Market potential for smart growth neighbourhoods in the USA: A latent class analysis on heterogeneous preference and choice

Abstract

With data from a 2011 National Association of Realtors community preference survey, we examined individuals’ preferences and the resulting market potential for smart growth neighbourhoods in the USA. Using a latent class choice model, we discovered four classes of individuals that reveal distinctive behaviours when choosing smart growth neighbourhoods, based on the interplay between aspects of community design, socioeconomic characteristics and personal attitudes. Based on these results we estimated the demand for smart growth neighbourhoods given the way they are planned and built. By linking the results of the latent class choice to a market diffusion model we were able to evaluate the effectiveness of a proposed smart growth neighbourhood design in inducing less sprawling development.

Keywords

choice simulation heterogeneous preference latent class analysis market diffusion smart growth

Introduction

Sustainable urban development is critical to mitigate human impacts on the local and global environment (Brown and Southworth, 2008). Unfortunately, sprawling, low-density development, especially in North American cities, is raising an increasing number of environmental concerns (Gonzalez, 2005; Martins, 2012). These concerns include a loss of land resources and biodiversity, the increase in extreme heat events (Stone et al., 2010), growing traffic congestion and an added burden on materials and energy use (Ewing, 1997; Johnson, 2001). In response to these concerns, urban planners and policy makers have been investigating more sustainable urban forms (Jabareen, 2006). In particular, ‘smart growth’ represents an integrated concept of creating compact, walkable, transit-oriented and mixed land-use neighbourhoods with a variety of housing choices (Environmental Protection Agency (EPA), 2012b). The benefits of smart growth by locating new homes closer to public transit, stores and other amenities are often significant, including an increase in the local tax base, savings in infrastructure costs and the reduction of environmental impacts when compared with conventional suburban developments (EPA, 2012a).

Although the literature shows some criticisms of smart growth associated with higher prices, traffic congestion and privacy (Bruegmann, 2005), there is an increasing need for the transformation to smart growth with a growing population living in urban areas. One of the driving factors for transformation of urban form is the public’s preference for and choice of smart growth (Levine et al., 2005; Lewis and Baldassare, 2010). Consumer surveys in the last decade indicate a growing proportion of households in the USA prefer smaller-lot and multi-family housing in exchange for locational accessibility and financial savings (Litman, 2009). However, most of the interpretations of such survey responses are descriptive, rather than predictive, and thus not as instructive for policy making and investment decision-making as would be more quantitative responses related to smart growth. There is a need here to capture more quantifiable evidence on home buyers’ behavioural patterns as they relate to the design of housing and neighbourhoods, rather than basing conclusions on somewhat vague concepts of ‘smart growth’ or ‘sprawling community’ (Ewing and Cervero, 2010; Giles-Corti et al., 2013). Urban planners, policy makers and developers need to gain a more accurate and comprehensive understanding of the public’s preference and demand for smart growth.

Discrete choice models (DCMs) have been widely used to analyse how individuals make choice decisions from a set of alternatives. The theoretical framework for the DCMs is derived from random utility theory (RUT), in which decision makers are assumed to pursue utility maximisation (McFadden, 1986; van Putten et al., 2011). To date, a variety of DCMs has been substantially developed, including the multinomial logit (MNL) model, nested logit (NL), generalised extreme value (GEV) model, and mixture logit model (Hoyos, 2010). In this study, we selected the latent class (LC) model, belonging to the class of mixture logit models, to account for the heterogeneity in people’s choices (Boxall and Adamowicz, 2002; Greene and Hensher, 2003; Magidson and Vermunt, 2002). Preferences for housing, neighbourhood and accessibility vary from homebuyer to homebuyer. Accounting for the preference heterogeneity can contribute to improving the modelling performance and understanding the diverse demands on smart growth (Birol et al., 2006). In the LC model, parameters, weights and choices are distinctive across the unobserved subgroups, or latent classes. Belonging to one class membership is probabilistic, depending on an individual’s socioeconomic characteristics and attitudes towards the choices offered. For instance, Rid and Profeta (2011) identified three segments in the private homebuyers’ market with distinctive ‘environmental awareness’ and preferences for sustainable houses in Germany.

In this study, we applied the LC choice analysis to data from a 2011 National Association of Realtors (NAR) community preference survey of 2071 American adults, to understand US citizens’ preferences and choice behaviour towards smart growth neighbourhoods versus conventional sprawling communities (National Association of Realtors, 2011). On the basis of this LC choice model, we observed the difference in the weights of design properties on the choice model’s utility function across different classes. We identified the impacts of socioeconomic status and attitudes on an individual’s class membership. We also applied the choice model to evaluate the market potential of a hypothetic smart growth neighbourhood (i.e. the probability of an individual choosing the smart growth neighbourhood versus a conventional sprawling community). We further evaluated the impact of the model’s implied market segmentation and potential for smart growth neighbourhoods through a market diffusion model. This market diffusion model predicts the macro-level diffusion of smart growth neighbourhoods by simulating the individual-level adoption and demand-supply rules. The results of these individual choice and market diffusion models should be helpful to urban planners, policy makers and developers in further exploration of strategies for investing and promoting smart growth.

Literature review

The literature review covered three topics: (1) residential location choice modelling; (2) preference for smart growth neighbourhoods and the factors that affect the preferences; and (3) the use of the residential location decision-making process in a bottom-up approach to urban growth modelling.

Residential location choice modelling

There are several utility-based choice models, such as MNL, NL, GEV and mixture logit model (Hoyos, 2010; Yang et al., 2013). The MNL assumes that individuals are homogenous and the relative odds of any two selection outcomes are irrelevant to the change in the rest of the alternatives. The violation of these assumptions can lead to biased estimation (Baltas and Doyle, 2001). In reality, an individual’s preferences vary with his or her socioeconomic status. The modification of one alternative can change the probability of the substitutes (i.e. substitute goods which can replace each other in use) as well as the relative odds of the substitutes being selected.

The MNL has been improved by allowing model parameters to vary randomly over individuals. However, it still does not well explain why individuals are different from each other (Boxall and Adamowicz, 2002). The NL is another alternative of the MNL, which groups those with similar alternatives into a nest and allows the alternatives within the same nest to share a similar choice sensitivity (Yang et al., 2013). The NL is useful for modelling individuals making decisions in a sequential fashion. For example, households may first decide whether to move or stay, then decide where to move (Lee and Waddell, 2010). However, the estimation of the NL is sensitive to the formation of nests (Yang et al., 2013). Within the same nest, the relative odds of any two selection outcomes remain invariant across individuals making the choice and despite a change in the rest of the alternatives. The GEV model allows for a more flexible substitution pattern among the alternatives (Bekhor and Prashker, 2008). But the GEV can still inadequately represent the variation in individual preferences (Baltas and Doyle, 2001).

The mixture or mixed logit model is a highly flexible DCM, which accounts for both substitution effects among alternatives and heterogeneous preference levels among individuals (Baltas and Doyle, 2001; Boxall and Adamowicz, 2002). The LC choice model used in this present research is a specific form of mixture logit models. In modelling residential location choice, the latent classes represent different segments in the population, each of which shows its own preferences and choice behaviour. Olaru et al. (2011) identified two classes with significant heterogeneity in the population of Perth, Western Australia in terms of their evaluations of transit-oriented development. Liao et al. (2014) identified two classes that show distinct preferences for compact development in the Wasatch Front, Utah. Both studies employed socioeconomic characteristics and personal attitudes to interpret the preference heterogeneity.

The estimation of the LC choice model requires data from a discrete choice experiment (DCE) (Birol et al., 2006). The DCE asks consumers to state their preferences for a set of hypothetical scenarios comprised of a set of attributes and different levels on these attributes (Kjaer, 2005). The advantage of the DCE is its ability to approximate an individual’s trade-offs or balancing between alternatives. Also the DCE allows the elicitation of preferences for attributes associated with similar products and guarantees these attributes are orthogonal for the LC choice model.

Considering the flexibility of the LC choice model and the availability of DCE results in the NAR survey, we adopted the LC choice model to evaluate US citizens’ preference for smart growth neighbourhoods.

Preference for smart growth neighbourhoods

Residential location choice modelling is a way of approximating the tradeoffs among a set of communities comprising different design attributes. Influential attributes that affect people’s choice of where to live include the type of houses (e.g. single-family house, apartment), locations (e.g. commute time to work) and amenities (e.g. quality of schools, access to food and public transportation) (Kiel and Zabel, 2008).

Preference for smart growth neighbourhoods varies among households depending on socioeconomic characteristics (e.g. lifestyle, life stage, housing tenure and education). Lifestyle refers to a certain desirable way of living and drives the decision of where to live. Walker and Li (2007) identified three lifestyle segments in Portland, Oregon in 1994, including suburban dwellers (43% of total households), urban dwellers (30% of total households) and transit-riders (27% of total households). Suburban dwellers prefer larger residences while transit-riders prefer lower travel time to work by transit. Life stage refers to the division of life into different stages, such as young and single, married, married with young children, married with grown up children and retired. In different stages, people exhibit different demands and preferences for living spaces and neighbourhood environments (Smith and Olaru, 2013). Households with children appreciate the value of green space and recreational opportunities, while those without children prefer locations with ease of access to services (Kim et al., 2005). Home ownership and education are the other two common indicators for preference heterogeneity. Renters tend to consider residences close to workplaces while owners focus more on the quality of the residence (Liao et al., 2014). Many well-educated households in recent years have chosen to live in multi-functional, high-density environments and larger cities (Frenkel et al., 2013).

Psychological factors are also included in the residential location choice model to explain the source of preference heterogeneity. These psychological factors account for individual perceptions and attitudes towards quality of life, neighbourhood and environment (van Putten et al., 2011). For example, households with high environmental awareness are more likely to choose more sustainable and compact houses (Rid and Profeta, 2011). One typical method of measuring individual attitudes is asking whether something (e.g. a mix of people from various income levels) is important when deciding where to live. People also often have distinct opinions about various aspects of community design, which can also help planners to understand an individual’s housing location choice behaviour. In this study, we refined and updated the US citizens’ preferences for smart growth neighbourhoods using the 2011 NAR community preference survey data. We explicitly interpreted the estimated preference and the corresponding behaviour of choosing smart growth neighbourhoods from the LC choice model.

Bottom-up urban growth modelling

Urban growth modelling is an important tool for understanding and managing the development of cities. One of these modelling techniques is called bottom-up simulation, which predicts urban growth as an emergent property from thousands of decisions and interactions (Filatova et al., 2009; Xie et al., 2007). These decisions include but are not limited to residential (re)location choice, traffic activities and new development (Waddell et al., 2003). These decision models are the foundation of the bottom-up modelling and they should properly reflect the change of behaviour in response to the environment change. In this study, we estimate a residential location choice model to predict individual adoption of smart growth neighbourhoods, comprising a certain set of design features. We also develop a market diffusion model which includes individual adoption of smart growth neighbourhoods and demand-supply rules. We show the market growth for smart growth neighbourhoods and how the designs of smart growth neighbourhoods can influence the emergent market diffusion.

Material and methods

Latent class choice model

In this study, a LC choice model is developed to predict the probability of an individual choosing a smart growth neighbourhood over a conventional sprawling community. In the LC choice model, the probability of individual i choosing alternative m equals the joint probability of the individual i belonging to class x and choosing alternative m. In the NAR survey, the DCE section presented seven choice sets to the respondents, with each set consisting of one sprawling community and a smart growth neighbourhood (Table 1). The design attributes involved in describing the community contain lot size and design, accessibility, commute time to work and the availability of public transportation. Individual respondents were asked to choose the preferred one out of the two alternatives in each set.

Table 1.

The seven choice sets in the DCE and the eighth choice set for evaluation in this study: each set has two alternatives, one is conventional sprawling community and the other one is smart growth neighbourhood; each set has different features that can distinguish the two community types.

Set ID	Conventional sprawling community	Smart growth neighbourhood
1	Houses are built far apart on larger lots and you have to drive to get to (schools, stores and restaurants/parks, playgrounds and recreation areas).	Houses are built close together on smaller lots and it is easy to walk to (schools, stores and restaurants/parks, playgrounds and recreation areas).
2	Houses are built far apart on larger lots and you have to drive to get to schools, stores and restaurants.	Houses are built close together on smaller lots and it is easy to walk to schools, stores and restaurants.
3	Houses are built far apart on larger lots and you have to drive to get to parks, playgrounds and recreation areas.	Houses are built close together on smaller lots and it is easy to walk to parks, playgrounds and recreation areas.
4	Houses are larger on larger lots, and you would have a longer commute to work, 40 minutes or more.	Houses are smaller on smaller lots, and you would have a shorter commute to work, 20 minutes or less.
5	The neighbourhood has houses only and you have to drive to stores and other businesses.	The neighbourhood has a mix of houses and stores and other businesses that are easy to walk to.
6	Own or rent a detached, single-family house, and have to drive to shops and restaurants and have a longer commute to work.	Own or rent an apartment or townhouse, and have an easy walk to shops and restaurants and have a shorter commute to work.
7	There are only single-family houses on large lots. There are no sidewalks. Places such as shopping, restaurants, a library and a school are within a few miles of your home and you have to drive to most. There is enough parking when you drive to local stores, restaurants and other places. Public transportation, such as bus, subway, light rail or commuter rail, is distant or unavailable.	There is a mix of single-family detached houses, townhouses, apartments and condominiums on various sized lots. Almost all of the streets have sidewalks. Places such as shopping, restaurants, a library and a school are within a few blocks of your home and you can either walk or drive. Parking is limited when you decide to drive to local stores, restaurants and other places. Public transportation, such as bus, subway, light rail or commuter rail, is nearby.
8	There are only single-family houses on large lots. There are no sidewalks. Places such as shopping, restaurants, a library and a school are within a few miles of your home and you have to drive to most. Places such as parks, playgrounds and recreational area are within a few miles of your home and you have to drive to most. There is enough parking when you drive to local stores, restaurants and other places. Public transportation, such as bus, subway, light rail or commuter rail, is distant or unavailable. Commute time to work is more than 40 minutes.	There is a mix of single-family detached houses, townhouses, apartments and condominiums on various sized lots. Almost all of the streets have sidewalks. Places such as shopping, restaurants, a library and a school are within a few blocks of your home and you can either walk or drive. Places such as parks, playgrounds and recreational areas are within a few blocks of your home and you can either walk or drive. Parking is limited when you decide to drive to local stores, restaurants and other places. Public transportation, such as bus, subway, light rail or commuter rail, is nearby. Commute time to work is less than 20 minutes.

In order to evaluate the impact of attitudes on choice behaviour, we applied a principal components analysis (PCA) using the SPSS software (IBM Corp., 2012) to extract a reduced number of interpretable principle components, or attribute metrics, from the original questions, ‘in deciding where to live, indicate how important each of the following would be to you: very important, somewhat important, not very important, or not at all important’. We extracted four principal components. The first component relates to the diversity of the neighbourhood environment, such as a mix of people, multiple housing types and green designs. The second component signifies the importance of big and sprawling lot design. The third component captures the weight of accessibility in deciding where to live. The fourth component concerns the privacy of the house. Together these four components explained around 75% of the total variation in responses to the original 16 questions. The scores on the four components for each respondent are included as variables (i.e. diverse neighbouhood environment, big and sprawling lot design, accessibility and privacy) in the LC choice modelling.

We used the Latent GOLD Choice 4.5 software for the estimation (Statistical Innovations Inc., 2012). The segmentation of classes was estimated according to the respondent’s socioeconomic and psychological variables (Appendix, Table A1). The choice behaviour of each class was estimated according to the probability formulation derived from RUT (Vermunt and Magidson, 2005). Accordingly, the conditional probability of the individual i choosing alternative m depending on class membership x has the form of a logistic probability (equation 1). The utility that individual i receives from alternative m depending on class membership x is determined by design attributes and the weights (equation 2). The estimation of the LC choice model is by means of Maximum Likelihood (ML) and the determination of optimal class number is based on Bayesian Information Criterion (BIC) and classification errors (Liao et al., 2014; van Putten et al., 2011). A smaller BIC indicates a better model fit with a smaller number of parameters to be estimated.

P (y_{it} = m | x) = \frac{\exp (η_{m | x}^{t})}{\sum_{m' = 1}^{M} \exp (η_{m' | x}^{t})}

(1)

where $η_{m | x}^{t}$ donates individual i’s utility associated with the alternative of m in choice set t, and:

η_{m | x, z_{t}}^{t} = β_{m x}^{c o n} + \sum_{p = 1}^{P} β_{x p}^{a t t} z_{t m p}^{a t t}

(2)

where $η_{m | x, z_{t}}^{t}$ donates the utility of alternative m in choice set t given individual i belonging to class x; $β_{mx}^{con}$ is the alternative-specific constant with the restriction $\sum_{m = 1}^{M} β_{mx}^{con} = 0$ ; $β_{xp}^{att}$ is the effect of attributes ( $z_{tmp}^{att}$ ). All the regression coefficients are class-specific.

Market diffusion model

We developed a dynamic diffusion model that predicts the rate of smart growth neighbourhood adoption in the housing market. The purpose of this diffusion model is to understand the impact of the potential market demand for smart growth neighbourhood living on the long-term land-use pattern. The diffusion model presents the interactive demand and supply process in the real estate market, following the idea of an agent-based housing market model developed by Lu et al. (Lu et al., 2013). We assume that there are two community types including one conventional sprawling community and one smart growth neighbourhood in the market. In year t, the probability ( $P_{t, s}$ ) of an individual home buyer purchasing a house in a smart growth neighbourhood is determined by equation (3):

\begin{matrix} P_{t, s} = \frac{f_{s} \exp (U_{s})}{f_{s} \exp (U_{s}) + f_{c} \exp (U_{c})} \\ = \frac{f_{s}}{f_{s} + f_{c} \times \frac{\exp (U_{c})}{\exp (U_{s})}} \\ = \frac{f_{s}}{f_{s} + f_{c} \times (\frac{\exp (U_{c}) + \exp (U_{s})}{\exp (U_{s})} - 1)} \\ = \frac{f_{s}}{f_{s} + f_{c} \times (\sum_{i = 1}^{4} P_{i} \frac{1}{P_{s | i}} - 1)} \end{matrix}

(3)

where, $f_{s}$ , $f_{c}$ is the market share of houses in smart growth neighbourhoods and conventional sprawling communities, respectively; $U_{s}$ , $U_{c}$ is the utility of living in houses locating in smart growth neighbourhoods and conventional sprawling communities; $P_{i}$ is the probability of the individual belonging to the class i (with a total of four classes from the LC choice model); and $P_{s | i}$ is the probability of choosing the smart growth neighbourhood given class i.

Accordingly, the sale of houses in smart growth neighbourhoods in year t ( $Sal e_{t, s}$ ) is calculated as follows (equation 4):

Sal e_{t, s} = min (N B_{t} \times P_{t, s}, S_{t, s})

(4)

where, $N B_{t}$ is the number of home buyers in year t; $S_{t, s}$ is the supply of houses belonging to smart growth neighbourhoods in year t.

Further, the projected demand for houses in smart growth neighbourhoods in year t+1 ( $D_{t + 1, s}$ ) is assumed to equal the $N B_{t} \times P_{t, s}$ . The supply of houses in smart growth neighbourhoods in year t+1 ( $S_{t + 1, s}$ ) is determined by equation (5):

S_{t + 1, s} = max (D_{t + 1, s}, S_{t, s} - Sal e_{t, s})

(5)

The starting conditions of the diffusion model include: (1) $N B_{t}$ constantly equals 1000 assuming the stable increase of home buyers; (2) the total supply of residential properties equals 1000 in year 0, for which 80% are built in conventional sprawling communities and 20% in smart growth neighbourhoods. The fraction of sold residential properties ( $f_{s}^{T}$ ) in smart growth neighbourhoods after year T is emergent as follows (equation 6):

f_{s}^{T} = \frac{\sum_{t = 1}^{T} Sal e_{t, s}}{\sum_{t = 1}^{T} Sal e_{t, s} + \sum_{t = 1}^{T} Sal e_{t, c}}

(6)

The $f_{s}^{T}$ value shows the time-dependent diffusion curves (i.e. adoption rate) of residential properties in smart growth neighbourhoods. In other words, the $f_{s}^{T}$ value and corresponding diffusion curve indicate the progress of land development following smart growth principles. We use the results for $f_{s}^{T}$ to assess whether the market potential for smart growth neighbourhoods can contribute significantly to less sprawling land development.

Results

LC choice model

We selected a four-class choice model as the optimal estimation because it had the smallest BIC and only 10% classification errors. The estimated class membership parameters and utility function parameters are shown in Table 2. All the coefficients in the utility function are significant at the 5% level. Also the differences of coefficients are significant across the four classes at the 5% level.

Table 2.

Four-class choice model estimates for community attributes, socioeconomic and attitudinal variables.

Model for choices
	‘Likely sprawling’	‘Conditionally sprawling’	‘Conditionally compact’	‘Likely compact’	Overall
R²(0)	0.70	0.33	0.17	0.98	0.59
Community attributes^a
Lot design	3.25	2.62	−0.63	−6.51
Accessibility	−0.41	0.21	0.25	0.87
Commute time to work	5.71	6.48	−1.72	1.00
Public transportation	11.41	10.97	−2.17	1.00
Model for classes
Respondent’s socioeconomic and attitudinal variables					p-value
Intercept	0.50	−0.47	0.11	−0.13	0.63
Gender	0.02	0.06	0.03	−0.11	0.78
Age	−0.03	0.04	0.02	−0.04	0.66
Region	0.12	−0.03	−0.03	−0.06	0.10
Education	−0.16	−0.07	0.06	0.17	0.01
Income	−0.06	0.04	0.07	−0.04	0.25
Home ownership	−0.64	0.25	0.05	0.33	0.00
Current residency	0.33	−0.03	−0.15	−0.15	0.00
Marital status	−0.66	0.40	0.15	0.11	0.00
Children under 18	0.14	0.12	−0.12	−0.13	0.34
Employed	0.99	0.08	0.13	−1.20	0.00
Commute mode	−0.84	0.27	−0.11	0.68	0.01
Head of household	0.58	−0.35	0.01	−0.24	0.01
Diverse neighbourhood environment	−1.01	0.17	0.21	0.64	0.00
Big and sprawling lot design	0.38	0.21	−0.02	−0.58	0.00
Accessibility	−1.15	0.22	0.23	0.70	0.00
Privacy	1.55	0.03	−0.38	−1.20	0.00

Notes:

All the coefficients are significant at the 5% level; also the coefficients are significantly different across the four classes.

We named the four classes as ‘likely sprawling’, ‘conditionally sprawling’, ‘conditionally compact’ and ‘likely compact’ according to the positive and negative impacts of design attributes on the utility value. In detail, for the ‘likely sprawling’ class, the coefficient of lot size and design is positive which means the utility increases as the lot size and design tends to be sprawling. In contrast, the negative effect of accessibility indicates the decline of the utility as the community gets close to recreational and commercial areas. Consequently, the ‘likely sprawling’ class mostly chooses a conventional sprawling community. However, if the commute to work becomes less than 20 minutes and public transit is available, there is a slight increase in the probability of choosing a smart growth neighbourhood for the ‘likely sprawling’ class. For the ‘conditionally sprawling’ class, both the lot size and design, and accessibility variables are positive. But the weight of accessibility is much smaller than the lot size and design given the same scaling of the value. Generally, this class has a relatively higher probability of choosing to live in a conventional sprawling community. A commute to work of less than 20 minutes and the availability of public transit can increase the adoption of a smart growth neighbourhood made by the ‘conditionally sprawling’ class. For the ‘conditionally compact’ and ‘likely compact’ classes, the coefficients of lot size and design are negative, which shows the decrease in utility as the community becomes sprawling. The coefficients of accessibility are positive, indicating the increase in utility as the community has easy access to the areas for recreational and commercial activities. Accordingly, these two classes tend to choose smart growth communities. In contrast, the ‘conditionally compact’ class is more likely to live far away from workplaces and public transit service.

Of the total respondents, 32% belong to the ‘likely sprawling’, 26% belong to the ‘conditionally sprawling’, 23% belong to the ‘conditionally compact’ and 19% belong to the ‘likely compact’ class. Table 2 shows the results of the effects of individual’s socioeconomic and psychological characteristics on class membership. Among these variables, the level of education shows a significant impact at the 5% level. Those who have a higher level of education have a higher probability of belonging to the ‘likely compact’ class, while people with a lower level of education tend to belong to the ‘likely sprawling’ class. Employment is another significant factor that influences the class membership. People who are employed are less likely to be in the ‘likely compact’ class than are unemployed people. Other significant socioeconomic factors include home ownership, head of household, commute mode, marital status and current residency (Appendix, Table A1). The impact of these variables on class membership can be understood according to the positive and negative contribution to the likelihood of the four classes.

The four attitudinal variables from PCA are included in the class segmentation model in order to understand the impact of personal attitudes on choice behaviour. The p-value indicates the effects of all the four attitudinal variables are explicit and statistically significant (see Table 2). Holding other variables constant, if a diverse neighbourhood environment is not seen by an individual as important, then we expect that there is more than 50% probability he or she belongs to the ‘likely sprawling’ class. If a person feels the big and sprawling lot design variable to be very important in deciding where to live, there is less than a 10% probability that he or she is a member of the ‘likely compact’ class. The effects of the other two attitudes associated with accessibility and privacy are illustrated in Figure 1.

Figure 1.

The impact of individual’s four attitudinal variables on class membership.

The estimated LC choice model can be applied to evaluate the proposed smart growth neighbourhood. For illustration, we created the 8th choice set, in which two more amenities including commute time less than 20 min and easy access to recreational opportunities were added to the features of the smart growth neighbourhood in the 7th choice set presented in the NAR survey (Table 1). We predicted that there is an 83% probability of an individual home buyer choosing the proposed smart growth neighbourhood in the 8th choice set (Figure 2). In contrast, only 56% of the total choose the smart growth neighbourhood in the 7th choice set. In the 8th choice set, around 90% of the smart growth neighbourhood selectors are in the ‘likely compact’, ‘conditionally sprawling’ and ‘likely sprawling’ classes. There are still 17% of home buyers choosing sprawling community and almost 100% of them belong to the ‘conditionally compact’ class. The reason for not choosing the smart growth neighbourhood is due to the negative impact of being close to workplaces and public transit. Thus, we expect that the potential market demand for smart growth neighbourhoods is influenced by the design details put forward by local urban planners and real estate developers. By adding more amenities, our results suggest that we may be able to achieve a higher-level market potential for smart growth.

Figure 2.

The estimated probability of an individual choosing smart growth neighbourhood over conventional sprawling community giving the 8th choice set (Table 1).

Diffusion of smart growth neighbourhoods

Building on the above-described diffusion modelling, we evaluated the impact of the potential demand for smart growth neighbourhood living on the long-term land-use pattern. We selected the 7th and 8th choice set (Table 1) for the comparison. In the 7th choice set, there is 56% probability of an individual home buyer choosing the smart growth neighbourhood. In other words, assuming the same probability of considering houses in a smart growth neighbourhood and a conventional sprawling community, 56% of home buyers choose the smart growth neighbourhood. Thus, the percent of purchased houses that belong to smart growth neighbourhoods reaches 56% at market equilibrium. However, we fail to observe the same market share according to the diffusion curve. In contrast, we see the decline of market share for smart growth neighbourhoods (Figure 3). This result appears to be caused, at least in part, by the market inefficiency in providing sufficient smart growth neighbourhoods (only 20% of properties belonging to smart growth neighbourhoods) and hence limiting the consideration of houses in such neighbourhoods. This market inefficiency might be overcome to some degree once there is a stronger demand for smart growth neighbourhood living. We observe an increasing adoption rate of smart growth neighbourhood living from the 8th choice set, where there is an 83% probability of an individual choosing a smart growth neighbourhood when he/she compares the two community types (Figure 3). The housing supply limitation is overcome in this case because of the high demand for the proposed smart growth neighbourhood. This leads us to conclude that, as we might expect, market potential is an important variable in driving a movement towards smart growth development, and that this market potential is in turn influenced by the way smart growth neighbourhoods are planned and built.

Figure 3.

Market diffusion patterns of the two designs of smart growth neighbourhoods in the 7th and 8th choice set (Table 1).

Discussion

In this study we have applied latent class choice analysis to data from a NAR 2011 community preference survey. We identified four classes of respondents, termed ‘likely sprawling’, ‘conditionally sprawling’, ‘conditionally compact’ and ‘likely compact’, where class membership is based on an individual’s socioeconomic and attitudinal characteristics. The corresponding behaviour of choosing a smart growth neighbourhood to live in was also estimated according to a set of class-specific utility functions. Compared with previous studies, our analysis demonstrates a more comprehensive classification of US citizens that have distinctive preferences for smart growth. At the regional level (i.e. Northeast, Midwest, South Atlantic, Inland South and West), the distribution pattern of the four classes does not show a significant difference (p = 0.10). However, the distribution of the four classes may vary across individual metropolitan areas because of differences in local-specific economic (e.g. unemployment rate, housing and transportation cost) and social (e.g. environmental awareness) conditions. On average, Atlanta residents are less interested in transit-oriented and walkable neighbourhoods than those of Boston (Levine et al., 2005), which implies a higher percentage of ‘likely sprawling’ class and a lower percentage of ‘likely compact’ class in Atlanta than Boston.

The primary limitation of using the NAR data is the lack of understanding of the market effect on the adoption of smart growth neighbourhoods. Even asking for the willingness to pay for a particular property does not objectively reflect the deal price. The optimal solution is to combine the data of market sales and surveys to assess the price effect and preferences for smart growth neighbourhoods (Phaneuf et al., 2013). Here, we assume that the impacts of market price and community attributes on the adoption are independent. As a result, the price effect is included as the unobserved utility in the error term. Overall, we believe our analysis, built upon the 2011 NAR community preference survey, offers some useful quantitative insights for urban planners, policy makers and real estate developers looking to promote smart growth.

Another limitation in this study is the lack of the understanding of the impact of K-12 school quality on community choice (Kiel and Zabel, 2008). The DCE does not include school quality as an important variable, despite the fact that 45% of the respondents in the NAR survey stated the quality of K-12 schools is very important in deciding where to live. In reality, the issue of the uneven school quality distribution between urban and suburban areas has been raised since land development became sprawling (Gould, 1969). Although the urban area has a higher incentive for smart growth than suburban areas, the poor schools in urban areas might hinder the economic prosperity resulting from smart growth. Therefore, policies that help improve the quality of school systems in urban areas may prove essential to the success of smart growth.

The development of the LC choice model allows the estimation of individual-level adoption of smart growth neighbourhood living compared with locating residence in a conventional sprawling community. In this study, we find that the high market potential can be achieved by adding more amenities (e.g. shorter commute time to work, easy access to recreational places). Meanwhile, we point out that not 100% smart growth neighbourhoods can be built as described in our most appealing (the 8th) choice set. One reason is the shortage of land supply, in particular, in the area close to commercial business districts. The shortage of smart growth neighbourhoods as described in our 8th choice set can lead to a higher sale price, which may induce the substitution with the conventional sprawling community. The average market potential should be less, therefore, than our prediction based on the 8th choice set.

Land development is a complex and adaptive process, which is far more sophisticated than the mechanism of the diffusion model (Kim, 2011). In the diffusion model, we only compare two neighbourhood types which cannot represent 100% of neighbourhood types in reality. We also simplify the complex process involving the planning, location, design and financing of smart growth. Therefore, the diffusion model is not intended to predict the market share of smart growth neighbourhoods precisely. In fact, the emerging systematic approach that accounts for the complexity of land development is indeed necessary to plan smart growth that can ultimately meet people’s preferences and needs (Nasiri et al., 2013). Part of the reason for the failure of smart growth initiatives has been the lack of systematic planning approaches for smart growth. Systematic planning should allow urban planners to figure out the optimal solution to integrating smart growth into the current land and infrastructure configuration at different levels of geographic coverage from the micro neighbourhood up to the macro city, allowing the location and design of smart growth neighbourhoods to be better optimised and customised.

Conclusion

This study contributes to the limited understanding on heterogeneous preference and behaviour of Americans choosing to live in smart growth neighbourhoods. The results indicate that there is considerable heterogeneity in preference and the corresponding choice of where to live by Americans. The impacts of both socioeconomic and attitudinal characteristics are found to be significant in neighbourhood choice. The analysis of heterogeneous preference and choice behaviour allowed us to demonstrate the potential market demand for smart growth neighbourhoods. According to our diffusion model, the market potential for the smart growth neighbourhood turns out to have a significant impact on land development. Thus, both heterogeneous choice behaviours and market potential should be considered when designing the form and location of smart growth neighbourhoods. Lastly, a more systematic approach to planning for smart growth is needed in order to effectively realise this market potential.

Footnotes

Appendix

Table A1.

The attributes and levels of community design and socioeconomic features used in the DCE and LC choice model.

	Item	Level of the attributes	Value assigned
Design attribute	Lot design	House	0
		Large lot	1
		Large lot, large house	2
		Single-family house	3
		Large lot, single-family house	4
	Accessibility	Housing only	0
		Close to recreational sites	1
		Close to school, stores, etc.	2
		Close to recreational sites and school, stores, etc.	3
		Mixed land use	4
	Commute to work	More than 40 minutes	0
		Less than 20 minutes	1
	Public transportation	Drive only	0
		Public transit is nearby	1
Socioeconomic attributes	Gender	Male	0
		Female	1
	Age	18–29	1
		30–39	2
		40–49	3
		50–59	4
		60+	5
	Region	Northeast	1
		Midwest	2
		South Atlantic	3
		Inland South	4
		West	5
	Education	<HS/HS	1
		Some college	2
		College graduate	3
		Post graduate	4
	Income	<US$25K	1
		US$25–50K	2
		US$50–75K	3
		US$75–100K	4
		US$100K+	5
	Home ownership	Own	0
		Rent	1
	Current residency	City	1
		Suburban – mixed	2
		Suburban – housing only	3
		Small town/rural	4
	Marital status	Married	1
		Single	2
		Divorced/separated/widowed	3
	Children under 18	No	0
		Yes	1
	Employed	No	0
		Yes	1
	Commute	By car	0
		Other mode	1

Acknowledgements

The authors wish to thank the anonymous reviewers for their valuable comments and Joseph Molinaro and National Association of Realtors for the survey data.

Funding

This research was supported by the Brook Byers Institute for Sustainable Systems, Hightower Chair, and the Georgia Research Alliance at the Georgia Institute of Technology. The authors are thankful for a grant (#0836046) from the National Science Foundation programme for Emerging Frontiers in Research and Innovation (EFRI).

References

Baltas

Doyle

(2001) Random utility models in marketing research: A survey. Journal of Business Research51(2): 115–125.

Bekhor

Prashker

(2008) GEV-based destination choice models that account for unobserved similarities among alternatives. Transportation Research Part B – Methodological42(3): 243–262.

Birol

Karousakis

Koundouri

(2006) Using a choice experiment to account for preference heterogeneity in wetland attributes: The case of Cheimaditida wetland in Greece. Ecological Economics60(1): 145–156.

Boxall

Adamowicz

(2002) Understanding heterogeneous preferences in random utility models: A latent class approach. Environmental & Resource Economics23(4): 421–446.

Brown

Southworth

(2008) Mitigating climate change through green buildings and smart growth. Environment and Planning A40(3): 653–675.

Bruegmann

(2005) Sprawling: A Compact History. Chicago, IL: University of Chicago Press.

Environmental Protection Agency (2012a) Residential construction trends in America’s metropolitan regions: 2012 edition. Available at: http://www.epa.gov/dced/construction_trends.htm.

Environmental Protection Agency (2012b) Smart growth and economic success: benefits for real estate developers, investors, businesses, and local governments. Available at: http://www.epa.gov/dced/pdf/economic_success.pdf.

Ewing

(1997) Is Los Angeles-style sprawl desirable?Journal of the American Planning Association63(1): 107–126.

10.

Ewing

Cervero

(2010) Travel and the built environment. Journal of the American Planning Association76(3): 265–294.

11.

Filatova

Parker

van der Veen

(2009) Agent-based urban land markets: Agent’s pricing behavior, land prices and urban land use change. Jasss – the Journal of Artificial Societies and Social Simulation12(1): 29.

12.

Frenkel

Bendit

Kaplan

(2013) Residential location choice of knowledge-workers: The role of amenities, workplace and lifestyle. Cities35: 33–41.

13.

Giles-Corti

Bull

Knuiman

. (2013) The influence of urban design on neighbourhood walking following residential relocation: Longitudinal results from the RESIDE study. Social Science & Medicine77: 20–30.

14.

Gonzalez

(2005) Urban sprawl, global warming and the limits of ecological modernisation. Environmental Politics14(3): 344–362.

15.

Gould

(1969) Quality of inequality – Urban and suburban public schools – Daly, CU. Cornell Law Review55(1): 145–151.

16.

Greene

Hensher

(2003) A latent class model for discrete choice analysis: Contrasts with mixed logit. Transportation Research Part B – Methodological37(8): 681–698.

17.

Hoyos

(2010) The state of the art of environmental valuation with discrete choice experiments. Ecological Economics69(8): 1595–1603.

18.

IBM Corp. (2012) IBM SPSS Statistics for Windows, Version 21.0. Armonk, NY: IBM Corp. Available at: http://www-01.ibm.com/software/analytics/spss/.

19.

Jabareen

(2006) Sustainable urban forms – Their typologies, models, and concepts. Journal of Planning Education and Research26(1): 38–52.

20.

Johnson

(2001) Environmental impacts of urban sprawl: A survey of the literature and proposed research agenda. Environment and Planning A33(4): 717–735.

21.

Kiel

Zabel

(2008) Location, location, location: The 3L Approach to house price determination. Journal of Housing Economics17(2): 175–190.

22.

Kim

(2011) Linking land use planning and regulation to economic development: A literature review. Journal of Planning Literature26(1): 35–47.

23.

Kim

Horner

Marans

(2005) Life cycle and environmental factors in selecting residential and job locations. Housing Studies20(3): 457–473.

24.

Kjaer

(2005) A review of the discrete choice experiment with emphasis on its application in health care. University of Southern Denmark. Available at: http://static.sdu.dk/mediafiles/Files/Om_SDU/Centre/c_ist_sundoke/Forskningsdokumenter/publications/Working%20papers/20051pdf.pdf.

25.

Lee

BHY

Waddell

(2010) Residential mobility and location choice: A nested logit model with sampling of alternatives. Transportation37(4): 587–601.

26.

Levine

Inam

Torng

(2005) A choice-based rationale for land use and transportation alternatives – Evidence from Boston and Atlanta. Journal of Planning Education and Research24(3): 317–330.

27.

Lewis

Baldassare

(2010) The complexity of public attitudes toward compact development survey evidence from five states. Journal of the American Planning Association76(2): 219–237.

28.

Liao

Farber

Ewing

(2014) Compact development and preference heterogeneity in residential location choice behaviour: A latent class analysis. Urban Studies DOI: 10.1177/0042098014527138.

29.

Litman

(2009) Where we want to be: Home location preferences and their implications for smart growth. The Congress for New Urbanism Transportation Summit. 4November 2009, Portland, Oregon. Available at: http://www.vtpi.org/sgcp.pdf.

30.

Noonan

Crittenden

. (2013) Use of impact fees to incentivize low-impact development and promote compact growth. Environmental Science & Technology47(19): 10,744–10,752.

31.

McFadden

(1986) The choice theory approach to market research. Marketing Science5(4): 275–297.

32.

Magidson

Vermunt

(2002) Nontechnical introduction to latent class models. Available at: http://www.statisticalinnovations.com/articles/lcmodels2.pdf.

33.

Martins

(2012) Urban compaction or dispersion? An air quality modelling study. Atmospheric Environment54: 60–72.

34.

Nasiri

Savage

Wang

. (2013) A system dynamics approach for urban water reuse planning: A case study from the Great Lakes region. Stochastic Environmental Research and Risk Assessment27(3): 675–691.

35.

National Assoication of Realtors (2011) The 2011 Community Preference Survey, What Americans are Looking for When Deciding Where to Live. Available at: http://www.realtor.org/articles/nar-2013-community-preference-survey.

36.

Olaru

Smith

Taplin

JHE

(2011) Residential location and transit-oriented development in a new rail corridor. Transportation Research Part A – Policy and Practice45(3): 219–237.

37.

Phaneuf

Taylor

Braden

(2013) Combining revealed and stated preference data to estimate preferences for residential amenities: A GMM approach. Land Economics89(1): 30–52.

38.

Rid

Profeta

(2011) Stated preferences for sustainable housing development in Germany – A latent class analysis. Journal of Planning Education and Research31(1): 26–46.

39.

Smith

Olaru

(2013) Lifecycle stages and residential location choice in the presence of latent preference heterogeneity. Environment and Planning A45(10): 2495–2514.

40.

Statistical Innovations Inc. (2012) Latent GOLD Choice. 4.5. Available at: http://statisticalinnovations.com/products/lgchoice.html.

41.

Stone

Hess

Frumkin

(2010) Urban form and extreme heat events: Are sprawling cities more vulnerable to climate change than compact cities?Environmental Health Perspectives118(10): 1425–1428.

42.

van Putten

Jennings

Louviere

. (2011) Tasmanian landowner preferences for conservation incentive programs: A latent class approach. Journal of Environmental Management92(10): 2647–2656.

43.

Vermunt

Magidson

(2005) Technical guide for Latent Gold Choice 4.0: basic and advanced. Belmont Massachusetts: Statistical Innovations Inc, 11–12. Available at: http://www.statisticalinnovations.com/products/choicemanual.pdf.

44.

Waddell

Borning

Noth

. (2003) Microsimulation of urban development and location choices: Design and implementation of UrbanSim. Networks & Spatial Economics3(1): 43–67.

45.

Walker

(2007) Latent lifestyle preferences and household location decisions. Journal of Geographical Systems9(1): 77–101.

46.

Xie

Batty

Zhao

(2007) Simulating emergent urban form using agent-based modeling: Desakota in the Suzhou-Wuxian region in China. Annals of the Association of American Geographers97(3): 477–495.

47.

Yang

Zheng

Zhu

(2013) Cross-nested logit model for the joint choice of residential location, travel mode, and departure time. Habitat International38: 157–166.