Integrated Transportation and Land Use Models (2018)

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.

63 This chapter (1) summarizes the state of the practice, (2) describes future directions for inte- grated land use/transport modeling, (3) discusses barriers to widespread applications, and (4) iden- tifies research needs. After providing some recommendations for future research in Section 8.1, Section 8.2 discusses benefits and limitations of microsimulation, and Section 8.3 offers some concluding remarks. 8.1 Research Needs Since Lowry (1964) published his seminal Model of Metropolis more than 50 years ago, sub- stantial progress has been made in integrated land use/transport modeling. Microsimulation discrete choice and spatial input-output models have added theoretical foundation to model land use decisions and augmented the possibilities in modeling detailed scenarios. Despite some failed attempts to integrate land use and transport models (Wagner and Wegener 2007), which were to be expected in a complex study field, many integrated models work well. Such models have proven to be useful as planning instruments for decisionmakers. Based on the survey findings, literature review, and ongoing discussions with colleagues, several important challenges of integrated land use/transport modeling have been identified as future research needs. Data As identified in Section 1.1, there is a trend toward microsimulation in integrated land use/ transport modeling. To improve collaboration, common ways to store and share micro data should be defined. For geospatial data, the ESRI Shapefile has been established as the stan- dard format to store data. A white paper (ESRI 1998) provides specifications that allow sharing shapefiles across platforms and software. For matrices, a similar standard was defined using the open-source OMX matrix specification (Stabler 2017). There is no such standard for storing and sharing micro data or synthetic populations. Developing such a standard would facilitate the ability to work collaboratively with micro data across different study areas and software. Openness An open-source modeling framework is needed for integrated land use/transport modeling. MATSim (Horni, Nagel, and Axhausen 2016) is a framework for traffic assignment applied in more than 40 study areas. MATSim offers an open-source license and flexible implementa- tion in code. Open Platform for Urban Simulation (OPUS) (Waddell, Borning, Ševíková, and Socha 2006) is a similar framework for land use, activity-based, and dynamic traffic assignment models, but is not widely used by the modeling community. ActivitySim (http://udst.github.io/ C H A P T E R 8 Conclusions

64 Integrated Transportation and Land Use Models activitysim/) is another open-source framework for activity-based models, as is VisionEval (https://gregorbj.github.io/VisionEval/) for strategic planning tools. Those are still in their infancy, but could become the standard for activity-based and strategic planning models, respec- tively. The modeling community should strive to develop and use a similar framework for land use modeling that seamlessly integrates with various transport models. Integration All modeling frameworks explored during the in-depth interviews in Chapters 4, 5, and 6 either manually transfer data or loosely couple land use and transport models. Brandmeyer and Karimi (2000) defined a hierarchy of model integration from manual data transfer, loose coupling, shared coupling, joined coupling, to tool coupling. Each of these approaches has its benefits and limitations, and the selection of the most suited approach depends on data flow directions and frequencies (Shahumyan and Moeckel 2017). Manually transferring data is error- prone and does not make good use of computer time if one model finishes outside of working hours. Loose coupling overcomes these two shortcomings, but wastes runtime due to reading and writing of data. Although more tightly coupled models exist in academia (Wegener 2011), such models have not been applied in practice yet. Tighter model integration is particularly relevant for models that exchange data more frequently (such as the integration of a land use model with a transport model run every 5 or 10 years). Tight integration is less crucial for sketch planning models that tend to transfer data between land use and transport models less often. Documents For most model applications, it has been found that the only successful model implementa- tions were done or advised by the original model developer (or someone working closely with the model developer). As a consequence, most models are not used once the model developer retires or no longer maintains the model. Many models are too complex, documentation is sparse, and model design tends to be too inflexible to account for different modeling require- ments without major revisions of the source code. Tools to visualize model output (and thereby, to better understand model performance) are rare. The responsibility of the modeling com- munity will be to develop and support models proven to be agile and transferable. Well-written model documentation is a key prerequisite to allow new users to quickly become familiar with the modeling system. Education Closely related is the responsibility of universities to educate graduates who are capable of understanding integrated models and interested in working with them. A recent unpublished survey among 70 U.S. transportation planning agencies revealed that 64% of those agencies that operate an integrated land use/transport model are concerned that they do not have sufficient staff with adequate training to work with these models. Universities must help educating future champions of integrated modeling. Agencies, on the other hand, have the responsibility to give champions the opportunity to move application of integrated models to the next level. Extensions It is widely accepted that models need to be designed for the scenarios they will analyze. A model used predominantly for environmental analysis has different model specifications, a dif- ferent spatial resolution and requires different input data than a model used for highway expan- sion analyses. As such, most models will never be able to handle every scenario. An unresolved challenge for model developers is to build models tailored to the task at hand but allowing flexible adjustments to new requirements. Only a few years ago, most transportation models

Conclusions 65 were asked to model the effect of autonomous vehicles, despite the fact that this scenario was not even considered possible when many of these models were built. Microsimulation has the promise to facilitate flexible adjustments of the model to future (and currently unknown) chal- lenges. Such models allow agents to be modeled at the finest resolution possible and thereby allow aggregation to any level needed for future model extensions. This will be an important strategy to keep models relevant in a changing environment. Although important progress has been made in integrated land use/transport models over the last few decades, this field of research has a fair share of attempts to integrate land use with transport models that did not succeed in creating operational models. This is not uncommon in research, as scientific progress can be achieved with both successful implementations and models that do not reach maturity. The latter may have relevance in two regards. First, scien- tific progress may have been achieved in submodels, even if the entire model system has not become operational. Second, a model implementation that did not work out well may provide lessons learned for future model development. However, models that become operational serve not only the model developer communities but also scenario analyses. Lee Jr. (1973) wrote an article on problems associated with large-scale models. Although scientific progress can be made with research that falls short of its original goals, researchers working on inte- grated land use/transport models will continuously need to proof to be relevant. Having an eye on operationality will help to validate the value of integrated land use/transport modeling both to the scientific community and to those who use models. 8.2 Benefits and Limitations of Microsimulation As discussed in Section 1.1, there is a tendency toward microsimulation in both land use modeling and transport modeling. In travel demand modeling, activity-based models have become common for the largest MPOs, with some smaller MPOs following this trend. In land use modeling, the development appears to be more diverse, with PECAS (in part) and UrbanSim (completely) using microsimulation, while the up-and-coming CUBE Land model as well as all sketch planning models follow the aggregate modeling paradigm. Microsimulation offers several benefits. Foremost, microscopic models are more flex- ible. Although aggregate land use models commonly define a few household types used in all modeling steps, microscopic models work with a synthetic population that stores the socio- economic data as individual records with a great level of detail. Therefore, microscopic land use models may aggregate households by different attributes for every modeling task. For example, auto ownership is an important variable to be considered in household relocation, but auto ownership is (mostly) irrelevant for the decision to have a child. Although aggregate models need to carry through the same household categorization throughout the entire model, micro simulation or agent-based models have the ability to redefine household types in every modeling step. If agents are defined explicitly, they may even learn from one simulation period to the next. For example, an agent who experiences severe highway congestion in one simula- tion period may decide to change the mode in the next simulation period, or another agent may decide to move back into a neighborhood where she or he has lived previously. However, learning has so far only been implemented in academic projects, such as ALBATROSS (Arentze and Timmermans 2000) and ILUTE (Salvini and Miller 2005). Last but not least, many practi- tioners have reported that they like agent-based approaches because they are easier to explain to lay persons and upper management. An aggregate (4-step) transport model may simulate fractional trips, such as 0.8 trips. An aggregate land use model may relocate fractional house- holds (such as 0.35 households). Although this is perfectly legitimate for aggregate models, it is inherently difficult to explain to lay persons why this is the case. Agent-based models tend to be simpler to explain to lay persons, because they explicitly represent the unit of decision making (such as household or person) and they tend to resemble more closely how decisions are made in real life.

66 Integrated Transportation and Land Use Models Microsimulation also comes with some additional costs. Microsimulation models have the tendency to result in longer runtimes. This issue can be overcome with efficient programing algorithms, multi-threading and modern hardware. Furthermore, microsimulation depends on a random number generator to simulate decisions between alternatives based on their utilities. Therefore, every model run will be slightly different (Wegener 2011). This is unproblematic when results are analyzed at a high level of aggregation, such as population by city or mode choice by employment status. However, detailed analyses may become invalid, such as auto ownership of high-income households in one particular neighborhood. Modeling larger firms microscopically cannot be done because those firms may locate in two different zones in two model runs, only because the random number generator provided a slightly different value. Small differences in scenario results also require special attention in microscopic models. If the difference of two scenarios is smaller than the differences caused by stochastic variation, it would be invalid to interpret the difference as a response to a given scenario. To overcome this issue, both the base scenario and the policy scenario need to be run multiple times, and the average difference between base and policy scenario can be used to describe the response to scenario settings. Given the longer runtimes of many microscopic models, running the same scenario multiple times is a challenge if deadlines need to be met. 8.3 Concluding Remarks The relevance of integrating transport models with land use models has been proven both theoretically and empirically. The location choice of households, firms, and developers largely defines origins and destinations of travel, and thereby, is the main explanatory vari- able for travel demand. Travel times, on the other hand, influence location choices by house- holds, firms, and developers. Although many more location factors are considered in location choice, travel times—or more generically accessibilities—have been shown to be an important factor for finding a new location. For example, Conder and Lawton (2002) showed convinc- ingly that new housing north of the Columbia River will not be attractive for Portland, OR, workers despite the lower costs of living, because the limited number of bridge crossings across the Columbia River severely limits accessibilities north of the river. Hansen (1959) has shown how transit stops may attract new development. Cervero et al. (2004) found in a literature review that home prices near transit stations are 6 to 45% higher than for otherwise equiva- lent sites, reflecting the high demand for living near transit. For most transport analyses, this feedback between travel behavior and location choice can no longer be ignored. This report presented successful examples of how land use modeling can be integrated in trans- port analyses. Data The survey showed that the most common concern for integrated land use/transport model- ing is data availability. Some agencies hesitated engaging in land use modeling because of data limitations. This is a problem that deserves attention, but sometimes the data argument appears to be used as an excuse to avoid dealing with this topic rather than a true restriction. Nowadays, many jurisdictions provide land use data at the parcel level. Although these data are not always up-to-date and need careful review, parcel-level datasets are useful because they can easily be aggregated to any TAZ system, if the land use model does not use the parcel level directly. In addition, new crowdsourcing data provided by internet sites, such as Foursquare, Google, OpenStreetMap, Rome2rio, or Twitter, may add important input data, often free of charge. Without a land use model, cumbersome manual disaggregation of future forecasts to the trans- port model’s TAZ system is required. Although setting up a land use model commonly is more

Conclusions 67 effort initially than this manual disaggregation, the land use model provides a long-term strategy that makes preparation of socioeconomic data for the transport model repeatable, transparent and—most importantly—sensitive to alternative scenarios. Developer Working with the model developer is key. If the model developer is not accessible, someone who has closely worked with the model developer may work equally well. The in-depth interviews showed two examples (ARC and MTC) in which agencies did not work with the original model developer personally, but rather with another person who had worked with the model developer. How easily a person who is very familiar with the model may be accessible should be part of the decision concerning which land use model is chosen. No case is known where a microsimula- tion discrete choice or spatial input-output land use model was applied successfully without engaging the model developer in a meaningful way. Although a common expectation by model users is that models should be self-explanatory, well-documented, and easy to implement, reality shows that this is rarely the case. These models are complex, often need very specific adjustments for a given study area, and usually do not carry enough years of development to work in a “plug-and-play” fashion. Although sketch planning models have been implemented without direct involvement of the model developer, this is not the case for microsimulation discrete choice or spatial input-output land use models. User Group Many regions have implemented model user groups. The organization of these groups is dif- ferent in every region, but commonly they meet several times a year to present modeling prog- ress and issues and discuss solutions and ideas for collaboration. Many agencies have reported that such user groups help to share experiences and overcome model usage issues. Often, con- sultants and researchers like to join such user group meetings as well to stay in touch with actual model application issues. Cooperation between state, MPO, county, and local governments may also provide strong support for model development, review, and applications. This may also include agencies out- side the transportation domain, such as housing, planning, water and sewage supply, or parks planning agencies. Adding interdisciplinary viewpoints to integrated land use/transport plan- ning may improve the modeling framework and provide added credibility within the region. Panel Closely related to the support of user groups are expert panels that have been successfully installed by several agencies interviewed for this report. Such panels commonly include experts in land use and transport modeling, real estate brokers, transit agency managers, bankers, aca- demic experts, economists, among others. Expert panels of about a dozen members may provide valuable advice in model development, data collection, and model application. Agencies that work with expert panels have reported that their modeling efforts gained credibility, because model design and results had been peer-reviewed. Openness Several model developers, including ActivitySim, Silo, and UrbanSim, have moved toward providing their models as open-source software. In such cases, model users have access to the model code and may change code for their own application as needed. The benefits for users include opening the black box and understanding the exact formulation of the model as well as

68 Integrated Transportation and Land Use Models the possibility to revise or extend the model to their own needs. This is particularly true if users become savvier in programming and wish to work with the model code themselves. Also, an open-source model allows for tighter integration with other models. For the developer, provid- ing the model open source adds the benefit that users may provide useful feedback on improve- ments in code, and those models tend to be applied more widely. Some model developers are concerned that they lose the benefit of selling software licenses when providing code as open source. Those who do provide open-source models either have no commercial interest (such as universities), or they have moved to a business model where they do not sell licenses but rather support for model implementation and application. It is likely that more and more users will move toward open-source products. Scenarios Land use scenario analyses are not as common among the agencies interviewed. Each agency created a base year scenario for calibration and at least one future year scenario, in most cases for the year 2040. MTC, SACOG, and WFRC developed numerous scenarios. ARC and TJCOG have generated one alternative future scenario, and each plans to create a second one. The Metro- politan Council and Ohio DOT have not engaged in land use scenario analyses yet. In part, this limited experience with scenario development might be biased by the fact that all agencies interviewed started operating land use models as little as 5 years ago. In another 5 years, it might be much more common for these agencies to run land use scenarios. If one goes through the effort of implementing a land use model, it seems prudent to make full use of the model. All agencies reported that they were able to conduct scenario analyses without consultant support, which should keep costs of additional scenario analyses relatively low. ARC reported 1.5 days of one person staff time for setting up a scenario on TOD. This is not an insignificant cost factor, but is seems reasonable in comparison to the additional benefit of better understanding the effects of planning on travel behavior. One might hope that agencies that spent the costs and efforts of implementing integrated land use/transport models are going to make good use of such a powerful modeling suite to the full extent possible. Complexity A question that tends to be very difficult to answer at the outset is how complex models should be. The generic answer is unspecific: On the one hand, the model should be complex enough to be sensitive to scenarios that need to be tested, and the other hand, the model should—following Albert Einstein’s words—be as simple as possible. In other words, higher complexity is only warranted if it helps answering new policy questions. However, there are other good reasons why agencies may decide to keep models simpler: Availability of data, staff time, funding for licenses and consulting fees as well as modeling experience may be key drivers to choose simpler over more complex models. Given that sketch planning models tend to be easier to implement and maintain, such models might be a good alternative for agencies with limited experience in modeling or limited resources. Such agencies might decide to switch to behavioral models once experience with land use modeling has been built up and scenario requirements demand moving toward more complex models. Validation Validating an integrated land use/transport model is inherently difficult. A transport model commonly is built with survey data and validated against traffic counts and transit ridership data. For a land use model, comparable data that describe the precise change from one year to the next rarely exist. Therefore, some model implementations start the land use model in

Conclusions 69 the past, such as the year 2000, and run the model until today. If population and employment numbers of today match the model results, the model is assumed to be sufficiently validated. Although matching observed data is no proof that the underlying model is valid, a reasonable match increases confidence that the model is performing reasonably well. Another confirma- tion for a reasonably well specified and calibrated model is a sensitivity test. The user may define sensitivity scenarios, such as exogenously change travel speeds up and down, and observe the effect on model results. If changes in model results are in line with expected sensitivities, the model user gains additional confidence that the model is representing behavior reasonably well. Although sensitivity tests strictly speaking do not count as validation, a realistic response to scenarios provides confirmation that important behavioral responses are represented in the model in a plausible way. The most advanced method of model validation is comparing different models that have been used to address the same problem (Wegener, Mackett, and Simmonds 1991). This method, however, is also the most expensive one, as different models need to be implemented and calibrated for the same study area. For agencies that transition from one model to another, however, such a comparison offers an excellent opportunity to confirm model reasonability or justify the benefits of moving to a different model. Because validation is very difficult for integrated land use/transport models, quality control and feedback by experts in the region is particularly important. As mentioned by several agencies interviewed for this report, such reviews often have led to important corrections of input data or revisions of model specifications. This is another reason why expert panels can be very useful. Accuracy At the same time, one should not be overly concerned about validation and model accuracy. In the end, models are not only used to replicate existing conditions. Rather, models are used to test the effect of alternative policies. Should the model slightly over-predict auto ownership in the base scenario, the model is likely to over-predict auto ownership in a policy scenario by a similar share. This does not mean that calibrating the model to represent the right auto owner- ship rates was unimportant. The model should represent the base year conditions as reasonably as possible. However, being overly concerned about accuracy in the base scenario may distract from the power of applying a model to scenario analysis, even if the model is not calibrated perfectly yet. Some models have implemented so-called k-factors to improve the accuracy of the model. K-factors force the model to provide a certain result expected by the user. Although k-factors help successfully to match observed data in base year implementations, k-factors also limit the model sensitivities. For example, if a suburban location is under-forecasted by the model, a k-factor could help the model to match population in this suburban location. In a scenario with higher transportation costs, however, this suburban neighborhood may become much less attractive. A model that has been adjusted by k-factors would continue to allocate many house- holds to this suburban location, because that location choice is not explained by utilities of this location but rather by a hard-coded k-factor. For this reason, k-factors should be avoided for most part in modeling. It is generally preferable to accept small deviations from observed data and preserve the full model sensitivity for scenario analysis. Agile Maybe the most important lesson learned that was mentioned by several agencies is the benefit of implementing a simple model at first and gradually increasing complexity as needed. In computer science, this approach is called agile development (Martin, 2003). At any point in time, the agency ought to maintain an operational model. At the beginning, the model

70 Integrated Transportation and Land Use Models would not be very useful as it poorly reflects reality. By identifying the parts of the model that perform the poorest and fixing those parts first, the overall model will gradually improve and soon be useful for at least simple scenario analysis. As more complex questions are asked by decision makers, the model can be improved gradually. Instead of designing a complex model upfront that takes years to develop and may not ever become fully operational (see for an example Wagner and Wegener, 2007), models are only built with the level of complexity currently needed. This also facilitates an environment where questions asked of the models may change quickly. A few years ago, pricing studies were the predominant concern of most transport modelers. This topic was overtaken by reliability studies, which were followed by big data analyses. Shortly thereafter, autonomous vehicles dominated the discussion. Although previous topics did not disappear entirely, models will always be asked to address new topics currently of particular relevance for decision makers. It is impossible to build our models to be prepared to tackle all of these questions, nor do we know what the questions asked next year will be. The agile model development scheme allows quickly adopting to new require- ments and provides the framework for early successes in the development of integrated land use/transport models.