network distance to destination, distance (destination) to the central business district (CBD), travel time, and log of jobs at the destination. • Non-motorized accessibility index (NMT AI): similar to the SOV AI as a logsum value, based on network distance to destination, land use mix at destination, and log of jobs at the destination. The following deficiencies were targeted, along with a description of the approach used to enhance the process. (The actual models are too voluminous to present here, but are avail- able for viewing, along with the corresponding elasticity esti- mates in Table A-2 in Appendix A.) Figure 4-8 illustrates where in the process the enhancements were made and what variables were used in each. Vehicle Ownership: Although vehicle ownership is not one of the official steps in the four-step model, it has an important role in both trip generation and in mode choice (in many mod- els). Because research shows that households residing in set- tings more transit and walk friendly own fewer vehicles, more Figure 4-7. Modifications to four-step trip-based model to improve non-motorized travel estimation. Standard Approach Auto Ownership Trip Generation Trip Distribution Mode Choice Assignment Enhanced Approach Auto Ownership Motorized Trip Generation Intra Zonal Mode Choice Assignment B/W Trips Discarded Inter Zonal NMT Trip Generation Mode Choice = New step, LU Sensitive

48 regional models are beginning to incorporate context factors when predicting household vehicle ownership (see Atlanta, Austin, Los Angeles, and Portland examples in Table 4-2). The Seattle research developed (1) a new vehicle owner- ship model using a Poisson regression approach (well suited for modeling “counts”) to predict the number of household vehicles based on household sociodemographic characteris- tics (number of members, workers, drivers and income) and (2) the following land use context measures: • Rural home location • Non-motorized mode accessibility index • Distance to nearest bus stop and bus stop density • Home TAZ population density • 4-way intersection density. Non-Motorized Trip (NMT) Generation: Most trip-based models have improved their procedures for estimating trip productions and attractions, moving from simple cross- classification procedures to models more tied to important context factors. Examples were listed earlier in Table 4-2 (Atlanta; Austin; Portland; Durham, NC; and Buffalo). How- ever, even the best of these models still makes a rough esti- mate of NMT productions by trip purpose and then removes those trips from consideration in the remaining steps of the modeling. The Buffalo approach (Wang, et al., 2010) is an exception and has similarities to the approach used in Seattle. For Seattle, the research team used a two-step approach using a binary logit model to first predict whether a household would make any NMT trips at all, followed by a negative bino- mial model that then predicted the number of NMT trips for households that make them. Both models incorporate land use variables, including parking availability and cost, intersection density, home TAZ density, bus stop density, and both auto and NMT accessibility indices for the home TAZ. The two-step approach was used in lieu of calculating NMT productions as part of the base trip generation process, which was left to focus on motorized trip generation. Vehicle Ownership Enhance to account for Built Environment SEDs: HH Size, workers, drivers, income Res BE: Rural resid, NMT accessibility (TAZ), density (TAZ), dist to bus stop, number bus stops (1/4 mi.), 4 way nodes (1/2 mi.) NMT Trip Generaon Expand exisng trip generaon calculaon SEDs: HH vehicles & income, Indiv gender and drivers license Res BE: SOV accessib & NMT accessib (TAZ), entropy (TAZ), dist to bus stop, hourly parking $ (1/2 mi), number bus stops (1/4 mi.), dead ends & 3 way nodes (1/2 mi.) Intra vs. Inter zonal Dest Choice Pre mode choice to factor role of distance SEDs: HH Size, workers, vehicles, drivers, income Res BE: SOV accessib & NMT accessib (TAZ), density (TAZ), free parking spaces &, hourly parking $ (1/2 mi), number bus stops (1/4 mi.), 3 and 4 way nodes (1/2 mi.) Intra zonal Mode Choice (DA, SR, WK) HBO NHB SEDs: Indiv gender, age and student Res BE: NMT accessibility (TAZ), density & entropy (TAZ), hourly parking $ (1/2 mi), dead ends, 3 and 4 way nodes (1/2 mi.) SEDs: Indiv gender, age and student Res BE: free parking spaces & hourly parking $ (1/2 mi), 4 way nodes (1/2 mi.) Inter zonal Mode Choice (DA, SR, TR, WK, BK) HBW HBO NHB SEDs: Indiv gender, age and student Res and Dest BE: free parking spaces & hourly parking $ (1/2 mi), 4 way nodes (1/2 mi.) SEDs: Indiv gender, age and student Res and Dest BE: free parking spaces & hourly parking $ (1/2 mi), 4 way nodes (1/2 mi.) SEDs: Indiv gender, age and student Res and Dest BE: free parking spaces & hourly parking $ (1/2 mi), 4 way nodes (1/2 mi.) Figure 4-8. Sociodemographic (SED) and built environment (BE) variables used to enhance Seattle trip-based model.

49 Intra- Versus Interzonal Destination Choice: The typi- cal trip-based model does not carry non-motorized trips beyond trip generation. With the downsizing of TAZs, greater opportunity exists to begin to include non-motorized trips into the destination choice and mode-choice determinations. To exploit this opportunity, the Seattle research inserted a procedure to predict whether non-motorized trips would be made to destinations in the same zone as the origin (intra- zonal) or to other zones (interzonal). The new model predicts whether an NMT trip production will travel to a destination within the origin zone or to another using a binary logit model incorporating sociodemographic characteristics (i.e., vehicle, income, drivers, and gender), an array of land use measures (i.e., SOV and NMT AI indices, land use mix, distance to bus stop, bus stop density, dead ends and 3-way intersections, and parking price), and trip purpose (commute) and time of day (i.e., mid-day, PM peak, evening). The earlier mentioned Buffalo study uses a similar approach. Mode Choice: Non-motorized modes typically do not prog- ress to the mode-choice step, but with the separation into intra- zonal and interzonal trip types this becomes possible. Because of very low bike and transit shares, the intrazonal model includes only three modes: drive-alone, shared-ride, and walking. The interzonal model has a similar specification, but includes five modes (i.e., drive-alone, shared-ride, walking, transit, and bicy- cling). Both interzonal and intrazonal models include indi- vidual models for home-based work, home-based other, and non-home-based travel. Key land use context variables in the intrazonal mode-choice models were NMT Accessibil- ity, intersection density, land use mix, density, and parking availability and price. For interzonal, the key context vari- ables were SOV accessibility at the origin and both NMT and SOV accessibility at the destination; density at both ori- gin and destination; intersection type and density at both origin and destination; bus stop density at origin and desti- nation; and land use mix at the destination. Destination Choice models for interzonal home-based work, home-based other, and non-home-based trips were esti- mated using a multinomial logit approach, with the weighted logsums across the five modal alternatives (aggregate accessi- bility) serving as key explanatory variables along with inter- sections, density, land use mix, distance to transit and transit stop density, and parking availability. The importance of the land use variables was further articulated in relation to key demographic segments (i.e., gender, senior citizen, income, and licensed driver). These enhancements and the methods used to create them will be useful for planners and agencies that want to make near-term improvements to existing trip-based models. The equations and elasticities are provided in Chapter 5 to assist users who wish to explore these methods further. As with any of the models offered by this research, however, important caveats should be observed in working with these tools: • The model coefficients and elasticities were derived using data from Seattle, and so should be used with caution in terms of direct transferability. • When calculating the non-motorized accessibility measure, walk and bike are combined into a single mode, which may exaggerate the level of accessibility, given the longer range of bicycle travel. • Although reduced TAZ size was an important factor enabling this analysis, the research models do not directly account for TAZ size in the models or measures, when it is likely that zone size could be an important contributing factor in determining the extent to which a trip is intra- versus interzonal. Despite these caveats, practitioners working with trip- based models may wish to build on or emulate this approach. Pedestrian Trip Generation and Flow Models The NCHRP Project 08-78 research team reviewed existing models that estimate pedestrian trip generation and assign those trips to facilities. Although they are not full choice- based models in the sense of deriving walk trips from a com- prehensive trip generation and mode-split process, they do offer an approach that employs accessibility principles to account for the combined effects of land use and network connectivity. Two models in this group—MoPeD and PedContext—have common lineage. The original model was the PedContext tool, developed under contract for the Maryland Department of Transportation (State Highway Administration) by the Uni- versity of Maryland for estimating pedestrian flows to support safety analyses (Urbitran Associates, 2004). http://smartgrowth. umd.edu/assets/cliftondaviesallenraford_2004.pdf The model was applied and validated in downtown Baltimore and Langley Park in suburban Washington, DC. MoPeD, described next, is a descendent of PedContext, and carries many of its characteristics but at a much reduced level of detail, which may offer a simpler option for some users or applications. The models have a structure familiar to the four-step trans- portation models, performing trip generation, distribution, and network assignment. However, these models concentrate solely on pedestrian trips and do not attempt mode choice. They also operate at a pedestrian scale of detail, substituting block-size pedestrian analysis zones (PAZs) for TAZs. Neither of these tools addresses bicycle travel, although except for suf- ficient bicycle data from travel surveys, there appears to be no obvious reason why the structure of either model could not accommodate bike as a mode.

50 PedContext Model PedContext is the more detailed of the two pedestrian models. It features a land use allocation step, pedestrian travel generator, a distribution module, and a stochastic assignment procedure to allocate the estimated pedestrian trips to the walk network. The steps in setting up the model are as follows: • Networks: A detailed street network was created from Cen- sus TIGER files, enhanced to account for sidewalk coverage (using aerial data) and characteristics important to walking (i.e., functional class of roadway, speed limits, volumes, and traffic control devices). Each link is assigned nodes at the end points, plus one in the center to serve as a mid-block crossing (i.e., jay-walking) opportunity, subject to various conditions. These nodes are later treated as “load points” when assigning trips to the pedestrian network. • Land Use Allocation: Parcel-level land use data available through Maryland’s “Property View” GIS database was coupled with Census data to reflect land use activity at each block face. • Trip Generation: Walk trip productions and attractions for seven different trip purposes were estimated for each block face from a set of equations developed using travel survey data from the New York metropolitan area; the trip genera- tion models were distinct in including innovative, purpose- specific land use accessibility measures. • Trip Distribution: Walk productions and attractions were converted to trips by purpose using a gravity-based trip distribution model in which a distance-decay relationship based on a gamma function was used to calculate the travel time impedance. • Trip Assignment: The estimated walk trips for each block face were associated with the nodal load points (described above) and then assigned to the sidewalk network by pur- pose and time of day using the weighted impedances and a stochastic, multi-path assignment algorithm (see Fig- ures 4-9 and 4-10). The PedContext model was developed using a combination of tools, including ArcGIS and CitiLabs CUBE and VIPER transportation planning software, with specialized routines written by the model development consultant to coordinate the various elements. This model has many features that could make it attractive to pedestrian planners; however, it is not in the public domain. Interested parties can contact the Maryland State Highway Administration or the University of Maryland National Center for Smart Growth for more information on its potential availability and use, either through acquisition of the actual software or by attempting to emulate the methods, which are detailed in Table A-4 of Appendix A. Model of Pedestrian Demand (MoPeD) The MoPeD model was also developed by the University of Maryland National Center for Smart Growth, as a some- what less complex and computationally demanding version of the PedContext model, but with an open-source software program. MoPeD also can estimate pedestrian activity levels at intersections at a subarea scale using readily available data in a GIS framework. Figure 4-9. PedContext multi-path assignment. Figure 4-10. Assigned pedestrian volumes.

51 MoPeD is similar to PedContext in the following ways: • Census TIGER network line files were enhanced to represent the full pedestrian network, accounting for the connectivity and impedances associated with sidewalks and crosswalks. • The spatial units of analysis are PAZs, which are in the form of blocks and block faces. • Like PedContext, MoPeD directly estimates pedestrian trips, rather than deriving them from a mode-choice analysis. Important differences with PedContext are as follows: • MoPeD focuses only on home-based and non-home-based walk trips, versus PedContext’s inclusion of six purposes. • Relationships were drawn exclusively from local data, using NHTS add-on travel survey data collected from the Baltimore region. • Unlike the detailed equations used in PedContext, trip gen- eration is a simpler function of vehicle ownership, street con- nectivity, residential development, and commercial mix. For non-home-based walk trips, correlated variables were retail, service, and other employment and housing within ¼ mile buffer of the trip end. • Walking trips were distributed and routed among produc- ing and attracting PAZs via a walk-distance gravity model and shortest-path assignment (i.e., not a detailed stochas- tic multi-path assignment as in PedContext). • MoPeD runs on a GIS platform with open-source analysis routines intended for use by planners and analysts without proficiency in regional travel models. Both MoPeD and PedContext can be used for various plan- ning purposes, to estimate walking under different land use and pedestrian network configurations. They support impact analysis of new or infill development and changes to the pedes- trian network (e.g., adding sidewalks, improving connectivity, or removing access). The features in MoPeD for creating and editing networks, processing land use and population and employment data into block-size units, and performing trip generation, distribution, and assignment are complete and well documented. Many bike/pedestrian and land use planners may find this model useful (see http://kellyjclifton.com/MoPeD/DemandModel Protocol07_08.pdf). Figure 4-11 illustrates the Baltimore City study area to which MoPeD was applied and the estimated 24-hour pedestrian counts by intersection. Portland Pedestrian Model A third pedestrian demand estimation model is included here—because of its lineage with PedContext and MoPeD and because it offers another potentially useful approach for enhancing the capabilities of regional trip-based models. Researchers at Portland State University (PSU) were contracted by the regional MPO, Metro, to develop a procedure to improve the pedestrian mode-choice capabilities in Metro’s existing trip- based model. The lead researcher also led development of the MoPeD model at the University of Maryland. The resulting procedure can either be used as an enhancement to the regional model or as a stand-alone pedestrian planning tool. Like MoPeD, the Portland pedestrian model approach uses PAZs as the analysis unit. The Portland PAZs were formed by disaggregating the regional TAZ system into 1.6 acre (264 × 264 feet) grid cells. The steps in the modeling procedure are pictured in Figure 4-12. The procedure first estimates total person trip generation for each PAZ, using Metro’s existing trip generation procedure. Metro estimates only trip productions, because attractions are identified through a destination choice model. Next, a set of binary logit walk models is used to separate the estimated trip productions into walk and non-walk for three purposes: home-based work, home-based other, and non-home-based. An important variable in these equations is a pedestrian index of the environment (PIE), a weighted sum of six different contextual variables. Metro developed its “Context Tool” to represent land use and other urban form contextual variables in its regional model. The standard Context Tool consists of the following measures: • Bicycle Access—Density of bicycle network links within a 1-mile radius, weighted by classification (e.g., off-street paths/trails, main bikeways, bike lanes, low/moderate/high traffic streets with no bicycle facilities). • Block Size—Block-size density within a ¼ mile radius. • Activity Density—Population and employment density within a ¼-mile radius. • Sidewalk Density—The percentage of road segments with sidewalks, weighted by continuity, within a ¼-mile radius. • Transit Access—The density of bus, light rail, and com- muter rail stops, weighted by service frequency, within a ¼-mile radius. • Urban Living Infrastructure—Grocery stores, cafes, restau- rants, clothing and other retail stores, schools, dry cleaners, and entertainment venues within a ¼-mile radius. Compiling all the measures into a single index proved an effective strategy for overcoming multicollinearity problems when using these variables. Developers of the pedestrian model augmented the standard Context Tool by applying weights to the individual components to reflect their dif- ferential importance in impacting the walk decision. Binary logit models were used to establish the importance levels pre- sented in Table 4-8. The analysis determined activity density

52 to be the highest-weighted attribute, followed by transit access. If each attribute were to realize its maximum value (5) in the given setting, the maximum weighted value would appear as shown in the final column. The PIE variable was found to be an important indicator of environmental context in the pedestrian mode share model. PIE values for all grid cells in the Portland region demonstrated the highest values in Downtown Portland, followed by other major neighborhood centers, then suburban centers, with the lowest values in isolated industrial, rural, and undeveloped areas. Figure 4-13 shows the values predicted for different areas in the region, accompanied by a picture conveying the “feel” of these areas in relation to the PIE score. The third step in the pedestrian model process was to match the pedestrian trip productions into origin-destination trips across the study area of PAZs, which was done using Metro’s destination choice model instead of distribution. The result- ing trip tables (by purpose) were then assigned to facilities in the network, although the current model does not perform that task. Boundary of Balmore City Study Area Predicted Intersecon Volumes Using MoPeD Model Figure 4-11. Application of MoPeD model in Baltimore City.

53 Once the pedestrian trip tables were determined, non- pedestrian trips could then be aggregated up to the TAZ-level trips and passed to the regional model for further analysis. This is very similar to how the walk-accessibility model developed for Arlington operates. Facility-Use Estimation Models This group of tools predicts user volumes or activity lev- els on bicycle or pedestrian networks for purposes of network design, assessment of sufficiency or potential improvements, or crossing volumes for safety studies. The difference in this group from the PedContext and MoPeD approaches is that they are not fully integrated approaches that estimate demand from a top-down process, but attempt to explain existing activ- ity levels or patterns with characteristics of the existing envi- ronment and then project changes in activity based on changes in the context factors. Three types of tools are included in this category: • Route choice models, • Network simulation models, and • Direct demand models. Route Choice Models Bicycle Route Choice There has been considerable research on quantifying the factors underlying bicyclist choice of route, resulting in insights on how physical factors (e.g., directness, facility type, slope, and traffic exposure) influence choice of route. By quantify- ing the importance of these characteristics in relation to travel time (or distance), it becomes possible to express the utility of choosing alternative paths based on their packaging of these characteristics. The best examples of models created for this purpose are those developed by the San Francisco County Transporta- tion Authority (SFCTA) and Portland State University (PSU), which used GPS recording methods to obtain data on actual route selection behavior. This differentiates them from simi- lar research studies that relied exclusively on stated preference information, although those studies (e.g., Hunt and Abraham, Krizek, Menghini in Table 4-2) also provide interesting and useful insights on these values and tradeoffs. Other research can be reviewed in Appendix 7 of the Contractor’s Final Report. The SFCTA model, shown in Table 4-9, accounts for distance, turns, slope, wrong-way links, path size, and proportion of Figure 4-12. Portland pedestrian model. Component Possible Values Weight MaximumWeighted Value Bicycle access 1 to 5 2.808 14.04 Block size 1 to 5 3.086 15.43 Acvity density 1 to 5 4.615 23.07 Sidewalk density 1 to 5 2.842 14.21 Transit access 1 to 5 3.529 17.65 Urban living infrastructure 1 to 5 3.120 15.60 Total 100.00 Table 4-8. Estimated importance weights for PIE index.

54 Class I, II, and III facilities in explaining choice of route. It also accounts for different trip purposes (work versus non-work) and gender in explaining the importance of particular features, which earlier research has shown to be fairly important in understanding bicyclist behavior. Also shown are the marginal rates of substitution (MRS), signifying the relative importance of each characteristic in relation to trip length. For example, the average cyclist would avoid a turn if it costs no more than 0.17 km and will avoid climbing a hill 10 m tall as long as the detour is less than 0.59 km. Similarly, a cyclist will not travel the wrong way down a one-way street unless doing so saves more than four times the distance (or its equivalent in turns or hill climbing) elsewhere. On the other hand, the average cyclist is willing to add a mile on bike lanes in exchange for only ½ mile on ordinary roads. The PSU model uses similar explanatory variables, but includes a provision to account for the effects of adjacent vehicular traffic volumes, as well as cyclist wait times at cross- ings. The PSU research obtained information on user type, but these factors were not found to be significant in the esti- mated models. The full PSU model is presented in Table A-7 of Appendix A, and the relative value of the route character- istics (similar to the SFCTA marginal rates of substitution) is provided in Table 4-10. In considering use of these models, it is important to be aware of their application limitations. They provide invaluable Figure 4-13. Illustration of pedestrian index of environment (PIE) in Portland region.

55 information on how these different characteristics are weighed by the traveler by converting those preferences into quantita- tive factors influencing perceptions of travel time or distance, so, if the planning question is to determine what improve- ments would make one path better than another, these tools would be directly relevant. However, these tools do not attempt to predict whether a bike trip will be made, which destination will be chosen over another, or whether the bike mode will be chosen over another for that destination. Network Simulation Models Another approach used to project route choice is through a spatially-driven network simulation procedure (e.g., Space Syntax). Space Syntax was developed in London in the 1980s and has been widely used in Europe for pedestrian planning. It has been used only marginally in the United States, for at least two reasons: (1) the software is proprietary, hence there is not a lot of freely available information on how it works; and (2) its process is not instantly intuitive to traditional transportation planners. Space Syntax does not cast travel flows in the context of trip generation and distribution in a conventional sense, but uses spatial characteristics and relationships to try to explain how particular paths will be chosen. The underlying assump- tion is that travel patterns in a network are not necessarily determined by individuals minimizing travel time or distance, Aribute Coefficient t stat. Length (km) 1.69 11.8 Turns per km 0.13 12.15 Proporon wrong way 13.5 19.87 Proporon bike paths 1.89 6.17 Proporon bike lanes 2.15 17.69 Cycling freq. < several per week 1.85 44.94 Proporon bike routes 0.35 3.14 Average upslope (m/100m) 0.50 6.35 Female 0.96 4.34 Commute 0.90 8.21 Path size (log) 1.07 26.38 Number of observaons 2.678 Null log likelihood 10,006 Final log likelihood 7,123 Adjusted rho square 0.23 Marginal Rate of Substuon (MRS) MRS of Length on Street for Value Units Turns 0.17 Km/turn Length wrong way 4.02 None Length on bike paths 0.57 None Length on bike lanes 0.49 None Length on bike routes 0.92 None Total rise 0.59 km/10 m Table 4-9. SFCTA bicycle route choice model and marginal rates of substitution. Aribute Distance Value (% distance) Non Commute Commute Turns per mi. 7.4 4.2 Proporon upslope 2 4% 72.3 37.1 Proporon upslope 4 6% 290.4 120.3 Proporon upslope > 6% 1106.6 323.9 Traffic signal exc. right turn (per mi) 3.6 2.1 Stop sign (per mi) 0.9 0.5 Le turn, unsig. AADT 10 20k (per mi) 16.2 9.1 Le turn, unsig. AADT 20k+ (per mi) 43.1 23.1 Unsig. cross AADT > 10k right turn (per mi) 6.7 3.8 Unsig. cross AADT 5 10k right turn (per mi) 7.2 4.1 Unsig. cross AADT 10 20k right turn (per mi) 10.4 5.9 Unsig. cross AADT 20k+ right turn (per mi) 61.7 32.2 Prop bike boulevard 17.9 10.8 Prop bike path 26.0 16.0 Prop AADT 10 20k w/o bike lane 22.3 36.8 Prop AADT 20 30k w/o bike lane 137.3 140.0 Prop AADT 30k+ w/o bike lane 619.4 715.7 Bridge w/ bike lane 29.3 18.2 Bridge w/ separate bike facility 44.9 29.2 Table 4-10. PSU bicycle route choice model—relative rates of substitution.

56 but in terms of “transitions” from one space to another. The approach requires coding of a detailed network, which is then treated as a “graph.” Topological methods are used to charac- terize the properties of the network (graph) through such mea- sures as connectivity (number of other nodes that connect to each node), depth (average number of steps between nodes), and integration (ease of access from other nodes). Integration is the key variable, whose formula compares an ideally con- nected graph with the one in question to determine a measure of accessibility for each node in the network. The quantified measures of accessibility and connectivity are then used to gen- erate movement “potentials,” which are then correlated with counts. The correlations are then used to predict volumes on a street-by-street basis for the defined study area. Illustrative tests of Space Syntax in the United States have occurred in the City of Oakland, CA, for pedestrian planning (Raford and Ragland, 2003) and in relation to bicycle travel in Cambridge, MA (McCahill & Garrick, 2008). In the McCahill & Garrick example, the correlation of Space Syntax measures and observed bicycle volumes in the Cambridge, MA, bicycle net- work was tested. The “choice” segment indicator was used as the means of predicting relative cyclist volumes on facilities, using road centerline maps in place of the traditional “axial maps,” and ArcGIS to compile information on segments from spatial analysis and census statistics. A linear regression was developed to reveal the best correlation between existing bike volume counts at 16 intersections, census population, and employment data to serve as productions and attractions, plus various Space Syntax measures. The researchers determined that the method was useful in predicting bike volumes in a network and could be useful in designing more efficient networks. In the City of Oakland, Raford and Ragland used Space Syntax to forecast pedestrian volumes for safety analysis in the City’s pedestrian master plan. Space Syntax was used to leverage existing count data from a sample of 42 inter sections into forecasts of pedestrian volumes at 670 intersections city- wide. However, because Space Syntax assumes an even popu- lation distribution, the researchers supplemented the model by using Census population and employment data to allow for distortions caused by major generators. Discrepancies in forecasting accuracy (remaining after the adjustments) included a tendency to underestimate volumes on high- volume streets and on streets connecting to three Bay Area Rapid Transit (BART) stations. However, the researchers believed that additional enhancements (e.g., including auto volumes and speeds and using more specific land use charac- teristics) could help improve accuracy. Because of the lack of clarity in how Space Syntax works and that it is proprietary, it has not been possible to fully evalu- ate Space Syntax’s capabilities, so it is not included in the best- practice recommendations. However, users can investigate further if the features of the tool seem interesting or useful. Direct Demand Models Direct demand models have been the accepted practice for estimating pedestrian or bicycle facility demand for some time. The NCHRP Project 08-78 background review recorded use of these methods back in the 1970s (Benham & Patel, 1977). Their structure is to explain observed levels of bicycle or pedes- trian activity on facilities (links) or at intersection (points) as recorded through counts, using a range of factors that describe local context. This is usually done using regression modeling techniques, with the calibrated models then applied back on all or a subset of the sampled system of intersections or links to assess their accuracy in replicating choices. Variables often used to represent context in these types of models include the following: • Population or employment densities, sometimes differenti- ated by type (e.g., populations differentiated by age, gender or income, or employment categorized as office or retail). • Population or employment activity levels within a nominal buffer distance of ¼ or ½ mile from the intersection. • Land use mix, measured either through an index (e.g., entropy) or implicitly through corresponding buffered activity levels. • Characteristics of the facility, including type of bike path and sidewalk existence and sufficiency. • Interaction with vehicle traffic (e.g., adjacent speeds or vol- umes, intersection approaches with crosswalks, sidewalk widths, on-road versus off-road bike facilities). • Transit availability (e.g., transit frequency and stop density). • Major generators (e.g., proximity to universities, schools, recreation, neighborhood shopping, major transit centers, and civic centers). Numerous examples of models in this genre are cited in Table 4-2 and documented in Appendix 7 of the Contractor’s Final Report under the Aggregate Demand Methods discussion. Because each is unique, it is difficult to name one or two that are exemplary; however, among those that have undergone the most development and had access to the best data resources are the Seamless Travel pedestrian and bicycle models developed by Alta Planning & Design in San Diego (Jones, et al., 2010) and the Santa Monica pedestrian and bicycle demand models (Fehr & Peers, 2010). Seamless Travel Models In the Seamless Travel study, pedestrian and bicycle models were developed to predict approach volumes at intersections during the 7 to 9 A.M. period on weekdays. Manual counts from a sample of 80 intersections supported the analysis. Counts were supplemented with traveler intercept surveys at 25 locations to obtain additional data, although the surveys did not iden- tify the type of trip in progress.

57 The Seamless pedestrian model is of the following form: P 1.555 0.723 ED 0.526 PD –1.09 R R 0.516AM 2( )= + + = where PAM = Morning peak pedestrian count ED = Employment density within 0.5 mile PD = Population density within 0.25 mile R = Presence of retail within 0.5 mile So the model predicts that A.M. peak-period walk trips will increase in proportion to adjacent employment and popula- tion density and decrease in the presence of retail activity. Even though these are probably work-related trips, given the time of day, it is not immediately clear why retail activity would have a negative effect on walk trip levels. Employment density carries a higher coefficient than population density, again presumably related to these being primarily work trips, although the buffer radii are different for population and employment and elastici- ties were not provided. The Seamless bicycle model has the following form: B 4.279 0.718 C 0.438 ED R 0.439AM 2( )= − + + = where BAM = Morning peak bike trips C = Footage of Class I bicycle path within 0.25 mile ED = Employment density within 0.25 mile This bicycle model predicts an increase in bike trips based on higher employment density and greater presence of Class 1 bikeways within ¼-mile of the count site. Santa Monica Models The pedestrian and bicycle models developed by Fehr & Peers for Santa Monica predict volumes for the 5 to 6 PM peak hour. The pedestrian model has the following form: P 222.18 0.00321 ED 3.675 BF 82.695 SDP – 0.00685 DO – 5.699 SL R 0.584 PM PM 2( ) = + + + = where PPM = Evening peak pedestrian volume ED = Employment density within 1⁄3 mile BFPM = PM bus frequency SDP = Intersection is within shopping district DO = Distance from ocean SL = Average speed limit on approaches This equation predicts that PM peak-period walk trips will increase in proportion to adjacent employment, with higher rates of PM bus service, and if the intersection lies within a shopping district. This equation predicts that PM peak-period walk trips will decline with increased distance from the ocean and with higher adjacent auto speeds. In contrast to the Seam- less Travel pedestrian model, this model sees a positive effect from retail proximity, which may be due to a higher proportion of non-work trips occurring during the PM peak. The Santa Monica bicycle model has the following form: B 1.317 0.120 Ln ED 1.632 MXD 0.431 BN 0.523 INT-4 R 0.401 PM 2( ) = + + + + = where BPM = Evening peak hour bike trips Ln ED = Log of employment density within 1⁄3 mile MXD = Land use mix within 1⁄3 mile BN = Proximity to bike routes (intersection is along a bike route or at the intersection of two bike routes, with higher weighting going to better classes of bike facilities) INT-4 = Four-legged intersection This equation predicts an increase in bike trips based on higher employment density, mixed land use, proximity to bike routes, and if the intersection is four-way. The appeal of these models lies in their simplicity and cus- tom quality. Although not easy to construct, they do not require advanced transportation modeling skills and are fairly easy to understand and apply. Aside from the activity counts, most of the data used to construct the context variables are generally available, and model builders are often resourceful in designing the models to use the data that they have. The caveat with these models is that they trade directness and simplicity for behavioral structure. In effect, they try to explain/ predict an aggregate quantity—activity counts in a particular time period—with factors descriptive of the surrounding envi- ronment. What results are relationships that may display strong correlations with the activity variable, but cannot be readily shown to “cause” the behavior represented in the counts (which is itself an amalgam of travel activity). What the NCHRP Project 08-78 research has shown is that accessibility is the most significant determinant of choice, par- ticularly for non-motorized travel, and representing accessi- bility requires a deliberate effort to simultaneously account for both the opportunities presented through the land use and the ease and efficiency with which the modal networks connect the traveler with these opportunities. It is difficult to apply this relationship in count-based models given that the modeled intersection or link is neither a trip production nor attraction. Therefore, this guidebook suggests that use of these models should be judicious in how they are developed and when they are used. The following guidelines are suggested: 1. None of these models should be construed as transferrable. Their coefficients are unique to how the models have been

58 specified (variables included) and the specific location for which they were developed. If an existing model presents an appealing structure, the user is advised to re-estimate the model(s) using identical data for the new study area. 2. The user needs to be aware of the uncertainties associated with modeling “count” data. In almost all cases, the models are blind to the travel behavior represented by the counts (e.g., the purpose of the trip, the sociodemographic char- acteristics of the traveler, the origin-destination of the trip, and the existence of alternatives). Focusing the counts and models on a particular time period (e.g., A.M. weekday peak for work or mid-day weekend for recreation) can nar- row the uncertainty as to the types of trips being observed, but, for other time periods, the mix of trips being modeled may be difficult to surmise. 3. Once the models are calibrated, the user should test their reliability in predicting activity at individual locations and overall for the study area. Although most of the models reviewed have R2 values of 0.5 or better, they may not be particularly accurate at the level of the individual inter- section or link. The Seamless Travel study experimented with methods to adjust the base estimates to account for unusual circumstances (that cannot be directly included in the model), and it may prove worthwhile to review and consider emulating these methods (see http://www.altaplanning. com/caltrans+seamless+study.aspx). 4. Be judicious in the types of applications or decisions to be supported by the models. For example, if measures of net- work connectivity are not included in the model structure, it would be misleading to estimate demand for a new or improved facility without recognizing that some portion of the new demand predicted may simply be a diversion from some other facility. At the same time, a network improve- ment that contributes to overall network connectivity may well induce new travel on other portions of the network. Given the above, it is recommended that the direct demand tools be reserved for either quick estimates or screening in advance of more comprehensive analysis, or for incremental extrapolations from an existing situation. Regardless, the fore- cast effort should be within the bounds of the explanatory variables in the model and not be used for forecasting new demand or changes within a network. For these types of appli- cations, the user is advised to apply one of the earlier choice- based tools (e.g., the GIS-Accessibility, MoPeD, PedContext, or even the Portland Pedestrian model approach).