Toward a Second Generation of Land-Use/Travel Models: Theoretical and Empirical Frontiers
Abstract and Keywords
The lack of a systematic behavioral framework for empirical studies of the influence of land use on travel raises questions regarding potential statistical bias. This article presents and assesses this debate in the language of economics to clarify the terms of the arguments and their underlying analytics. It also offers a discussion of modern empirical challenges and strategies. The article develops a general integration of land use into a travel demand framework, where land use potentially affects not only out-of-pocket and time costs of both travel and goods but also can affect travel utility directly. Thus it models built environment features that affect travel and consumer goods cost, in time and money, as well as quality with a view to study these issues in a great variety of settings. Finally, it discusses associated statistical challenges of this work and then develops an empirical model of path choice.
Among the most actively researched questions in urban planning the past two decades are if, how, and how much urban design influences individual and aggregate travel behavior. Put another way, what is the scope for using land-use form and regulation as transportation policy tools?
This line of research was initially motivated by fairly traditional transportation problems, such as traffic congestion, that seemed resistant to treatment with traditional transportation planning tools, such as infrastructure investment and even modest pricing strategies. More recently, related quality-of-life concerns regarding sprawl and excess automobile dependence have been directly linked to land-use planning, if often only by anecdote or intuition. Although much urban planning of the post–World War II period—if not earlier—explicitly favored travel by car, more contemporary planning thought finds some fault with that approach. Today, such (p. 523) efforts are also increasingly motivated by a pressing desire to promote physical activity, to address an “obesity epidemic,” and to mitigate climate change, by reducing the carbon footprint associated with fuel use associated with vehicle miles traveled (VMT).
As the role of transportation in each of these problems has become more apparent, and as the problems themselves become more pressing, interest in using urban design or other features of the built environment has grown markedly. Increasing the density of development, segregating land uses less, and opening up the street and walkway circulation patterns have all been advertised as intuitive means to reduce VMT and promote alternative modes. For example, the 2008 study “Growing Cooler” reports that such land-use practices would reduce VMT an average of 35 percent, in turn translating into a reduction of greenhouse gases by as much as 12 percent (Ewing et al. 2008).
The related and more developed theoretical and empirical literatures on consumer or labor behaviors have evolved along several lines. Early studies tended to focus on simple correlations. Workers responded to higher wages, for example, generally by increasing the amount they work. Drivers responded to higher gasoline prices, or longer trips, by reducing VMT. Such studies often confirmed intuition as to the direction of the effect, while providing data on its magnitude.
If results were consistent regarding both sign and magnitude, then measurement of these relationships could be considered robust and well understood. However, there were also cases where results were not robust, and where the underlying behavior margins were revealed by theorists to be more complex than first thought. In a familiar example, while the pure price effect of a wage increase is to make leisure more expensive, and thus stimulate labor supply, a higher wage also increases real purchasing power. This income effect makes leisure more affordable, reducing labor supply. Thus a wage increase will increase or decrease labor supply depending on whether the price effect is larger or less than the income effect. From a policy perspective, wage taxes follow the same pattern.
Thus empirical strategies for determining how taxes will affect labor supply must be modeled carefully. Further, the development of statistical models of association has a long and storied history clarifying how the structure of the data (e.g., measurement errors) and the specification of the model variables (e.g., excluded variables, multi-collinearity, simultaneity, self-selection) introduces biases or otherwise influences the credibility of either point estimates or hypothesis tests about intervals around those estimates. In some cases, the empirical literatures are more concerned about such methods questions than reporting estimated magnitudes, because of the importance of the former in establishing the credibility of the latter.
By contrast, most attention in the empirical land-use/travel research literature concerns measuring magnitudes, with relatively little attention to establishing the soundness of data structures or study methods. One explanation is that the underlying empirics typically lack a systematic behavioral framework, such as something comparable to the analytics of labor supply. The reason for that may be how well simple intuition seems to apply.
(p. 524) A common approach is to regress a travel outcome, VMT say, on a number of urban design variables. When the coefficient on population density is negative and significant, say, that point estimate is represented as the average influence of population density on VMT. It is not unheard of that the coefficient is used this way even if insignificant, if the sign corresponds to expectations, as in Ewing and Cervero (2010). That is, the standard error is not used to test a hypothesis; rather, the point estimate of the correlation is taken at face value as a precise or imprecise metric of causal influence.
The lack of a systematic behavioral framework for empirical studies of the influence of land use on travel naturally raises questions regarding potential statistical bias; it is also surprising given the long and well-known literature on travel demand. This chapter presents and assesses this debate in the language of economics, per the themes of this volume. The idea is to clarify both the terms of the arguments and their underlying analytics. No effort here is made to resolve these issues, so much as to indicate which parts are normative and which positive, and of the latter which are empirical. A discussion of modern empirical challenges and strategies is also offered. The contribution of this exercise is to indicate which empirical strategies are more likely to illuminate. The idea is to make the structure of such issues more transparent, and thus more amenable to study.
Crane (1996b) developed the microeconomic theory of land use and travel by introducing land-use variables to a simple model of trip demand as trip time costs. He then performed basic comparative statics to determine which land-use policies have theoretically unambiguous effects on trips and vehicle miles traveled, given the assumptions of the model, and which are, by contrast, empirical questions that vary with the data at hand. For example, if an urban design feature effectively reduces trip cost, such as by shortening trip length, then it will unambiguously increase trip frequency if trip demand is downward sloping in time—that is to say, if reducing trip time encourages people to make more trips. On the other hand, the resulting VMT is unknown without reference to data, as that depends on how much trips shorten compared with how many more trips are taken. Boarnet and Crane (2001a, 2001b) refined this framework slightly and then considered its implications for econometric issues, such as residential self-selection.
This chapter makes two further contributions. It develops a more general integration of land use into a travel demand framework, where land use potentially affects not only out-of-pocket and time costs of both travel and goods but also can affect travel utility directly. Thus it models built environment features that affect travel and consumer goods cost, in time and money, as well as quality. This model is used to discuss the theoretically sound construction of models of how land use is expected to influence the traditional travel outcomes of trip generation, mode choice, and VMT—and also the choice of a possibly multimodal route. Few studies have discussed route or path choice in any detail, although it clearly is a major aspect of urban travel. The final section of the chapter discusses associated statistical challenges of this work and then develops an empirical model of path choice.
(p. 525) The focus of this chapter is on how to study these issues in a great variety of settings rather than on presenting specific results from a particular data set. Yet as discussed in the conclusion, it is difficult to usefully analyze the influence of land use on travel with data without a firm grounding in the nature of the behavioral questions of interest.
The Influence of Land Use on Travel: First-Generation Studies
As mentioned, the “first generation” of statistical models of land use/travel frequently use specifications that attempt to explain travel with various elements of the built environment, such as circulation patterns and density, along with various controls entered in more ad hoc fashion. They might, as in most of the studies reviewed in the meta-analysis of Ewing and Cervero (2010), regress a travel behavior on a list of available land-use measures as well as other individual socioeconomic controls. The functional model is thus , where a is the travel outcome measure, such as automobile trips or VMT, Lk is land-use feature k, and z is a list of demographic terms. There typically is no story that suggests the respective roles of each land-use feature, such as whether they enter additively, or linearly, and so on. Indeed, the empirical specification is most frequently linear, as with:(23.1)
where ε is the stochastic error term. The estimated coefficients on each feature L are then interpreted as the marginal influence of that land-use characteristic on travel behavior. (For a relatively complete bibliography, see National Research Council  and a list of estimated coefficients from many of these studies, statistically significant and not, in elasticity form in Ewing and Cervero .)
This approach has the advantage of data availability and simple interpretation, which explains its popularity, but there is a risk of biases due to both specification and estimation issues. On the specification side, the narrative of these studies is fairly quiet about whether each land-use feature is expected to affect behavior in the same or different manner. For example, does the presence of a sidewalk influence behavior in a like or different manner than whether the street pattern is a grid? Put another way, if the travel story is based on one of consumer demand, then the size and accuracy of estimates of the influence of urban design are potentially affected by whether prices and resources are included as explanatory variables.
Further, estimation issues sometimes follow from specification issues. If the street pattern influences trip length or speed, it might be interpreted as something akin to a price variable. However, route characteristics are often the choice of the traveler, generating something like an endogenous price structure.
(p. 526) Other estimation problems might be attributable to the structure of the data. A good example in this context is the so-called self-section problem, where the mapping of travelers to land-use features is not entirely random with respect to travel behavior. For example, individuals with an above average propensity to walk or bike, for leisure or work, might choose to live or work in neighborhoods (or with associated commuting or leisure routes) that facilitate such travel. This bias in the sample would show up in a simple regression of walking on density as an overestimate of how the average person might respond to higher densities. The good news is that the literature is well aware of this potential bias and has examined several means for identifying it, testing for it, and then controlling for it in different data sets (e.g., Cao, Mokhtarian, and Handy 2009; Cao, Xu, and Fan 2010). Still, the issue is not always well understood by researchers in the field; indeed, it has not infrequently been mischaracterized as an alternate explanation for cross-sectional correlations between land-use features and travel behavior, rather than simply a potential source of sample bias (e.g., Lund 2003).
More generally, any risk of misspecification is the risk of systematically biased estimated coefficients. Indeed, this is the same critique of what is known as the “first generation” of travel demand studies, which is distinguished from the second generation by their lack of behavioral content (e.g., Lerman and Ben-Akiva 1976; Train 1978; Williams and Ortuzar 1982). Sometimes these land-use studies do contain variables that serve as proxies for travel prices, such as distances and travel cost (as in Shen 1998). There is of course the argument that the land-use elements themselves play the role of prices, by setting and changing the incentives facing potential travelers (Crane, 1996b; Boarnet and Crane 2001a, 2001b; Cao, Mokhtarian, and Handy 2009). It still remains rare, however, that land-use/travel studies begin with a careful model specification based on microeconomic demand. New research on land-use/travel connections today remains almost exclusively in the first generation of such models.
The departure for the second generation of land-use/travel studies is their foundation in the microeconomics of consumer demand, as in the second generation of travel demand studies. The latter is an extremely mature literature, so far as it goes; it does not, however, address the issues associated with urban design or treat land use in a systematic manner, if at all.
In this chapter, we contribute to the literature on extending travel demand modeling to the roles potentially played by urban design. Thus we start with a model of how choices are made over scarce consumption goods with well-defined prices. The plainest model of demand stipulates that consumers make choices based, fundamentally though not exclusively, on three sets of factors: tastes, (p. 527) resources, and prices. An individual buys more or fewer tomatoes, say, in a manner explained by how much she likes tomatoes (as compared to alternatives), her budget (over the alternatives), and how much tomatoes cost (as compared to alternatives). This kind of model works best to explain average behavior, as individuals can be idiosyncratic in ways that we hope wash out in the aggregate. It is incomplete to the extent it ignores systematic elements of consumer psychology, endogenous tastes, imperfect information about these data, and other challenges to rationality and consistent decision making in choice making, though extensive literatures exist on each of these fronts.
Thus this is a model of how choices are made over scarce consumption goods with well-defined prices. One wrinkle for our purposes is that travel is normally considered only a means to an end rather than an object of desire itself, or what is sometimes called derived demand. Perhaps the easiest way to incorporate travel is as another cost of consumption. The consumer not only has to pay the farmer for tomatoes but also has to drive to pick them up.
Travel models have an illustrious history in policy analysis as successful ways to accurately measure and then anticipate how changes in resource and costs will affect travel decisions. Daniel McFadden was awarded the Nobel Prize in Economics in 2000 in part for his work developing behavioral-based and accurate predictions of ridership on the yet-to-be-built BART transit system in the San Francisco Bay Area. Related travel models are used in every local community and interstate highway project to project expected use. The methodology is roundly considered sound and valuable.
Yet these models rarely include land use or urban form, the bread and butter of urban planning analysis and policy. This is one explanation for why studies of the influence of land use on travel did not historically use a demand framework; another may be that it was not considered particularly useful. Engineering and design approaches to travel, probably the most common basis for transportation planning studies, rarely use economic models of choice and behavior, so there is little tradition in applying them to the role of land-use aspects in travel outcomes.
Space limitations do not permit a complete argument for how the demand framework might be useful in this setting beyond the appeal of an internally consistent logic of how observed choices depend on the interaction of relative prices with preferences over the goods in question. In this case, the goods are all consumer goods obtained by travel and the means of travel themselves. The prices are goods prices, plus transportation costs. Additionally, travel decisions can often turn on time, so we include that in our story as well. Thus, our story presumes that travel is a function of resources and travel-related costs, plus other factors intended to capture tastes (such as sociodemographic variables).
To introduce notation, say individuals consume goods to obtain benefits, which we summarize by the relation,(23.2)
(p. 528) where U is a benefits metric and x is a vector of goods consumed. The value added of the demand framework is that it formalizes this choice as the solution to maximizing U subject to the available budget y and the relative prices of goods x, which we denote by p, giving the demand functions x(p, y). For most (substitute) goods and persons, these fall in relative prices and rise in income. Using observed behavior, researchers can then measure the average sensitivity of demands to changes in price (i.e., the price elasticity of demand) and income (i.e., the income elasticity), using these metrics to forecast the effect of changes in either—due to taxes or wages, for example—on final consumption.
In empirical studies, where individuals face the same prices but have different preferences, this is often modeled as x(p, y, z), where z is a vector of characteristics capturing taste differences—such as age, sex, and other socioeconomic variables. In a linear multivariate statistical model explaining the consumption of some good xi this would suggest regressing the quantity of xi purchased on its relative price, income, and individual taste variables, such as:(23.3)
where ε is the stochastic error term, pi is the price of good i, and the βs are the parameters to be estimated.
What about, as in (23.1), putting transport on the left-hand side of (23.3) and land use on the right, but also include demand variables such as relative prices and trip resources? The first challenge is to model the demand for an activity that is not normally considered a consumption good in itself, the trip or some other travel measure, such as trip length or mode choice. The second is to model land-use and urban design features in a manner consistent with the demand framework.
In standard travel demand theory, both (23.1) and (23.3) would thus be considered theoretically underspecified and thus statistically suspect due to missing variable bias. In the first case, it is probably missing prices and budgets (unless an argument can be made that land-use features play those roles exogenously), and in the second it is missing land use (if land use plays a systematic behavioral role, especially as prices or resources). In an effort to make either set of models consistent with the demand theory frame, Crane (1996b) treated trips as basically an ordinary consumer good, where the time duration of the trip stood in for trip price. The reasoning was that as a first approximation, each trip could represent the goods purchased on a trip. He thus specified the demand for a trip by all modes as the demand functions , where Ta is the trip time by mode a (as in a = auto, b = bus, w = walking and so on), and y is the time budget in this formulation.
Adding this theoretical framing to the transportation/land-use models in use in the 1990s (and still) permitted elementary comparative statics to clarify which hypotheses concern the model, and which the underlying empirical questions. For example, a shortening of the trip by any mode would be expected to increase the number of trips by that mode, much as a price decrease moved consumers down their demand curve for that good. That is, the demand for trips is expected to be (p. 529) downward sloping in trip length and time. This simple point extended the analysis of a family of computer simulation models, such as by Kulash, Anglin and Marks (1990) and McNally (1995), to include trip generation. Those models had calculated that a more open street circulation pattern effectively reduced trip lengths, and thus VMT—but under the extremely restrictive assumption that trip frequencies did not change with changes in either trip distance or speed.
Yet, if trip frequencies were permitted to adjust to shorter distances, they will likely rise by the law of demand. Since VMT is the number of trips times their average length, the unambiguous simulation result that VMT would fall in neighborhoods with shorter trips becomes, more credibly and reasonably, ambiguous. The other comparative statics results of this particular model indicated that many design features had similarly ambiguous travel outcomes a priori, with the exception of features that changed trip speeds: slower speeds, such as from traffic calming, lead to both fewer trips and lower VMT (Crane, 1996b).
Crane and Crepeau (1998) and Boarnet and Crane (2001a, 2001b) implemented this framework empirically, as has Cao (2006). Their focus has been on nonwork travel, since that is more discretionary, and on design features that changed trip lengths, since length is an easy measure of trip cost. In these applications, it is appropriate to model both money and time costs, and doing so often resulted in built environmental features having less statistical influence on travel decisions than in models where these are absent.
However, these treatments, fuller though they may be as choice frameworks, remain unsatisfactory for several reasons. First, land-use measures have almost always been applied to the neighborhood of the trip origin, with only a few extending such measures to route-option scales. This approach dismisses the role of the built environment at the destination or, more notably regarding choice options, along the way. In addition, it is incomplete to model trips as ultimate consumption goods. For Crane (1996b), this was a simplification to proxy for the goods purchased in those trips.
Say utility is determined by consuming things obtained through travel, which takes time and money. In addition, the act of traveling may have other benefits. So the consumption problem is to choose goods g to maximize(23.4)
(p. 530) and(23.6)
where x are consumption goods not obtained by travel, is the fixed plus variable money travel cost of consuming goods g, FCij is the fixed travel cost of purchasing good i by mode j (the part that does not depend on distance), MCij is the per mile cost of purchasing good i by mode j, and AD is the number of miles traveled along the path between trip origin A and destination D. In addition, y is the budget for goods obtained traveling, T is the travel time associated with those goods, 24 is the number of hours in the day, and h is the nontravel time in the day. (Even better, we could deconstruct T into its fixed and variable components much as how we treat out-of-pocket travel costs c. We could also model time as a argument of utility, where the user cost of travel would then also include the disutility of time, as discussed in the time allocation literature, as in Bates 1987.)
This formulation is fairly general, as it permits travel to have both utility and disutility (via costs). Land use could affect either scale, via c or g. Note that a number of subproblems are suppressed there to simplify the notation. For example, the cost of travel c is the solution to the problem of finding the least cost trip for every purchase of g between any two points A and D, that is, of choosing a path c such that:(23.7)
In turn, every path involves multiple mode options.
The solution to (23.4) subject to (23.5) and (23.6) can be represented by the indirect utility function(23.8)
and the associated demand functions, , and where m is mode and f is trip frequency. The optimization problems of choosing modes, trip frequency, and then VMT are thus summarized as functions of the fixed and marginal time and money costs of travel, as well as of other attributes of the trip.
Each of these choice margins has counterparts in different kinds of urban design features. Some, such as more open circulation patterns, lower the cost of each trip and do so differently by mode. Others, such as parking regulations, affect the fixed trip cost. And still others, such as the effort or pleasure associated with walking or biking, have direct consumption benefits.
Rather than explore these in detail, to save space we leave those exercises for other researchers. For example, as in Crane (1996b) and Boarnet and Crane (2001a, 2001b), these demand functions could be used to obtain comparative statics of the direction of influence of specific built form features on different travel behaviors. Some are theoretically unambiguous, such as how higher marginal automobile trip costs (i.e., ), such as from the lowered car trip speeds associated with speed (p. 531) bumps or other traffic calming measures, will lower trip frequency. Others are theoretically indeterminate, such as how VMT will respond to shorter trip lengths (which in turn increase trip frequency). Instead, the next section further develops the empirical basis for explaining how land-use variables affect path and mode choice.
As a preliminary step, note that modes and paths are discrete. In a discrete choice framework, where ADk is a choice among K alternatives, the probability that person n will choose alternative k is(23.9)
The corresponding logit model (Ben-Akiva and Lerman 1985) can then be written as(23.10)
where the denominator sums across all choices. The next section develops a path choice model using this approach.
Path choice investigates travelers’ decisions over multiple path options between the same origin and destination. It has been widely used in transportation planning to assign flows within a transportation network but has rarely been used in a land use/travel study (Lee and Moudon 2006). Following the general analytical framework presented in the earlier part, we develop a modeling structure to test the amenity impact of urban design features along a pedestrian path on the probability of that path being chosen by pedestrians. The utility of a walking trip is a function of path length (proxy of travel time and cost since walking normally does not involve monetary cost), urban design amenities (sidewalk width, intersection design, etc.), and trip and personal characteristics. Here urban design features may affect walking travel time/cost (e.g., people can walk faster on wider sidewalk), but the main effect is defined here as psychological. For example, pedestrians are willing to walking on indirect paths instead the shortest one if the former provides more amenities along the way. In addition to the sound behavioral base, a path choice approach is also able to mitigate the self-selection problem that perplexes the land use-travel research due to its unique setup.
For example, this approach is able to avoid some typical measurement issues often seen in the first generation land-use/travel model, such as zone-based measure at either trip origin or destination. The zone-based approach often measures (p. 532) the pedestrian environment incompletely and disproportionally: only a portion of the path is measured, and short-distance trips are weighted more than long-distance trips if the zone size is fixed. In other words, the zone-based approach measures the “treatment” differently for different “subjects,” which violates the principles of experimental design.
More importantly, the path choice approach offers a new perspective to understand the self-selection problem. Self-selection occurs because residents are often grouped based on their social status, ethnic background, lifestyle preferences, and other factors. Distinct groups tend to live in different neighborhoods, and distinct neighborhoods often comprise different social groups. Therefore, comparisons across distinct neighborhoods raise the question of self-selection, while comparisons among similar residents, who often live in neighborhoods with a similar land-use pattern, result in a lack of “treatment” in research design (Guo 2009). The first comparison can easily identify a correlation between land-use and travel behavior, but proving it causal is a challenge. The second comparison may not even find a correlation, but if it does, such a correlation is more likely to be causal.
Which method is better depends on whether the first can effectively control for unobservable personal differences, and whether the second can secure sufficient variation in the treatment (land use; Guo 2009). While most prior studies on self-selection in land-use/travel models used the first method, the empirical outcome suggests that the second method tends to produce a better result. The path choice approach follows the second method. In the following part, we compare the two methods in detail, explain why a path choice situation with careful design can represent a quasi-experimental design, and proposed a suitable modeling structure.
Treatment-First versus Traveler-First Design
Following the treatment-fist design, studies tend to target neighborhoods with distinct land-use patterns, such as urban versus suburban neighborhoods, transit-oriented versus auto-dependent neighborhoods, conventional versus neotraditional neighborhoods, or diverse neighborhoods in a metropolitan area (table 23.1).
Various methods have been used to control for unobservable personal differences and self-selection. Cervero and Duncan (2002) used a joint location and travel choice model and found that self-selection accounts for 40 percent of the probability of making a decision to commute by rail. Using a different joint-choice model, Bhat and Guo (2007) did not find any evidence of self-selection and confirmed that land use affects automobile ownership. Using instrumental variables, Greenwald and Boarnet (2001) concluded that certain characteristics of land use do promote walking, even while taking into account the possibility of self-selection. Khattak and Rodriguez (2005) had comparable findings using survey data from Chapel Hill, North Carolina.
(p. 533) (p. 534) (p. 535) However, using a similar method, Boarnet and Sarmiento (1998) found that predicted land-use measures were not significantly related to individual nonwork auto trip frequency after taking into account the influence of self-selection. Using structural equations, Bagley and Mokhtarian (2002) found that when attitudinal, lifestyle, and sociodemographic variables are accounted for, neighborhood type has little influence on travel behavior. Based on the same method for the same region, Cao, Mokhtarian, and Handy (2007) found that land use still affects auto ownership, even after controlling for self-selection. After incorporating personal attitudes and preferences into their analysis, Handy, Cao, and Mokhtarian (2006) showed that self-selection at least partially explained walking behavior, but that land use still had an impact. Following a similar method, however, Chatman (2005) found self-selection to be a negligible factor, as those with strong mode preferences seemed to be less sensitive to land-use variables, while those with weak mode preferences seemed to be more responsive to differences in land use.
Table 23.1 Summary of Prior Studies on Self-Selection and Causality Investigation
Greenwald and Boarnet
ODOT Travel Survey in 1994
Yes (only at neighborhood, not zip code, level)
Khattak and Rodriguez
Chapel Hill and Carrboro, NC
Self-conducted household travel survey in 2004
Modal choice, length, frequency
Yes (self-selection exists)
Boarnet and Sarmiento
Los Angeles, CA
Panel Study of S. CA Commuters from 1990 to 1994
Number of auto trips and VMT
No (except for nonwork auto trip frequency)
Cervero and Duncan
San Francisco, CA
Bay Area Travel Survey in 2000
Car ownership, modal choice
Yes (self-selection exists)
Bhat and Guo
San Francisco, CA
Bay Area Travel Survey in 2000
Yes (self-selection does not exist)
Handy et al.
San Francisco, CA
Self-conducted household travel survey in 2003
Walking and biking frequency
Yes (self-selection exists)
San Francisco, CA San Diego, CA
Travel Survey by Chatman in 2004
Modal choice and VMT
No (self-selection does not exist)
Bagley and Mokhtarian
San Francisco, CA
Travel Survey sponsored by CA Air Resources Board in 1992
Mileage by car, transit, and walk/bike
No (self-selection exists)
Cao et al.
San Francisco, CA
Self-conducted household travel survey in 2003
Yes (self-selection exits)
PSTP Panel Survey from 1989 to 1998
VMT and frequency
Meurs and Haaijer
Dutch Time Use Study from 1990 to 1999
Number of trips by modes
Boarnet et al.
Field survey by authors
Schwanen and Mokhtarian
San Francisco, CA
Travel Survey by authors in 1998
Distance traveled overall and by mode
Such inconsistency is especially striking given that these studies are based on similar methods and sometimes the very same data set, as well as on the same metropolitan region during similar times (see table 23.1). It is hard to believe that such inconsistency reflects different self-selection styles at the metropolitan level. Rather, the inconsistency is likely caused by inefficiency in the studies’ methods to control for unobservable personal differences, especially given the limits of current land-use measurement, available data, and statistical models (Guo 2009).
In contrast, the traveler-first design is (1) to find travelers with similar attitudes and preferences toward travel, and (2) to secure sufficient variation in the land use experienced by travelers. A few studies following this method have produced consistent results showing that causality exists despite the “threat” of self-selection. Such research designs include longitudinal research, intervention design, and matched attitude (see table 23.1).
Longitudinal research examines the land-use impact on movers’ travel behavior before and after a move. Targeting the same person before and after a treatment is surely an advantage for a causality study, but the longitudinal design also raises two issues. First, if a move is caused by a preference change, the mover self-selects and is therefore no longer essentially the “same person” before and after moving. Second, many movers relocate close to their initial locations (Krizek 2003), meaning that the variation of land use before and after the move might not be large enough to allow a statistical analysis. Sometimes researchers must combine movers and nonmovers together in their model estimation, which essentially becomes a cross-sectional research design (Krizek 2003).
Intervention design investigates residents’ behavioral change before and after a major investment in a neighborhood. Because most residents remain the same, this method solves the residential self-selection problem but still raises similar questions as in longitudinal design. First, the investment might be self-selected. For example, a neighborhood that is more favorable to pedestrian activities might be more likely to request an improvement, and thus more likely to get it. Studies on the Safe Routes to School projects in California indicated that investments were more effective in (p. 536) school areas with already moderate or high levels of walking but were insufficient to affect travel modes in schools with previously low levels of walking or bicycle travel (Boarnet et al. 2005). Second, such improvements are often marginal (such as improved sidewalks and crosswalks, bike paths, traffic signals, speed bumps, and other improvements), and thus may not be strong enough to induce behavioral changes (Meurs and Haaijer 2001).
The matched attitude method finds individuals with similar preferences toward travel but who live in neighborhoods with distinct land-use patterns. If they travel differently, such a difference should be largely caused by a land use difference. This approach is best represented by Schwanen and Mokhtarian's (2004, 2005) study of dissonant residents. They compared dissonant urban residents with consonant suburban residents and found the former used cars less frequently than did the latter but more frequently than did the consonant urban residents. However, the method assumes that dissonant urban residents share similar travel and living preferences to consonant suburban residents, which is not always true. Being unsatisfied with urban living does not mean these residents want to live in suburban neighborhoods (Guo 2009).
Despite these various methods, traveler-first design tends to produce consistent results that confirm the causal impact of land use on travel behavior. This design's merit might not be an accident. The same logic is used in other study areas with similar methodological concerns. For example, when researchers investigate how the environment affects children's behavior and achievement, researchers often use twins (Horwitz et al. 2003) or siblings (Aaronson 1998) from the same household to control for unobservable personal and household attributes.
The comparison between these two methods suggests that controlling for unobservable personal differences is more difficult than is finding land-use differences between similar travelers. However, empirical studies often choose the first method that enlarges, instead of minimizes, the personal differences. In other words, they sacrifice causality for correlation (Guo, 2009).
Path Choice and Self-Selection
The path choice approach meets the requirements of the traveler-first design in several ways. The biggest advantage of path choice is that it is less likely to correlate with residential location decisions. Other travel decisions, like mode choice, VMT, trip frequency, and car ownership, are long-term lifestyle decisions and are therefore associated with housing location choice. However, path choice is a sublevel decision given location choice—fixed origin and destination. For example, when pedestrians’ walking paths from a subway station to their office are targeted, their path choice decision is unlikely to correlate with their job location decision. Workers are unlikely to choose their workplace based on the attractiveness of the walking path to the location.
The same argument might not be true for shopping trips since shoppers may choose their stores based on the attractiveness of a walking path. Second, it is reasonable to assume that subway commuters to an urban center share a similar (p. 537) attitude and preference toward travel, at least comparing to the samples used in prior studies. Third, when the pedestrian environment is measured along a path, it is more sensitive to pedestrian environment differences experienced by pedestrians than is the traditional measure at the zone level (e.g., neighborhood, census tract, Traffic Analysis Zones). In other words, this path choice set up, based on subway commuters’ egress walking path in urban center, reduces the heterogeneity in travelers while is able to detect even small variation in the treatment (pedestrian environment).
Figure 23.1 illustrates such a situation. A, B, and C are stations on two separate subway lines, and D is the trip destination. Suppose a traveler enters the area on line 1 from the south. When the traveler reaches station A, he or she has two options to get to destination D, which is closer to station C. The traveler can leave the system at A and walk to D, or can continue traveling on line 1 to line 2, transfer at B, exit at C, and then walk a shorter distance to D. Therefore, the traveler has two possible egress subpaths: ABCD or AD. Which path is better depends on four path attributes: extra in-vehicle time spent on path ABC, transfer convenience at station B, the street walking time saved (AD − CD), and the pedestrian environment along AD and CD. Therefore, what we present here is a subpath choice situation where the path decision is made at the egress segment, conditional on the mode and service selected by the traveler for the trunk portion.
Such a path choice situation is not typical but offers both methodological and technical advantages over a traditional path choice situation that covers the entirety from origin to destination. Besides the mitigated self-selection problem, path choice modeling also becomes simple. Path choice is normally difficult to model for two reasons. First, we know only the chosen path, not alternative paths considered by a traveler when the decision is made (Bovy and Stern 1990). These alternative paths must be generated based on various decision rules, which is very difficult and hard to validate (Hoogendoorn-Lanser 2005). Second, these multiple paths often overlap because they begin at the same origin and end at the same destination, violating the assumption of independent and identical distribution (IID) for discrete choice models (Ben-Akiva and Lerman 1985). The overlap can be very complicated and makes the correcting effort extremely difficult. In the path situation set up in Figure 23.1, we do not need to define the attributes of the full path from the origin to the destination. Alternative paths are easy to identify because they are attached to a few transit lines. The problem of path correlation is greatly reduced because ABCD and AD are unlikely to overlap with one another.
Additionally, such a setting benefits path choice modeling because of the high ridership in an urban center guarantees a large sample size; concentrated transit stations and job locations provide ample egress path options; the pedestrian environment is diverse; and there is no competing egress mode other than walking. A subway survey in Boston reveals that multiple egress options exist, and passengers indeed made the trade-off between the two subpaths, ABCD and AD. Among the 6,500 subway trips ending in downtown Boston, half of them have the ABCD option, and one-third of this group finally chose the subpath ABCD over AD to get to their (p. 538) destinations (Guo and Wilson 2004). Most subway commuters (98 percent) left the system and walked to their final destinations in downtown Boston.
Readers should be aware that we apply the path choice method to the pedestrian environment and walking because walking is an indispensible part of most modes of travel; it is universally available to almost everybody; it exposes people directly to the built environment; and it is the most common and preferred form of physical activity for the general population (Badland and Schofield 2005). However, it can be applied to other land use/travel relationships as well.
A subway commuter always chooses the path with a higher utility between ABCD and AD. Their path utilities are determined by the transit and walking experiences (for path ABCD), or solely the walking experience (for path AD). The walking experience is further determined by two factors: path length (walking time) and the pedestrian environment along the path. If the pedestrian environment affects the (p. 539) path choice, it does so by affecting the utility of walking along either AD or CD, and then the causal relationship between the pedestrian environment and walking is justified.
The follow-up question would be how much the utility of walking is affected by the pedestrian environment because, if the magnitude is small, such a causal relationship does not make sense for policy intervention, even if it is statistically significant. Therefore, the investigation results in two questions:
Question 1: Does the pedestrian environment affect the utility of walking? Alternatively, are pedestrian environment variables statistically significant in the path choice model after controlling for other variables?
Question 2: Is the causal effect significant enough to justify policy intervention? Or, what is the percentage change in walking utility caused by the pedestrian environment?
The two questions are elaborated in mathematical form in the following equations. Suppose paths ABCD and AD are recoded as paths 1 and 2, respectively, and path 2 (AD) is treated as the base, then the modeling structure takes a binary logit form (Ben-Akiva and Lerman, 1985):(23.11) (23.12) (23.13)
where Pn(1) is the probability of person n selecting path ABCD; V1n and V2n are the systematic components of the utility for paths ABCD and AD for person n; µ is the positive scale parameter; C is the alternative specific constant for path ABCD; S1n is a vector of transit characteristics for path ABCD for person n; T1n and T2n are egress walking times for path ABCD and AD for person n; PE1n and PE2n are vectors of pedestrian environment variables along the path of CD and AD for person n; K1n is a vector of all other attributes for path ABCD for person n; and α, β, γ and δ are the coefficients to be estimated. All these coefficients are vectors, except β.
When the pedestrian environment effect is not considered, the utility of walking is solely determined by walking time (since cost is normally not presented). The initial utility of walking may either increase or decrease after the pedestrian environment is included, depending on pedestrian environment quality. The utility change can be measured as the equivalence of walking time by calculating the coefficient ratio between walking time and the pedestrian environment variables. Denoting the initial utility of AD or CD for street segment i as U0i, and the combined utility as U1i, then the mathematical forms are(23.14) (p. 540) (23.15) (23.16)
where Ti is the walking time on pedestrian path i, and PEi are the pedestrian environment variables along path i. Ri is the percentage of walking utility change due to the pedestrian environment effect, indicating the magnitude of the impact for each pedestrian path i.
In estimation, Ri could be either positive or negative. A negative sign indicates that the pedestrian environment is poorly maintained and adds negative utility to walking. A positive sign indicates a pleasant pedestrian environment that adds extra utility to walking. When Ri is positive and greater than one, it means that the positive utility conferred from the pedestrian environment out-weights the negative utility of walking time (time is always negative since travel is viewed as a derived demand). U1i becomes positive instead of negative. In other words, the pedestrian environment is so enjoyable and walking itself possesses a utility. Pedestrians may increase the consumption by walking longer paths and staying longer in the environment.
This modeling structure has direct policy implications to pedestrian infrastructure planning and investments. By measuring the utility of the pedestrian environment in monetary terms at the street segment level, this approach helps planners identify which streets should be improved, what improvements are needed, and how investments could be justified in a quantitative way. The current assessment method on pedestrian investments focuses on head count (e.g., accident reduction or new users) whereas the mainstream assessment method in transportation is based on monetary gains (through time savings) for all users. The head count–based method tends to underevaluate the benefits of investments in pedestrian infrastructure. The proposed method could correct the bias and enhance the efficiency and effectiveness of decision making in pedestrian infrastructure planning and investments.
Several concerns about the approach remain. First, self-selection may still exist at the path level. For example, if a group of people work in a particular neighborhood, or work in a particular downtown district served by a particular subway line, where ABCD is more convenient than AD due to the network configuration, then these people will always choose path ABCD. In this case, the path choice decision is correlated with housing and job location choices, which would either over- or underestimate the pedestrian environment effect on walking. This is a case-by-case condition and needs to be tested in empirical studies.
Second, the shortest-path assumption from a subway station to a destination may not be true. One way to relax this assumption is to select a group of destinations that do not involve street directional changes from a station (such as D in (p. 541) figure 23.1). Another assumption is the uniform walking speed of three miles per hour for all pedestrians. Unless walking speed is correlated with the pedestrian environment, which we know little about empirically and theoretically, this assumption is unlikely to result in a systematic bias in the estimation result.
Finally, there are two issues regarding the generalizability of findings. On the one hand, the adopted research design tends to underestimate the pedestrian environment impact on walking: the pedestrian environment effect would have been stronger if the comparison were across different neighborhoods or covered non-commuting trips. On the other hand, the pedestrian environment impact is likely to be weaker in a modal choice than in a path choice. Case studies in multiple metropolitan regions for multiple travel types are necessary to provide a complete picture of the causal land-use/travel relationship.
The Transportation Economics of Urban Design
Few policy problems have been as energetically studied in recent years as how to use land to accomplish transportation planning ends. Yet the sharpest critiques of this body of work are its common lack of rigor, much as found in the first generation of travel demand studies, and its pervasive failure to examine how land use and mode options along the trip matter for the traveler.
Rigor refers here to our confidence in using the results of any one study, of any one data set, to draw useful and reliable conclusions about the feasibility and usefulness of applying urban design toward transportation goals elsewhere. To make progress toward that end, this chapter considered several respects in which the behavioral foundations of statistical models, and the specification and estimation of those models, might be improved. It first developed a more general model of the demand for travel with built environment characteristics, where travel influences consumption cost as well as direct utility. Second, it further investigated the empirics of path choice, where the land-use characteristics of the path matter.
Neither of these improvements displaces the work that came before. Indeed, the first generation of these analyses was extremely useful in identifying questions and generating complex land use metrics. This chapter does, however, challenge the use of simple urban form/travel correlations, significant or not, and land uses measures surrounding trip origins and destinations only, to address policy issues. The sooner this body of research makes regular use of behavioral- and path-based statistical strategies, the sooner we can agree as a field about the roles the built environment might play in solving transportation problems.
Aaronson, D. 1998. “Using Sibling Data to Estimate the Impact of Neighborhoods on Children's Educational Outcomes.” Journal of Human Resources 33:915–946.Find this resource:
Badland, H., and G. Schofield. 2005. “Transport, Urban Design, and Physical Activity: An Evidence-Based Update.” Transportation Research Part D 10:177–196.Find this resource:
Bagley, M., and P. Mokhtarian. 2002. “The Impact of Residential Neighborhood Type on Travel Behavior: A Structural Equations Modeling Approach.” Annals of Regional Science 36:279–297.Find this resource:
Bates, J. 1987. “Measuring Travel Time Values with a Discrete Choice Model: A Note.” Economic Journal 97 (386):493–498.Find this resource:
Ben-Akiva, M., and S. R. Lerman. 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. Cambridge, MA: MIT Press.Find this resource:
Bhat, C., and J. Guo. 2007. “A Comprehensive Analysis of Built Environment Characteristics on Household Residential Choice and Auto Ownership Levels.” Transportation Research Part B 41:506–526.Find this resource:
Boarnet, M., and R. Crane. 2001a. “The Influence of Urban Design on Travel: Specification and Estimation Issues.” Transportation Research A 35(9):823–845.Find this resource:
———. 2001b. Travel by Design: The Influence of Urban Form on Travel. New York: Oxford University Press.Find this resource:
Boarnet, M., K. Day, C. Anderson, T. McMillan, and M. Alfonzo. 2005. “Can Street and Sidewalk Improvements Enhance Walking and Bicycling to School? Evaluating California's Safe Routes to School Program.” Journal of the American Planning Association 71:301–317.Find this resource:
Boarnet, M. G., and S. Sarmiento. 1998. “Can Land Use Policy Really Affect Travel Behaviour? A Study of the Link between Non-work Travel and Land-Use Characteristics.” Urban Studies 35:1155–1169.Find this resource:
Bovy, P., and E. Stern. 1990. Route Choice: Wayfinding in Transport Networks. Norwell, MA: Kluwer Academic.Find this resource:
Cao, X. 2006. “The Causal Relationship between the Built Environment and Personal Travel Choice: Evidence from Northern California.” Ph.D. dissertation, University of California, Davis.Find this resource:
Cao, X., P. L. Mokhtarian, and S. L. Handy. 2007. “Cross-sectional and Quasi-panel Explorations of the Connection between the Built Environment and Auto Ownership.” Environment and Planning A 39:830–847.Find this resource:
———. 2009. “Examining the Impacts of Residential Self-Selection on Travel Behavior: A Focus on Empirical Findings.” Transport Reviews 29:359–395.Find this resource:
Cao, X., Z. Xu, and Y. Fan. 2010. “Exploring the Connections among Residential Location, Self-Selection, and Driving: Propensity Score Matching with Multiple Treatments.” Transportation Research A 44 (10):797–805.Find this resource:
Cervero, R., and M. Duncan. 2002. “Residential Self-Selection and Rail Commuting: A Nested Logit Analysis.” Working Paper 604, University of California Transportation Center. http://www.uctc.net/papers/604.pdf. Accessed March 18, 2007.Find this resource:
Chatman, D. 2005. “How the Built Environment Influences Non-work travel: Theoretical and Empirical Essays.” Ph.D. dissertation, University of California, Los Angeles.Find this resource:
Crane, R. 1996a. “Cars and Drivers in the New Suburbs: Linking Access to Travel in Neotraditional Planning.” Journal of the American Planning Association 62:51–65.Find this resource:
———. 1996b. “On Form versus Function: Will the New Urbanism Reduce Traffic, or Increase It?” Journal of Planning Education and Research 15:117–126.Find this resource:
(p. 543) Crane, R. and R. Crepeau. 1998. “Does Neighborhood Design Influence Travel?: A Behavioral Analysis of Travel Diary and GIS Data.” Transportation Research D 3(4):225–238.Find this resource:
Ewing, R., K. Bartholomew, S. Winkelman, J. Walters, and D. Chen. 2008. Growing Cooler: The Evidence on Urban Development and Climate Change. Washington, DC: Urban Land Institute.Find this resource:
Ewing, R., and R. Cervero. 2010. “Travel and the Built Environment: A Meta-analysis.” Journal of the American Planning Association 76:265–294.Find this resource:
Guo, Z. 2009. “Does the Pedestrian Environment Affect the Utility of Walking? A Case of Path Choice in Downtown Boston.” Transportation Research D 14:343–352.Find this resource:
Guo, Z., and N. H. W. Wilson. 2004. “Assessment of the Transfer Penalty: A GIS-Based Disaggregate Modeling Approach.” Transportation Research Record 1872:10–19.Find this resource:
Greenwald, M., and M. Boarnet. 2001. “The Built Environment as a Determinant of Walking Behavior: Analyzing Non-work Pedestrian Travel in Portland, Oregon.” Transportation Research Record 1780:33–42.Find this resource:
Handy, S., X. Cao, and P. Mokhtarian. 2006. “Self-Selection in the Relationship between the Built Environment and Walking: Empirical Evidence from Northern California.” Journal of the American Planning Association 72:55–76.Find this resource:
Hoogendoorn-Lanser, S. 2005. “Modeling Travel Behavior for Multi-modal Transport Networks.” TRAIL Thesis Series T2005. TRAIL, UT Delft, Netherlands.Find this resource:
Horwitz, A., T. Videon, M. Schmitz, and D. Davis. 2003. “Rethinking Twins and Environments: Possible Social Sources for Assumed Genetic Influences in Twin Research.” Journal of Health and Social Behavior 442:111–129.Find this resource:
Khattak, A. J., and D. Rodriguez. 2005. “Travel Behavior in Neo-traditional Neighborhood Developments: A Case Study in USA.” Transportation Research Part A 39:481–500.Find this resource:
Krizek, K. 2003. “Residential Relocation and Changes in Urban Travel: Does Neighborhood-Scale urban Form Matter?” Journal of the American Planning Association 69:265–281.Find this resource:
Kulash, W., J. Anglin, and D. Marks. 1990. “Traditional Neighborhood Development: Will the Traffic Work?” Development 21 (July/August):21–24.Find this resource:
Lee, C., and A. Moudon. 2006. “Environmental Correlates of Walking for Transportation versus Recreation Purposes.” Journal of Physical Activity and Health 3 (1): S99–S117.Find this resource:
Lerman, S. and M. Ben-Akiva. 1976. “Disaggregate Behavioral Model of Automobile Ownership.” Transportation Research Record 569:34–55.Find this resource:
Lund, H. 2003. “Testing the Claims of New Urbanism: Local Access, Pedestrian Travel, and Neighboring Behaviors.” Journal of the American Planning Association 69 (4):414–429.Find this resource:
McNally, M. and S. Ryan. 1993. “Comparative Assessment of Travel Characteristics for Neotraditional Designs.” Transportation Research Record 1400:67–77.Find this resource:
Meurs, H., and R. Haaijer. 2001. “Spatial Structure and Mobility.” Transportation Research Part D 6:429–446.Find this resource:
Mokhtarian, P. L. and X. Cao. 2008. “Examining the Impacts of Residential Self-Selection on Travel Behavior: A Focus on Methodologies.” Transportation Research B 43 (3):204–228.Find this resource:
National Research Council. 2009. Driving and the Built Environment: The Effects of Compact Development on Motorized Travel, Energy Use, and CO2 Emissions, Transportation Research Board Special Report 298. Washington, DC: National Academy Press.Find this resource:
Rodríguez, D., A. Khattak, and K. Evenson. 2006. “Can New Urbanism Encourage Physical Activity?” Journal of the American Planning Association 72:43–54.Find this resource:
(p. 544) Schwanen, T., and P. Mokhtarian. 2004. “The Extent and Determinants of Dissonance between Actual and Preferred Residential Neighborhood Type.” Environment and Planning B: Planning and Design 31:759–784.Find this resource:
———. 2005. “What If You Live in the Wrong Neighborhood? The Impact of Residential Neighborhood Type Dissonance on Distance Traveled.” Transportation Research Part D 10:127–151.Find this resource:
Shen, Qing, 1999. “Transportation, Telecommunications, and the Changing Geography of Opportunity.” Urban Geography 20:334–355.Find this resource:
Train, K. 1978. “The Sensitivity of Parameter Estimates to Data Specification in Mode Choice Models.” Transportation 7:301–309.Find this resource:
Williams, H.C.W.L. and J. de D. Ortuzar. 1982. “Behavioural Theories of Dispersion and the Misspeciﬁcation of Travel Demand Models.” Transportation Research 16B:167–219.Find this resource: