Activity-Based Travel Demand Models: A Primer (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.

17 2.1 DEVELOPMENT PROCESS Modeling managers must consider numer- ous factors when developing an activity-based model system. They must evaluate how model designs and specifi cations can address policies and projects being considered and make trade- offs between model capabilities and develop- ment costs and schedule. In addition, they must confront data development challenges and ad- dress model implementation and maintenance requirements. The purpose of this chapter is to provide an overview of activity-based model development approaches that agencies have followed and to summarize the primary steps in the activity-based model development process, which include model design, data development, model implementation, and model application and maintenance. 2.1.1 Approaches Regional and state transportation planning agencies have generally followed one of three model development trajectories: (1) upfront devel opment, (2) incremental development, and (3) transfer and refi nement. Each approach offers different advantages and disadvantages, and the identifi cation of the best development path is highly dependent on each individual agency’s analysis needs and available resources. In addition, model development often com- bines elements of these three different trajec- tories. For example, an agency may choose to transfer an existing model in order to leverage the results of other agencies and then develop substantial new features in order to achieve their own important objectives. 2.1.1.1 Upfront Development This approach refers to an implementation process in which an agency develops an en- tirely new activity-based model system, largely independent of any prior trip-based or other travel demand model efforts by the agency. When developing a new activity-based model system upfront, an agency may gather new household survey data and other data describ- ing travel behavior, design and estimate new component models such as destination choice or mode choice, develop new network integra- tion processes and information such as skims (measures of network impedances) by detailed mode or detailed time of day, and calibrate and validate the model system. The result is a 2 TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM

18 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER model system designed specifically to address the concerns and analytic needs of the partic- ular region. Advantages of this approach are that the entire activity-based model system is designed and implemented as a single effort; the coherence of the overall model system design is enhanced and the development can include features important to the agency; and the overall amount of time required to have an operational, calibrated mode may be reduced. The primary disadvantage is that this approach requires significant upfront resources. Many of the first activity-based model efforts followed this development process. 2.1.1.2 Incremental Implementation Rather than design, estimate, and implement an entirely new model system as part of a single effort, an incremental implementation process involves the gradual development of activity- based model system components, either in concert with or parallel to trip-based model maintenance and enhancement efforts. A com- mon first incremental step is to implement a population synthesis component. A synthetic population is a key input to an activity-based model that can also be aggregated and used for input to a traditional trip-based model. Sub- sequently, a region might also implement an activity generator that uses the synthetic pop- ulation and replaces a typical trip generation process. Advantages of this approach are that it allows agencies to make gradual investments in developing an activity-based model and pro- vides an opportunity for agency staff to be- come familiar with disaggregate activity-based model inputs, outputs, and operation. It may also provide agencies with the opportunity to prioritize model investments that best address regional analysis needs. The primary disadvan- tage of this approach is that it may lengthen the schedule and increase the budget required for overall model development. 2.1.1.3 Transfer-and-Refine Implementation Recently a number of agencies have pursued a transfer-and-refine implementation strat- egy. This strategy involves taking an existing activity-based model developed for another region and reconfiguring it to a new region. The primary tasks required to transfer an exist- ing activity-based model are to develop new activity-based model inputs, such as develop- ing new socioeconomic input files and updated network supply information and procedures, reconfigure the software to use these new in- puts in conjunction with model parameters transferred from other regions, and recalibrate and revalidate the model system to the new region’s observed data. Advantages of this ap- proach are that the activity-based model can be implemented within 6 months, and there is no immediate need to re-estimate new model parameters. Also, the agency can get experience working with a fully functional activity-based model system and subsequently decide to refine it based on hands-on experience. Recent expe- riences and research suggest that this approach is both practical and statistically defensible ( Bowman 2004, Picado 2013). A drawback of this approach is that many of the coefficients in the model components are not based on local data. 2.1.1.4 Stakeholder Acceptance Developing an activity-based model requires the involvement of more than just agency tech- nical staff and consultants. It must also involve the stakeholders who will be using the model to assist in decision making. The stakeholders in- clude agency board members, executives, man- agers, planners, and other nontechnical staff, as well as outside constituencies. It is critical that the stakeholders have confidence in the tool, which can be most easily accomplished by providing transparency about model design

19 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM and inputs and outputs and by demonstrating the usefulness of the tool for analyzing real projects and policies. 2.1.2 Design Activity-based model design starts with an assess ment of an agency’s analysis needs in rela- tion to the policies and projects that are expected to be considered by the agency. This assess ment should consider not only the types of projects and policies to be evaluated using the model but also the specific performance measures to be produced by the model and a determination of the level of accuracy and precision that is re- quired. Of course, practical considerations such as the model development costs and schedule are also key design determinants. The design process should involve the technical managers who will guide model development and the con- sumers of the model outputs, such as planning project managers. 2.1.2.1 Analysis Needs Activity-based models can provide useful assess ments of the effects of different transpor- tation investment and policy alternatives, and the alternatives being considered help define an agency’s analysis needs. Required policy and project sensitivities are a primary concern. For example, activity-based models can provide robust insights into projects and policies that involve willingness to pay, such as road pric- ing and parking costs. Similarly, activity-based models can provide the ability to analyze poli- cies that involve coordination between individ- uals and time-sensitive scheduling constraints, such as telecommuting and compressed work schedules. Activity-based models are typically linked with static network assignment models and can produce the network-based perfor- mance measures that are used for project and plan development. But, activity-based models also produce a much broader set of measures than are possible using traditional aggregate approaches and can be used to support much more detailed and complex analyses, such as the effects on communities of concern. 2.1.2.2 Cost Cost is a critical and practical model devel- opment consideration and is influenced by a number of factors. If an agency wishes to intro duce new features not found in exist- ing activity-based model implementations, in order to address an important analytic need, it will likely result in higher development costs. Implementing new features often necessitates conducting applied research and developing new software code, which can be expensive. Conversely, adopting an existing activity-based model structure and software can reduce devel- opment costs. However, even if an agency de- cides to adopt an existing activity-based model platform, it may be necessary or desirable to estimate new parameters based on local data, which can increase costs. In addition, new costs may arise from the need to collect and maintain data, such as household travel behavior survey data, parcel or business data, or transporta- tion networks by detailed time of day. Potential costs associated with model application may also be considered. Finally, the costs for devel- opment may be significantly affected by the use of external consultants instead of, or in addi- tion to, agency staff. 2.1.2.3 Schedule Schedule is also an essential model develop- ment consideration that is influenced by many factors. If a new model system is required in order to fulfill immediate analytic needs, devel- opment schedules may be accelerated. Data availability influences model development schedules. Although some of the data from an agency’s traditional trip-based model may be repurposed for use in an activity-based model,

20 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER additional data such as household survey or traffic counts by detailed time-of-day may be required, and assembling this information takes time. Funding also influences schedules. Although activity-based model development costs have dropped significantly, agencies may not wish to or may not be able to fund large- scale model development efforts over short periods of time, which may lengthen the model development schedule. Anticipated model ap- plications may also affect the schedule. 2.1.3 Data Development A key stage in the activity-based model devel- opment process is assembling all of the data re- quired to implement the model system. These data may include household survey data, land use information, demographic information, transportation network data, and other urban form indicators. Depending on the specific model implementation approach selected, not all data items may be required. Development of a data collection plan or strategy, in coordi- nation with a model development plan, can be useful to ensure consistent data development and standards. Such a plan would inventory existing data sources and identify where addi- tional data or better data are required and may also consider how the timing of data collec- tion relates to the overall model development schedule. 2.1.3.1 Household Survey If an agency pursues a transfer-and-refine pro- cess, household travel survey data collected to support trip-based model development can usually be used to support calibration of the activity-based model. Additional analysis of the survey data is required to chain the trips into tours and to classify the sequence of tours into relevant descriptors of a full-day activity and travel pattern, but after that the use of the expanded survey data to compare against and calibrate the relevant coefficients in the models is analogous to their use in calibrating a 4-step model. The survey data are used pri- marily to modify relevant alternative-specific constants and various impedance parameters in the models so that the results of applying the activity-based model system match the ob- served, expanded choice distributions from the survey to an acceptable degree, as a precursor to model validation on external data. If an agency pursues an upfront develop- ment or incremental implementation pro- cess, the travel survey data requirements can be more stringent than for typical trip-based model work. Because activity-based models tend to consider a wider variety of socio- demographic variables and types of choice alter natives than in most trip-based models, the sample size requirements tend to be somewhat larger. A recent study of the transferability of activity-based model parameters ( Bowman et al. 2013) concluded that a survey sample of at least 6,000 households may be adequate for a medium-to-large region. As the behavioral de- tail of the model increases, larger sample sizes are required. Smaller samples may support the development of calibration targets for transfer- and-refine implementation processes but may not support the estimation of new coefficients. Another special case is if an agency requires the consideration of joint travel decisions and coordinated activity scheduling across house- hold members. For such models, it is critical that survey data contain complete data across all household members, and that the survey methodology successfully captures instances when household members traveled together and performed activities together, including internally consistent data on trip arrival and departure times and locations. Household surveys increasingly include in- formation derived from the collection of GPS samples. In most cases, these GPS samples have been collected only for a subset of households,

21 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM although in some cases GPS data have been collected for the entire household sample. GPS data can be used to understand underreporting of stops and tours, as well as the misreporting of activity locations and travel times. 2.1.3.2 Land Use Land use data needed for activity-based models are similar to that needed for trip- based models, including the number of house- holds, number of jobs by sector, and school enrollment by school level. This information is primarily used to influence activity genera- tion and location choices. If an activity-based model is specified to use the same or similar basic travel analysis zone (TAZ) spatial units as a trip-based model, then there is little or no difference in the data required for an activity- based model as compared with a trip-based model. However, recent activity-based models have taken advantage of the flexibility of the methods to use basic spatial units that are smaller than standard network TAZs. For land use data, such models either use microzones, which are similar to U.S. Census Bureau blocks, or indi vidual parcels. The use of parcels can entail some challenges, such as obtaining accu- rate and up-to-date parcel data for all jurisdic- tions in a region, addressing heterogeneity in how parcels are defined for different land uses (e.g., a large university may be a single parcel), and specifying forecast year land use at the parcel level. Census block–size microzones are an attractive compromise because they still of- fer a good deal of local spatial detail for defin- ing land use, walk-access times to transit, dis- tances for very short trips, and more, but they do not present the same challenges in terms of acquiring consistent data. Basic population data at the block level are available from the U.S. Census Summary File One (STF1) data, and detailed employment data at the block level are available from U.S. Census Longitudinal Employer–Household Dynamics (LEHD) data. Such data can be used to distribute TAZ-level land use data down to the microzone level, so that the control totals within each TAZ still match the original data, but the more detailed spatial distribution best matches available information. Alternatively, if parcel data are available, that data can be aggregated up to the block or microzone level, and the Census data can be used for quality assurance and quality control comparisons on the results. The primary goal of including finer geo- graphic detail is to model travel behavior at a level of detail closest to what decision makers experience in order to avoid statistical aggrega- tion bias in the models, which can be very large when measuring travel impedance for tran- sit access and short trips, or attractiveness of potential destinations. Aggregating the model outputs rather than the inputs helps to avoid the ecological fallacy that one can model rela- tionships at an abstracted level and trust that the results will be consistent with the underly- ing behavior of individual travelers. 2.1.3.3 Demographic Activity-based models are implemented within a microsimulation framework. This micro- simulation framework requires a record for each household and person rather than aggre- gate totals for each TAZ. These household and person records are referred to as a “syn- thetic population” that represents the socio- demographic characteristics of the residents of the modeled area. This demographic informa- tion is a key influence on travel choices. Soft- ware tools generate synthetic populations using information about distributions of key regional demographic attributes and detailed samples. The types of demographic controls used for generating synthetic populations typically in- clude the following:

22 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER • Household size; • Household composition and life cycle (e.g., age of householder by presence of own children); • Number of workers per household; • Household income category; • Age and gender of each person; and • Employment and student status of each person. Additional variables such as housing type and owning or renting housing could be used, but it is important to remember that any con- trols used in the base year should also be pos- sible to forecast in the future year, or at the minimum one should be able to make a reason- able assumption that the distribution will not change substantially from the base year. Using this variety of controls produces a synthetic population that is representative of the actual population along all of these dimensions, and thus allows all of these variables to be used as explanatory variables in the model. With all these combinations of variables, the model may be able to consider literally thousands of differ- ent types of persons and households, rather than just a few different demographic segments as is typical in a trip-based model. This is another way that activity-based models can avoid poten- tially significant aggregation bias, and thus pro- vide more likely predictions of choice behavior. 2.1.3.4 Network In most cases activity-based models use the same types of network procedures and vari- ables to represent automobile, transit, and nonmotorized travel routes that are used in trip-based models. These data include mea- sures of travel times, costs, distances, and other network-related attributes that influence deci- sion making, such as the number of transit transfers. This network information is typically generated at more spatially aggregate levels such as TAZs, even in activity-based models where finer spatial resolutions such as micro- zones are used for other model inputs, in order to avoid having to produce and store very, very large skim matrices. Activity-based models can incorporate additional network detail, such as more time periods or modes than trip-based models, and many activity-based models have been moving to more detailed network data as well, often using more than 3,000 TAZs to represent the region, and using 5 or more different time periods in the day (e.g., a.m. peak, midday, p.m. peak, evening, night/early morning) to represent different service levels for transit and different congestion levels for automobiles. With activity-based simulation models the amount of random-access memory (RAM) needed for software can increase with the number of TAZs and time-of-day periods, as does the time required by the network model components of the system. But, in contrast to trip-based models, the run time of the activity- based demand component of the model system does not depend on such considerations; its run time depends primarily on the size of the syn- thetic population. In recent activity-based model imple- mentations, some variations on the standard TAZ-based network approaches have been implemented. In order to provide more accu- rate representation of pedestrian, bicycle, and walk-to-transit travel impedances, a number of recent activity-based models have used a supplementary all-streets network to create shortest-path distances that consider every street in the region. The potential matrix size for such information can be huge, so the in- formation is typically only used to provide dis- tances for short distances of up to 2 miles or so. This radius will include the range of most actual walk and bike trips, as well as walks to and from transit stops, and even a large per- centage of noncommute automobile trips as well. Beyond a distance of 2 miles or so, the

23 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM TAZ–TAZ skims are likely to be as accurate as, or more accurate than, all-streets network- based skims, especially since those skims also include the effects of traffic congestion. Another recent variation, used in models in San Diego and other regions, is to use separate zone systems for the automobile and transit networks. The automobile network uses the existing TAZ system, while the transit network is based on transit stop–to–transit stop service levels, so each stop or cluster of stops (stop area) essentially forms a new type of zone to use in generating the transit network skims. This approach can improve accuracy in the service level information and the walk-access times to the transit stops. 2.1.3.5 Calibration and Validation Calibration and validation involve comparing estimates of travel demand and related choices output by the model to observed real-world data and making changes to individual model components to improve model system perfor- mance. Calibration and validation data and observed target data vary based on the model design and the specifics of each region. For example, if transit ridership forecasting is a critical model capability, then transit ridership validation should be a primary focus. Agency and consultant staff should work together to identify critical calibration and validation mea- sures and data sources for each model com- ponent as well for the overall model system. Calibration of the individual demand model components is primarily based on household travel survey data, which can provide neces- sary information describing observed activity patterns, destination choices, mode choices, and time-of-day choices. For some model system components, such as automobile availability models or usual work location choice models, household survey data may be augmented with information derived from other sources such as the U.S. Census. In addition, required calibration and validation data also include information such as roadway counts and speeds by detailed time of day and vehicle class. Transit data are typically also re- quired and may include on-board surveys as well as transit operators’ data such as transfer rates, ridership by submode, route, and major stop or station. Bicycle count data may also be necessary if the model includes a bicycle route choice model. It should be noted that different types and sources of calibration and validation data often conflict, despite the fact that these data are all “observed.” Significant cleaning or manipulation of data involving subjective judg- ment by agency or consultant staff may be re- quired. Table 2.1 presents activity-based model data items along with their uses and sources. 2.1.3.6 Data Challenges Model managers face many challenges when addressing activity-based model data needs. Key concerns include the identification of appro priate levels of spatial, temporal, and typo logical detail for model inputs and outputs, and the collection and maintenance of informa- tion at the selected resolutions. For example, extremely fine-grained spatial information at the parcel level can improve model sensitivity to short-distance trips, but such detailed data must be also be prepared for all future-year or alter- native scenarios. Managers must consider issues such as data acquisition costs, data detail and reliability, and the level of effort required for ongoing model data updates and maintenance. 2.1.4 Implementation 2.1.4.1 Component Design To determine whether the next model update should be to an activity-based model and to ensure that the activity-based model will be appropriately sensitive and able to generate the

24 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER information required by decision makers, it is necessary for an agency to carefully consider the critical questions and analysis needs over a time horizon of at least the coming 10 years. Agencies should consider whether there are policies or investments of interest to decision makers that are better evaluated using activity- based modeling approaches. For some policies, an activity-based model may be able to provide entirely new model sensitivities and capabili- ties, while for other policies, an activity-based model may simply provide enhanced model sen- sitivities and capabilities. These desired model sensitivities should be prioritized in order to allow decision makers to consider trade-offs between required capabilities and resource and schedule constraints. 2.1.4.2 Estimation Estimating the component models for an activity-based model system involves using local data to identify the variables that are most important to activity and travel decision making and quantifying the relative importance of these variables in addition to processing survey data and attaching relevant data from network skim matrices and land use data. It also includes specifying the model utility func- tions and alternative availability constraints in a model estimation software package and then carrying out the estimation in an iterative pro- cess. This process requires expert judgment to determine what variables to include and some- times involves constraining some coefficients to typical values in cases where the data for esti mation are inadequate. The estimation task typically has been carried out by consultants rather than by the agency staff. Recently, some of the activity-based model software packages have been designed to make it easier to mod- ify and re-estimate activity-based component models when starting from a known model specification (e.g., from another region) rather than starting from scratch. Estimating new models depends on having robust travel behavior data that include a suf- ficient number of samples across key market TABLE 2.1. ACTIVITY-BASED MODEL DATA ITEMS, USES, AND SOURCES Data Item Use Source Household survey • Model estimation • Calibration targets • Local data collection of the National Household Travel Survey Land use • Synthetic population generation • Activity generation • Location choice • U.S. Census • Business databases • Tax assessors data • Regional land use data • School departments Demographic • Synthetic population • All component models • U.S. Census • Regional demographic forecasts Network • Transportation network geometries • GIS databases • Transit agencies • Public works agencies Calibration and validation • Model calibration and validation • Count databases • Highway performance monitoring • Transit agency reporting

25 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM segments and choices and involve the use of model estimation software. This software is used to describe the relative importance of variables factors affecting travel-related deci- sion making. Model estimation is discussed in greater detail in this chapter and in Chapter 3. 2.1.4.3 Software Development There currently are three or four software platforms that are used for the large major- ity of activity-based models implemented in the United States; most of these platforms are governed by open source licensing agree- ments. These platforms have been developed by the consultants who have created the origi- nal models, and then the models are adapted and improved over time as they are imple- mented for new regions. The software field for activity-based models may change substan- tially in the future, however, as the market for activity-based models matures. 2.1.4.4 Transferability Given the extensive work involved in designing a new activity-based model system, estimating the component models, and creating the soft- ware to implement them, it can be expected that many activity-based model implementa- tions in the future will start with models and software transferred from another region. Starting from this base, features can be added if necessary, and coefficients can be re-estimated and/or re-calibrated. Within this context, esti- mating model coefficients from robust local data is the preferred approach. However, there is evidence that it is better to transfer model coefficients from another similar region that have been estimated from a large data set, than to estimate local coefficients using a very small or poor quality local data set (Bowman et al. 2013). 2.1.4.5 Relationship of Convergence to Equilibration Convergence to an equilibrated or stable solu- tion is critical with both trip-based as well as activity-based model systems. The network performance indicators, such as travel times and costs that are output by the final model system assignment process, must be consistent with the times and costs used as input to the model system. In order to achieve this model system-level convergence, it is first necessary to establish network assignment model conver- gence. Most activity-based model systems are linked with static user equilibrium roadway network assignment models and transit assign- ment models. Activity-based model managers must ensure that the activity-based model sys- tem is configured to pursue convergence to an equilibrium or stable solution and that such a solution can be achieved within reasonable run times. The model system configuration also should consider issues such as the stochasticity of the activity-based model and how this effect can be reasonably managed. 2.1.4.6 Calibration and Validation Calibration and validation of the entire activity-based model system is an iterative pro- cess in which changes are made to individual model components in order to ensure that the model matches observed data describing travel behavior and network performance reasonably well. Calibration also must include sensitivity testing to ensure that the model responds plau- sibly to changes in model inputs and that these changes are reasonably consistent with real- world outcomes. As with a trip-based model, the outputs of each individual component of an activity-based model are compared to observed external data, and overall network indicators, such as link volumes and speeds, are compared to observed traffic counts and speeds. The level

26 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER of effort to calibrate an activity-based model system may not be significantly greater than to calibrate a trip-based model system because all the choice components in the activity-based model system are more closely linked. 2.1.4.7 Auxiliary Demand Like trip-based models, activity-based models represent the trips made by residents of the modeled area when these residents are travel- ing entirely within the modeled area. Typically these trips make up about 80%–90% of the total demand. Auxiliary demand refers to trips that are not represented in the activity-based model system, such as commercial vehicle trips, truck trips, visitor trips, internal-external and external-external trips, and special generator trips. In most cases the auxiliary trip models used in conjunction with activity-based models are similar to those used in conjunction with trip-based models, although additional temporal or spatial detail may be included. In some cases more sophisticated auxiliary demand models have been implemented in conjunction with activity-based models, but such models are not required. Examples of more sophisticated aux- iliary models include activity-based models for seasonal residents and/or overnight visitors, and tour-based commercial vehicle demand models. 2.1.5 Application and Maintenance Model maintenance and application are neces- sary ongoing activities. New data—such as sur- veys, demographic assumptions, networks and other spatial data, and validation data—need to be incorporated and the model potentially adjusted as a consequence. In addition, as the model is increasingly applied to a variety of project and policy evaluations, refinements or enhancements to the model often are desired. Sometimes these can be transferred or adapted from other areas. Model application also re- veals bugs in software code that must be fixed. 2.2 DESIGN 2.2.1 Policy and Investment Analysis Needs To determine whether the next model update should be to an activity-based model and to ensure that, if selected, the activity-based model will be appropriately sensitive and able to generate the information required by deci- sion makers, it is necessary for an agency to carefully consider its analysis needs. Agencies should consider whether there are policies or investments of interest to decision makers that are better evaluated using activity-based model- ing approaches. For some policies, an activity- based model may be able to provide entirely new model sensitivities and capabilities, while for other policies, an activity-based model may simply provide enhanced model sensitivities and capabilities. These desired model sensi- tivities should be prioritized in order to allow decision makers to consider trade-offs between required capabilities and resource and schedule constraints. Examples of some of the poten- tial sensitivities and capabilities that may be of interest are • Long-range transportation plans. By in- cluding accessibility measures throughout, most activity-based models are better able to capture effects such as induced demand. By functioning at a fully disaggregate level, activity-based models can support com- prehensive equity analyses. Incorporating more detailed representations of indi vidual choice contexts renders activity-based models more appropriately sensitive to a wider variety of policy and investment strategies. • Conformity and air quality analyses. Most activity-based models are linked to tradi- tional static network assignment models, which provide the primary inputs to many air quality and emissions models. However,

27 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM activity-based models typically are used to provide more detailed information, such as network assignment model results by de- tailed time of day, and can also provide esti- mates of start and stop emissions by time of day. Because of their disaggregate nature these models can provide measures such as emissions by household, and, due to their flexible structure, activity-based models can easily accommodate new models, such as automobile ownership, that include in- formation on vehicle type choice. • Pricing. Activity-based models have sig- nificantly more robust capabilities with respect to representing pricing strategies and effects than trip-based models. These enhanced sensitivities result from factors such as the activity-based models’ greater sociodemographic, purpose, and time- period detail. In addition, activity-based models can explicitly represent parking pricing and subsidies. • Reliability. Reliability is a key focus of fed- eral transportation policy highlighted in the Moving Ahead for Progress in the 21st Cen- tury Act (MAP-21). Recent second Strate- gic Highway Research Project (SHRP 2) research projects have provided evidence about the trade-offs that travelers make between cost, usual travel time, and travel time reliability. These research efforts also have identified and investigated innova- tive methods for incorporating travel time reliability into both trip-based and activity- based travel demand models, as well as into network assign ment models. The detailed scheduling capabilities of activity-based models can potentially exploit network assign ment model-based reliability mea- sures, though additional investigation into methods for quantifying this information on an O-D basis is required because of the non additive nature of travel time reliability. • Travel demand management and transpor- tation systems management. Activity-based models can better represent the effects of effective travel demand management strat- egies, such as flexible work schedules, because they represent the entire set of activities that individuals and households participate in during the day. In addition, these models include time-of-day choice models, which are absent from the vast majority of trip-based models, and can be sensitive to peak-spreading effects. • Transit. By using tours (a series of linked trips that begins and ends at a single home or work location) as a fundamental or- ganizing structure, activity-based models have more realistic sensitivities to transit investments and policies. For example, transit options on the return “half tour” influence the choice of transit on the out- bound half tour, or tolls may be only in one direction or may vary by time of day and direction. Activity-based models also pro- vide the ability to test a wider range of fare policies, including person-level influences such as transit pass-holding or targeted discounts. Activity-based models also typi- cally provide better sensitivity to the in- fluence of urban form, accessibility, and demographics on auto ownership choices. • Land use. Activity-based models typically include more travel purposes than trip- based models, providing more sensitivity to detailed land use information. In addi- tion, they can better capture the effects of transit-oriented investments because of the ability to more easily incorporate short- distance travel factors that influence transit choice. • Active transportation. As with transit, activity-based models more easily incorpo- rate short-distance travel factors that influ- ence active transportation modes and are

28 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER better able to represent true modal avail- ability and constraints. • Equity. Because they work at the disaggre- gate levels of individuals and households, activity-based models provide more oppor- tu nities to report effects by any socio- demographic characteristic included in the synthetic population file. 2.2.2 Design Considerations When designing a new activity-based model or choosing an existing activity-based model to transfer, there are several different consider- ations, including the following: • The level of detail to include in terms of spatial resolution, number of different time periods, number and type of market seg- ments, number and type of modes to con- sider, and number and type of different activity purposes. • The range of different types of model com- ponents to include in the system. • Internal component linkages, including how outcomes of higher-level models influ- ence the choices available in the lower-level models, and how the accessibility vari- ables, representing the total accessibility across lower-level alternatives, affect the higher-level choice decisions. • External linkages with other model compo- nents, such as network assignment models and models of auxiliary markets such as freight, special generators, and external trips. Each of these considerations is described as follows. 2.2.2.1 Resolution and Detail There are several different types of resolution to be considered in an activity-based model design. When these resolutions are being con- sidered, there are two important aspects that distinguish activity-based models from trip- based models. First, in contrast to the zone-based loop- ing structure of most trip-based models, adding more zones, more time periods, more demo- graphic segments, and/or more trip purposes does not greatly increase the run time of activity- based model components. Adding detail along all of those dimensions has not been practical in the past because the run time and data stor- age requirements in trip-based models increase with the square of the number of zones, times the number of population segments, times the number of trip purposes, and times the num- ber of time periods. In an activity-based model, however, the run time depends primarily on the number of different households and per- sons simulated. The amount of detail used in the various dimensions may add to memory re- quirements but will not substantially influence run times. This fundamental difference is what has made it possible to include more detail in activity-based models. Second, additional detail is incorporated into the model system to provide more accurate aggregate forecasts, not to report the detailed forecast for a particular household, or parcel, Census block, or 15-minute period of the day. With an activity-based model, one simulates behavior in more detail and then aggregates the model outputs, rather than aggregating the inputs from the outset. This feature has the statistical advantage of avoiding likely aggre- gation bias in the forecasts. The feature also allows the model to represent a wider range of aspects of travel behavior, and thus the model can be used to study a wider range of policies and scenarios. 2.2.2.1.1 Representation of Space and Accessibility Activity-based models can use the same size TAZs as those used in a typical aggregate

29 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM zone-based model. However, it has become more common to include more detail to repre- sent key spatial data such as employment and school enrollment, distance to transit stops, urban design, and local street infrastructure. In some activity-based models, individual par- cels or points are the basic spatial unit. This approach requires robust parcel data for the modeled area as well as a means of forecast- ing future-year land use at the parcel level. An intermediate approach that is becoming more common is to use spatial units that are roughly the size of the Census blocks. A typical model might include 30,000–150,000 microzones, an order of magnitude more than the typical number of TAZs but also an order of magni- tude less than a typical number of parcels in a region. These microzones are often based on Census blocks in order to exploit data available at this geographic level. Using blocks or parcels as the basic spa- tial unit does not require using automobile and transit network skims at that level of detail because such skim matrices would be imprac- tically large. Instead, each block or parcel is associated with a network TAZ and uses TAZ- level skims. For most trips longer than a mile or two, this level of network path detail is ade- quate. For shorter trips, these TAZ skims can be augmented with short-distance impedance measures that use finer level street network de- tail, such as the block-to-block shortest-path distance along an all-streets network between any pair of Census blocks. This flexibility in the structure of activity-based models makes it possible to incorporate multiple levels of spa- tial resolution of skims. 2.2.2.1.2 Representation of Time Most activity-based models contain model components that simulate scheduling of trips and activities consistently across the day in re- sponse to congestion levels and other factors. Therefore, it is advantageous to distinguish multiple time periods, particularly during peak periods. As with spatial detail, the temporal detail does not need to be a perfect match for the number of different time periods used in the accompanying network assignment model. The earliest activity-based models only used four or five time periods across the day, both in their scheduling models and in network assignment. Recently, most new activity-based model systems have used time periods as small as 15, 30, or 60 minutes for scheduling models within the activity-based system. The tempo- rally detailed information is aggregated up to only 5 periods or so for the network assign- ment model. This aggregation has been done mostly for purposes of minimizing network assign ment run times. Ideally, more detailed in- formation on network congestion during peak and shoulder time periods can be fed back into the activity-based model. Thus, recently imple- mented activity-based model systems tend to use more detailed time periods for assignment, including as many as 10 to 15 time periods, with separate assignments and skims for 30-minute or 60-minute time periods during the a.m. and p.m. peaks. This provides a closer match to the length of time periods used inside the activity- based model scheduling components. As time goes by, both activity-based models and network assignment models are moving toward a more continuous representation of time across the day, using periods as short as 5 or 10 minutes for scheduling models within the activity-based components and integrating with dynamic network assignment models [col- loquially referred to as DTA, or dynamic traffic (or network) assignment, models, although they may consider transit in addition to vehicu- lar traffic] on the supply side to simulate and feedback measures of traffic congestion at that same level of detail. However, such detailed temporal resolution may not be necessary or appropriate for all regions. The SHRP 2 C10

30 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER projects provided a proof of concept for inte- grating demand-and-supply models with high levels of temporal (and spatial) detail, although significant methodological and practical issues associated with these advanced models remain. 2.2.2.1.3 Market Segmentation Demographic Segmentation. Market segments are used in trip-based models to reflect the in- fluences of attributes on different populations. In contrast to trip-based models, most activity- based models do not necessarily need to pre- define a set of market segments. Each house- hold and/or person is simulated individually, and any characteristics that are known about that household and/or person can be used in the models. Variables typically used to define market segments are household size, number of household members in different age catego- ries, number of employed adults, number of students, household income, person age, per- son gender, person employment status, and student status. These are standard variables available in household travel surveys and in the synthetic population that is used as an input to the activity-based model. When one considers the number of different combinations of these household and person variables that can exist across the population, the number of demo- graphic segments used in the activity-based model can be virtually unlimited. Of course, for reporting purposes, results are aggregated for specific districts, income groups, and other seg- ments. Note that it is also possible to use other demographic variables such as race/ ethnicity, housing type and/or ownership status, and length of residence in the neighborhood. Using race/ethnicity has often been avoided because of the difficulty of separating those effects from income effects and predicting future land use residential patterns of different race/ethnic groups. Other variables such as housing type and own/rent status may become more com- monly used in the future in cases where the activity-based model is integrated with a land use model that predicts such outcomes. Activity Purpose Segmentation. Early activity-based models tended to include only three or four distinct activity purposes, such as work, school, other, maintenance, and dis- cretionary. Recently, as many as 7–10 activity purposes have been included in activity-based models. “Escort” activities, also referred to as “chauffeur ing” or “serving passenger,” tend to have different characteristics than other ac- tivities, particularly in terms of mode choice, since these activities tend to involve automobile shared-ride tours. Meal activities can usefully be separated from other types of maintenance activities, because they tend to take place dur- ing certain periods of the day at locations where food service employment is located. Shopping is another type of maintenance activity that can be tied to specific attraction variables, such as retail employment, and tends to happen during store opening hours. Medical visits are another activity purpose that can be tied to a specific attraction variable (medical employment). On the discretionary side, it can be useful to sepa- rate social visits as a separate activity purpose, as they often occur at residential locations and outside of working hours. Outdoor recreation can be another useful activity category, as it can be tied to open space/parks/sport fields, as attraction variables. In general, if the land use data have sufficient detail that will allow prediction of where specific types of activities are likely to take place, then the data can also be useful to distinguish those types of activities in the activity-based model components. The model then can better predict which types of people tend to visit certain types of locations during certain periods of the day. Mode Segmentation. In general, the set of modes used in an activity-based model is simi- lar to the set that would be used in a trip-based

31 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM model. A minimal number of modes to include in an activity-based model would be automo- bile, transit, and nonmotorized. With the in- creased interest in walking and biking, these two nonmotorized models are considered as separate modes in most recent activity-based models. If park-and-ride is a relevant choice in a region, it can be separated from walk- access transit as a separate mode alternative. Some models treat different transit submodes (e.g., bus, light rail, heavy rail, ferry) as sepa- rate modes in mode choice, while other models leave it to the transit network path building software to determine the best overall transit path and its attributes. Different activity-based models use differ- ent patterns to segment the auto submodes. In some model designs, only the automobile occupancy is important [e.g., different mode choices for single-occupancy vehicle (SOV), high-occupancy vehicle (HOV, HOV 2, or HOV 3+)], without predicting exactly which person is the driver and who are the passen- gers in the HOV tours and trips. In other model designs, the driver–passenger distinction is pre- dicted explicitly. All designs tend to include an explicit choice between different occupancy levels, however, which is important for obtain- ing the correct number of vehicle-trips for as- signment purposes. In addition, some model designs also treat the choice between a tolled path and a nontolled path as essentially two different types of automobile alternatives in mode choice, while in other models, that choice is done completely in the automobile network path building process. In general, the specific set of modes used is probably the least consistent across the vari- ous activity-based models that have been de- signed and used in practice, partly because different regions have different types of modes available currently and different emphases in adding certain modes in the future. One de- sign issue to consider is the ease in adding new types of modes into the choice set for future- year scenarios, rather than being restricted by the model design to a fixed set of modes in all scenarios. 2.2.2.2 Subcomponents Table 2.2 provides an overview of the types of model components that have been included in advanced model systems as they have evolved from simple tour-based models (in the late 1980s and early 1990s) to the most advanced activity-based models today. In order to pro- vide concise information, Table 2.2 does not consider every detail of every model used in practice, so there may be exceptions to this typology. This table is meant to cover the major variations in model designs, as an aid to dis- cussing the various components in more detail. The earliest tour-based models were simi- lar to trip-based models, but generated tours instead of trips, and typically used regression models and logit models for all components, including generation and distribution (destina- tion choice). Complexities such as time-of-day choice and intermediate stop-making on tours were typically dealt with using simple factor- ing approaches, rather than detailed choice models. Because the structure of such models is similar to trip-based model structures, they could be applied in a traditional aggregate zone-based framework, without simulating individual households. More advanced tour- based models replaced the simple factoring with explicit time-of-day choice models and explicit models of the generation and location of intermediate stops on tours. This structure, which essentially adds a third spatial dimension (stop location depends on both the tour origin and destination locations), made traditional aggregate zone-based application infeasible, so these models moved to the microsimulation approach, or some combination of aggregate and micro simulation approaches. The shift

32 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER TABLE 2.2. COMPONENTS INCLUDED IN VARIOUS TYPES OF TOUR-BASED AND ACTIVITY-BASED MODELS Simple Tour- Based Advanced Tour-Based Day-Pattern- Based Day Pattern with Longer- Term Choices Day Pattern with Longer- Term and Mobility Choices With Explicit Intra- Household Interactions Population segmentation Population synthesis Population synthesis Population synthesis Usual work (and school) locations Population synthesis Usual work (and school) locations Population synthesis Usual work (and school) locations Automobile ownership Automobile ownership Automobile ownership Automobile ownership Automobile ownership Transit pass and parking pass ownership Automobile ownership Transit pass and parking pass ownership Joint household day pattern and joint (half) tour generation Tour generation Tour generation Day pattern (tours and some aspects of intermediate stops) Day pattern (tours and some aspects of intermediate stops) Day pattern (tours and some aspects of intermediate stops) Remaining individual tours and some aspects of intermediate stops Tour-level models—mode and destination choices Tour-level models — mode, destination, and time-of-day choices Tour-level models — mode, destination, and time-of-day choices Tour-level models — mode, destination, and time-of-day choices Tour-level models — mode, destination, and time-of-day choices Tour-level models — mode, destination, and time-of-day choices Simple postfactoring Intermediate stop generation Intermediate stop location Intermediate stop generation Intermediate stop location Intermediate stop generation Intermediate stop location Trip-level mode and departure time choices Intermediate stop generation Intermediate stop location Trip-level mode and departure time choices Intermediate stop generation Intermediate stop location Trip-level mode and departure time choices Usually aggregate application Aggregate and microsimulation applications Person-based microsimulation applications Person-based microsimulation applications Person-based microsimulation applications Household and person-based microsimulation applications

33 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM to a micro simulation approach required the introduction of population synthesis in order to provide records for individual households and persons in a proto typical, representative population. The first practical use in the United States of what we now recognize as activity-based models came with the introduction of the day-pattern approach (Bowman 1995). The day-pattern approach assumes that tours for different activity purposes are not generated independently of each other, but are gener- ated jointly to take account of substitution and complementarity effects of different types of tours and activities across a single day. In some model designs, some aspects of intermediate stops on tours are predicted at this day-level as well, although the exact numbers of stops and allocation to tours are always predicted at the lower levels of the model system. Because there are so many different possible day-long patterns of tours and stops, all models of this type have been applied using a microsimula- tion approach at the person-day level, using single stochastic (Monte Carlo) choices for each model component. Although the person- day is used as the basic unit of simulation, effects of other household members are typi- cally incorporated through using variables on household characteristics and/or simulating the household members in a specific order and making some household members’ choices con- ditional on the choices of previously simulated members. An advance to the day-pattern approach that was put into practice around the year 2000 was to include longer-term location choices such as workers’ usual work locations and stu- dents’ usual school locations at the upper level, instead of leaving these choices to be predicted at the tour level only. This provides more con- trol for matching work and student commute flows to control data, and allows those choices to influence other aspects of the model system (e.g., people may be more likely to own a car and/or to telecommute from home on some days if their usual work location is a long dis- tance from their residence). Recently, some activity-based model systems have included other longer or medium-term mobility choice models such as the decision to own a transit pass or whether or not free parking is provided at the workplace. The most advanced activity-based models explicitly link joint travel and activities across different household members. This feature, termed “explicit intra-household interactions,” has evolved to include complex interactions such as parents taking children to school. The inclusion of such features requires a more com- plex microsimulation framework and software to account for the joint scheduling and loca- tions of different persons’ activities across the simulated day. 2.2.2.2.1 Population Synthesis Population synthesis is a procedure to create a set of household and person records that match relevant demographic control totals for a recent base year or for a forecast year. For a base year, the control totals are typically from the U.S. Census and/or American Community Survey (ACS), and the individual household and person records are drawn from public use micro data sample (PUMS) files, such as from the Census and/or ACS. Early population syn- thesizers were fairly simple and tended to use only household-level marginal controls such as distributions on household size, household workers, income, and/or age of head of house- hold. More recent procedures and software provide users with more flexibility in using a wider range of person-level and household- level controls, as well as more sophisticated methods of drawing samples into the popula- tion to match controls more closely. The avail- ability of sociodemographic forecasts for a

34 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER region is one aspect that should be considered when selecting which control variables to use in population synthesis. An alternative approach to synthesizing a population separately for each forecast year is to synthesize the base-year population and then use a population evolution model to evolve that population over time to represent phenomena such as aging, births and deaths, marriage and divorce, and immigration and emigration. At the time of writing, these types of evolution models have not yet been used in practice along with activity-based models. Because the population is synthesized at the TAZ or Census block group level because of the availability of marginal control data at this geography, it may also be necessary to use methods to allocate synthetic households to more detailed spatial units used in the model system such as microzones or parcels. In prac- tice, this allocation has often been done ran- domly, although modeling approaches have been used to associate households with spatial units based on the households’ characteristics, such as size and income, and the locations’ at- tributes, such as type of available housing and distance from transit. 2.2.2.2.2 Long-Term Models All tour-based and activity-based models have included a model component to predict house- hold auto availability, as this is one of the most important variables in subsequent models such as tour generation and mode choice. As men- tioned, more recent activity-based model sys- tems also predict the usual work location for workers and often predict the usual school location for students. Those locations tend to be key aspects of how workers and students plan their days spatially and temporally. Also, specifically in the case of work locations, the user may wish to doubly constrain the choice to make sure that the number of workers pre- dicted to have jobs in specific zones or micro- zones closely matches the number of jobs that are located there. In activity-based models, this type of constraint is typically implemented using an iterative shadow pricing procedure, where the attractiveness (utility) of specific work locations is iteratively revised until the match between predicted workplaces and available jobs in each spatial area is considered acceptable. In most U.S. model systems, work and school locations are predicted before automo- bile ownership on the assumption that one can more easily buy or sell an automobile to fit the commuting needs than one can find a different job to match the car ownership level. In reality, these two choices are interdependent and have sometimes been modeled that way. It should be noted that all choices modeled in an activity- based system are interdependent to some de- gree, but it would not be possible to estimate a model along all dimensions simultaneously. One of the aspects that differentiate model de- signs, even within the same family of designs, is which model components are modeled and applied jointly versus which model components are modeled and applied sequentially. An additional consideration regarding longer-term models, such as usual work location and usual school location, is that some land use forecasting models also predict these choices, so it may not be necessary to include them as part of the activity-based model system. In that case, the usual work and school location is just passed to the activity-based model as a vari- able on the synthetic population file, the same as residence location, income, and so forth. A possible advantage of including workplace loca tion as part of a land use simulation model might be that the residence location and work- place loca tion can be predicted simultaneously, or in a naturally occurring temporal sequence through the course of the land use simulation, since in many cases people will choose their res-

35 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM idence location based on where their workplace is l ocated, rather than vice versa. This is par- ticularly relevant for multiworker households. Finally, consideration should be given to the policy or project contexts for which it is appro- priate to run long-term choice models. 2.2.2.2.3 Mobility Models (Medium Term) A recent trend in activity-based model design is to explicitly model mobility modifiers, such as the ownership of a transit pass or the avail- ability of free parking at the workplace. These variables can significantly influence the relative attractiveness of competing modes in the subse- quent lower-level models such as mode choice. A key consideration is how these modeled out- comes are used in the lower-level models of the system, and how the relative attractiveness of the modes available to usual destinations is used to influence the upper-level choices as described in the “Accessibility Measures” sec- tion (see Section 2.2.2.3) that follows. A per- son is likely to own a transit pass only if there is reasonably attractive transit service to that person’s workplace or other usual destinations. This aspect needs to be included in order for a transit pass ownership model to predict reason- able behavior. 2.2.2.2.4 Daily Models (Short Term) In addition to the long-term and medium-term choices, activity-based models also consider short-term or day-level choices. Activity-based models consider these short-term choices at three main sublevels: • The full-day level; • The tour level; and • The trip level. The Full-Day Level. The representation of day-level choices is the aspect of activity-based models that first distinguished them from earlier tour-based and trip-based models. Modeling an entire day makes it possible to fully integrate time-of-day models and temporal constraints into the model system, and the modeling of a day pattern at the day level is an important way of incorporating the trade-offs that people make when faced with the limits of a 24-hour day. Not surprisingly, it is at the full-day level that the existing activity-based models imple- mented in practice vary the most. There are a number of different design considerations as to how the day-level is modeled, and different combinations of these aspects are used in prac- tice. These include the following: 1. Are the numbers of tours made during the simulated day for the various pur- poses modeled independently of each other or modeled jointly (allowing for substitu- tion between tours for different purposes)? 2. Are the mandatory (work and school) tours generated before generating the tours for the other (nonmandatory) purposes, or are these tours all generated jointly in a single model? 3. If work and school tours are generated first, are they also scheduled before gener- ating any nonmandatory tours, so that the time left over in the day can influence the likelihood of making individual tours? 4. Are all activity episodes generated first and then allocated to tours or, at the oth- er extreme, are only tours generated first, with any additional activities at inter- mediate stops generated later at the tour level? Several models used in practice are somewhere between these two extremes, where certain information about interme- diate stop activities is generated at the day level at the same time as tour generation, but indi vidual stop activities are generated and allocated to tours at the tour level, conditional on the day-level predictions.

36 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER In reality, some intermediate stops are planned in advance, and some are made on the spur of the moment for convenience sake, so there is no completely correct way to structure such models. 5. For workers, is the decision to work at home during the day (telecommuting), or to have a nonwork day, modeled explicitly? 6. Are the types of day patterns (go to work or school, go to other types of destinations, stay home all day) coordinated jointly across different household members? If so, is this coordination done in a sequential model fashion or in a single joint model across all household members, and how do travel conditions affect these choices? 7. Are household maintenance tours (e.g., shopping, errands, dropping off and pick- ing up passengers) generated at the house- hold level? Are individual maintenance tours generated for individuals? 8. Are completely joint tours (where differ- ent household members travel to all activ- ity locations together for the entire tour) modeled explicitly as joint travel deci- sions, or are all tour generation decisions modeled separately for each household member? 9. Are partially joint tours for work and school purposes (where one family mem- ber chauffeurs or accompanies another one to work or school) modeled explicitly or as separate decisions for different household members? 10. Are partially joint tours for other pur- poses (where one family member chauf- feurs a child to a sporting event or to play at a friend’s house) modeled explicitly or as separate decisions for different household members? 11. Is joint travel between members of different households (where two workers from dif- ferent addresses carpool to work together) modeled explicitly as a joint decision? 12. Are full social networks, including non- family members, taken into account when generating and scheduling tours for a day? Aspects 6–10 are what are commonly re- ferred to as “explicit intra-household interac- tions,” and have been incorporated in the de- signs of some of the most recent activity-based models. These explicit intra-household inter- actions can be quite complex to model, par- ticularly the partially joint tours in Aspects 9 and 10, because the simulated travelers travel together for part of the tour but can then do different things for the remainder of the tour, so there are many possible permutations of such tours and many ways to model those per- mutations. To date, partially joint tours have been included in practical models for work and school activities (Aspect 9), but not yet for other activities (Aspect 10), where the possibili- ties are even more numerous. Aspects 11 and 12 address modeling joint travel decisions between members of different households. Although such modeling has been done as part of university research, it has yet to be incorporated into practical activity-based models, in part because the methods for iden- tifying travel partners across the region would be very complex. Overall, among these aspects there is a trade-off between model complexity and be- havioral accuracy of the results. On the one hand, models that do not represent explicit intra-household interactions may be too sen- sitive to some travel policies because these models do not fully portray the constraints on travel choices imposed by household mem- bers having to synchronize their schedules and vehicles to travel together. On the other hand, models that pose such joint behavior as inflex- ible constraints may over-constrain the model predictions in the long term, as household

37 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM members can adjust their joint travel patterns if conditions change enough. This issue brings up the importance of including travel accessi- bility effects at all levels of the model system, as the convenience of traveling together and the options for traveling by alternative modes can influence whether or not household mem- bers choose to travel together. As this is one of the newer and continually developing aspects of activity-based model design, these issues are likely to be the subject of future research. The Tour Level. Given the day-level outcomes, the tour level is where many of the key behav- ioral outcomes in activity-based model systems occur. Once tour decisions are made, models at the trip level add important details, but the trip-level models are constrained by the out- comes at the tour level, so they are not as influ- ential as the tour-level models. In most cases, three main decisions are modeled at the tour level: 1. Destination choice; 2. Mode choice; and 3. Time-of-day choice. Destination choice involves choice of par- cel (or point), microzone (block), or TAZ, de- pending on the level of spatial resolution used in the model. When the spatial resolution is more detailed than TAZ, destination-choice models use sampling of alternatives, identify- ing only a subsample of spatial units as choice alternatives rather than the full set. Sampling is done for reasons of computational efficiency and, since we are modeling choices for many thousands or even millions of tours per day in a region, each possible choice alternative will still be included in the choice sets for many tours in the course of a simulation. Typically, impor- tance sampling is used (Ben-Akiva and Lerman 1985) where the destination sampling prob- abilities are based on a simpler function with a form analogous to a gravity model, so that the more attractive alternatives are more likely to get into the choice set and the sampling is more statistically efficient. In general terms, logit destination-choice models are very similar to, but more flexible than, a singly-constrained gravity model. To also add controls at the des- tination end for a doubly constrained model, shadow pricing methods are often used, but typically only for work or school destinations. Tour-level mode choice models are very sim- ilar to their trip-level counterparts, but they con- sider both halves of the round trip tour. Because people tend to use the same mode for an entire tour in more than 90% of cases, it is logical to model this as a single, tour-level decision. Subse- quent trip-level models are used to represent the infrequent cases of multimodal tours. Alterna- tives such as park-and-ride are also modeled at the tour level, because people have to return to the same park-and-ride lot to retrieve their cars on the way home. The typical modes included in activity-based models are as follows: • Car drive alone (sometimes distinguished by toll status); • Car shared ride 2 (sometimes distinguished between driver and passenger and by toll status); • Car shared ride 3+ (sometimes distin- guished between driver and passenger and by toll status); • Transit-walk access; • Transit-drive access (sometimes distin- guished between park-and-ride and kiss- and-ride); • Nonmotorized (usually distinguished between walk and bike); and • School bus (only for school tours in areas where such service is provided). Tour mode choice models typically use nested logit modeling either using a pre- determined nesting structure from previous

38 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER stated preference research or letting the esti- mation data decide which nesting structure performs best. The nesting structure may vary depending on activity purpose, available modes, or other local characteristics. Tour time-of-day models are also typically done for the round trip. In some cases, these models predict the time the person leaves home to begin the tour and simultaneously the time the person arrives back at home. In other model designs, the model predicts the time arriving at the primary tour destination and the time de- parting from the tour destination again. The number of different time periods used in the tour time-of-day models can be as low as 4 or 5 (e.g., a.m. peak, midday, p.m. peak, evening, night), or as many as 24 (one-hour periods) or 48 (half-hour periods). An important design option at the tour level is the hierarchy to use for the three types of models: destination choice, mode choice, and time of day. Nearly all activity-based models used in the United States have included destina- tion choice above mode choice, but for some ac- tivity purposes alternative sequences may make sense. There is no clear consensus on where in this hierarchy the tour time-of-day models should be placed. Both research and practice have indicated that the more detailed the time periods used in the time-of-day models, the more likely that travelers are to shift time periods, and thus the lower down in the choice hierarchy that the time-of-day models should be. An alternative to estimating complex, multidimensional, nested models is to assume a nesting structure and estimate and apply the models as sequential nested models. This may mean estimating a mode/time-of-day log- sum across all possible mode and time-of-day combinations for use in the destination-choice model. The disadvantage of calculating all the accessibility logsum variables across all times of day is that this calculation can greatly in- crease the run time of the model, particularly if there are many different time periods used. These design considerations affect the ver- tical integrity of the models—the idea that al- though each choice in the hierarchy is condi- tional on the choices simulated above it, each choice alternative also receives information about the expected utility across all of the re- maining choice alternatives at all levels below it, if it were to be chosen. As more different types of choices and choice levels are incorpo- rated in the design, it becomes more difficult, yet no less important, to maintain the ideal ver- tical integrity of a fully branched tree of nested models from top to bottom. The Trip Level. Below the tour level, the trip- level component models tend to include the following: • Work-based subtour generation; • Intermediate stop generation; • Intermediate stop location; • Trip mode; and • Trip departure time. As mentioned previously, in some model designs detail has already been simulated at the full-day level about what types of stops and/ or how many stops must be simulated on the tours that occur during the day. This informa- tion is used to condition the choices simulated in the intermediate stop generation model. In other designs, there are no such prior con- straints from the day-level predictions. The intermediate stop location model can be one of the most complex models in an activity-based system because it is anchored by two different locations—the current location (the destination of the previously simulated trip or tour) and the location where the particular half tour is ultimately heading, as shown in Figure 2.1. The latter is either the tour origin or the tour destination, depending on whether the model design simulates each half tour in-

39 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM ward toward the tour destination, or outward toward the tour origin, which is the home loca- tion for home-based tours. Just as at the tour level, the relative order- ing of the trip mode and departure time models can vary from one model design to the other. In this case, however, it is not as critical because the trip-level models do not have as much in- fluence on the model results as the tour-level models, since they simply provide more detail based on the tour-level choices. For model systems where the tour time-of-day model already models the main tour arrival and de- parture times at the most detailed level, often 30 or 60 minutes, the trip-level departure time model only needs to be used to model the de- parture time from any intermediate stops, typi- cally at that same level of temporal detail. In some recent model systems, however, even more detailed time periods have been used at the trip level, with periods as detailed as 5 or 10 minutes. There are two reasons for moving toward a design with more detailed time periods, as small as 5 or 10 minutes. The first is that the output is more compatible with DTA methods that use similar time slices, if one is planning to move from static assignment to DTA in the near future. Second, some of the most complex logic in the activity-based model software is that of scheduling activities and travel in avail- able time windows for each simulated person in the simulated day. The more accurate the trip departure and arrival times and the activ- ity start and end times can be simulated in the system, the more accurately the time-window accounting can be simulated. Such precision may not be necessary for most model applica- tions, and model users should be cognizant of the difference between the precision of these times and predictive accuracy. 2.2.2.3 Accessibility Measures Accessibility measures are critical to ensur- ing reasonable policy sensitivity at the various levels of the model design to changes in infra- structure or land use, or both. In general, four types of accessibility variables are included in the models: 1. Direct measures of travel times, distances, and costs from modeled network paths; 2. Detailed logsums calculated across alter- natives in models that include direct mea- sures; 3. Aggregate (approximate) logsums calcu- lated across alternatives in models that in- clude direct measures; and 4. Buffer measures representing the activity opportunities and urban design surround- ing each parcel or microzone (e.g., Census block). The direct measures are used in all mode, destination, and time-of-day choice models wherever possible. Often, however, the model 2014.11.18 C46 Primer FINAL for composition.docx 72 Figure 2.1. Intermediate stop locations. Just as at the tour level, the relative ordering of the trip mode and departure time models can vary from one model design to the other. In this case, however, it is not as critical because the trip-level models do not have as much influence on the model results as the tour-level models, since they simply provide more detail based on the tour-level choices. For model systems where the tour time-of-day odel already models the main tour arrival and departure times at the most detailed level, often 30 or 60 minutes, the trip-level departure time model only needs to be used to model the departure time from any intermediate stops, typically at that same level f temporal detail. In some recent model systems, however, even more detailed time periods h ve been used at the trip level, with periods as detailed as 5 or 10 minutes. There are two reasons for moving toward a design with more detailed time periods, as small as 5 or 10 minutes. The first is that the output is more compatible with DTA methods that use similar time slices, if one is planning to move from static assignment to DTA in the near future. Second, some of the most complex logic in the activity-based model software is that of scheduling activities and travel in available time windows for each simulated perso in the simulated day. The more accurate the trip departure and arrival times and the activity start and end times can be simulated in the system, the more accurately the time-window accounting can be simulated. Such precision may not be necessary for most model applications, and model users Tour Destination Tour Origin Intermediate Stop Figure 2.1. Intermediate stop locations.

40 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER hierarchy makes it impossible to use a direct measure because it depends on a yet- unmodeled outcome. This would be the case, for example, for travel time in a destination-choice model that is higher in the hierarchy than mode and/ or time-of-day choice, since in order to measure travel time directly it is necessary to know the mode and time of day. In such cases, detailed logsums can be calculated from the lower- level choice models and used instead of direct measures in the upper level model. A typical example in practical activity-based models is the use of tour mode choice model logsums in higher-level models such as tour time-of-day choice, tour destination choice, and workplace location choice. There are cases when it is not practical to apply fully detailed versions of the logsums that are calculated on the fly during the simulation every time a logsum is needed. The number of logsum measures required and the time to generate, store, and access these measures may simply be infeasible. More generalized logsums that reflect overall perceived accessibility may be more useful. To address this issue, a com- mon approach is to precalculate more aggre- gate accessibility logsums to be used in models in which using the more impractical ones would not be computationally or conceptually feasible. For example, some model systems use aggregate accessibility logsums calculated from each origin TAZ or microzone to all possible destinations via all possible modes. Aggregate logsums are typically calculated for each com- bination of up to four or five critical dimen- sions, including the following: 1. Origin TAZ or microzone; 2. Tour purpose; 3. Household income group, or value-of-time (VOT) group; 4. Household automobile sufficiency (auto- mobiles owned compared with driving-age adults); and 5. Household residence distance from transit service. Aggregate measures are used most often in the day-level models and some of the longer- term models, where the model is not yet con- sidering a tour to a specific destination, but is considering, for example, how many tours to make for a given purpose from the home loca- tion during the day. Finally, buffered measures represent the ac- cessibility to very nearby destinations, as could be visited by walk, bike, or very short car trips. The typical measures that are buffered include the number of nearby • Households; • Jobs of various types (as proxies for activ- ity locations); • School enrollment places of various school types; and • Transit stops. Clearly, these measures are most relevant when the spatial units themselves are much smaller than the radius of the buffer area. Thus, using buffer-based measures is really only use- ful when the spatial unit of the model is parcels or, at the largest, Census blocks. One way to make the buffer measures more accurate and relevant is to use on-street, shortest-path dis- tance to measure the distance to the edge of the buffer, rather than using straight line (as the crow flies) or Euclidean distances. Additionally, use of distance-decay functions can mitigate the “cliff effects” that are sometimes associated with the use of simple buffer measures.

41 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM 2.2.3 Other Model Component Design Considerations The preceding sections have described the issues that need to be considered when de- signing the core components that form the activity-based model system and configuring the interactions and exchanges of information among these components. However, the overall activity-based model system design must also consider the interactions with other tools that, in conjunction with the core activity-based model, make up the entire model system. These additional tools include the synthetic popula- tion model, the network assignment model, and potentially other tools such as a land use model. 2.2.3.1 Synthetic Population Characteristics The synthetic population is key input to the activity-based model, as it provides informa- tion about the fundamental decision-making units in the model system: households and persons. There are two fundamental aspects that must be considered to ensure that the synthetic population is compatible with the activity-based model. First, it is essential that the synthetic population include any attributes that may be included in the specifications of any of the model components that compose the activity-based model, or implement a submodel that provides these attributes. For example, if income variables are used, the synthetic pop- ulation must include relevant information on income that can be used in the model. If new models are to be estimated using local house- hold survey data or other information, the in- cluded variables in the survey can be compared to those available in the synthetic population. If models are to be transferred, the specifications of these existing models should be reviewed to ensure consistency. Second, it is also important that the design of the synthetic population re- flect the policy analysis needs of the region. For example, if it is anticipated that a region wishes to perform analyses of communities of concern based on ethnicity, race, income, or age, it is necessary to ensure that these attributes will be included in the synthetic population. 2.2.3.2 Network Assignment Model Design Considerations Activity-based models are travel demand models that are influenced by the performance of transportation network—travel times, dis- tances, costs, and other attributes by different travel modes and different times of day. Net- work assignment models, which assign travel demand to individual transportation network links and transit routes, produce estimates of transportation network performance. Thus, an activity-based model system is typically formed by two main components: the activity-based demand component and the network assign- ment model component. The network assignment model produces network performance indicators, such as travel times by time of day and mode, which are in- put to the activity-based model. In turn, the activity-based model provides estimates of travel demand from origins to destinations by time of day and mode that are input to the net- work assignment model. Most activity-based models have been linked with static network assignment models, although a limited number of models also have been linked with more ad- vanced DTA models. 2.2.3.2.1 Policy Analysis Needs and Consistency with Other Model Components Just as it is necessary for agencies to carefully consider their analysis needs before initiating an activity-based travel demand model devel- opment effort, agencies must similarly consider how such needs may affect the design of the network assignment model. The policy analy- sis needs of the models are intrinsically related

42 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER to the issue of maintaining consistency with other model system components. The spatial, temporal, and typological resolution of the activity-based model must inform and align with these same dimensions in the network assign ment model. For example, if the activity- based model design incorporates the ability to represent policy changes such as network pric- ing by time of day, then ideally network infor- mation such as times and costs at a temporal resolution consistent with this pricing scenario can be fed back from the network assignment model to the activity-based model. If time- of-day information is used when predicting demand, or if truck volumes on roadways are of particular concern, then the network assign- ment model should be configured to exploit this time-of-day information and should assign the truck classes that are output by the auxil- iary truck model component. Spatial Resolution. Activity-based models can more easily incorporate more detailed spatial resolution than traditional trip-based models because they are not constrained by the limita- tions imposed by the use of matrices for gener- ating estimates of travel demand. As a result, many activity-based models incorporate en- hancements to use smaller spatial units, such as microzones or parcels, or improved methods for representing short-distance travel imped- ances. Including more accurate short-distance travel can be particularly important for accu- rately predicting pedestrian, bicycle, and tran- sit travel modes. The network assignment model design should be consistent with the spatial resolution of the activity-based model. When using tradi- tional static network assignment models to de- velop measures of automobile impedances, also known as “skimming,” or to assign automobile trips, this typically means aggregating more spatially detailed outputs to the level of TAZs and using less detailed planning-level network. Transit network skimming and assignment is also typically performed using a spatial resolu- tion of TAZs, although some regions are now using alternative transit spatial schemes such as transit access points. Detailed all-streets net- works are being used in some regions to de- velop pedestrian and bicycle network imped- ances, although very few actually assign these nonmotorized trips to these spatially detailed networks. Temporal Resolution. One of the distinguish- ing aspects of activity-based models is that they include an explicit representation of time of day. Many trip-based models generate esti- mates of daily trips and incorporate peak and off-peak assignment models by using fixed time-of-day factors. In contrast, activity-based models explicitly predict tour and trip arrival times, departure times, and activity dura- tions. The temporal resolution of these more detailed times of day can vary from as broad as 3 hours or more, to as detailed as 15 min- utes or less. The network assignment model design should be consistent with the tempo- ral resolution of the activity-based model. If the activity-based model produces more tem- porally aggregate demand, such as multihour periods, then the network resolution must reflect the activity-based model resolution. However, if the activity-based demand model produces more temporally detailed demand, then users may have tremendous flexibility in the network assignment model design to in- corporate this detail. This detail can provide better estimates of network performance by time of day and potentially provide more sen- sitivity to phenomena such as peak spreading. Ideally, the temporal resolution of the activity- based demand and network assignment model would be exactly aligned. This alignment is usually not possible in practice, because most activity-based demand models are linked with static network assignment models, which are

43 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM incapable of generating reasonable measures of link volumes and network performance in- dicators for small time periods less than 1 hour in duration. Typological (Including Auxiliary) Resolution. The overall demand for travel can be consid- ered to comprise the sum of a set of smaller market segments. For example, one could con- sider a distinction between the transit travel demand market segment and auto travel de- mand market segment. It is highly desirable that the market segmentation in the network assignment model maintain consistency with the market segmentation in the activity-based demand model. If the activity-based model in- cludes transit submode choices, such as bus and light rail, then ideally the network assignment model will include a parallel segmentation. Par- allel segmentation makes it possible to produce submode-specific network performance indica- tors that are used as input to the activity-based model as well as to assign the travel demand forecast by the activity-based model to mode- specific transportation networks. An important feature of activity-based models is that they incorporate significantly more detail by func- tioning at the disaggregate level of individual persons and households. This disaggregation allows for the flexible definition of market seg- ments from activity-based model outputs. 2.2.3.2.2 Types of Network Assignment Models Activity-based models, like trip-based models, can be linked with either static network assign- ment or DTA models. Static network assignment models have been used widely in practice for decades, and their properties are well under- stood. However, static network assignment models are limited in their ability to capture changes in network performance by detailed time of day and to represent many network operational attributes. DTA models have only begun to be used more widely in practice and can provide detailed temporal information used as input to activity-based models, but they have a number of properties that complicate their broad use in practice, such as long run times, and an inherent stochasiticity. All activity-based models implemented in the United States have been linked with static network assignment models, although only a few activity-based models have been linked with DTA models. Static. Static network assignment models are the most widely used roadway network assign- ment models. They are used typically with the input travel demand generated for longer time periods, such as multihour peak periods or en- tire days. As a result, they can generate only estimates of average network travel times and link volumes representing longer time periods for input into the activity-based travel model. While the behaviors of static network assign- ment models are well known and they have rel- atively fast run times, they are limited by their insensitivity to many operational attributes and by their inconsistency with traffic flow theory (Chiu et al. 2011). Most transit network as- signment models are also static, although they vary in their level of complexity. Most activity- based models are linked with simple shortest generalized path transit assignment models, although some transit assignment models used include the ability to distribute flows across multiple competing routes and even to reflect the effects of transit crowding. Dynamic. In contrast to static network as- signment models, DTA models capture the changes in network performance by detailed time of day and can be used to generate time varying measures of this performance for in- put to an activity-based model. This temporal resolution can be flexibly defined, and recent implementations have tested resolutions as fine as 10 minutes. The network performance indicators derived from DTA models such as congested travel times arise from the dynamic

44 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER interaction of individual vehicles or packets of vehicles being simulated or calculated using ex- tremely fine-grained temporal resolution, such as seconds or fractions of seconds. DTA models are sensitive to operational attributes and are founded on traffic flow theory, but their wide adoption has been hindered by long run times and by their inherent stochasticity (Chiu et al. 2011, p. 4). 2.2.3.2.3 Modes The modes defined in the network assignment model should be consistent with those em- ployed in the mode choice component of the activity-based model system. Activity-based models have the flexibility to incorporate addi- tional modes and relevant modal attributes but generally include the same modes as advanced trip-based models. Roadway. Most network assignment models linked with activity-based models include the traditional vehicle modes by occupancy class. In some cases, the network assignment model has also included assignment classes by value of time. The roadway network assignment model design also must consider any truck as- signment classes generated by a commercial or truck auxiliary model component. Transit. There are two basic options for rep- resenting transit submodes: explicitly modeling submodes in the activity-based model using sub- modal skims produced by the transit network model or allowing the transit network model to select the submodes, and passing composite transit mode skims back to the activity-based model. If the first option is chosen, the net- work assignment model design must reflect any transit submodal detail included in the activity-based model mode choice component. Activity-based models usually include some ad- ditional submodal detail such as distinguishing local buses, premium buses, light rail, and com- muter rail. Recent activity-based models have included enhancements that eliminate the need to generate separate transit choices by access mode (i.e., walk-access transit alternatives and drive-access alternatives), which simplifies the transit network assignment modeling process. Active. Activity-based mode choice models are increasingly sensitive to active transportation models such as bicycling and walking. These sensitivities are predicated on having network assignment modeling processes that can pro- duce enhanced active transportation network impedance measures. In the limited number of model systems that have incorporated enhanced active transportation sensitivities, the active transportation network assignment modeling tools used are distinct from the roadway and transit network assignment modeling tools. 2.2.3.2.4 Network Performance Measures The roadway, transit, and active transporta- tion network assignment models provide two primary types of information: network imped- ance measures and network flows. The net- work impedance measures are indicators of travel times, costs, distances, transfers, and boardings to and from different places. The network flows are indicators of total volumes on roadway, transit, and active transportation network links by mode and time of day. The information generated by the network assign- ment models to support activity-based models is generally consistent with the information generated to support traditional trip-based models, although there may be some additional segmentation or details generated. 2.2.3.3 Land Use Model Design Considerations Although land use models have existed for years, in many ways land use modeling is still an emerging practice, and the underly- ing theories, policy sensitivities, and software

45 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM capabilities vary widely by land use model. Some models may represent the actions of dis- aggregate households, individuals, firms, and developers; may incorporate more complex interactions within a regional economy; or may address long-term demographic changes. Transportation system performance indicators derived from activity-based or traditional travel demand models are key inputs to many land use models, and the locations of households and employment produced by a land use model are key inputs to activity-based and tradi tional travel demand models. The design of the land use model should consider the policy analysis requirements of an agency, the measures that can be produced as output from the activity- based model for input to the land use model, and the information required as input to the activity-based model that is produced by the land use model. 2.2.3.3.1 Policy Analysis Needs The motivation for linking a travel demand model with a land use model is a recognition of the complex interactions between trans- portation system performance, land use, and regional economics (Johnston and McCoy 2006). An agency may desire to understand and evaluate the influences of transportation network performance and accessibility on land use development patterns as well as to under- stand how these land use development pat- terns, in turn, influence transportation system performance (Fehr & Peers 2007). Agencies have developed linked land use and travel de- mand models in order to represent the effects of land use planning and zoning constraints, to generate better long-range estimates for input to air quality and emissions models, to evaluate transit-oriented development scenarios, to pro- vide new regional indicators such as housing affordability, and to address many other policy questions. Because of the tremendous variety of land use model structures, policy sensitivities, data requirements, and output capabilities, agencies must carefully consider which policy or analysis questions are of greatest interest or concern as well as what data are required for both model development and application. 2.2.3.3.2 Consistency with Other Model System Components The earlier discussion of linkages between activity-based models and network assignment models proposed that the policy analysis capa- bilities of the overall model system are intrinsi- cally related to the issue of maintaining con- sistency among model system components. The spatial and typological resolution of the land use model must inform and align with these same dimensions in the activity-based travel demand model. For example, if the activity- based demand model design uses information about the locations of housing and employ- ment at the spatial resolution of TAZs, then ideally the land use model can produce hous- ing and employment information at this same spatial resolution. Typologically, if the activity- based model requires as input information on employment by detailed industrial sector, then ideally the land use model should produce in- formation using the same level of detail. How- ever, the flow of information is not simply from the land use model into the activity-based model. Information on network performance output by the activity-based model and the network assignment model is also key input to a land use model. The land use model design must consider this information flow as well. 2.2.4 Model Integration Considerations All activity-based model systems comprise a series of subcomponent models that exchange information in a systematic way to simulate an outcome. In addition to the issues of pol- icy sensitivity and information consistency,

46 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER agencies developing integrated models must also specifically consider how these different model components are configured to interact to achieve an overall solution. Some compo- nents may be interacted to equilibrate to a con- vergent solution or at least a stable solution, while other components may be interacted in a more path-dependent manner. This design must consider the interactions among sub- components of the activity-based model sys- tem (such as the destination-choice models and mode choice models), the interactions between the activity-based model and the network as- signment model, and the interactions between the linked activity/network assignment model and the land use model. Implementing an inte- grated model system involves configuring indi- vidual model components as well as configur- ing an overall model system flow. 2.2.4.1 Activity-Based Model Components Linkages Activity-based models include a number of subcomponent models that interact and which are intended to provide behavioral realism by addressing numerous choice dimensions such as activity generation, destination choice, mode choice, and time-of-day choice. These subcomponent models are linked and executed in a manner that is intended to realistically represent the interaction of the various impor- tant dimensions of choice that individuals and households face in carrying out their daily ac- tivities and travel, as discussed in an earlier sec- tion (Bowman 1998). Typically a set of multi- nomial logit and nested logit choice models is estimated and implemented (Bowman 1998). The activity-based model components do not equilibrate explicitly, although measures of accessibility from lower models such as mode choice are fed back up to higher-level models such as automobile ownership. Most activity- based models are implemented using Monte Carlo simulation, which means that they are subject to some degree of simulation variation. In some regions, multiple activity-based model simulations are executed and averaged before being used as input to the network assignment model. 2.2.4.2 Activity-Based Model–Network Assignment Model Linkage There is essentially a two-way exchange of in- formation between the activity-based model and the network assignment model. In this ex- change, the activity-based model provides esti- mates of travel demand that are used as input to the network assignment model. In turn, the network assignment model uses this travel de- mand information to generate estimates of net- work performance that are then used as input to the activity-based model. In order to facili- tate this exchange of information between the activity-based model and the network assign- ment model, it is essential that there be some basic typological, spatial, and temporal consis- tency between the two models and their inputs and outputs. In addition to this data consistency, it is also essential that the model system incor- porate a systematic means of iteratively execut- ing the model system components in order to achieve a convergent, or at least a stable solu- tion. These iterative feedback strategies vary in their complexity. Some employ averaging pro- cedures in which network impedances, demand trip tables, or link volumes are averaged and delays recomputed. Rarely will naïve strategies, in which there is a direct exchange of informa- tion, result in a model system that achieves a stable or converged solution efficiently. 2.2.4.3 Activity-Based/Network Assignment Model–Land Use Model Linkage If a regional agency has a land use model that it wishes to link with the integrated activity- based/network assignment model system,

47 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM there is typically a two-way exchange of in- formation between these model system com- ponents. The land use model often provides information on location of housing units and employment (typically with some detail by in- dustrial sector), although in some cases land use model outputs may need to be translated into the basic units that are used as input to the activity-based models. The land use model may also be used to provide some estimates of long-term choices such as work locations. The integrated activity-based/network assignment model provides estimates of network imped- ances (travel times, costs, accessibility mea- sures) that influence the location of housing and employment location choices predicted by the land use model. The predictions about lo- cations are then used as input to the integrated activity-based/network assignment model. Unlike activity-based model/network assign- ment model linkage, which is configured to achieve either a convergent or stable solution, most integrated travel demand/land use models are path-dependent. There is no assumption of a convergent solution across time. 2.2.5 Performance Metrics Activity-based model systems produce a broader range of performance metrics than traditional trip-based models. Activity-based models can provide more explanatory power to decision makers because they consider the interrelated nature of the activities and travel that individuals participate in, as well as how individuals within a household may coordinate their activities. These metrics can be more in- tuitive than those generated by a trip-based model, which can help facilitate communica- tion with decision makers. Two main types of activity-based model system performance metrics can be considered: (1) metrics derived directly from the activity-based model compo- nent and (2) metrics derived from the network assignment model that is linked to the activity- based model component. 2.2.5.1 Activity-Based Model A distinguishing feature of virtually all activity- based models is that the primary outputs from the model are estimates of travel demand in the form of disaggregate lists of individual tours and trips rather than aggregate trip tables. The list of tours and trips output from the activity- based model is essentially a travel diary that is similar to that provided by a detailed house- hold travel survey, with full spatial and tem- poral consistency within each person’s daily travel for all persons in the population. This disaggregate approach to representing travel choices provides the ability to track important measures, such as daily vehicle miles traveled (VMT) by household (Resource Systems Group 2012b), and to perform detailed and targeted equity analyses (Castiglione et al. 2006). Activity-based models can produce all the types of performance measures that are typically generated by trip-based model systems, such as mode shares, trip length frequency distribu- tions, and trips. But activity-based models also have the ability to report a broader range of per- formance metrics. For example, activity-based models can produce detailed information about tour and trip making, capturing the trade-offs between making multiple tours or linking to- gether activities into fewer tours but with more trips. Tour and trip rates can be calculated from activity-based model outputs rather than be used as fixed inputs to a trip-based genera- tion model. As a result, an activity-based model can indicate how tour and trips rates may vary as a result of different levels of accessibility. Be- cause activity-based models explicitly represent time-of-day choices, often using a temporal res- olution of half-hours or even finer, measures of trip making by detailed time of day and activity durations can be easily generated.

48 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 2.2.5.2 Network Assignment Model When linked with an activity-based model, static network assignment models can produce all the metrics that are typically produced when such network assignment models are linked with traditional trip-based models, such as link volumes and congested link travel times. Because activity-based models include explicit time-of-day choice models, the static net- work assignment models that are linked with activity-based travel models often include more time periods than those linked with trip-based models. In addition, because the dis aggregate structure of the activity-based model system can more easily support additional market segmentation such as toll and value-of-time classes, static network assignment models linked to activity-based models may have addi- tional assignment classes. The types of perfor- mance metrics that can be derived from a DTA model are considerably more extensive than the types that can be derived from a static as- signment model. Their more detailed metrics may include system delays and volumes at fine- grained temporal intervals, queues, and other more detailed operational statistics. 2.2.5.3 Reporting and Visualization As previously described, the disaggregate nature of the activity-based model facilitates more detailed reporting, such as the ability to support detailed analyses of effects on commu- nities of concern. The primary output formats of activity-based model components are nota- bly different from the output formats from tra- ditional trip-based models. Whereas trip-based models employ matrices indexed by origin and destination zones as the primary format for travel demand data, activity-based models use lists of trips and tours. Use of tour and trip lists is more efficient than using matrices, because matrices frequently have large numbers of O-D pair cells that are either empty or have small fractional values of trips. In contrast, all the tour and trip lists generated by the activity-based model represent discrete tours and trips, with no inefficient storage devoted to representing no-demand or fractional-demand zone pairs. Tour and trip lists, as well as other disaggre- gate inputs and outputs from an activity-based model system, can easily be used to perform re- lational queries and provide flexible summaries that can easily be visualized using standard sta- tistical and geographic information software. 2.2.5.4 Uncertainty Analysis Many activity-based models employ Monte Carlo simulation in order to realize discrete outcomes from the probabilities predicted by the model components. Use of Monte Carlo simulation represents a fundamental shift in model implementation, from more determin- istic approaches to forecasting demand, which employs fixed trip generation rates and ap- plies mode shares to aggregate estimates of demand (Rasouli et al. 2012). When using Monte Carlo simulation, the same probability distributions may result in different outcomes based on the random numbers used to select choices. Therefore, an important consideration when applying activity-based models that em- ploy Monte Carlo simulation is identifying the number of runs necessary to have confidence in the model outputs. Generally, more runs are required when seeking stability for rarer choices, such as information for smaller geo- graphic areas or travel modes that compose a small market share. Empirical investigations of activity-based models have shown that all activity-based models were generally stable and that only a relatively small number of runs are typically required to address simulation error (Castiglione et al. 2003). In practice, only a few existing activity-based model systems explicitly address simulation variation.

49 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM 2.3 DATA DEVELOPMENT Activity-based model system development, like trip-based model development, requires assem- bling a diverse set of data. These data reflect travel behavior, regional demographics, land use, network configuration, and network per- formance, and many of the required data items must be available for all base-year, future-year, or alternative scenarios. 2.3.1 Activity-Based Model The data required for development of the activity-based model system come from ex- ogenous and endogenous sources. Exogenous information sources include items such as a household travel survey, base and future-year land use, and demographic and economic as- sumptions. Endogenous information sources include synthetic populations and network performance indicators generated by ancillary tools. 2.3.1.1 Survey Depending on the specific development path pursued by an agency, household travel survey data collected to support trip-based model de- velopment can usually be used to support the estimation and calibration of the activity-based models. Additional analysis of the survey data is required to chain the trips into tours and to classify the sequence of tours into relevant descriptors of a full-day activity and travel pattern. Household survey data requirements are greater if new activity-based model com- ponents are to be estimated from the survey data, as is the case with upfront development or incremental implementation activity-based model development. Activity-based models in- corporate a wider variety of sociodemographic variables and types of choice alternatives than trip-based models and, as a result, the sample size requirements tend to be larger (Bowman et al. 2013). A recent study of the transfer- ability of activity-based model parameters (Bowman et al. 2013) concluded that a survey sample of at least 6,000 households may be adequate for a medium-to-large region. If an existing activity-based model is to be trans- ferred and refined, the household survey data requirements may be less because the data will primarily be used to derive calibration targets. Another consideration is whether the activity- based model will consider joint travel across multiple members within a household. If so, it is essential that survey data contain complete data across all household members and that the survey successfully captures instances when household members traveled together and per- formed activities together. 2.3.1.2 Employment and Land Use Land use data needed for activity-based models are similar to data needed for trip-based models, including the number of households, number of jobs by sector, and school enroll- ment by school level. If an activity-based model is specified to use the same or similar basic TAZ spatial units as a trip-based model, then there is little or no difference in the data required for an activity-based model as compared to a trip- based model. However, recent activity-based models have begun to use basic spatial units that are smaller than standard network TAZs, such as microzones and parcels. The use of parcels can entail some challenges, such as the ability to obtain accurate and up-to-date parcel data for all jurisdictions in a region, heteroge- neity in how parcels are defined for different land uses, and challenges in specifying forecast year land use at the parcel level. The primary goal of including finer geographic detail is to model travel behavior at a level of detail closest to what decision makers experience, in order to avoid statistical aggregation bias in the models.

50 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 2.3.1.3 Synthetic Populations and Demographic A key input to most activity-based models is a synthetic population that is used as the basis for forecasting the behavior of the households and persons in the modeled area. The specific data required to generate a synthetic population are influenced by the design of the synthetic popula- tion, which is itself reflective of the specification of the activity-based model system components. Although there are a vari ety of population syn- thesis approaches and tools and the data re- quirements of these tools vary, in general there are two primary data inputs to most synthetic population generation tools (Bowman et al. 2013).The first primary input is control data. These control data represent the attributes that are being explicitly accounted for in the gen- eration of the synthetic population. The more attributes that are explicitly controlled for in the synthetic population design, the greater the data requirements in both the base year and for any alternative or horizon year. In addition, the control data information must be provided at relatively detailed geographic levels, and infor- mation at multiple geographic levels are often combined. These data are typically derived from Census data and from external demographic forecasts. These demographic forecasts may be based on growth factor models, land use models, or other methods. The second primary input is sample data. After the control data have been used to identify a multidimensional distribution of households and population at a fine-grained spatial level, it is necessary to then sample households and persons to create a list of households and persons that matches these distributions for input to the activity-based model system. In the United States, Census PUMS data are often used as the source for this sample, although it is also possible to use a regional household survey as the basis for the disaggregate sample (Resource Systems Group 2012d). 2.3.1.4 Network Performance Indicators Network performance indicators, often called network skims, are a fundamental input to the activity-based model. Information about travel times, costs, and other metrics by time of day and market segment are used in both the es- timation of activity-based model parameters, as well as in the application of the final activity-based model system. Network skims are endogenous to the overall activity-based model system because they are produced by the network assignment model component that is linked to the activity-based model component. The network skims should be carefully con- structed to provide unbiased information re- quired by the activity-based model component. For example, if the activity-based model com- ponent includes transit submode alternatives in the tour or trip mode choice components, then detailed information on transit network performance by submode should be generated by the network assignment model component. However, it is not always practical to ensure complete consistency between the network performance indicators and the activity-based model component structure. For example, the activity-based model component may use half- hours as the fundamental temporal unit, but it is typically not realistic for a static network as- signment model to generate performance indi- cators at this resolution. Note that the network assignment procedures used in the model sys- tem are consistent with the network skimming procedures. Accessibility indicators are also frequently included in activity-based model component specifications and must also be endogenously generated by the overall activity-based model system. These accessibility indicators represent the combined effects of land use and transpor- tation system performance. Simple accessibil- ity indicators include buffer measures, such as the number of jobs within a fixed travel time,

51 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM while more complex accessibility indicators can include combination mode-choice and destination-choice logsums. 2.3.1.5 Calibration and Validation Calibration refers to the process of adjusting model parameters to better match some base case observed conditions, while model valida- tion involves the application of the calibrated model and the comparison of the results to observed data that have not been used in the model estimation or calibration process. Each component of the activity-based model system is individually evaluated relative to observed data source. Automobile ownership model re- sults are usually compared to household sur- vey data or to Census data. Activity genera- tion and time-of-day model results are initially evaluated relative to targets derived from the household survey, although these targets are sometimes adjusted to be consistent with ob- served traffic and transit counts by time of day. Work location choice models may be compared to either Census journey-to-work data or to household survey data, while the destination- choice models for all other activity purposes are compared to household survey informa- tion. Mode choice results are evaluated relative to calibration targets derived from the house- hold survey but may also be adjusted in order to ensure consistency with observed traffic and transit counts. The network assignment model component and overall model system calibra- tion and validation area primarily evaluated by comparing the estimated traffic and transit volumes to observed count data. Speeds may also be evaluated, although static assignment models often produce unreliable speed esti- mates. Comparisons to regional VMT statistics derived from highway monitoring systems as well as other measures, such as transfer rates derived from transit on-board surveys, are also often used. 2.3.1.6 Base Year Versus Forecast Year In order for activity-based models to generate estimates of future travel demand, future-year inputs to the model are required. Most critically, these inputs include future land use, demo- graphic, and economic assumptions, as well as future network configuration assumptions. The availability of future input assumptions should be a key consideration when designing the model system specifications. For example, if detailed estimates of employment by industrial sector are unavailable for forecast years, then it may be ad- visable to employ a simpler destination-choice specification that reflects only the available data. 2.3.2 Network Assignment Model The data required for development of the net- work assignment model are determined by the overall model system design and are primarily derived from existing exogenous data sources. Information about market segmentation such as modal detail as well as temporal detail sig- nificantly influences the type of information re- quired to build the network assignment model. There are many robust sources for information for building network assignment model inputs, although extensive checking and cleaning of the network are often required to ensure the integrity of the network. 2.4 IMPLEMENTATION 2.4.1 Estimation Estimating the component models for an activity-based model system involves using local data to identify the variables that are most important to activity and travel decision mak- ing and quantifying the relative importance of these variables. In addition to processing sur- vey data and attaching relevant data from net- work skim matrices and land use data, which are required whether or not the implementation

52 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER involves model estimation, it includes specify- ing the model utility functions and alternative availability constraints in a model estimation software package, and then carrying out the estimation in an iterative process. This process requires expert judgment to determine what variables to include and sometimes involves constraining some coefficients to typical values in cases where the data for estimation are in- adequate. The most typical example of this is in the travel time and cost coefficients of mode choice models, where the data for estimation do not give reasonable results in terms of im- puted time and cost trade-offs (value of time). The need to constrain time and cost coefficients often arises because there are very few non- automobile observations in the data, so there is very little observed time and cost trading be- tween modes. In such cases, one may use coeffi- cients from previously revealed preference and stated preference models or from project docu- ments such as the SHRP 2 C04 report (Parsons Brinckerhoff et al. 2013). The estimation task has typically been car- ried out by consultants rather than by the re- gional planning agency staff. Recently, some of the activity-based model software packages have been designed to make it easier to modify and re-estimate activity-based component models when starting from a known model specification (e.g., from another region) rather than when starting from scratch. So, rather than simply transferring another model and calibrating a few specific variables to observed data (a process discussed in Section 2.4.3), it may be possible for local agency staff to re- estimate all parameters of the model system based on local survey data, provided that the data are sufficient in quality and sample size. That process would also typically benefit from some guidance from outside consultants, unless the agency has the required expertise in-house. 2.4.1.1 Data Requirements and Preparation The data required to estimate the model com- ponents are, for the most part, the same as required to apply the model in the base year, including • A land use database with all variables that will be used in model, at the level of spatial detail desired for the ultimate model (TAZ, microzone, or parcel); • TAZ-to-TAZ network skim matrices for all travel time and cost variables, for all modes and time periods that will be used in the model; • Additional data for distance along an all- streets network for short trips, if that is in- cluded in the model design; and • Data from a household travel diary (or wearable GPS-based) survey, containing all trips and activities for all household mem- bers for at least one full day. The last data component, the survey, is the only one that is not used in model application, but it should contain data on all of the impor- tant household and person characteristics that will be included in the synthetic population to which the models will eventually be applied. The survey data, when expanded to represent the full population, are also typically used in the later model calibration stage. The data preparation stage typically re- quires processing the survey data to create files at all the different levels at which the model will be applied: • The household and person levels (to model longer-term and mobility choices); • The day level (to model tour and trip gen- eration patterns at the full-day level); • The tour level (to model tour-level choices); and • The trip level.

53 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM Many survey data sets already are struc- tured into household-level, person-level, and trip-level records, so two of the main steps in the data processing are tour formation and day-pattern formation. Tour formation is the process of combining home-based and work- based trip chains into tours and writing records with the relevant tour attributes. Day-pattern formation is the process of combining infor- mation on all the tours made by persons and households across the day, to be classified into types of full-day activity patterns, depending on the model design. There are no standard procedures for tour formation or day-pattern formation, as they match the model design, although the hierar- chies used across travel modes (to determine the main mode of a tour) and activity purposes (to determine the primary destination of a tour and/or the primary tour of the day) tend to be quite similar across different model designs. The most complex logic tends to arise in model systems that explicitly model joint tours and joint half-tours across household members, as this adds another layer of logic that must be included in the data processing, and the logic can be very complex in survey datasets where the travel records of household members who actually travel together do not match up very well, leaving the analyst to do a fair amount of guesswork regarding whether household mem- bers actually performed certain trips and activi- ties together or not. 2.4.1.2 Model Design and Testing The specification of the component models of an activity-based model system is, along with the overall model system structure, of funda- mental importance in determining how the model system performs in forecasting. A model component’s specification includes its struc- ture (e.g., alternatives and nesting), as well as the variables and coefficients included in each alternative’s utility functions. If a model speci- fication lacks aspects of how conditions affect behavior, or if it specifies them incorrectly, then the model will not behave correctly when those conditions change; rather, it will either not re- spond at all or it will respond with bias. The ability to correctly specify a model is limited by the quantity and quality of the data, as well as by the modeler’s understanding of the be- havior being modeled and model development expertise. On the one hand, it is easy to over- specify a model through a process known as data mining, which tries to find all possible statistically significant effects and ends up in- cluding effects that represent idiosyncrasies in an imperfect sample. On the other hand, it is easy to underspecify a model by excluding all coefficients that fail a certain significance test, thereby excluding an important variable—one that should be there but is insignificant merely because of a small data sample—and caus- ing the resulting model to be biased. In prac- tice, there are many possible ways to specify a model, and no two modelers do it the same way. It is important that the modelers respon- sible for specifying and estimating models have a solid understanding of model development theory and well-developed skills in model de- velopment practice. 2.4.2 Software Development There are currently three or four software platforms that are used for the large majority of activity-based models implemented in the United States. These have been developed by the consultants who have created the original models and then adapted and improved over time as they are implemented for new regions. The software field for activity-based models may change substantially in the future, how- ever, as the market for activity-based models matures.

54 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 2.4.2.1 Components Not surprisingly, the components of the soft- ware system tend to mirror the components of the model system itself. Typically, a software platform uses objected oriented code (e.g., in C#, Java, or Python) to create general model classes that can then be customized to accom- modate each separate model component. 2.4.2.2 Process Flow The most complex part of the software design tends not to be in the programming of each sep- arate model component but in how the infor- mation is passed from each model to the next, and from the user to the model and back to the user again. This is typically done in the context of an iterative feedback loop with network as- signment software, so that the final levels of travel demand and congestion will be consistent between the activity-based model outcomes and the network assignment model outcomes, at least for the highway assignment. When considering software implementa- tions, the main factors that a potential user may be most concerned with are • Model documentation and clarity: How easy is it to understand how the various models are coded and how the data flow through the model? • Front-end data input: How easy is it to specify and create the specific types of data that go into the model and to detect errors in the data before they influence the model results? • Code efficiency: What are the memory and hardware requirements for using the soft- ware, and how fast will the model run on various types of hardware? • Back-end data output: Are there ways of analyzing and visualizing the data that are produced from the model system (other than via the network assignment model software package that the system is inte- grated with)? • Configurability: How difficult is it to turn certain features in the model on or off, or to change certain aspects of the model such as the number of modes used, the number and definition of time periods used, and so forth? • Re-estimation and calibration: Does the software platform contain features that automate some of the work necessary to re-estimate models on new survey data, or to calibrate the model to match external data? 2.4.3 Transferability Given the extensive work involved in design- ing a new activity-based model system, esti- mating the component models, and creating the software to implement the models, it can be expected that most activity-based models implemented in the future will be models trans- ferred from another region, starting with the model design, software platform, and model coefficients from that previous model and then re-estimating and/or re-calibrating some of the coefficients. In general, one expects more be- haviorally detailed models to be more transfer- able across regions than very simple models, because many of the variables that are different across regions—different socio demographics, different land use patterns, different mode availability, and more—have been incorpo- rated more fully into explicitly modeled effects rather than left as part of the error components that are captured in alternative-specific con- stants and the overall scale of the coefficients. A recent Federal Highway Administra- tion (FHWA) study on model transferability ( Bowman et al. 2013) recommends full re- estimation may only be beneficial in cases where there are new survey data available with

55 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM a very substantial sample size. Otherwise, it is likely to be more accurate to use the coefficients from a model system estimated on a large sur- vey in a comparable region elsewhere, and use the local survey data to simply recalibrate cer- tain key model coefficients such as alternative specific constants. 2.4.4 Synthetic Population Earlier sections have identified a synthetic pop- ulation as a key input to most activity-based models. This synthetic population is used as the basis for forecasting the behavior of the households and persons in the modeled area. The synthetic population process uses both aggregate information about the distributions of households and persons along key demo- graphic dimensions, as well as detailed dis- aggregate household and person records, in order to create a full enumeration of regional households and persons. In order to implement a synthetic population process, it is necessary to establish a population synthesis design that outlines the demographic attributes that will be controlled for, to gather the information re- quired to implement the population synthesis, and to validate that the process is producing reasonable results. 2.4.4.1 Design and Controls The specific data required to generate a syn- thetic population are influenced by the design of the synthetic population, which is itself re- flective of the specification of the activity-based model system components. The first input is marginal “control data,” and the second input is “sample data.” 2.4.4.1.1 Marginals The control data represent the attributes that are being explicitly accounted for in the gen- eration of the synthetic population. The more attributes that are explicitly controlled for in the synthetic population design, the greater the data requirements in both the base year and for any alternative or horizon year. In addition, the control data information must be provided at relatively detailed geographic levels, and infor- mation at multiple geographic levels are often combined. The types of marginal controls used for generating synthetic populations, and for which base-year data and future-year data should be generated, often include attributes such as household size, household income, and household workers. A more detailed list can be found in the earlier synthetic population discussion. Data sources for marginal controls may come from a variety of sources. Ideally, this information is derived from a demographic forecasting model or method that can provide base-year and future-year distributions for all of the marginal controls. However, it is more common that agencies may have information on only a limited number of these marginal controls. In these cases, base-year distributions may be assumed to remain fixed, although such an assumption will influence the distribution of other marginal controls. 2.4.4.1.2 Samples The second data input is sample data. After the control data have been used to define a multi- dimensional distribution of households and population at a fine-grained spatial level, it is necessary to then sample households and per- sons to create a list of households and persons that matches these distributions for input to the activity-based model system. The samples used to generate the synthetic population must have the detailed information corresponding to the marginal controls. Data sources for samples are typically more limited than sources for marginal con- trols, because these samples ideally contain detailed information. The samples may include

56 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER both controlled for and uncontrolled for attri- butes. In the United States, Census PUMS data are often used as the source for this sample, although it is also possible to use a regional household survey as the basis for the disaggre- gate sample. 2.4.5 Feedback, Convergence, Equilibration Linked demand-and-supply model systems such as activity-based model systems typi- cally include processes of iterative feedback. These feedback processes are implemented in the network assignment model and between the network assignment model and the activity- based model components. Iterative feedback is used to ensure that the models are achieving convergence to an equilibrium, or at least a sta- ble condition. Convergence is important within the context of activity-based model systems because it provides confidence in the integrity of the model system and helps ensure that the model will be a useful analytic tool. 2.4.5.1 Network Assignment Model Convergence In the context of an activity-based model sys- tem, an essential precondition for pursuing overall model system convergence is establish- ing network assignment convergence. In static network assignment models, convergence to a user equilibrium condition is typically pursued. User equilibrium is described as a condition when travelers cannot reduce their travel costs by unilaterally changing their route. There are well-established procedures for pursuing convergence to a user equilibrium condition. As convergence is approached, estimated link volumes and impedances stabilize, as do the aggregate network impedance skims that are produced by the network assignment model component and input to the activity-based model component. Although less stringent convergence may be acceptable in early model system iterations, later model system iterations should incorporate stricter network conver- gence criteria. 2.4.5.2 Linked Activity-Based Network Assignment Model System Convergence When convergence is achieved, the network performance measures that are used as input to key activity-based model components are consistent with the network performance mea- sures output by the network assignment model when the activity-based demand is assigned. This consistency is important for establishing that the activity-based model system will be a useful tool for analysis. The stability of model outputs is essential to support planning and engineering analyses, and changes to demand or supply should lead to reasonable changes in model outputs. 2.4.6 Integration with Auxiliary Model Components Activity-based models typically provide esti- mates of weekday travel demand made by regional residents when traveling within the region, but this demand makes up only a por- tion of the total travel demand that uses re- gional transportation networks. In addition to the internal demand generated by the activity- based model system, the overall model system also must incorporate auxiliary demand, such as trips in which one or both trip ends is out- side the regional (often referred to as external trips), truck and other commercial demand trips, trips made by nonresidents or visitors traveling within the region, and trips associated with special purposes or events. Often, this auxiliary demand is represented in a region’s existing trip-based model system but must be refined or enhanced to facilitate integration with the more detailed demand generated by the activity-based model. Note that estimates of auxiliary demand are combined with the de-

57 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM mand estimates generated by the core activity- based model components prior to the network assignment stage in the overall model system. 2.4.6.1 Internal-External and External-External Travel Internal-external and external-external de- mand represents travel in which one or both of the trip ends is outside the modeled area. In activity-based model systems, this travel demand is often generated from two sources. Most of this demand is generated by separate internal-external models. These models provide estimates of flows to and from critical external stations that are located at the boundaries of the modeled area. These estimates may be de- rived from data such as existing traffic counts or from estimates of base-year and future-year traffic volumes at these gateways. The level of typological segmentation can vary, with some internal-external models providing detailed estimates of flows by vehicle occupancy class and truck type and others providing aggre- gate estimates of total vehicle flows. In either case, these segments must be aligned with the classes used in the network assignment model. Also, it should be noted that in some activity- based model systems at least a portion of the internal-external demand can be generated by the activity-based model, which may predict which regional residents are commuting out- side the modeled area or which regional jobs may be filled by people commuting in from outside the modeled area. 2.4.6.2 Freight and Commercial Vehicles Modeling freight and commercial vehicle travel demand is an important component of the over- all model system, as these vehicles can represent between 10% and 20% of the volumes on re- gional roadway networks. Precise estimates are difficult to obtain because vehicles being used for commercial purposes are not always iden- tifiable. Some regional model systems include components that distinguish between freight and nonfreight (commercial) vehicle move- ments, while others use a set of more generic truck classes, often segmented by size. More advanced practice includes the development of tour-based models, as well as the linkages be- tween freight demand and regional land use models. When integrating estimates of freight and commercial vehicle travel demand, agencies should pay particular attention to the temporal aspect of this demand, as the timing of this de- mand is often different from the diurnal pattern observed with passenger travel demand. 2.4.6.3 Special Purpose Models In addition to internal-external travel demand and freight and commercial travel demand, con- sideration should be given to other unique travel markets that may affect overall travel demand and network performance. Many regions have developed airport models, which predict travel demand associated with non regional resident visitors arriving at the airport and accessing re- gional destinations, as well as regional residents traveling to the airport to depart for nonregional destinations. Some regions that have significant numbers of tourists have implemented visitor models that estimate the travel demand of visi- tors to the region, using information on hotel room locations and key regional visitor desti- nations. Other special purpose models have in- cluded border crossing models and models that generate demand associated with special events such as sports competitions. In general, the de- mand from these special purpose models can be easily integrated in the activity-based model system, provided that attention is paid to ensure temporal, typological, and spatial consistency. The regional household survey typically does not include sufficient information about these special purpose markets, so additional data col- lections are often required.

58 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER 2.4.7 Calibration and Validation Calibration refers to the process of adjusting model parameters in order to better match some base case observed conditions. Model validation involves the application of the cali- brated model and the comparison of the results to observed data that have not been used in the model estimation or calibration process. Ad- justments to other model assumptions and in- puts are also often made during the calibration and validation process (Cambridge Systematics 2010). It has been argued that it is necessary to be able to match base conditions before using a model for any future-year application ( Cambridge Systematics 2010). Calibration and validation of activity-based models bears some similarities to the calibra- tion and validation of traditional trip-based model systems in that the individual model sub components are individually calibrated, and then the overall model system is validated. Activity-based model calibration and validation efforts are different from trip-based model ef- forts primarily because there are more compo- nent models to evaluate and adjust. However, the overall level of calibration and validation effort for activity-based models is usually not significantly different from trip-based models, because fewer adjustments to any individual model component are typically required. The calibration and validation process should be systematic. A model calibration and validation plan should ideally be established early in the model development process, and required data identified and collected. Each component of the activity-based model, as well as the network assignment model, is then evaluated and adjusted as necessary, ultimately leading to the validation of the overall inte- grated model system. For each model system component, key metrics are identified, and comparisons between model estimates and ob- served data are made, which inform the adjust- ment of model parameters. The strategies used to systematically adjust model parameters vary by model component. Some components such as automobile availability models may be cali- brated by simply adjusting alternative specific constants based on the ratio between observed and estimated values, while other components such as tour destination-choice models may require more complex calibration strategies in order to ensure that tour length frequency distributions are reasonable. Calibration and validation of an activity-based model system is an iterative process due to the activity-based model components being interrelated. Adjust- ments to upstream model components affect downstream model components and, ulti- mately, the network performance measures. These network performance measures, in turn, affect the upstream model components, neces- sitating the re-evaluation and adjustment of all model component calibrations until no signifi- cant changes occur. Finally, the calibration and validation process should be documented, iden- tifying adjustments made to model parameters and assumptions and summarizing final results (Cambridge Systematics 2010). 2.4.7.1 Activity-Based Model Model calibration and validation efforts usually begin after the entire model system has been im- plemented, including any estimation or transfer of model coefficients, development of any auxil- iary model components, and completion of any software code or scripts required to implement model components or to facilitate the linkage with the network assignment models or to con- trol the overall model system flow. 2.4.7.1.1 Components The results of each individual component of the activity-based model system are compared to observed data in order to ensure that all components are producing reasonable results.

59 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM Automobile ownership model results are usu- ally compared to household survey data or U.S. Census data. Activity generation and time- of-day model results are initially evaluated relative to targets derived from the household survey, although these targets are sometimes adjusted to be consistent with observed traffic and transit counts by time of day. It is not un- common for household surveys to underreport nonmandatory activities or to miss short trips, which often occur during off-peak travel times. Work location choice models may be compared to either Census journey-to-work data or to household survey data, while the destination- choice models for all other activity purposes are compared to household survey informa- tion. Mode choice results are evaluated relative to calibration targets derived from the house- hold survey, but may also be adjusted in order to ensure consistency with observed traffic and transit counts. 2.4.7.1.2 Model System The overall model system is primarily evalu- ated by comparing the estimated traffic and transit volumes to observed vehicle count data. Speeds also may be evaluated, although static network assignment models, which are the pri- mary type of network assignment model linked to activity-based models, often produce unreli- able speed estimates. Comparisons to regional VMT statistics also are often used. 2.4.7.2 Network Assignment Model The calibration and validation of the network assignment model component are performed in conjunction with the calibration of the activity- based model component because of the linked nature of these two primary model system components. However, considerable additional effort is typically expended in the final network assignment model calibration and validation to ensure that the network assignment model results look reasonable not only at aggregate regional levels, but also at the level of detailed individual links, and that differences in vol- umes between scenarios are also reasonable. Link volumes by time of day are one of the main metrics produced by the model system of interest to analysts and decision makers. 2.4.7.2.1 Static Network Assignment Two types of modifications are usually per- formed as part of the calibration and valida- tion of static network assignment models: (1) adjustments to the volume delay functions that control the relationship between forecast volumes and speeds and (2) adjustments to network assumptions. These adjustments are analogous to those that would be made if the static network assignment model were linked to a trip-based model. Volume delay func- tions are often stratified by facility type (e.g., freeways, major arterials, minor arterials, and collectors), and the volume delay function parameters are adjusted until a reasonable dis- tribution of flows by facility type is achieved. Network assumptions, such as hourly per lane capacities, freeflow speeds, network loading points, and transit transfer penalties, also are adjusted as necessary to achieve reasonable es- timates of traffic and transit volumes. As with trip-based models, making fine-grained adjust- ments to network assumptions in an activity- based model system can be time-consuming. 2.4.7.2.2 Dynamic Network Assignment Although dynamic network assignment models, often referred to as DTA models, have begun to be more widely used in practice in the past decade, there are very few examples of the inte- gration of dynamic network assignment models with activity-based models. The overall process for calibration and validation of a dynamic network assignment model is in many ways similar to the process for static network assign-

60 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER ment model, involving the adjustment of model parameters and input assumptions. However, calibration of a dynamic network assignment model is significantly more involved because of the additional details included in the model, such as explicit inclusion of traffic flow models and the vastly more detailed network input as- sumptions that affect results such as lane con- figurations and signal timings. The inherent stochasticity of dynamic network assignment models, which may represent network per- formance more realistically, also can lengthen model calibration and validation efforts. 2.4.7.3 Sensitivity Testing Sensitivity testing should be considered an in- trinsic part of the model calibration and vali- dation process. Matching observed base condi- tions is necessary, but not sufficient, to ensure that the model will provide reasonable results when used to evaluate a future-year or alter- native scenario. It is possible that adjustments made to match observed base conditions may, in fact, distort and compromise the sensitivity of the model. Sensitivity testing should not only confirm the model’s sensitivity to changes in policy and investment inputs but also confirm that these changes are reasonably consistent with real-world outcomes. 2.4.8 Application Model application refers to the use of the model to simulate alternative scenarios and to produce meaningful performance metrics in a timely manner that can inform decision mak- ing. Model application requires software that has been developed to implement the model de- sign and the availability of computing resources or hardware that can efficiently run the model software. Model application also involves ex- tracting performance metrics from the model system and evaluating whether these measures are reasonable. 2.4.8.1 Hardware and Software Hardware and software issues need to be con- sidered simultaneously in model design, imple- mentation, and application. The model design has hardware implications, such as the amount of memory that is required to ensure that data required for model application can be accessed quickly or the number of processors that are re- quired in order to ensure that model run times are reasonable. Distributed processing is a key issue. The software and hardware should be implemented in such a way as to use multiple processing cores on a single server or to pro- vide the ability to exploit processors across a cluster of multiple machines. Few of the general purpose, commercial travel demand forecasting model packages that are widely used by agencies are able to efficiently implement core activity-based model compo- nents and, as a result, much of the software that implements activity-based models are cus- tom products created by activity-based model system developers. In many cases, this software is governed by open source licenses, which, in theory, allows others to acquire, modify, and run the software. Until now, no one other than the original software developers has managed activity-based model code (Resource Systems Group 2012a). However, some agencies have recently initiated collaborative, activity-based model software development efforts. Although it is general purpose software, commercial travel demand modeling software has not been widely used to implement the core activity-based model components. This type of software is usually used in the overall model system typically to provide roadway and transit network assignment capabilities and to perform associated matrix manipulations. The core activity-based model components and the commercial travel demand modeling software should ideally be implemented to allow for the efficient exchange of network skim infor-

61 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM mation from the commercial package to the activity-based model and exchange of demand information from the activity-based model to the commercial package. 2.4.8.1.1 Configuration Options Activity-based model systems should provide the ability to configure individual model com- ponents as well as the overall model system. This ability to configure allows for the most efficient use of available computing resources and ensures that the model sensitivities can be fully exploited. Most activity-based model sys- tem components require some degree of con- figuration. For example, it may not be neces- sary to run a full sample of all households and people in the synthetic population in order to analyze every type of alternative scenario, or on all iterations within the equilibrating model system, so the ability to flexibly configure the sample rate is important. Activity-based model components ex- change information with network assignment models in a systematic way to create an overall model system that produces equilibrated solu- tions. Users may want to configure this overall model system equilibration process in order to achieve different levels of equilibration or stability or to ensure reasonable overall model system run times. It can also be useful to be able to configure or specify inputs associated with different alternative scenarios. 2.4.8.1.2 Flexibility and Extensibility Flexibility is a desirable quality in the software design. This flexibility can encompass many aspects of the model and software design. At the most basic level, model component soft- ware should allow users to update model as- sumptions that may need to be revised without fundamentally altering the underlying func- tionality of the model components or making changes to the underlying software code base, such as key coefficients and calibration con- stants. An intermediate level of flexibility may provide model users with the ability to redefine more fundamental aspects of the model system design, also without making changes to the underlying software code base. For example, some activity-based model system software has been designed to allow users to redefine the alternatives that are included in a model, or to revise the manner in which network skims are used. Advanced flexibility in software involves designing an overall model software architec- ture that allows the software code base to be easily extended to include new types of models or data structures. 2.4.8.1.3 Run Times Model run time is an important issue. Activity- based model software has matured significantly in the past 15 years, and the availability of in- expensive distributed and multithreaded com- puting resources has reduced activity-based model run times significantly. However, model run times may still vary depending on the spe- cific design aspects of the model components, the size of the population being simulated, and the extent of computing resources available. In addition, overall model system run times may be significantly affected by the run times associ- ated with the network supply component. Run times for the network assignment model com- ponents linked with activity-based models may be longer because of the inclusion of additional temporal, modal, or typological detail in the network assignment and skimming procedures. Long model system run times can limit the ability of agencies to evaluate multiple alternatives and to resolve any inconsistencies or errors that are discovered in the course of performing model runs. Long model system run times also may result in longer model de- velopment, calibration and validation times, as model development requires the iterative run-

62 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER ning and re- running of the model components and the overall model system. Model run times are typically reduced by providing additional computing resources and through software engineering and optimization. Consideration should be given not only to the model run times but also to the entire time required for the anal- ysis process. A model that runs quickly may not be helpful if it takes significant amounts of time and effort to extract meaningful statistics. 2.4.8.1.4 Software Transferability Cost, schedule, and data limitations are often cited as factors in developing activity-based model ing systems (Picado 2013, Resource Systems Group et al. 2014). Transferring activity-based model software from one region to another can be an effective means of addressing these concerns. There are numerous examples of activity-based model software being success fully transferred from one region to another. This transferability has taken different forms. In some cases, the software and coefficients have been transferred from one region and adapted for another. This can involve making minor adjustments to the activity-based model software and undertaking coefficient recalibration and revalidation effort (Resource Systems Group et al. 2014) to reflect local conditions. In other instances, more sub- stantive changes have been made to the model specifications and software code in order to provide enhanced model capabilities and to be more sensitive to critical local concerns (Picado 2013). 2.4.8.2 Data Extraction, Analysis, and Interpretation Activity-based model systems produce far more detailed and voluminous outputs than a traditional trip-based model. Although these detailed outputs facilitate more targeted analyses, they also require more skills to iden- tify and extract relevant measures. Simple spreadsheet-based methods used for trip-based model analyses are insufficient for analyzing the information-rich outputs of activity-based models. Many agencies store model output in relational databases or highly compressed and efficient data structures and then use a vari- ety of off-the-shelf tools such as GIS software, as well as custom analysis tools, to summa- rize and interpret model outputs. It should be noted that agency staff may spend significant amounts of time analyzing model outputs be- cause activity-based models can support many dimensions of analysis. Some agencies have implemented more complex systems to view and analyze activity- based model outputs; these systems are made up of a database component and a visualization and query interface. These interfaces may pro- vide the opportunity to view either predefined analyses or perform custom queries. Data visu- alizations may include traditional tables, charts, and maps, as well as other unique representa- tions that exploit the detailed activity-based model outputs (Atlanta Regional Commission 2010). One powerful data visualization tech- nique is to simultaneously illustrate multiple dimensions of model outputs. Agencies may also consider means of facilitating model anal- ysis and disseminating model outputs through the use of Internet-based reporting and visual- ization tools. At a minimum, agencies should develop standard outputs that allow agency staff to quickly determine if model results ap- pear reasonable before performing any more detailed analysis. 2.4.8.3 Simulation Variation Many activity-based models employ Monte Carlo simulation in order to realize discrete outcomes from the probabilities predicted by the model components. When using Monte Carlo simulation, the same probability dis- tributions may result in different outcomes

63 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM based on the random numbers used to simulate choices. In application, one important issue is controlling for how the random number seeds and sequences are used to choose outcomes. A second important issue is exploiting this fea- ture to provide more robust performance in- dicators. A third important issue is how these performance indicators are transmitted and explained to decision makers. The ability to control the random seeds and sequences in model application is useful because it provides confidence that the model implementation runs consistently and that it produces the same outputs, given the same inputs. Once the random numbers are con- trolled, users are able to exploit this model fea- ture to run the model system more efficiently and to produce better performance measures. Activity-based models should be run multiple times in order to account for simulation vari- ation (also referred to as simulation error). The disaggregate outputs can be used to pro- duce distributions and confidence intervals for core activity-based model measures, in addi- tion to average values that can be used as in- puts to tradi tional static assignment models. If regional-scale DTA models are adopted, use of disaggregate outputs to produce multiple network simulations may be desirable. Empiri- cal testing of the number of runs required to produce results with confidence is also desir- able, although only a limited number of regions have implemented this in practice. The number of runs is dependent on the spatial, temporal, or typological detail that is of interest. Analysis of smaller spatial, temporal, or typological seg- ments requires more runs. Traditional trip-based models do not pro- duce distributions of outcomes given fixed in- puts. Decision makers are most familiar with the single-point forecasts generated by these models. Thus, the communication and inter- pretation of activity-based models that include ranges of potential outcomes present new op- portunities and challenges. Distributions or ranges of outcomes provide the advantage of illustrating the existence of uncertainty around different outcomes, but if not properly pre- sented, may be misinterpreted by decision makers. However, it should be noted that these distributions of outcomes, while informative, are also contingent on the assumption that decision-making processes and elasticities re- main fixed over time, that future land use and transportation inputs are known with cer- tainty, and a number of other dynamic factors are unchanged over time. All of these assump- tions may be unreasonable. 2.4.8.4 Reasonableness Checks Reasonableness checks are an important prac- tical consideration. Travel patterns, whether for a single scenario or when comparing across scenarios, should be plausible. Reasonableness checks for activity-based models may include an analysis of activity durations by purpose, the amount of time spent in out-of-home activities, and the share of people who choose to make no travel, either overall or for a particular pur- pose (Resource Systems Group 2012b). Addi- tional checks may include an analysis of tours by activity purpose, stop generation rates, and possibly, joint activity generation (Cambridge Systematics 2010). The tour-based nature of activity-based models necessitates the review of tour-level statistics, such as tour mode choice and tour destination-choice, and the inclusion of explicit time-of-day models requires the re- view of temporal measures. Although these measures are considered in the calibration and validation process, subjecting the model system to a set of sensitivity tests and evaluating these metrics can ensure confidence that the model responds appropriately to investment and pol- icy alternatives. In addition to the reasonableness checks that are unique to activity-based models,

64 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER there are additional checks that are similar to those performed for trip-based models, such as evaluations of trip length frequency dis- tributions, trip flows by geography, and trip mode choices. In addition, because, at pres- ent, activity-based models are typically linked to static network assignment models, many of the reasonable checks appropriate to trip-based static network assignment models are equally applicable to static network assignment models linked to activity-based models. These checks may include assessments of regional traffic vol- ume errors, regional VMT, transit boardings and transfer rates, screenline and cordon vol- umes, and many other measures (Cambridge Systematics 2010). 2.4.8.5 Logging and Archiving To ensure the defensibility of model system re- sults, it is essential that the assumptions used to generate alternative scenario results be, at a minimum, logged, documented, and traceable. Version control of networks, inputs, control files, parameters, and software executables is mandatory. The model results for every alter- native scenario evaluated using a model system are the result of a specific set of inputs to the model system, the versions of the model system component software, scripts, configurations, and potentially even the computing resources on which the model run was performed. Alter- native scenario results should generally be ar- chived to allow for a more detailed review of these assumptions, if necessary. Archiving runs requires that agencies establish protocols for which alternative scenario model runs should be stored, how long these runs should be main- tained, and what media should be used to store these runs. Model applications often result in the proliferation of model runs as alternative scenarios are tested and refined, and activity- based model inputs and outputs can require significant amounts of data storage capacity. Ideally, model results could be duplicated by third parties running the model system at a later date or on different computing resources, although practically this can be challenging because of the dynamic evolution of hardware and software external to the model system, such as hardware operating systems. 2.5 ADMINISTRATION AND MAINTENANCE 2.5.1 Inputs Activity-based models require a rich set of input information, including detailed information and input assumptions about regional socio- demographics, employment, multimodal trans- portation networks, urban form, and other key influences on travel behavior. These input as- sumptions are developed for both the base year used for model calibration and validation, as well as for all required forecast years. Both the base-year assumptions and future-year assump- tions are dynamic. Over time, it is necessary to update base-year model assumptions to ensure that they reflect the most up-to-date informa- tion about existing conditions. It is also neces- sary to update future-year model assumptions to reflect changing expectations of these inputs. 2.5.2 Computing Platform An important challenge to agencies developing activity-based models is how to maximize the opportunities provided by recent hardware and software improvements, such as distributed and multithreaded software and the availabil- ity of large numbers of computing processors, while minimizing the expenses and risks asso- ciated with purchasing and maintaining these resources in-house. Broadly speaking, there are two main approaches that agencies may con- sider. They are acquiring and maintaining local computing resources, or using remote com- puting resources. At present, the majority of agencies are using local resources, although it

65 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM seems very likely that in the near future many agencies will start to employ remote computing resources. Local computing resources are typically in-house servers or workstations. Use of local resources provides agencies with complete control over software and hardware and may avoid many concerns about data control and data transfer issues. However, maintaining local computing resources requires agencies to make up-front as well as ongoing invest- ments in computing hardware and mainte- nance. Remote computing resources are often either externally managed servers, or “cloud computing” resources. If the activity-based model software has been designed to efficiently use large numbers of processors, then either of these may offer compelling performance improvements. In addi tion, both remote com- puting options can reduce the burden on agen- cies to make ongoing investments in hardware and maintenance, although any savings would be balanced against costs for the use of these remote resources. One potential drawback of these servers is that they may offer less control if additional commercially licensed software is required for model system application. 2.5.3 Licensing The core components of many activity-based model systems in the United States are imple- mented using custom software. Typically, this software is not commercially licensed but rather is governed by open source agreements, and long-term software maintenance and en- hancement is often funded on an as-needed basis through on-call or project-specific consul- tant contracts. These core activity-based components are used in conjunction with network assignment model components to make up the overall equilibrating model system. In all activity-based models systems currently used by agencies in the United States, the network assignment model component employs static assignment methods implemented using commercially li- censed travel demand forecasting software. In addition, these licensed travel demand fore- casting software packages are also typically used to perform network-related data process- ing, such as manipulating matrices of travel demand input to the static assignment proce- dures or generating matrices of network per- formance skims for input to the activity-based components. These licensed software packages may also provide other capabilities, such as distributed computing capabilities or optional DTA capabilities. Although some advanced DTA software is available on an open source basis, agencies should expect that licensing of some commercial travel demand forecasting software will be a necessary to complement the activity-based model components. 2.5.4 Schedule Inevitably, as advanced model systems have become more widely adopted, the amount of time required for model development has shortened. This shortening of model develop- ment time is likely the result of the increasing maturity of the software used to implement these models, the broader use of transferred models, the improved knowledge and sophis- tication of both agency staff and consultants, and the ability to acquire and manipulate data and perform model runs more quickly. This trend can be observed across all types of ad- vanced models, including activity-based model systems, DTA models, and land use models. However, there are still aspects of model development that are largely insensitive to these changes. For example, the amount of time required to collect new household survey data in order to estimate region-specific models can still be significant, although automated data processing and analysis has reduced these

66 Part 1: ACTIVITY-BASED TRAVEL DEMAND MODELS: A PRIMER schedules as well. Agencies must carefully con- sider the schedule trade-offs between develop- ing new capabilities or region-specific models, project or analysis schedule constraints, and agency staff and resource availability. New activity-based models have been im- plemented in as little as 6 to 9 months when the model structure, software, and coefficients have been transferred from one region to another (Resource Systems Group et al. 2014). The main tasks in these types of simple model transfer efforts are developing the input data from existing sources, and calibrating and vali- dating the model to regional conditions. More involved activity-based model development ef- forts can often take longer, however, especially if new capabilities are being developed or if additional data collection is involved. For ex- ample, the development of a new activity-based model system with enhanced capabilities for a large region may easily take 4 years and, in some cases, has taken much longer (Resource Systems Group 2012a). 2.5.5 Personnel No two agencies are exactly alike in terms of their institutional responsibilities, modeling resources, or staff knowledge and availabil- ity. As a result, the type and level of involve- ment of agency staff, and the reliance on ex- ternal consultants required to develop, apply, and maintain an activity-based model varies significantly. However, the active involvement of agency staff in all aspects of model devel- opment improves the ability of agency staff to apply and understand the model and to com- municate model results effectively to decision makers. Staff involvement may also reduce the dependence of an agency on consultant help. Agencies using activity-based models have as few as 2 people devoted to modeling to as many as 10 supporting model data, develop- ment, and application. 2.5.5.1 Staff Skills and Training Specific technical skills useful for activity-based modeling include discrete choice modeling, knowledge of the activity-based modeling prin- ciples and process, statistical analysis, and data- base or data management. Ideally, agency staff also has software development and scripting experience (Resource Systems Group 2012a). In addition, depending on the specific features of the model, it may be useful for staff to have other skills as well. For example, if a detailed spatial resolution is used in the model, such as microzones or parcels, familiarity with GIS is essential. Similarly, if the modeling system in- cludes a DTA model component, then knowl- edge of traffic engineering principles is critical. In general, common sense and critical thinking, as well as a willingness to work through the complexities of software development are es- sential skills for activity-based modelers. In addition to all these technical skills, it is essential that agency staff have robust com- munication skills and knowledge of the overall planning context in which travel forecasting is performed. This knowledge is especially impor- tant when complex tools, such as activity-based models, are being used, because the practical advantages of such tools need to be conveyed in a logical and compelling way. 2.5.5.2 Development Agencies have often, but not always, relied on consultants to lead model development and enhancement efforts; this reliance on consul- tants has allowed advanced models to be im- plemented more quickly than would have been possible if relying solely on agency staff. How- ever, staff participation and training during model development are very important because this investment in staff builds familiarity with the model system, improves the ability of staff to communicate effectively about the model,

67 Chapter 2: TECHNICAL ROAD MAP FOR DEVELOPING AN ACTIVITY-BASED MODEL SYSTEM and facilitates immediate testing and applica- tion by agency staff. 2.5.5.3 Application and Maintenance In contrast to activity-based model system de- velopment, agency staff have typically played a much more significant role in activity-based model system-level applications. Active in- volvement in model applications provides agency staff with the opportunity to understand model capabilities and sensitivities, increases familiarity with model inputs and outputs, and helps agency staff identify potential model im- provements. Agency staff have also played a leadership role in overall activity-based model system-level maintenance, especially with respect to updating model input assumptions. Institutional knowledge about model system data has been identified as a challenge (North Central Texas Council of Governments 2013). Consultants have also played a significant role in model maintenance, particularly with respect to software code and scripting. 2.5.5.4 In-House Versus Consultant Expertise As stated previously, agencies have often used consultants to provide activity-based model development expertise that they may not have in-house or to supplement the capabilities of agency staff. Consultant services provide agen- cies with the ability to effectively spend limited model development resources on critical model development tasks for limited periods of time. However, reliance on consultant expertise has some risks. Some agencies have reported that the use of consultants to develop advanced models can result in situations in which the agency staff are not able to effectively under- stand and apply the new tools; these situations may be problematic (North Central Texas Council of Governments 2013). It is essential that documentation of all model development and implementation efforts be prioritized. 2.5.6 Funding Agencies have pursued a variety of different strategies for funding activity-based model system development. Broadly speaking, most agencies have pursued strategies that can be characterized along two primary dimensions: program-based versus project-based and in- cremental versus comprehensive strategies. Program-based strategies involve funding activity-based model system development and enhancement as a core element of agencies’ work programs, while project-based strate- gies involve funding model development and enhancement using project-specific resources. For practical reasons, agencies often use a mix of these strategies to fund model efforts. In- cremental strategies involve the gradual devel- opment of an activity-based model system by funding a set of discrete successive tasks over a longer period of time. Comprehensive strat- egies involve developing an initial complete model more quickly as part of a single effort. Activity-based model systems have been suc- cessfully implemented using a variety of dif- ferent funding strategies. Decisions about how to fund activity-based model efforts should be primarily informed by agencies’ requirements and constraints. In recent model development efforts, agen- cies have spent as little as $250,000 for the transfer of an activity-based model in a small region over a relatively short schedule, to as much as $1.2 million for a large region over a 4-year time period, excluding household survey data collection. Other large regions have spent comparable amounts for model development, but spread this expenditure over longer time periods (Resource Systems Group 2012c).