An integrated framework for managing the complex interdependence between infrastructures and the socioeconomic environment: An application in metropolitan Atlanta

Abstract

Chinese

Keywords

agent-based model compact development complexity latent-class choice model preference sustainable urban infrastructure

Introduction

Urban systems are both complex and adaptive (Baynes, 2009), and the development of a more sustainable urban infrastructure requires both an understanding and an ability to manage this complexity (Pandit et al., 2015). Complexity results from the millions of decisions and interactions of diverse adaptive entities (i.e. citizens, firms, developers and governments). These decisions and interactions drive the dynamic and evolving interdependence between the urban physical infrastructure and the socioeconomic environment through which it operates. This interdependence leads to the emergence of, among other things, specific land-use arrangements, quality of life issues and carbon footprints. To manage this complexity, we should first start with a better understanding of people’s preferences and demands for more sustainable infrastructure designs. If we provide a suitable combination of more sustainable features and develop suitable policies to meet their preferences, we should be able to increase people’s adoption of more sustainable physical infrastructures.

In this study, we introduce an integrated framework for managing the complex interdependence between urban infrastructures and the socioeconomic environment within which it evolves, in pursuit of sustainable and environmentally cleaner urban living (see Figure 1). The framework starts with a preference-based demand analysis for amenities served by more sustainable infrastructures. The first step is citizen engagement, which involves the development of a survey that asks individual residents to make a choice from a set of options. We selected Mechanical Turk, a crowd-source platform operated by Amazon (https://www.mturk.com/mturk/welcome) for this citizen engagement and data collection. Previous studies that compared Mechanical Turk with laboratory experiments and traditional web studies have shown that the self-selection bias of the sample (i.e. people who are not interested in the topic of the survey may not answer the survey) is smaller using Mechanical Turk than traditional web-based surveys (Paolacci et al., 2010). The data quality from Mechanical Turk has also been shown to be comparable with that from laboratory experiments (Amir et al., 2012; Crump et al., 2013). The second step is the development of a choice model that estimates the impact of different amenities on people’s choice using both the choice responses and socioeconomic data collected for each respondent. Specifically, we choose a latent-class choice modelling technique that accounts for the heterogeneity in preference (Liao et al., 2015; Lu et al., 2015; Rid and Profeta, 2011). In the third step, we feed the results of this latent-class choice model into an agent-based model that allows us to simulate homebuyers, developers and the interactions between them (Chu et al., 2009; Parker et al., 2003; Schwarz and Ernst, 2009). With the agent-based model, we predict the adoption of more sustainable infrastructures and the emerging urban growth pattern. Moreover, we can explore possible policy incentives to drive the adoption of more sustainable development. The fourth and final step in the modelling process is to assess the results of adopting land-use policies that support environmentally less damaging, more sustainable urban infrastructures.

Figure 1.

The framework for managing the complex interdependence between infrastructures and the socioeconomic environment.

To both test and also explain this integrated framework in more detail, we used metropolitan Atlanta as a case study. During the past half a century, low-density residential development (notably single-family and semi-detached houses in suburban areas) has dominated the land-use pattern of metropolitan Atlanta. In 2014, Atlanta was, according to at least one study, the most sprawling big metropolitan region (i.e. regions over 1 million in population) in the USA (Ewing and Hamidi, 2014). Although low-density development provides comfortable living spaces, the negative impacts of such low-density development are becoming more obvious and serious (Bereitschaft and Debbage, 2013; Stone et al., 2010; Zhao et al., 2010), encouraging longer driving distances for work, food, shopping and entertainment (Cervero and Murakami, 2010). Greater material and energy investments are also required to support transportation, water and utility infrastructures. These needs come at the same time that significant resources are needed to address the challenge of aging urban infrastructures (e.g. water/wastewater networks, gas pipelines and highways). To make matter worse, property tax revenues are often insufficient to maintain the quality of infrastructures in the low-density development areas (Katz, 2013).

To solve these urban sprawl-exacerbated issues, alternative urban growth patterns are required. Two often-referenced growth patterns, and the ones explored in this present research effort, are termed low-impact development (LID) and transit-oriented development (TOD). LID is an alternative green strategy for conventional stormwater management (Chau, 2009; Cook, 2007). LID strategies include rainwater harvesting, greywater reclamation, bio-retention facilities, rain gardens, vegetated rooftops, rain barrels and permeable pavements. LID not only has the ability to control stormwater runoff and to provide non-potable water, it also provides additional amenities including the creation of green space, reduced heat island effect, improved air and water quality (Roy et al., 2008). When well-designed, these amenities increase walkability, livability, property value and taxes (Schmitz, 2008; Williams and Wise, 2009). The increased tax revenues make LID more economically viable (Raucher, 2009; Schmitz, 2008). Transit-oriented development (TOD) is the creation of compact, walkable, mixed-use communities centred on high quality public transit services (Litman, 2012). It can provide a high quality of life and reduce automobile use and fuel consumption (Suzuki et al., 2013). In this study, we examined the potential for LID and TOD as an integrated alternative solution to support more sustainable water, transportation and land-use patterns for metro Atlanta. Using our framework, we first determined residents’ preferences for LID and TOD, and how their preferences relate to socioeconomic characteristics. There has been a wealth of literature on analysing people’s preference for both LID and TOD separately (Bowman et al., 2012; Lund, 2006; Olaru et al., 2011; Olorunkiya et al., 2013). In this study we treat LID and TOD as an integrated solution to support more compact urban development, and use our agent-based modelling results in the third step to explore the sort of policies that might encourage higher compact development adoption rates. Finally, we estimate the reduction in environmental impacts as an example of a sustainability assessment of more compact development.

Materials and methods

Following the four-step process introduced above, this section of the paper describes each step in turn, i.e.: (1) survey design and data collection for measuring residential preferences; (2) latent-class residential community choice modelling; (3) scenario exploration and policy analysis using an agent-based market diffusion model; and (4) an example environmental impact assessment, in the form of a carbon footprint calculation, of more compact development.

Survey design and data collection for measuring preference

We conducted a survey to measure people’s preference for LID and TOD. Our survey includes four parts: (1) demand for housing square footage, number of bedrooms, number of bathrooms and the affordability (i.e. either price or rent); (2) personal attitudes on LID and TOD; (3) discrete choice experiment (DCE); (4) respondent’s socioeconomic statistics (see Supporting Information (SI) for Design of the Survey). The DCE includes 14 choice sets generated using JMP 10 (SAS Institute Inc., Cary, NC, USA) (SAS, 2012). In each choice set, there are four options comprising different levels of community amenities and price (Table 1). Respondents were asked to choose the most and least desirable communities for each set. Our survey follows the structure of the community preference survey developed by the US National Association of Realtors.

Table 1.

The variable definition and levels of amenities in the DCE and choice model.

The community amenities
Variables	Definition	Value
Categorical variable		Level of attributes
House design	House design	Single-family: single-family house, 2 storeys, 5000 ft² lot size, 2400 ft² habitable area
		Semi-detached: semi-detached house, 2 storeys, 3000 ft² lot size, 2400 ft² habitable area
		Row house: row house, 3 storeys, 1800 ft² habitable area
		4-storey: 4-storey stacked flats elevator & corridor access around 1-level parking podium, 1200 ft² habitable area
		8-storey: 8-storey midrise stacked flats over 2-level parking podium, 1200 ft² habitable area
		16-storey: 16-storey, 160 ft high high-rise stacked flats over 3- to 4-level parking podium, 1200 ft² housing square footage
Green building	Green building design	None
		RWH: rainwater harvesting for single-family and semi-detached houses
		GRWR: green roof and wastewater reclamation for row house and 4- to 16-storey flats
Numerical variable	Level of attributes	Values to run the estimation of choice model
Public LID	Rain garden, porous pavement and native vegetation for public open space	No	0
Public LID		Yes for public green space	1
Private LID	Rain garden, porous pavement and native vegetation for private yard	No	0
Private LID		Yes for private yard	1
Accessibility	Accessibility to food, shopping and entertainment	Driving in 30 minutes	0
		Biking in 20 minutes	0.5
		Walking in 10 minutes	1
Green space	Size of public green space	A quarter acre	0.25
		Half an acre	0.5
		One acre	1
		Five acres	5
Commute	Commute mode and time to work/school	Car only, more than 40 min	0
		Car only, within 40 min	1
		Public transportation (e.g. transit, bus) within 40 min	2
		Public transportation (e.g. transit, bus) within 30 min	3
		Biking within 20 min	4
		Walking within 15 min	5
School quality	Great school rating	Low	0
		Middle	0.5
		High	1
Transit	Walking distance to transit station	Less than 10 minutes	1
		Less than 20 minutes	0.5
		More than 20 minutes	0
Increase in price	Increase in price (or rent) as compared with the budget (%)	−20	−0.2
		−10	−0.1
		0	0
		+10	0.1
		+20	0.2

We designed our survey on SurveyMonkey (https://www.surveymonkey.com/). The SurveyMonkey generated a web link that allows people to open the survey. We published a task containing the link of the survey on Mechanical Turk. In fact, Mechanical Turk provides the interface to design the survey or other tasks (e.g. translations from the images). But it requires adequate knowledge of HTML5 for complex survey design. In this study, Mechanical Turk only played its role as a cost-effective solution for collecting survey responses in a short time period. People on Mechanical Turk can find our task of completing the survey immediately on the top of the task list when we published the task. Because we focused on residents in the metro Atlanta, respondents were asked to report the county and the zip code of where they are currently living. The respondents reporting an incorrect zip code were excluded in this study. We paid US$1 for each respondent who followed the instructions and completed the survey. The US$1 was determined through several pilot tests. As we increased the payment, we got responses more quickly. Also the quality of responses was improved as people spent more minutes on the survey.

Latent-class residential community choice model

In the DCE section of the survey, respondents were asked to provide the best and worst choices in each choice set. As compared with a traditional best choice experiment, the best–worst choice experiment significantly enhances data quantity/quality without much extra response effort. The best–worst choice data also have a technical advantage that provides a reference attribute level (i.e. a certain level of one attribute that occurs most often in all worst choices) that allows comparison of utilities across all levels of all attributes (Marley and Louviere, 2005; Najafzadeh et al., 2012).

The modelling of best–worst choices is built on a sequential choice process, which assumes people first select the best option out of the choice set and second select the worst option out of the remaining alternatives. The conditional probability of the individual i choosing alternative m as the best option depending on class membership x is determined by equation (1).

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

(1)

where, $(U_{m' | x}^{t})$ denotes the utility associated with the alternative m" in choice set t of individual i belonging to class x.

The conditional probability of the individual i choosing alternative m as the worst option depending on class membership x is determined by equation (2).

P (y_{i t} = m | x) = \frac{\exp (- U_{m | x}^{t})}{\sum_{m' \in A_{i t}} \exp (- U_{m' | x}^{t})}

(2)

where, $A_{i t}$ are the set of remaining alternatives of individual i excluding the best option in choice set t.

The value of $(U_{m' | x}^{t})$ in equations (1) and (2) is determined by the variables in equation (3). $(U_{m' | x}^{t})$ is a hypothetical internal state that explains the relationships between the best–worst choices and community amenities. The definition of the variables associated with community amenities and levels are listed in Table 1. All the coefficients are class-specific. The modelling of class membership is built on individual’s socioeconomic characteristics and personal attitudes on LID and TOD (Table S1, Supplementary material, available online). We used Latent Gold Choice 5.0 Software (Statistical Innovations Inc., Belmont, MA, USA) and Maximum Likelihood (ML) method to estimate the segmentation of classes and the class-specific best–worst choice model (i.e. the latent-class residential community choice model) (Vermunt and Magidson, 2012).

\begin{array}{l} U_{m | x} = β_{0, x} + β_{1, x} \times House Design + β_{2, x} \times Increase in Price + β_{3, x} \times Green Building + \\ β_{4, x} \times Accessibility + β_{5, x} \times Green space + β_{6, x} \times Public LID + β_{7, x} \times Privtae LID \\ + β_{8, x} \times Schoolquality + β_{9, x} \times Commute + β_{10, x} \times Transit \end{array}

(3)

Scenario exploration and policy analysis using an agent-based market diffusion model

The purpose of an agent-based market diffusion model was to predict the adoption of more compact development (using an apartment community as an example of multi-family designs, Figure S1, Supplementary material, available online) in the metro Atlanta in the next 20 years from 2014 to 2034. The probability of homebuyers choosing new apartment homes is determined by equation (4).

\begin{array}{l} P_{a p t, t} = \frac{f_{a p t, t} \exp (U_{a p t})}{f_{a p t, t} \exp (U_{a p t}) + f_{s i g, t} \exp (U_{s i g})} \\ = \frac{f_{a p t, t}}{f_{a p t, t} + f_{s i g, t} \times \frac{\exp (U_{s i g})}{\exp (U_{a p t})}} \\ = \frac{f_{a p t, t}}{f_{a p t, t} + f_{s i g, t} \times (\frac{\exp (U_{s i g}) + \exp (U_{a p t})}{\exp (U_{a p t})} - 1)} \\ = \frac{f_{a p t, t}}{f_{a p t, t} + f_{s i g, t} \times (\sum_{i = 1}^{4} P_{i} \frac{1}{P_{a p t | i}} - 1)} \end{array}

(4)

where, $f_{a p t, t}$ , $f_{s i g, t}$ is the market share of new apartment homes and new single-family houses in year t, respectively; $U_{a p t}$ , $U_{s i g}$ is the utility of living in apartment homes and single-family houses; $P_{i}$ is the probability of the homebuyer belonging to the class i (with classes from the latent-class residential community choice model); and $P_{a p t | i}$ is the probability of choosing apartment homes given class i.

The demand for new apartment homes in year t is shown in equation (5).

D_{a p t, t} = p_{a p t, t} \times (N B_{t} + S H R_{t})

(5)

where, $D_{a p t, t}$ is the demand for new apartment homes in year t; $N B_{t}$ is the new buyers from immigration in year t; $S H R_{t}$ is the existing single-family householders to relocate in year t. We considered the relocation of existing single-family households who settled in the metro Atlanta before 2013, because the relocation of this group from old single-family houses to new apartment homes creates the demand for new apartment homes. We should also include the relocation of new households who moved after 2013, the relocation of existing apartment households to single-family houses, and the second-time relocation of existing single-family households who moved once between 2014 and 2034. However, we need to model the affordability of these agents and the competition among these agents in the housing market if we include these relocations as well. Unfortunately, we did not have the information and thus we only considered the relocation of existing single-family households who settled in the metro Atlanta before 2013.

Correspondingly, the sale of new apartment homes in year t is shown in equation (6):

S a l e_{a p t, t} = \min (D_{a p t, t}, S_{a p t, t})

(6)

where, $S a l e_{a p t, t}$ is the sale of new apartment homes in year t; $S_{a p t, t}$ is the supply of new apartment homes in year t.

As existing single-family house residences relocate to apartment homes, their houses enters the housing market for resales ( $S_{o l d, s i g, t}$ , equation 7).

S_{o l d, s i g, t} = \min (D_{a p t, t} \times \frac{S H R_{t}}{N B_{t} + S H R_{t}}, S a l e_{a p t, t} - D_{a p t, t} \times \frac{N B_{t}}{N B_{t} + S H R_{t}})

(7)

The demand for new single-family house is determined by equation (8).

D_{s i g, t} = N B_{t} - \min (D_{a p t, t} \times \frac{N B_{t}}{N B_{t} + S H R_{t}}, S a l e_{a p t, t}) - S_{o l d, s i g, t}

(8)

where, $D_{s i g, t}$ is the demand for new single-family houses in year t, which equals the number of new buyers ( $N B_{t}$ ) minus those choosing new apartment homes and old single-family houses.

The sale of new single-family houses in year t is shown in equation (9).

S a l e_{s i g, t} = \min (D_{s i g, t}, S_{s i g, t})

(9)

where, $S a l e_{s i g, t}$ is the sale of new single-family houses in year t; $S_{a p t, t}$ is the supply of new single-family houses in year t.

In year t+1, the supply of new apartment homes and new single-family houses is determined by equations (10) and (11), respectively.

S_{a p t, t + 1} = \max (N B_{t + 1} \times \frac{D_{a p t, t}}{D_{a p t, t} + D_{s i g, t}}, S_{a p t, t} - S a l e_{a p t, t})

(10)

S_{s i g, t + 1} = \max (N B_{t + 1} \times \frac{D_{s i g, t}}{D_{a p t, t} + D_{s i g, t}}, S_{s i g, t} - S a l e_{s i g, t})

(11)

The fraction of new apartment homes sold ( $f_{a p t}^{T}$ ) up to year T is emergent as follows (equation 12):

f_{a p t}^{T} = \frac{\sum_{t = 2014}^{T} S a l e_{a p t, t}}{\sum_{t = 2014}^{T} S a l e_{a p t, t} + \sum_{t = 2014}^{T} S a l e_{s i g, t}}

(12)

The $f_{a p t}^{T}$ value shows the time-dependent diffusion curves (i.e. adoption rate) of new apartment homes. We use the results for $f_{a p t}^{T}$ to assess whether LID and TOD for apartment communities can contribute significantly to less sprawling land development in the metro Atlanta.

The starting conditions of the agent-based diffusion model include: (1) $N B_{t}$ constantly equals 29,038 assuming the stable increase of new households from 2014 to 2034 in the metro Atlanta (Atlanta Regional Commission (ARC), 2009); (2) the number of existing single-family households to relocate ( $S H R_{t}$ ) in 2014 is 231,370, which is 12% of total households in 2013 (see Supplementary material: Mover Rate in the Metro Atlanta, available online); and $S H R_{t + 1} = 0.88 \times S H R_{t}$ because in year t+1 the remaining existing households who settled down in the metro Atlanta before 2013 is 12% less than that in year t after the relocation; (3) the total supply of new apartment homes and new single-family houses equals 29,038 in year 2014, for which 92% are single-family houses and 8% in apartment homes (see Supplementary material: Housing Types in the Metro Atlanta, available online).

The market diffusion model predicts the adoption of more compact communities under three policy scenarios (see Table 2). The first scenario is a ‘Business As Usual (BAU)’ scenario, where LID and TOD are unavailable. In the second and third scenarios, TOD is provided assuming that an investment fund is available from federal and state governments and only apartment communities are permitted in the TOD zone, in an effort to make TOD profitable. In 2013, the Department of Watershed Management in the City of Atlanta revised the Post-Development Stormwater Management Ordinance to promote the use of green infrastructures (i.e. LID) on new and redevelopment projects in the city. According to the revision, new multi-family housing communities are required to treat the first 1.0 inch of storm runoff on a community basis. New single-family homes are required to manage the first 1.0 inch of runoff on their sites. We defined the second scenario as a ‘TOD plus home-based LID’. For comparison, the third scenario is defined as a ‘TOD plus community-based LID’ that requires both new multi-family housing community and new single-family homes to manage and treat the first 1.0 inch of runoff on a community basis. The controversy and battles have been going on over the issue of site-level versus community-based stormwater requirement (United States Environmental Protection Agency (US EPA), 2009). The impacts of TOD and LID on residential community choice under three scenarios is summarised in Table 2. The comparison among the three scenarios demonstrates whether more compact development has a greater potential to emerge by implementing LID and TOD, and which policy for LID can best contribute to the emergence of more compact development in the metro Atlanta.

Table 2.

Summary of the scenarios and the modelling of the effects of TOD and LID in the household community choice.

	TOD		LID
	Single-family house	Apartment home	Single-family house	Apartment home
Scenarios
Business As Usual (BAU)	No	No	No	No
TOD plus home-based LID	No	Yes	New single-family homes are required to manage the first 1.0 inch of runoff on their sites independently	New apartment communities are required to treat the first 1.0 inch of storm runoff on a community basis
TOD plus community-based LID	No	Yes	New single-family homes are required to manage and treat the first 1.0 inch of runoff on a community basis	New apartment communities are required to manage and treat the first 1.0 inch of runoff on a community basis
Modelling of the effects of TOD and LID in the household community choice (equation 3)
Business As Usual (BAU)	Accessibility = driving in 30 minutes Commute = car only, more than 40 min Transit = more than 20 minutes	Accessibility = driving in 30 minutes Commute = car only, more than 40 min Transit = more than 20 minutes	Green building = none Private LID = no Public LID = no	Green building = none Private LID = no Public LID =no
TOD plus home-based LID	Accessibility = driving in 30 minutes Commute = car only, more than 40 min Transit = more than 20 minutes	Accessibility = walking in 10 minutes Commute = public transportation (e.g. transit, bus) within 30 min Transit = less than 10 minutes	Private LID = yes Public LID = no Green building = RWH	Private LID = no Public LID = yes Green building = GRWR
TOD plus community-based LID	Accessibility = driving in 30 minutes Commute = car only, more than 40 min Transit = more than 20 minutes	Accessibility = walking in 10 minutes Commute = public transportation (e.g. transit, bus) within 30 min Transit = less than 10 minutes	Private LID = no Public LID = yes Green building = RWH	Private LID = no Public LID = yes Green building = GRWR

Environmental impact assessment

To simplify the environmental impact assessment as an example of sustainability assessment, we followed the life-cycle environmental and economic assessment results of transit-oriented neighbourhood designs reported by Chester et al. (Chester et al., 2013). The details of transit-oriented neighbourhood and conventional suburban community for 3200 housing units are shown in Table S2 (Supplementary material, available online). The annualised life-cycle carbon emissions of the transit-oriented neighbourhood are 0.033 million tons/yr on average while the conventional suburban community emits 0.053 million metric tons/yr. The impact of LID on the reduction in carbon emissions of the sewer network can be ignored, which is 3% of that from more compact development (De Sousa et al., 2012). However, it should be noted that LID at the community scale does contribute to the reduction in carbon emission indirectly through its impact on the adoption of more compact development.

With the projected numbers of apartment homes and single-family houses from the agent-based market diffusion model, we first determined the number of transit-oriented neighbourhoods and conventional suburban communities that were built between 2014 and 2034 for the three scenarios. Then the carbon emissions were calculated as the sum of the carbon emissions of transit-oriented neighbourhoods and conventional suburban communities.

Results

Summary of respondent’s characteristics

There were 764 useful responses obtained from a total of 811 responses from Mechanical Turk within three weeks. Only 6% of respondents did not follow survey instructions, including those who did not complete the survey and those who gave the same answers to the choice of best and worst communities. On average, each respondent spent 16 minutes on the survey, which is about 45 seconds for each choice set plus about 6 minutes for other quick questions (e.g. income, education). About 5% of useful responses were completed within 5 minutes. These respondents covered the entire metro Atlanta and the top three counties where respondents came from are Fulton, Cobb and DeKalb (the core area of metro Atlanta). A summary of socioeconomic characteristics of these 764 respondents is provided in Table S3 (Supplementary material, available online). Variables such as sex, household income, ethnicity, employment and presence of children from Mechanical Turk are close to census statistics of metro Atlanta. However, respondents from Mechanical Turk are younger and have achieved a higher level of education. Also the majority of the respondents (66%) on Mechanical Turk rent the properties, which is significantly higher than that in the census. Spatial locations of the respondents matches the pattern where people live in metro Atlanta (Figure S2, Supplementary material, available online).

Latent-class residential community choice modelling

We used 648 responses to estimate the latent-class residential community choice model. We selected four classes as the optimal number to represent the preference heterogeneity. Although the Bayesian information criterion (BIC) of the four-class choice model is higher than five-or-more class choice model, the interpretation of classes is much clearer using the four classes. Meanwhile, there is no significant improvement in modelling accuracy in terms of percentage of choices that are modelled correctly as the number of classes is more than four (see Supplementary material: Selection of the Number of Classes, available online). Thus, we adopted the results of the four-class residential community choice model to describe the preferences of individuals in the metro Atlanta for LID, TOD and more compact development. A summary of the four-class residential community choice model is provided in Table 3, including the class-specific β associated with community amenities (Table 1) in the choice modelling and β associated with socioeconomic and attitudinal variables (Table S1, Supplementary material, available online) in the class membership modelling.

Table 3.

Summary of the latent-class residential community choice modelling estimates for community attributes, socioeconomic and attitudinal variables.

Model for best–worst choice
	‘Compact’	‘Sprawling’	‘School-dominate’	‘Price-sensitive’	Overall
R ²	0.065	0.416	0.384	0.258	0.217
Community design attributes					p *	p(=)**
Constants	−0.197	−0.159	−0.284	0.051	0.000	0.049
House design
16-storey	−0.037	−1.110	−0.446	−0.520	0.000	0.000
8-storey	0.012	−1.155	−0.180	−0.061
4-storey	0.088	−0.506	0.016	0.043
Row house	0.055	−0.318	−0.158	−0.179
Semi-detached	−0.039	0.870	0.192	0.263
Single-family	−0.078	2.219	0.577	0.455
Green building
None	−0.172	−0.140	−0.358	−0.186	0.000	0.000
GRWR	0.139	0.175	−0.007	0.014
RWH	0.033	−0.035	0.366	0.172
Increase in price	−0.567	−0.435	−1.660	−6.490	0.000	0.000
Accessibility	0.624	0.088	−0.016	0.090	0.000	0.000
Greenspace	0.037	0.044	0.085	0.056	0.000	0.080
Public LID	0.084	0.192	0.086	−0.042	0.000	0.070
Private LID	0.126	0.480	0.093	0.098	0.000	0.000
Commute	0.153	0.039	0.147	0.095	0.000	0.000
School quality	0.328	0.440	3.184	0.773	0.000	0.000
Transit	0.092	−0.018	−0.032	0.206	0.002	0.066
Model for classes
	‘Compact’	‘Sprawling’	‘School-dominant’	‘Price-sensitive’	p-value
Intercept	1.403	0.616	−1.651	−0.368	0.007
Covariates
Time to complete	−0.024	−0.032	0.027	0.029	0.000
TOD	0.956	−1.108	0.260	−0.108	0.000
LID	0.561	0.615	−0.074	−1.101	0.024
Housing	−0.601	0.382	0.516	−0.298	0.000
Kids	−1.215	−0.368	0.960	0.624	0.000
Grouping rent	−0.072	−0.740	−0.046	0.858	0.049
Education	0.071	−0.385	−0.129	0.443	0.027
Income	0.105	−0.135	0.222	−0.192	0.017
Employed	−0.346	0.914	−0.258	−0.310	0.010
Age	−0.051	0.188	−0.015	−0.122	0.096
Ethnicity	0.103	0.036	−0.032	−0.108	0.660
People in the house	0.021	0.064	0.081	−0.166	0.530
Gender	−0.042	0.012	−0.077	0.107	0.950
Non-work travel	−0.001	−0.001	−0.001	0.003	0.290
Commute time	−0.063	0.123	0.214	−0.274	0.720

Notes: * The p-value shows the significance level of the coefficients. The p larger than 0.05 associated with a certain attribute indicates no significant effect of the attribute on individual utility. ** The p (=) value shows the significance level of the difference in coefficients among the four classes. The p (=) larger than 0.05 associated with a certain attribute indicates no significant difference in preference for this attributes among the four classes.

Preference of the four classes

Preference heterogeneity is revealed by the relative importance of community amenities in decision-making (Vermunt and Magidson, 2012). The relative importance is a maximum effect of each amenity on utility (equation 3) that is rescaled to sum to 1 across all amenities within a latent class (equations 13 and 14). Different ranking of the relative importance of the amenities in influencing decision-making shows the preference heterogeneity of people living in metro Atlanta. According to the ranking of different amenities, we named the four classes ‘compact’, ‘sprawling’, ‘school-dominant’ and ‘price-sensitive’. As shown in Figure 2, ‘Commute’, ‘Accessibility’ and ‘School quality’ are the top three amenities for the ‘compact’ class. To be more specific, the ‘compact’ class prefers the community that takes a shorter commute to work/school, and is closer to food, shopping and entertainment, and has better school quality. In contrast, the ‘sprawling’ class considers ‘House design’, ‘Private LID’ and ‘School quality’ as the top three important amenities. In other words, the ‘sprawling’ class prefers single-family houses, LID on private yards and better schools. The ‘school dominant’ class considers school quality to be the most important amenity in choosing where to live, while members of ‘price-sensitive’ class think that price is most important. Both ‘school-dominant’ and ‘price-sensitive’ classes prefer single-family houses over multi-family homes. The comparison of the demand for bedrooms, bathrooms, size of housing area and affordability among the four classes is provided in the Supplementary material, available online.

\max e f f_{x p} = \max (β a |_{x p}) - \min (β a |_{x p})

(13)

rel e f f_{x p} = \frac{\max e f f_{x p}}{\sum_{p} \max e f f_{x p}}

(14)

where, max $e f f_{x p}$ is the maximum effect of attribute p for latent-class x on the utility; α is a certain level of attribute p (Table 1).

Figure 2.

The relative importance of different amenities for the four classes: a list of community amenities and levels is also available in Table 1.

Spatial distribution of the four classes in the metro Atlanta

Considering the bias of sampling (Table S3, Supplementary material, available online), the ratios of the four classes in the sample cannot properly reflect the true percentages of the four classes in the metro Atlanta. In order to estimate the ratios of the four classes correctly, we used the Public Use Microdata Sample (PUMS) data to calculate the probability of individuals belonging to each class using the class membership model. The PUMS is a subsample of individual person and housing unit records (e.g. sex, education and employment status) from American Community Survey (ACS). The PUMS is considered as an unbiased sample of individuals in the metro Atlanta. It should be noted that the PUMS does not provide any personal attitude variables, which may still lead to a biased estimation of the percentage of the four classes.

Results show that the percentage of the ‘compact’, ‘sprawling’, ‘school-dominant’ and ‘price-sensitive’ classes is 22%, 62%, 9% and 7%, respectively. The ‘sprawling’ class is the dominant class, which imposes a great challenge to promote more compact development in the metro Atlanta. The PUMS provides the locations of sample individuals in the Public Use Microdata Area (PUMA) level. PUMAs are non-overlapping areas that partition each state into areas containing about 100,000 residents. Ratios of the four classes for each PUMA are presented on the PUMA map that covers the metro Atlanta. As shown in Figure 3, a higher percentage of the ‘compact’ class is found in the central urban area of the metro Atlanta while a higher percentage of the ‘sprawling’ class live in the suburban areas. A higher percentage of ‘school-dominant’ class is found in where have the best schools in terms of school test scores (Whitfield, 2013). The percentage of the ‘price-sensitive’ class is higher in places where the household income is lower or where students are living. Spatial visualisation indicates that stated preference in our survey can reflect the revealed preference in real community choices.

Figure 3.

Spatial distribution of the four classes in the metro Atlanta: percentages refer to the ratio of residents in the metro Atlanta belonging to each class; the breakdown of the percentages into ranges is based on the Jenks natural breaking method in ArcGIS 10 (Eris, Redlands, CA, USA) to optimise the visualisation.

Reliability test of choice modelling

With the latent-class residential community choice model, we are able to calculate the probability of the remaining 116 respondents choosing a certain option in each choice set. We considered the existence of some randomness in choosing the best one, which could be significant especially when the sample size of 116 respondents is small. We use a random number generator to determine the best choice. In each choice set, the four options are listed in a sequence with a corresponding probability interval for each option; these intervals collectively span from 0 to 1. The random number generator produces a number between 0 and 1. The respondent chooses the option with the probability interval into which this randomly generated number falls. The prediction varied each time and we repeated the prediction five times. The Chi-square test was used to distinguish whether the difference is significant between the predicted and actual distribution of people among the four options in each choice set. For each choice set, if the Chi-square test shows that one or more predictions have no significant difference with the actual number of people for each option to be selected as the best option, we consider that our prediction is validated to produce the real choice pattern. The Chi-square test shows that our choice model can predict the choice pattern that has no significant difference with the actual distribution pattern for the 9 out of 14 choice sets (see Figure S5, Supplementary material, available online, as an example). The result indicates the usefulness of our survey design and modelling approach in capturing people’s preferences. We consider that our latent-class residential choice model can be trusted to some degree.

Market diffusion of more compact development with LID policy intervention

We evaluated the implementation of LID and TOD in the metro Atlanta on future land-use patterns using the market diffusion model. The agent-based market diffusion model predicted the adoption of new apartment homes in three scenarios (Figure 4a). First, in the BAU scenario, the share of new apartment units sold from 2014 to 2018 increases slightly to 10%. It indicates that there is an under-supply of high-density communities in the metro Atlanta even in the BAU. After that, the share of total new apartment units sold starts declining because the demand for new apartment homes is not sustained, because the annual number of existing single-family households that choose to relocate decreases ( $S H R_{t}$ ). In the ‘TOD plus home-based LID’ scenario, the share of total new apartment units sold in new home sales increases in the first 8 years from 2014 to 2021. The increase is driven by the higher demand for new apartment homes with LID and TOD features. But the increase is not sustained after 2021. By the end of 2034, single-family houses still dominate new residential development. The major reason is that the implementation of LID for single-family houses (including rainwater harvesting and rain garden, porous pavement and native vegetation for private yards) leads to a significant increase in utility for single-family houses. The improvement in single-family house utility offsets the improvement in utility of the apartment home from LID and TOD. In other words, the improvement in utility of the apartment home is not enough to create a sustained demand for new apartment housing. In contrast, the share of total new apartment units sold increases consistently from 2014 to 2034 in the ‘TOD plus community-based LID’ scenario. By switching the requirement of treating the 1.0 inch stormwater runoff on individual lots to the community-based treatment using LID, we reduce the increase of utility for single-family houses. In addition, part of the demand for new single-family houses is met by old single-family houses, which are vacated as existing single-family households relocate to new apartment homes. As a result, the supply of new single-family houses in the house market decreases further, leading to a higher probability of homebuyer’s considering and adopting new apartment homes. Overall, we conclude that the implementation of LID and TOD can be expected to increase compact, more sustainable development. Meanwhile, the ‘TOD plus home-based LID’ should be switched to the ‘TOD plus community-based LID’ that requires new single-family homes to manage the first 1.0 inch of runoff on the community basis.

Figure 4.

(a) Share of apartment communities in newly built home units up to year T; and (b) annualised life-cycle carbon emissions of new development in the metro Atlanta from 2014 to 2034 under three scenarios.

Environmental impacts of more compact development

As shown in Figure 4(b), there is no significant difference in carbon emissions between BAU and ‘TOD plus home-based LID’ scenarios. In contrast, there is 28% of carbon emission reduction in the ‘TOD plus community-based LID’ scenario as compared with BAU. It is worth noting that apartment homes have no TOD feature in the BAU scenario. Thus, the potential in carbon emission reduction should be higher than 28%. But the carbon emission data do not accurately reflect the constructions and operations of new development in the metro Atlanta. The result here only provided a rough estimate of carbon emission reductions. More accurate estimates should be explored in the future to inform the final decision-making.

Discussion

Here, we will further explain both advantages and shortcomings of the detailed method in each step of our framework to study more sustainable infrastructure designs. In the step of survey design, we published our survey on Mechanical Turk, which is a cost-effective solution to engage with the cohorts and collect the data for decision modelling. Only 6% of respondents dropped out of the survey or gave wrong answers to the questions. Another benefit of using Mechanical Turk is the back and forth communication with the cohorts without much effort. It allows one to refine the survey from the feedback of respondents. We will also assess whether the cohort give the same answers at different times to address the reliability of survey design. Furthermore, the cohort on Mechanical Turk may serve as stakeholder advisory board to validate our research findings, for example, the switch from home-based LID to community-based LID policy. However, the size of the cohort is relatively small when we focus on a specific neighbourhood. Also, Mechanical Turk is not workable for cities in developing countries. More importantly, sample-selection bias still exists, which will affect the accuracy in preference estimation. Thus, we encourage the development of the crowd-source platform similar to Mechanical Turk for survey data collection and stakeholder engagement.

We employed the latent-class choice model to simulate people’s choice based on the responses in the choice experiment. The latent-class choice model is easy to interpret and it takes account of the heterogeneity in preference. Owing to the sample bias, we can overestimate/underestimate the preference by modelling the sample as a homogeneous group. However, the sample bias cannot be fully eliminated by using the latent-class choice model. The correction of the sample bias will be explored in future study. Other approaches such as hedonic price analysis can also help understand the effect of the amenities as revealed preference-based on transaction data (Bartholomew and Ewing, 2011). From the hedonic price analysis, we can estimate an individual’s willingness to pay. However, hedonic price analysis can only evaluate the amenities that have widely existed. For the emerging infrastructure techniques that have not been adopted yet, it is hard to conduct hedonic price analysis because of the lack of data. In the case that data are available for hedonic price analysis, we can combine the results from our approach and hedonic price analysis to develop a more reliable choice model plus an estimate of willingness to pay for the agent-based modelling in the next step.

The reliability of the agent-based model is critical to studying the complex urban system and predicting adoption of more sustainable infrastructures. We should mention one caveat, here, i.e. the accuracy of the choice simulation in the agent-based model is always a great challenge given the uncertainty of human decision-making. While the first two steps in our framework provide a means of generating a reliable choice model for agent-based modelling, we were not able to predict adoption of LID, TOD and more compact development with great accuracy. However, we did learn some general lessons on how to drive a higher adoption of LID, TOD and more compact development, and gained valuable insight into the sort of policies required in metro Atlanta to do so. In the future, careful designs and validation of the agent-based model will be required to predict accurate adoption rates. One approach here would be to compare historic growth rates with the outcomes from the agent-based model.

The last step of our sustainability assessment should account for social, economic and environmental aspects. In this paper we focused on one aspect of environmental assessment only, carbon emissions. Such assessments will benefit from more effort in estimating vehicle miles travelled, water treatment and conveyance costs, and building energy and material consumption for different land-use arrangements. In the future, we will focus on the agent-based model development with spatial visualisation of land use and transportation. This spatial information will allow the estimation of the change in traffic behaviour, energy use and water consumption. Further, we will build the inventory data on construction and operation of urban infrastructures systems for life-cycle assessment. Uncertainty and sensitivity analysis is required for modelling, which is not included in this study. Uncertainty analysis will quantify the variance of our prediction of adoptions and sustainability benefits to verify the investment risk. Sensitivity analysis will identify the most influential factors on the adoption and sustainability benefits in order to reduce the risk and uncertainties.

Overall, the main purpose of this paper is the demonstration of the framework for managing the complex interdependence between infrastructures and the socioeconomic environment. By applying our framework in the case of metro Atlanta, we showed the importance of how we develop policies and provide infrastructure amenities in affecting the sustainability outcomes of urban development. In the future, city planners and managers in the world can use this framework to explore their best practices in developing the infrastructures and creating more sustainable cities. Although the methods we presented in each step may not be the best practice, users can adaptively change our methods for each step under the framework, depending on the scope and uniqueness of each individual situation. This allows for exploring the best practices for each step and flourishing the methods for the framework.

Footnotes

Funding

Financial support from the National Science Foundation through grants (#0836046 and # 1441208), is gratefully acknowledged. This work was also partially supported by Brook Byers Institute for Sustainable Systems, the Hightower Chair and Georgia Research Alliance. The views and ideas expressed herein are solely of the authors and do not represent the ideas of the funding agencies in any form.

References

Amir

Rand

Gal

YaK

(2012) Economic games on the internet: The effect of $1 stakes. PLoS ONE 7: e31461.

Atlanta Regional Commission (2009) ARC’s 20-County Forecasts: What the Future Holds. Atlanta, GA: Atlanta Regional Commission.

Bartholomew

Ewing

(2011) Hedonic price effects of pedestrian- and transit-oriented development. Journal of Planning Literature 26: 18–34.

Baynes

(2009) Complexity in urban development and management. Journal of Industrial Ecology 13: 214–227.

Bereitschaft

Debbage

(2013) Urban form, air pollution, and CO₂ emissions in large U.S. metropolitan areas. Professional Geographer 65: 612–635.

Bowman

Tyndall

Thompson

et al . (2012) Multiple approaches to valuation of conservation design and low-impact development features in residential subdivisions. Journal of Environmental Management 104: 101–113.

Cervero

Murakami

(2010) Effects of built environments on vehicle miles traveled: Evidence from 370 US urbanized areas. Environment and Planning A 42: 400–418.

Chau

H-F

(2009) Green Infrastructure for Los Angeles: Addressing Urban Runoff and Water Supply Through Low Impact Development. Doctoral dissertation. Los Angeles, USA.

Chester

Nahik

Fraser

et al . (2013) Integrating life-cycle environmental and economic assessment with transportation and land use planning. Environmental Science & Technology 47: 12,020–12,028.

10.

Chu

Wang

Chen

et al . (2009) Agent-based residential water use behavior simulation and policy implications: A case-study in Beijing city. Water Resources Management 23: 3267–3295.

11.

Cook

(2007) Green site design: Strategies for storm water management. Journal of Green Building 2: 46–56.

12.

Crump

MJC

McDonnell

Gureckis

(2013) Evaluating Amazon’s Mechanical Turk as a tool for experimental behavioral research. PLoS ONE 8: e57410.

13.

De Sousa

MRC

Montalto

Spatari

(2012) Using life cycle assessment to evaluate green and grey combined sewer overflow control strategies. Journal of Industrial Ecology 16: 901–913.

14.

Ewing

Hamidi

(2014) Measuring Sprawl 2014. Washington, DC: Smart Growth America.

15.

Katz

(2013) The missing metric. Government Finance Review 29: 20–32.

16.

Liao

Farber

Ewing

(2014) Compact development and preference heterogeneity in residential location choice behaviour: A latent class analysis. Urban Studies 52: 314–337.

17.

Litman

(2012) Evaluating Public Transit Benefits and Costs. Victoria, BC: Victoria Transport Policy Institute.

18.

Southworth

Crittenden

et al . (2014) Market potential for smart growth neighbourhoods in the USA: A latent class analysis on heterogeneous preference and choice. Urban Studies 52: 3001–3017.

19.

Lund

(2006) Reasons for living in a transit-oriented development, and associated transit use. Journal of the American Planning Association 72: 357–366.

20.

Marley

AAJ

Louviere

(2005) Some probabilistic models of best, worst, and best-worst choices. Journal of Mathematical Psychology 49: 464–480.

21.

Najafzadeh

Lynd

Davis

et al . (2012) Barriers to integrating personalized medicine into clinical practice: A best–worst scaling choice experiment. Genetics in Medicine 14: 520–526.

22.

Olaru

Smith

Taplin

JHE

(2011) Residential location and transit-oriented development in a new rail corridor. Transportation Research Part A: Policy and Practice 45: 219–237.

23.

Olorunkiya

Wilkinson

Fassman

et al . (2013) Eliciting stakeholders’ preferences for low-impact design incentives: Conjoint analysis approach. Journal of Legal Affairs and Dispute Resolution in Engineering and Construction 5: 180–190.

24.

Pandit

Crittenden

(2015) Managing the complexity of urban systems. Journal of Industrial Ecology 19: 201–204.

25.

Paolacci

Chandler

Ipeirotis

(2010) Running experiments on Amazon Mechanical Turk. Judgment and Decision Making 5: 411–491.

26.

Parker

Manson

Janssen

et al . (2003) Multi-agent systems for the simulation of land-use and land-cover change: A review. Annals of the Association of American Geographers 93: 314–337.

27.

Raucher

(2009) The Triple Bottom Line Assessment of Traditional and Green Infrastructure Options for Controlling CSO Events in Philadelphia’s Watersheds. Proceedings of the Water Environment Federation. Session 91 through Session 100, 6776–6804(29). Alexandria, VA: WEFTEC.

28.

Rid

Profeta

(2011) Stated preferences for sustainable housing development in Germany – A latent class analysis. Journal of Planning Education and Research 31: 26–46.

29.

Roy

Wenger

Fletcher

et al . (2008) Impediments and solutions to sustainable, watershed-scale urban stormwater management: Lessons from Australia and the United States. Environmental Management 42: 344–359.

30.

SAS (2012) JMP. Cary, NC: SAS Institute Inc.

31.

Schmitz

(2008) Open space in residential subdivsions: Effect of low impact development and open space on single-family housing values. The Graduate College at the University of Nebraska. Omaha, NB: University of Nebraska.

32.

Schwarz

Ernst

(2009) Agent-based modeling of the diffusion of environmental innovations – An empirical approach. Technological Forecasting and Social Change 76: 497–511.

33.

Stone

Hess

Frumkin

(2010) Urban form and extreme heat events: Are sprawling cities more vulnerable to climate change than compact cities? Environmental Health Perspectives 118: 1425–1428.

34.

Suzuki

Cervero

Iuchi

(2013) Transforming Cities with Transit: Transit and Land-use Integration for Sustainable Urban Development. Washington, DC: World Bank Publications.

35.

United States Environmental Protection Agency (US EPA) (2009) Water Quality Scorecard: Incorporating Green Infrastructure Practices at the Municipal, Neighborhood, and Site Scales. Washington, DC: US EPA.

36.

Vermunt

Magidson

(2012) Latent GOLD Choice. 4.5 Edition. Belmont, MA: Statistical Innovations Inc.

37.

Whitfield

(2013) Atlanta School Test Scores Heat Map. The Atlanta & Northside Area Real Estate Finder. Available at: http://www.atlanta-real-estate-online.net/atlanta-real-estate-blog/atlanta-school-test-scores-heat-map/ (accessed 25 May 2016).

38.

Williams

Wise

(2009) Economic impacts of alternative approaches to storm-water management and land development. Journal of Water Resources Planning and Management-Asce 135: 537–546.

39.

Zhao

de Roo

(2010) Urban expansion and transportation: The impact of urban form on commuting patterns on the city fringe of Beijing. Environment and Planning A 42: 2467–2486.