Walking is underrepresented in large area models of urban behavior, largely due to difficulty in obtaining data and computational issues in representing land use at such a small scale. Recent advances in data availability, like the ubiquitous point-of-interest data collected by many private companies, as well as a worldwide dataset of local streets in OpenStreetMap, a standard format for obtaining transit schedules in GTFS, etc, provide the potential to build a scalable methodology to understand travel behavior at a pedestrian scale which can be applied wherever these datasets are available.
In addition, the recent invention of fast network algorithms like Contraction Hierarchies greatly reduce related computational issues, as most network computations in this work are computable in less than a second. This thesis is a presentation of such a scalable methodology, which uses widely available datasets wherever possible, with computations that run quickly to encourage exploration of nuance in urban behavior and transparency of outcomes.
Additionally, indexes like WalkScore have been widely studied in the literature recently, both to predict walking behavior and real estate home values. This dissertation takes the position that WalkScore does not sufficiently support the set of destinations it includes, the weights that are applied, the distance decay function, and most importantly does not account for variation in behavior based on the demographics of the traveler. It is also likely that the use of destinations like coffee shops and bookstores in the index measures a specific kind of walkability that embeds a certain kind of neighborhood into its definition.
This dissertation improves on similar indexes like WalkScore by estimating a model that represents the substitution of destinations around a location and between the modes of walking, automobile, and transit. This model is estimated using the San Francisco Bay Area portion of the 2012 California Household Travel Survey to capture observed transportation behavior, and accounts for the demographics included in the survey. These representations of travel behavior can then be used as right-hand side variables in other urban models: for instance, to create a residential location choice model where measures of accessibility and available demographics are used to understand why people choose to live where they do.
In all cases, location choice models - both destination choice and residential location choice - use a level of detail not common in the literature in order to accurately represent walkability. This dissertation proposes the concept of "street node geography" which uses the local street network to define the geography with which to perform aggregations in the city. In this conceptualization, land uses and other urban data are mapped to their nearest street intersections, and overlapping aggregations are performed along the street network up to a given horizon distance. This representation of urban space is equivalent to a voronoi diagram around the intersections of the local street network, and can be thought of as having automatically generated set of 226,000 micro-zones in the San Francisco Bay Area. Street node geography thus provides a novel compromise between detail and performance for the kinds of computations performed here.
This dissertation is organized into four topics, one for each of chapters 2-5. The first topic establishes a framework for measuring the network of destination opportunities in the city for each of the walking, transit, and auto transportation modes. Destinations in the form of parcels and buildings, businesses, population, and points of interest are tied to each network so that the distance from each location to every destination can be computed by mode. The use of a points-of-interest dataset as the set of public-facing destinations is novel in the context of a traditional travel demand destination model.
This chapter also creates a case study model of trip generation for home-based walking trips is the 2012 California Household Travel Survey. This model finds that WalkScore is predictive of walking trips, that residential density and 4-way intersections have an additional but small impact, and that regional access by the transit network has a synergistic effect on walking, but regional access by auto has no impact when controlling for regional access by transit.
The second topic engages with the question of the impact of accessibility to local amenities on home values. Although early research has found that the composite index WalkScore is positively correlated with home values, this dissertation unpacks the impact of each category of destination used in WalkScore (as well as several others) on home values. The model shows that some amenities are far more predictive of home values in the datasets used here; in particular, cafes and coffee shops tend to be the indicator of neighborhood-scale urban fabric that has the largest positive relationship with home values, where a one standard deviation increase in access to cafes is associated with a 15\% increase in home values.
Although the previous topic provides some evidence that walkable amenities are related to increased home values with the datasets analyzed here, it does not prove that households are valuing walking to these amenities; it is equally plausible that households are capitalizing short driving trips into increased home values. The third topic thus creates a nested mode-destination model for each trip purpose (with destinations nested into modes) so that the logsums of the lower nest give an absolute measure of the accessibility by mode for each purpose for each location in the region.
These logsums are then weighted by the number of trips made for each purpose, and segmented by income and weighted by the incomes of the people that live at each location in the city. The result is an index based only on empirically observed behavior (in this case, the primary dataset is the 2012 CHTS) which is an absolute measure of walking behavior, not just of walkability. The methodology from this chapter yields an index for all three modes, and all indexes are included in the hedonic model described above. The model shows that a one standard deviation change in the auto index has the largest impact on home values, but that the walking index is positive, statistically significant, and almost as large. Although part of the reason for this finding might be that these neighborhoods are undersupplied, where they exist they are clearly in high demand.
The fourth topic then engages with the question of how many people actually value walking when making the residential location choice decision. In this section, latent class choice models are used so that coefficients on the three mode-specific indexes (and other neighborhood descriptors) are allowed to change based on selection into unobserved classes. This can be thought of as a form of consumer preference segmentation for mode-specific accessibility.
The model shows that there are three large segments present in the Bay Area. One that is young and moderately high-income that selects into the walkable neighborhoods of San Francisco, Oakland, and Berkeley (13\% of households), one that is transit-oriented and selects into the relatively less-expensive neighborhoods near BART but outside the urban core (37\% of households), and one that is composed of middle class families that prefers the idyllic suburbs outside San Francisco (50\% of households). Apparently about 50\% of Bay Area households value transit access, likely because BART allows commute access to the thriving labor market in the urban core (e.g. the SOMA neighborhood which is the target of so much venture capital in the region).
The main research question explored by this methodology is the question of the size of the segment of the population that is positively affected by walking accessibility for the residential location choice and the results show that this segment exists but is of modest size. However, a major finding of this research is that for planning interventions that seek to increase travel by active modes, members of the transit-oriented segment might have the most latent potential to change their behavior. Perhaps creating denser and more walkable environments around the less expensive neighborhoods near BART stations in the region could relieve pressure on the San Francisco housing market as well as create walkable environments for the lower middle class that appear to be a major component of residential demand in the region.
A ripe area for future research is to perform a gap analysis that compares neighborhoods that are high probability areas for each of the three classes presented here to test for the impact of increases in transit service and pedestrian infrastructure on both the residential location choice and travel behavior. Taking into account the heterogeneity of preferences explored here, the result of such a study would target the locations that could have the highest impact on sustainable behavior for the smallest amount of public investment.