An Application of Urbansim to the Austin, Texas Region: Integrated-model Forecasts for the Year 2030

Land use patterns are key determinants of traffic conditions, as well as energy consumption and greenhouse gas emissions. This work describes the modeling of year-2030 land use patterns of the Austin, Texas region using UrbanSim, an open-source model for microscopic simulation of land development, location choices and land values, at fine spatial resolution (typically 150 m x 150 m grid cells). An accompanying travel demand model was run every five years, resulting in accessibility indices for use in UrbanSim location choice models. A business-as-usual trend scenario was compared to urban growth boundary (UGB) and added transport-cost-sensitivity scenarios (TCS), expanded highway capacity scenario (EXPAN), and added state highway 130 (SH 130) scenario in order the appreciate UrbanSim’s performance and the potential land use and travel impacts of such policies. As expected, several land use results (e.g., population densities), travel patterns and energy consumption results responded to scenario contexts. Local access variables (within 600-meter Euclidean distances) also enjoy significant relevance in this implementation of UrbanSim. While UrbanSim specification limitations are multiple and its data requirements are serious (and may be impossible for almost any planning agency to meet − even after substantial effort), the model does run reasonably fast and may make good sense over the longer term for interested regions with sophisticated planning staff on board to pursue. Enhancements in the newer version of UrbanSim rendered the model to be more user-friendly. INTRODUCTION With an annualized population increase of 3.5% per year over the past 10 years, Austin is one of the fastest growing mid-size regions in the U.S. Such shifts, coupled with major transportation investments (including several new tolled highways and a commuter rail line) and variations in transport, land use, and energy policies significantly impact the region’s future land use patterns, traffic conditions, greenhouse gas emissions (GHG), housing affordability, environmental encroachment, and other key facets of community life. Such changes are evident in most regions, as population pressures, rising incomes, and evolving markets necessitate change. This change regularly sparks strong interest in land use forecasting, in synch with travel demand modeling and transportation planning. The U.S. is the world’s leading producer of greenhouse gases (emitting over 6 billion metric tons of CO2-equivalents annually, and accounting for 22.2% of the world’s emissions) (EIA, 2007). Home energy use accounts for 21% of the nation’s GHG emissions, and transport accounts for 32% (EIA, 2006). Building energy use altogether accounts for 47% of the nation’s GHG emissions (EIA 2007), but transport sector emissions are rising faster than the total (24% versus 13%, between 1990 and 2003, according to Brown et al. [2005]). While UrbanSim does not yet output estimates of all building sizes, the addition of modeling equations to anticipate square footage of various home types allows this work to anticipate energy demands as a function of both transport and buildings. This study evaluated the ease of UrbanSim’s implementation in Austin, Texas and the model’s sensitivity to various transport and land use policy scenarios. URBANSIM OVERVIEW A variety of land use models (LUMs) now exist. Most are mathematically and behaviorally based, while others are more normative in nature. While most LUMs rely on aggregations of space and agents (such as traffic analysis zones [TAZs] and all low-income households) (Dowling et al. 2000), UrbanSim emphasizes relatively small grid cells (and, more recently, parcels) while tracking the cell locations of individual households and jobs. A 150 m x 150 m grid cell is 5.56 acres, while the average Austin TAZ is 1691 acres − or 300 times larger. Though various models ignore constraints on land use and built-space availability, UrbanSim emphasizes these key facets of urban form (Waddell et al. 2003). Of course, land use modeling is a complex endeavor, and the UrbanSim modeling system, like any abstraction of reality, exhibits many limitations. For example, households and firms do not evolve, workers are not linked to job sites, jobs are not linked to firms, economic interactions are neglected, and job growth and economic conditions are exogenous. Moreover, the travel demand modeling process is external (with only a relatively weak link, through regional accessibility 1 Relatively common LUMs include Putman’s (1983, 2005) gravity-based ITLUP, Waddell’s (2003) UrbanSim, and models based on spatial input-output specifications, like de la Barra’s (1989, 1984) TRANUS, Hunt and Abraham’s PECAS (2003), Kockelman et al.’s (2004) RUBMRIO, and Echenique’s (1990, 1969) MEPLAN. Others include, Simmond’s (2001, 1999) DELTA, Martinez’s (1996) MUSSA, and Landis and Zhang’s CUF (1998a, 1998b). Suitability models assign new development to highest scoring sites, and include What If? (Klosterman, 1996, 2007) and UPlan (Johnston et al., 2003). 2 Other microsimulation-based LUMs are Landis and Zhang’s (1998a, 1998b) California Urban Futures (CUF) model, Miller et al.’s (2001) Integrated Land Use, Transport and Environment (ILUTE) model, Timmermans’ (2000) ALBATROSS model, and portions of Hunt and Abraham’s (2003) PECAS model. terms that may or may not make it into the location choice specification), the land price model is not generally fully estimated (since dynamic land price and vacancy data are so difficult to come by), and jobs and households are assigned one by one and do not compete for space in any given year. In practice, UrbanSim’s location choice models are typically calibrated using crosssectional data sets (rather than those of recent movers) and population synthesis is up to the analyst. However, the package is evolving and many of these issues can be addressed in nearterm and future versions. There also are many computing challenges for new users of UrbanSim, and data set acquisition always poses a major issue for such detailed models. Nevertheless, UrbanSim pays attention to many important land market features; and, once it is running properly, it runs quickly. Moreover, UrbanSim permits estimation of most sub-model parameters within the Opus environment, rather than requiring that users enter the model with parameters in hand. UrbanSim uses independent logit models to simulate the relocation decisions of existing households and firms, place households and jobs in grid cells, and anticipate grid cell-level changes in development type (Waddell 2004). Ordinary least squares (OLS) methods for parameters of continuous logistic expressions provide estimates of residential land shares across zones, and land price estimates are based on a hedonic regression equation (estimated using OLS). New transport infrastructure and local use restrictions are coded in at proper time points over the multi-year modeling process. Monte Carlo techniques are then used to simulate future year forecasts of location choices (by developers, workers, and households). Figure 1 of Waddell (2007) illustrates UrbanSim’s sub-model and data set interactions, along with user-specified events (such as road building and changes in zoning policy) and scenario details. DATA SETS USED As described below, calibration and application of the UrbanSim model is a highly data intensive process, particularly for a large multi-county region. UrbanSim can simulate land use patterns at any resolution (Waddell, 2001); a typical resolution is 150 m × 150 m and so was used here. Many assumptions had to be made, in placing and defining individual households and buildings at the grid cell level, in order to get the model to run. Few regions are likely to have such data and will need to resort to some sort of reasonable rules for data generation. Household and Employment Data UrbanSim’s household data set consists of a list of all households, with current locations (by gridcell), household size (number of members), age of the household head, race, and number of workers, children and autos. Household data was synthesized using iterative proportional fitting techniques at the level of year-2000 Census block groups (as described in Lemp et al. [2007] and McWethy [2006]). 3 Market response emerges via a land price shift in the following year, based on prior year vacancy rates in each cell. 4 According to Waddell (2008), the number of sampled alternatives is user defined, and is available as an option in a new version of UrbanSim. They believe that allocation results generally stabilize by around 30 alternatives. However, in reality, locators are likely to consider alternatives far more strategically than simply 30 randomlydrawn sites (out of hundreds of thousands of cell alternatives). 5 UrbanSim uses the predicted means of land prices and residential shares (rather than random, simulated values) (Waddell, 2008). Exogenous regional household control totals were obtained from Capital Area Metropolitan Organization (CAMPO) and used as model inputs. Annual relocation probabilities for households and the vacancies in residential units were imported from Eugene data sets and assumed to hold for Austin. These range from 63 percent (for those under 24 years of age) to 2.4 percent per year. The employment data consists of a list of all jobs (by sector), their cell locations, and building type occupied. This data set was generated from a file of firm point locations provided by CAMPO. Annual relocation rates (for each job type) and non-residential vacancies also are required inputs, but had to be imported from Eugene data sets. Built Space and Transportation Data The Travis County Appraisal District (TCAD) provided residential unit locations (and year-built information), as well as square footage (and year-built information) of all commercial and industrial establishments within the region’s central county. These data were used to estimate the proportions of each type of residential unit, square footage, and age distributions for buildings in the Hays and Williamson counties (which hold what 19.6% and 27.3% of the region’s jobs and housing, respectively), using logistic and OLS regression equations UrbanSim requires network travel times to the region’s CBD and major airport from each TAZ centroid, along with Euclidean distances to the nearest arterial and freeway from TAZ and gridcell centroids. These values were computed using CAMPO’s 1997 network. Energy Data 2005 Residential Energy Consumption Survey data (RECS, 2001) and 2003 Commercial Buildings Energy Consumption Survey data (CBECS, 2003) were obtained from the Energy Information Administration (EIA) and used to estimate energy per square foot , to apply to UrbanSim’s year 2030 outputs. MODEL SPECIFICATION Future land use patterns depend on household and job location choices, which in turn depend on the supply, quality, and price of built space, access to jobs and other destinations, household income, industry sector, and so forth. The following discussion describes the model estimation process for key sub-models. Household Location Choice Model (HLCM) UrbanSim’s household location choice sub-models are based on a transition model and a relocation model. The transition model generates a list of households to be added to or subtracted from the set of existing households. These shifts can result from various demographic processes, like aging, marriage, divorce, births, and deaths, and are provided as exogenous control totals for 6 Point locations of employment data were obtained as GIS layers from the Texas Workforce Commission via the Texas DOT and then cleaned by CAMPO. 7 Interestingly, UrbanSim does not call upon network files directly; all distances must be computed externally and provided (or coded internally by the analyst). This is an opportunity for relatively easy model improvement. future years. The household relocation model generates vacant spaces when households are selected to move and adds all movers to the list of unplaced households. While the probability that a household moves should depend on the relative attractiveness of available alternatives, as compared to the current dwelling; this presently is controlled by exogenous relocation probabilities, as pre-defined by UrbanSim users. The set of unplaced households are placed into grid cells via a multinomial logit (MNL) model of household location choice, based on the composite utility of 30 alternative grid cells (each having at least one vacant housing unit at the time the household is allocated). Explanatory variables affecting location choice can reflect elements of urban economic theory and sociology (CUSPA 2006), including regional and local accessibilities, race, incomes, and land rents (Waddell 2002). Table 1 shows the HLCM’s estimation results. These suggest that increases in housing and land costs and the share of land in residential use negatively impact a cell’s residential location utility, everything else constant. In contrast, variables such as income of current cell residents, open space in the cell, proximity to arterials and highways, accessibility to regional jobs, local employment, population, and travel time to the region’s CBD are predicted to have a positive impact on a cells’ utility. Also, households with similar income levels and races tend to attract. The positive signs on distances to arterials and highways and on CBD travel times are not intuitive, but these are more than offset by the positive benefits of the regional jobs accessibility values of all almost all locations. Employment Location Choice Models (ELCMs) UrbanSim’s employment location choice model (ELCM) is analogous to its household location choice model. A transition model generates or removes the newly created jobs in each sector, depending upon the growth or decline of employment in that sector (as compared to the prior year). Such input assumptions are exogenously obtained from state economic forecasts and/or commercial and in-house sources (CUSPA 2006). An employment relocation model then determines which individuals will change jobs in any given year. Employment relocation probabilities are determined exogenously and given as inputs to UrbanSim. Removed and relocating jobs are noted in the database of job-site vacancies, and jobs created via the employment transition and relocation models are added to the database of unplaced jobs. The unplaced jobs are allocated into the gridcells by sector, based on the composite utilities of each gridcell. UrbanSim assumes that every job is independent of the other and moves separately. Though, this assumption simplifies the modeling procedure, it is an important limitation of UrbanSim since most job relocation is based on relocation of firms. The employment location choice model is also based on a multinomial logit specification. The only difference between household and employment location choice models is that the latter allocates jobs using sector-specific preference functions (MNL model estimates), while all households are allocated using a single MNL specification − with indicator variables for variables like race and income, to accommodate some forms of preference variation. The sectors considered in this study are industrial, commercial and home-based employment. The Industrial Employment Location Choice Model locates new and relocating industrial jobs based on the relative values of all grid cells’ composite utilities. Estimation results presented in Table 1 suggest that increases in home and work accessibility values, presence of industrial square feet, and the presence of high-income households have a positive effect on gridcells’ locational utilities. In contrast, increases in distance to the nearest highway and in travel time to the CBD have negative impacts. All parameters have intuitive signs except that for the presence of higher-income households (perhaps both high income households and industries are attracted to similar properties). The Commercial Employment Location Choice Model locates commercial jobs seeking locations. Table 1’s estimates suggest that increases in commercial and industrial square footage, distance to the nearest highway, number of service sector jobs within “walking distance” (600 meters, Euclidean from gridcell centroid), access to population, and number of high-income households positively impact a cell’s attractiveness. In contrast, the cost of land and the number of retail jobs have a negative impact. While the positive impact of highway distance and the negative impact of local retail jobs are not intuitive, these can be counteracted by semi-collinear variables (like square footage, service job access, and population access). UrbanSim requires data on home-based jobs in order to run. However, Austin does not yet have such data, so this data set had to be manufactured (using a two-percent-of-jobs-per-zone assumption). Table 1 shows the results of this Home-based Employment Location Choice Model, which indicates a preference for older, more expensive residences in denser, low-vacancy locations, close to other home-based jobs. While many parameters appear reasonable, others may be at odd with actual trends, because these job sites were randomly selected. This model should not affect the overall forecasts much because only 2% of the modeled population works from home (Bayles, 2002). Modeling Land Prices UrbanSim’s Land Price Model provides a key input to the land development, household and employment location choice models. Land price is modeled using hedonic regression on attributes such as land use type (including 8 mixed use types, 8 residential, 3 commercial, 3 industrial, 1 government use, vacant and undeveloped land uses), site characteristics, access variables, and neighborhood and zoning characteristics. UrbanSim implicitly assumes that households, businesses and developers are all price takers, and annual price and development adjustments help match aggregate supply and demand over time. (CUSPA 2006) Moreover, as cell-level vacancy rates fall below a user-specified long-term structural vacancy rate, land price increases and vice versa (DiPasquale 1996). Land price estimates are updated annually, after all 8 Buildings for workers are classified into industrial, commercial, governmental and home-based buildings. Industrial employment refers to jobs located in industrial buildings. Commercial employment refers to jobs located in commercial buildings. And home-based jobs are those where workers work from home. Government jobs are assumed to maintain their locations, so there is no choice model for these. the construction and development is undertaken and vacancy rates have been computed, according to the following equation: ilt s i c it s i ilt X V V V P β δ α + − + = ) ( (1) where Pilt is the land price per acre of development type i at location l and time t, Vi is the longterm structural vacancy rate, Vit is the current vacancy rate (at time t), Xilt is the vector of site attributes, and α, β, and δ are parameters to be estimated. However, grid-cell-level or buildinglevel vacancy rates can be quite difficult for analysts to avoid, particularly vis-à-vis property prices over time, in order to estimate a parameter like δ. Thus, users may often find themselves importing (or guessing at) this key parameter. This is a key concern for the model, though most users may not rely on UrbanSim for land price information. Either way, current year price estimates are used to estimate the next year’s market activities, including land development and location choices (Waddell et al. 2003). Land Price estimation results (Table 1) suggest that an increase in commercial square feet, total employment, population density, access to population, percentages of commercial and developed lands, and high-income households (within walking distance) are associated with higher land prices (as assessed by the Travis County Appraisal District), whereas the presence of industrial space, open space and residential land use within walking distance of a cell (600 meters) are associated with lower land prices, ceteris paribus. Such signs are consistent with behavioral expectations. UrbanSim’s Residential Land Share Model is used to compute the residential land share in a gridcell, as follows: ∈ + = ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ − β X y y 1 ln (2) where y is residential land use shares in each grid cell. Essentially, the model assumes a logistic specification for the fractional shares, but ordinary least squares is used for estimation of parameters (β). From Table 1, it can be observed that residential land share predictions tend to rise with the number of residential units in a cell, and job access, but fall with jobs counts, non-residential built space and travel time to the CBD, as expected. UrbanSim also uses MNL models to place new development of residential, industrial and commercial structures, based on composite utilities. These model parameters, for development location choices, can be found in Kakaraparthi (2009). 9 Due to a lack of occupancy rate data for two time points across zones, δ could not be estimated and was simply set equal to 1.0 here, to be able to run the model and evaluate performance. INCORPORATION OF A TRAVEL DEMAND MODEL (TDM) According to Opus (2006) workshop documents, UrbanSim has been used with TDMs based in TP+, MinUTP, EMME/2, and other systems. Here, Caliper’s TransCAD software was used, and outputs of both models manually transferred by the user. Using 1996/1997 Austin Travel Survey (ATS) data, Lemp (2007) estimated the parameters of and coded (in TransCAD’s GIS-DK) details of the fairly standard TDM employed here. Regression models are used for trip generation, at the household level for home-based trips and at the zonal level for non-home-based trips. An MNL model of destination choice is used for trip distribution, and includes a logsum parameter measuring the maximum expected utility achieved over all modes and times of day (TODs). A joint MNL model of mode choice (drive alone, shared ride, transit, and bike/walk) and four TODs is also used. Separate models were applied for each of four trip types (home-based work, home-based non-work, non-home-based work, and non-home-based non-work). Finally, deterministic network assignment routines were used in each TOD, and 25 feedback iterations (from network-equilibrium travel times and costs to trip distribution) were performed, in order to obtain estimates of interzonal travel times, trip distances and travel costs − specific to each of the four modes at each of the four TODs. Using this information, logsum accessibility indices were computed and then input into UrbanSim, in order to anticipate the next five years’ land use patterns. The logsum values are computed as follows (see, e.g., Ben-Akiva and Lerman [1985]): ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ = ∑ ∈C m nijm ij U L ) exp( ln (3) nijm n nijm n GC nm nijm TT VOTT Cost U + + = ) ( , 1 β β (4) And these are used in somewhat simplistic accessibility indices, as proscribed by UrbanSim documentation (CUSPA, 2006):