Contextual Density and US Automotive CO 2 Emissions across the Rural

Abstract

Using data from the 2009 National Household Travel Survey, we quantify the effects of settlement patterns on individual driving habits and the resulting automotive carbon dioxide (CO₂) emissions. We employ CO₂ emissions to capture this impact accurately, as it reflects both vehicle miles traveled and any spatial differences in vehicle fuel efficiency choices. While previous studies have compared automotive travel in urban and suburban areas, our approach characterizes emissions across the entire US rural–urban gradient, focusing on the effects of population density. Rather than using categorical measures of contextual density (city, suburb, town, etc.), we use a geographical information system to calculate continuous measures of contextual density, that is, density at different proximities to households. These measures of contextual density allow us to model travel effects induced by the gravitational pull of the population densities of urban cores. Further, our methodological approach frames location choice as an endogenous treatment effect; that is, residential locations are not randomly assigned across our sample and significantly alter driving behavior. We find that individuals living in urban cores generate the lowest per capita automotive CO₂ emissions, due to close proximities of population concentrations. Rather than attracting individuals who would likely have low CO₂ emissions anyway, urban location apparently mitigates the emissions of people who would otherwise tend to have high automotive CO₂ emissions. We find larger elasticities with respect to density than previous studies and also find that the attractive forces of population densities affect driving patterns at distances up to sixty-one kilometers outside of urban areas.

Keywords

urbanization travel behavior carbon emissions residential location endogenous treatment effects

Introduction

In the United States, as in much of the industrialized world, personal transportation by automobile is one of the major impacts people have on their environments. In 2012, the United States emitted an estimated 6,526 teragram (Tg) of carbon dioxide-equivalent (CO₂e) greenhouse gases from fossil-fuel combustion, with the transportation sector accounting for 34.4 percent and passenger cars and light trucks emitting 1,072 Tg CO₂e or 21 percent of total greenhouse gas emissions from fossil fuel (Environmental Protection Agency [EPA] 2014b). Personal automobile travel is a significant contributor to US greenhouse gas emissions and associated climate change.

This shapes our interest in the determinants of US driving habits. We focus on the impact of residential location choice on driving behavior, with an emphasis on the attractive forces of urbanization and nearby economic activity. Urban areas now account for 80.7 percent of the US population with the rate of urbanization outpacing growth (12.1 percent vs. 9.7 percent) of the overall population over the last decade (US Census Bureau 2010). Given the current rates of urbanization across the globe, the implications of our findings extend beyond the United States. Today about 52 percent of world population resides in urban areas, with that proportion expected to rise to 67 percent by 2050 (Heilig 2012). Urbanization’s impacts are clearly relevant to the problem of climate change and other problems of environmental sustainability.

Environmentalists have often viewed urban areas as problems: “… entropic black holes drawing in energy and matter from all over the ecosphere” (Rees and Wackernagel 1996, 237). Empirically, however, cities have positive environmental aspects in reducing automobile usage and building energy consumption (Kahn 2006; Owen 2009). The growing suburbanization of US urban populations is also a central motivator of our research. Suburbanization can increase distances between places of work and residences and thus increase commuting and associated greenhouse gas emissions.

The United States is characterized by substantial variation in population and job growth with regard to cities versus their suburban counterparts. In particular, the overall rate of population growth has reversed over the last decade with some cities experiencing growth that far outpaces their suburbs (Frey 2013). On the other hand, job growth is seemingly occurring the fastest in suburbs (Glaeser and Kahn 2001; Raphael and Stoll 2010). This US geographic and temporal variation in urbanization strengthens the motivation for our research.

In contrast to most prior studies, we examine differences in automobile use across the entire rural–urban continuum. Ewing and Cervero (2010: 267) observe that “the potential to moderate travel demand by changing the built environment is the most heavily researched subject in urban planning.”¹ Yet most studies compare different urban or suburban areas to each other, focusing on how differences in urban design affect travel behaviors (Crane and Crepeau 1998; Frank and Pivo 1994; Khattak and Rodriguez 2005; Marshall and Garrick 2012). By contrast, our primary interest is comparing urban to rural areas, that is, assessing the effects of urbanization as a whole, which is informative of policies regarding broader rural-to-urban migration. We also present results in terms of CO₂ emissions rather than vehicle distance traveled, testing for any spatial variation in vehicle fuel efficiency as well as spatial differences in travel patterns. Our study demonstrates one aspect of urbanization environmental impact in the United States and is also relevant for the world’s less-industrialized countries where large-scale urbanization is still occurring today.

Like Kim and Brownstone (2013), our main focus is on “contextual density,” which we define as the pattern of population density surrounding each household. While population density in one’s own neighborhood clearly affects automobile usage and emissions, density patterns of surrounding areas—contextual density—is likely important as well. In this study, we develop and test continuous measures of contextual density, using data developed with geographic information system (GIS) software as described in the fourth section, which provides more nuanced results than categorical variables such as “suburb” or “town.” There is great variation within general contextual density categories, in terms of both own-area density and the proximity and density of other areas to which household members may travel. Our continuous measures of contextual density capture such differences and describe residential spatial patterns in more detail than in past studies, providing a stronger foundation for conclusions about effects of spatial patterns.

We use these continuous measures of contextual density in a gravity model, which we posit will have explanatory power for automotive travel between areas of different population density in the United States. Amenities in close proximity to one’s residence may reduce the total automotive CO₂ emissions by reducing the need to travel—an effect of greater accessibility. Similarly, having attractions at a distance may increase the propensity to drive and thereby increase CO₂ emissions—an effect of lesser accessibility. We thus adapt the idea of gravitational pull to the attractiveness of nearby amenities, which we proxy with population density (Glaeser, Kolko, and Saiz 2001).

Our study diverges from Brownstone and Golob (2009) and others in our approach to self-selection of residential location. We view location decisions as endogenous treatment effects that directly cause nonrandom assignment of individuals to locations and find that such selection effects are present and significant. To correct for the likely endogeneity of density and driving, we utilize a full-information maximum likelihood (FIML) endogenous treatment regression model (Cameron and Trivedi 2005; Heckman 1976, 1978; Wooldridge 2010). We do so for a number of reasons. Primarily, we aim to control for the endogenous relationship between unobservables that affect both the treatment (residential location choice) and the potential outcomes (CO₂ emissions). Further, FIML has been proven superior to the two-step approach and has come to dominate the closely related estimation approach of the Heckman’s selection model (Puhani 2000). As described below, we use exclusion restrictions based on housing market characteristics as the variables affecting location choice. Our results indicate a strong justification for this approach.

We find that choice of urban location has a significant and negative effect on automotive CO₂ emissions. We also find significant negative elasticities of CO₂ emissions with respect to population density, with larger magnitudes than have been found in most previous studies. Our continuous measures of contextual density also show that population density has varying effects on driving behavior, depending on where a household is located in relation to dense areas.

In the next section, we briefly discuss the literature that sets the stage for our contribution. We then describe the range of automotive CO₂ emissions observed in different US settlement patterns, providing the background for our focus on CO₂ emissions and driving patterns in the third section. The fourth, fifth, and sixth sections present our empirical approach, results, and discussion. We briefly discuss the policy implications of our findings in the seventh section.

Literature Review

Urban density greatly predicts automobile usage. Denser areas have a number of attributes related to automobile use, including greater availability of public transit, greater proximity to employment and other points of interest, as well as driving deterrents like more traffic and less parking availability (Cervero and Kockelman 1997). In exploring global urbanization-transportation impacts, Newman and Kenworthy (1989) compare population density to automobile use in thirty-two global cities, finding a strong negative correlation between density and driving. This study has been criticized for the possible incomparability of the different countries studied (Gordon and Richardson 1989). Compared to the US cities, for example, Hong Kong has very high population density and very low automobile usage, yet the differences between Hong Kong and Houston likely go far beyond density. Others criticize Newman and Kenworthy for their method of aggregating observations at the national level (Ewing and Cervero 2010). Still, the worldwide examples provided by Newman and Kenworthy (1999, 1989) illustrate the range of possible urban automobile use and provide examples of well-functioning cities that are not highly dependent on automobiles.

Cervero and Kockelman (1997) describe specific ways in which the built environment can affect travel behavior, introducing the 3Ds: urban density, diversity, and design. Greater population density, a diversity of land uses, and design that accommodates pedestrian and bicycle trips are all thought to reduce the demand for automotive travel and increase nonautomotive travel modes such as using public transit and walking. Using the 1995 Nationwide Personal Transportation Survey (NPTS), the study finds evidence for small effects of the built environment: for example, nonwork vehicle miles traveled (VMT) decreases by 0.06 percent with respect to a 1 percent increase in the intensity factor, a derived variable with population density as one component.

Using the same NPTS data, Kahn (2000) estimates the effect of urban density on VMT in US metropolitan statistical areas (MSAs; nonmetropolitan areas excluded), with a particular interest in urban compared to suburban outcomes. Although VMT are significantly higher in suburbs, the elasticity of VMT with respect to population density is estimated to be only −0.06, while income elasticity of VMT is much higher at 1.39. Kahn finds significant VMT differences between US regions as well as between individual MSAs. Bento et al. (2005) also find evidence for small spatial effects on automobile use; the authors estimate the elasticity of VMT with respect to city shape, road density, rail supply, and jobs-housing balance to be not more than ±0.07 (sign depending on the variable). But these small effects sum to larger total differences between US metropolitan areas. The authors do not consider rural areas.

As illustrated by these studies, population density tends to be a significant factor in explaining automobile use, but the magnitude of the effects may be small. Demographic factors frequently affect VMT more than environmental factors—unobservable tastes and preferences clearly play a large role in deciding what kind of vehicles to drive and where to drive them. Furthermore, relatively low explanatory power has been found in these models (Bento et al. 2005; Cervero and Kockelman 1997; Kahn 2000), with cited R ² values of .15, .14, and .11, respectively. There is clearly much variation in automobile travel between people who are spatially and demographically similar.

Our focus on population density as a proxy for urban amenities is based on the well-documented emergence of cities as consumption hubs across the United States (Glaeser, Kolko, and Saiz 2001; Glaeser and Gottlieb 2006). Glaeser, Kolko, and Saiz (2001) identify the increase in demand for urban amenities as the core determinant of the rise of reverse commuting and the population growth of (high-amenity) cities. Additionally, high-amenity cities are found to exhibit housing cost growth that outpaces growth in urban wages. These factors taken together demonstrate a significant shift since the 1980s in the demand for urban amenities. While it has been hypothesized that this shift may be due to the decrease in urban crime, thereby shifting the trade-off between the costs and benefits of living in cities, Glaeser and Gottlieb (2006) find (using estimates from Cullen and Levitt [1999]) that the reduction in crime is not the driving force of the increase in demand for urban amenities. Instead, demand for urban amenities, such as museums, restaurants, movie theaters, and concert halls, is indeed rising as incomes increase across the country. Importantly, for our article, the availability of these social and leisure activities is directly related to the population, possibly due to the benefit of large population densities generating sufficient customers to cover fixed costs (Glaeser and Gottlieb 2006).

One significant complication in the exploration of this topic is the endogeneity of location and travel behavior, a residential self-selection problem. Individuals who enjoy driving may choose to live where driving is easier or in locations that necessitate driving (Cao, Mokhtarian, and Handy 2009). Estimates of the impact of density on distance traveled are biased if households systematically differ in their driving preferences across location-density alternatives. Further, driving may be an outcome of residential choice irrespective of personal preference. Thus, we must first account for residential choice and then model travel behavior.

In a study of new urbanist and conventional neighborhoods in North Carolina, United States, Khattak and Rodriguez (2005) use an instrumental variables approach to control for the endogeneity of neighborhood choice—those who prefer less driving may systematically choose to live in locations like the new urbanist neighborhood and estimates of differences in driving behavior would be biased without an appropriate correction. The authors first use a logistic regression model to estimate the dichotomous choice of neighborhood, and subsequently use predicted neighborhood choice in a regression of the number of household trips on neighborhood choice and controls for household characteristics. The new urbanist design is found to reduce both the frequency and the distance of automotive travel, without reducing the total household excursions.

Brownstone and Golob (2009) directly model the joint choice of residential density and vehicle usage to control for potential residential selectivity. Unlike other studies, they also explicitly model vehicle fuel consumption to account for the possibility that residents of high density areas choose more fuel-efficient vehicles. By adopting a weighting approach, they correct for the bias caused by systematic missing data problems. Using the California subsample of the 2001 National Household Travel Survey (NHTS), Brownstone and Golob (2009) find a statistically significant but quantitatively small impact of residential density on household vehicle usage and fuel consumption. Yet they could not reject the null hypothesis of no self-selection effects.

Cao, Mokhtarian, and Handy (2009) review thirty-eight studies which correct travel estimates for self-selection effects, classifying methodological approaches into nine general categories. Evidence for self-selection has generally been demonstrated. While the authors find that the built environment has significant effects on travel behavior, they also find that the “apparent influence of the built environment diminish[es] substantially once residential self-selection is taken into account” (Cao, Mokhtarian, and Handy 2009, 390). This provides a strong justification for using a model that accounts for self-selection, as we use in this study.

Using the 2001 NHTS, Kim and Brownstone (2013) model US household vehicle distance traveled and fuel consumption based in part on categorical measures of contextual density, for example, urban, suburban, second city, town, and rural. Such contextual measures describe a household setting better than a simple measure of population density in a household’s census block or census tract. The study shows that these contextual density attributes are significant predictors of automotive travel. For example, the authors estimate that moving a household from a rural location to an urban one reduces vehicle distance traveled by 34 percent and reduces fuel consumption by 37 percent (since urban locations are also associated with more fuel-efficient vehicles).

The literature on spatial effects of personal automobile use includes a minor but consistent theme on gravity effects, where personal travel decisions are thought to be driven in part by proximity to areas of greater density or diversity, that is, by contextual density (Levinson and Kumar 1997). This psychological gravity is thought to increase with the “mass” of the metropolitan area or other attraction, and thought to decrease with distance, time, or some other impedance measure. Handy (1993) uses an exponential form of the gravity model, where attraction is a function of employment within a zone, and attraction decays as a function of time or distance to reach that zone. Shen (2000) also uses an exponential impedance function in a different specification for a gravity variable. Both Handy (1993) and Shen (2000) find gravity variables to be significant in explaining travel behavior.

Automotive CO₂ Emissions: Spatial Variation across the Rural–Urban Continuum

Many US transportation studies utilize the Department of Transportation’s periodic NHTS, conducted in 2001 and 2009, or the earlier NPTS, last conducted in 1995. These surveys query a national sample of households about vehicle ownership and use, and the data sets provide population density of respondent’s census tract in addition to other demographic information. Our study relies primarily on data from the 2009 NHTS (US Department of Transportation 2009), which includes 150,147 household observations, weighted for best representation of the entire US population. The NHTS data include household gasoline-equivalent gallons consumed per year (equivalents for diesel and other nongasoline fuels). We convert these to automotive CO₂ emissions using a factor of 8.78 kilogram CO₂/gallon gasoline (EPA 2014a) and use CO₂ emissions as the primary dependent variable for our regressions. We express CO₂ and some other household variables in per capita terms, in this case automotive CO₂ emissions per household member, removing the direct effect of household size on results. We use the natural log of CO₂ emissions as the dependent variable since preliminary analysis revealed substantial outliers in the data, which we wanted to maintain. Further, (truncation at) the lower bound of zero emissions defeats any support for the normality of the data (Table 1 presents the percentage of households by subcategory with zero CO₂ emissions, which varies by category). Transforming the data to natural logs addresses both of these issues.

Table 1.

Per Capita Automotive Carbon Dioxide Emissions by Settlement Type.

	Country	Town	Second City	Suburban	Urban
Percentage population with no automotive emissions	3%	4.5%	7.8%	4%	14.8%
Percentage population who emitted < 1,000 kilogram	6.7%	10.2%	15.6%	10.6%	25%
Percentage population who emitted > 10,000 kilogram	9.5%	5.9%	4.2%	4%	2.7%
Average per capita emissions (kilogram per year)	5,315 (4,959)	4,380 (3,925)	3,739 (3,682)	3,942 (3,332)	2,963 (3,073)

Note: Urbanization categories are from NHTS and Census definition: Country calculated from Town-country = 1 and census Urban = 0; Town calculated from Town-country = 1 and census Urban = 1. Standard deviations in parentheses. NHTS = National Household Travel Survey.

As shown in Table 1, every urbanization category has substantial proportions of households with zero automotive CO₂ emissions and with low emissions of less than one metric ton per person per year (though these groups are largest in urban areas). Similarly, Table 1 shows that all urbanization categories have households with high emissions of more than ten metric tons of automotive CO₂ emissions per person per year (though this group is largest in country areas). The last row of Table 1 also presents the average per capita automotive CO₂ emissions per year and their standard deviations. Analysis of variance tests, accounting for unequal variances between the groups, find that the difference in averages is statistically significant, at the 0.1 percent level, between all groups.

Figure 1 presents the bivariate relationship between urbanization extent (categorical contextual density variables) and the natural log of per capita automotive CO₂ emissions. The urbanization categories are arranged in increasing order of population density, and the figure seems to indicate a negative relationship between population density and automotive CO₂ emissions. Second cities have only slightly lower median emissions than suburbs, though second cities are presumably more self-contained and might be expected to support lower carbon lifestyles than suburbs. Country areas exhibit both the highest median emissions and the highest values at the twenty-fifth and seventy-fifth percentiles. Figure 1 also depicts substantial overlap of interquartile ranges between the urbanization level categories. For example, over 10 percent of urban households have higher automotive CO₂ emissions than the median country household.

Figure 1.

Log (per capita automotive CO₂ emissions) by urbanization extent.

The relatively small differences between groups as compared to the large differences within groups suggest that while spatial variables are important, they are not the primary factor determining household automotive CO₂ emissions. We observe households with very low (and very high) emissions in all settlement categories. Given this reality, we develop an approach to analyze the variation in CO₂ emissions within and across geographic locations.

Regression Analysis: Data and Methods

As M. G. Boarnet and Sarmiento (1998) observe, in theory spatial variables including land use should affect travel behavior through the price of travel, which in turn depends both on direct travel expenditures (fuel, public-transit tickets, etc.), and the time required for travel. Increasing the price of travel effectively increases the price of amenities for which consumption includes a travel requirement (e.g., shopping, dining out, etc.). In this study, we measure the price of travel in part with accessibility variables reflecting distances to areas of greater population density (or the contextual density), as measured in circular bands around each household census tract (described below). In this case, we use population density² as a proxy for availability of amenities and associated land uses, including commercial and institutional land uses associated with consumption of goods and services. Controlling for as many confounding variables as possible, we hypothesize that population densities at close proximity to a household reduce automotive CO₂ emissions, that having population densities at somewhat greater distances increases emissions, and that at some distance, remote population densities are not a significant predictor of travel or associated emissions.

Our approach assumes that automobile travel is primarily a derived demand (M. Boarnet and Crane 2001; Cervero and Kockelman 1997). While we do acknowledge that driving itself may provide utility for some (Mokhtarian, Salomon, and Redmond 2001), we assume that these individuals comprise a small proportion of the general population. CO₂ emissions are an externality of driving, an inevitable consequence of travel, albeit with variation in the ratio of CO₂ emissions to travel distance. We thus estimate the impact on CO₂ emissions of variables directly associated with household consumption choices (Brownstone and Golob 2009). These include demographic variables which also proxy for unobservable tastes and preferences. Such variables include age, race, income, education, and age-groupings of children in a household. We also control for whether a household includes anyone not born in the United States—a reasonable control based on the significant differences in driving attitudes in other countries. Controlling for distance to work means that results generally reflect noncommuting automotive use.

We use 2009 NHTS data with household weights, omitting Alaska and Hawaii from the study, assuming that automotive travel in those states is fundamentally different than in the contiguous states. Our dependent variable is the natural log of automotive CO₂ emissions per capita, calculated from fuel consumption data provided in the NHTS. Independent variables from the NHTS control for household composition, type and ownership of the home, and household income. We express all relevant variables in household per capita terms.

We include regional binary variables that capture broadly invariant factors such as weather and urban design, assuming that these factors relate to automotive travel and its alternatives. Regional variables may also capture geographic differences in tastes and preferences. In this study, we use climatic regions from the National Oceanographic and Atmospheric Administration (NOAA).³ We employ these nine NOAA regions rather than the nine Census divisions since weather may be an important determinant of driving behavior, and Census divisions do not adequately reflect weather differences between regions. For example, the Census Mountain and Pacific Divisions extend north-south from the Canadian border to the Mexican border.

We calculate a number of variables at the household level. For example, household average driver age is the mean age of all drivers associated with a household. We also include driver age square to capture any nonlinear effects of average driver age on driving habits. Per capita household distance to work is based on the sum of distance for all workers in the household. We also calculate household per capita bicycle and walking trips in the last week. Mean household fuel price is the total household automotive fuel expenditure divided by gasoline-equivalent gallons (equivalents for diesel and other nongasoline fuels). We include all other demographic variables that are available in the NHTS data set (income, education, race, etc.), as proxies for unobservable tastes and preferences. Table 2 provides descriptive statistics for the variables used in the regression model.

Table 2.

Descriptive Statistics (Observations: 114,586).

	Mean	SD	Min	Max
Log carbon dioxide emissions	8.00	1.38	0	12.18
Household size	2.31	1.20	1	13
Number of drivers in household	1.82	0.70	1	8
Number of workers in household	1.01	0.92	0	6
Mean age of drivers	56.90	15.41	15	92
Life cycle: no children	0.32	0.47	0	1
Life cycle: youngest child 0–5 years	0.09	0.28	0	1
Life cycle: youngest child 6–15 years	0.13	0.33	0	1
Life cycle: youngest child 16–21 years	0.05	0.22	0	1
Life cycle: retired, no children	0.42	0.49	0	1
Homeowner	0.89	0.31	0	1
Home type: attached	0.15	0.36	0	1
Home type: apartment	0.05	0.23	0	1
Climate region: northeast	0.14	0.35	0	1
Climate region: southeast	0.37	0.48	0	1
Climate region: central	0.06	0.23	0	1
Climate region: south	0.16	0.36	0	1
Climate region: east north central	0.05	0.21	0	1
Climate region: west north central	0.03	0.16	0	1
Climate region: southwest	0.05	0.23	0	1
Climate region: west	0.14	0.35	0	1
Climate region: northwest	0.01	0.08	0	1
Household income: less than US$10 k	0.04	0.21	0	1
Household income: US$10 k–US$15 k	0.05	0.21	0	1
Household income: US$15 k–US$25 k	0.11	0.31	0	1
Household income: US$25 k–US$35 k	0.11	0.32	0	1
Household income: US$35 k–US$50 k	0.17	0.37	0	1
Household income: US$50 k–US$75 k	0.18	0.39	0	1
Household income: US$75 k–US$100 k	0.14	0.35	0	1
Household income > US$100 k	0.20	0.40	0	1
Respondent race: white	0.89	0.32	0	1
Respondent race: black	0.05	0.22	0	1
Respondent race: Asian	0.02	0.13	0	1
Respondent race: Hispanic	0.02	0.14	0	1
Respondent race: other	0.02	0.15	0	1
Household has immigrant	0.12	0.32	0	1
Highest household degree: < high school	0.04	0.19	0	1
Highest household degree: high school	0.20	0.40	0	1
Highest household degree: associate’s	0.28	0.45	0	1
Highest household degree: bachelor’s	0.24	0.43	0	1
Highest household degree: graduate	0.24	0.42	0	1
Mean household fuel price per gallon	3.07	0.14	1.56	4.52
Per capita internet deliveries	0.98	2.19	0	100
Kilometers to work per capita	8.24	17.51	0	1493.47
Times worked from home per capita	0.19	1.13	0	30
Number bike trips per capita	0.27	1.02	0	65.75
Number walking trips per capita	4.17	5.54	0	99
Household has self-employed person	0.15	0.36	0	1

To establish continuous measures of contextual density, we use the following gravity model specification:

accessibility = \sum_{i = 1}^{5} β_{i} ln ({population}_{i}),

where population is a proxy for amenities available in distance band i and calculated coefficients β _i are impedance factors, representing the decay of accessibility with distance. This is similar to specifications used in other studies, for example, Handy (1993). In this study, we base distance-band size on percentiles of distance to work for workers in the NHTS data set. The most proximate distance band is a three-kilometer radius circle around each grid cell, with three kilometers being the twenty-fifth percentile of distance to work for workers in the NHTS person file. Additional bands are annuli with outer radii at 13, 26, 61, and 106 kilometer, based on the 50th, 75th, 95th, and 99th percentiles of distance to work. Our objective is to estimate all induced travel effects from nearby land uses, and we assume that distances people drive for work are broadly reflective of willingness to drive for consumption. While 106 kilometer is not “nearby,” using such a large outer band allows us to establish a perimeter beyond which land uses are not statistically significant.

We develop the accessibility variables that proxy for amenity-related land uses in a GIS model. We first estimate 2009 census tract population as a linear function of 2000 tract population and 2010 population. We convert the census tract vector model to a 500-meter-square grid model, where each grid cell value is population, assuming that people are distributed equally throughout each census tract. We calculate population totals in each distance band for each US mainland grid cell. We then estimate distance-band population by census tract as the mean of cell values within each tract. Finally, we use a regression model to estimate coefficients β _i for the natural log of the population sums, as in equation (1).

We find the greatest values for three-kilometer radius population densities in urban centers (Table 3), as one would expect. The greatest values for other bands are found surrounding but not containing urban cores (as illustrated in Figure 2). For example, many city centers have small values of population in their twenty-six to sixty-one kilometer annuli, since areas between twenty-six and sixty-one kilometer from an urban center are often rural. Similarly, census tracts located between twenty-six and sixty-one kilometer from an urban center often have large values of twenty-six to sixty-one kilometer population, since at this distance, a twenty-six to sixty-one kilometer annulus includes population in the city center. For monocentric cities, the high-value distance bands are quite distinct (as illustrated in Figure 2 for Columbus, OH), while for polycentric or closely spaced urban centers, the high-value bands are overlapping and less distinct (as in Figure 2 for the Cleveland, OH, area).

Table 3.

Mean Values of Continuous Contextual Density Measures, by Urbanization Category.

Accessibility	Country	Town	Second City	Suburban	Urban
Log(Population in 3 km radius)	6.95	8.40	10.12	10.07	11.24
Log(Population in 3–13 km annulus)	10.12	10.85	12.00	12.62	13.61
Log(Population in 13–26 km annulus)	11.33	11.60	12.30	13.11	14.09
Log(Population in 26–61 km annulus)	13.06	13.15	13.45	13.95	14.56
Log(Population in 61–106 km annulus)	13.89	13.91	13.86	14.01	14.17

Note: Urbanization categories are from NHTS and Census definition: Country calculated from Town-country = 1 and census Urban = 0; Town calculated from Town-country = 1 and census Urban = 1. Km = kilometer.

Figure 2.

Values of log population for twenty-six to sixty-one kilometer bands in Ohio.

As noted above, we view location decisions as endogenous treatment effects: individuals make joint decisions about where to live and how much to drive in relation to their places of work, leisure interests, and so on. Thus, there are potentially unobserved variables that jointly determine the treatment (settlement type) and the outcome (CO₂ emissions). Such an omitted variable bias is typically referred to as a self-selection problem. For example, residential self-selection could occur due to socioeconomic constraints, if low-income households chose to live in areas with ample public transit (Glaeser, Kahn, and Rappaport 2008). Residential location choices may also reflect desires for particular characteristics of a neighborhood, such as school quality, or proximity to lifestyle amenities. Given this endogenous decision, a simple ordinary least squares regression model is inappropriate, as the “assignment” of individuals to settlement patterns (the treatment) is not random. The typical approach to addressing endogeneity is to identify suitable instrumental variables; in our case, we would need a proxy variable for individual choices of where to live. In addition, we face sample selection biases stemming from individual car ownership decisions: for an individual who chooses not to own a car, we do not observe any automotive CO₂ emissions. Such issues of simultaneous sample selection and endogeneity have been encountered in relation to estimating wage equations for women, especially with an additional confounding factor such as union membership (Amemiya 1985; Maddala 1983). In the travel behavior literature, the methods to deal with simultaneous sample selection and endogeneity have been exceptionally varied (Ewing and Cervero 2010; Mokhtarian and Cao 2008).

We use an FIML estimation approach (Maddala 1983), based on the work of Heckman (1976, 1978). More recent discussions can be found in Cameron and Trivedi (2005) where the models are discussed under the umbrella of a Roy model (Roy 1951). Rather than assuming we have missing data due to sample selection issues—as with Heckman’s sample selection models (Heckman 1976, 1979)—we instead construct an endogenous switching model where we observe different CO₂ emissions for individuals living in and outside of urban areas, and the urban residential location choice is one underlying determinant of CO₂ emissions. Thus, the choice to live in an urban setting is the binary endogenous treatment variable. Cameron and Trivedi (2005) and Wooldridge (2010) also discuss such models more broadly as standard treatment-effect approaches with binary endogenous variables. In our approach, a probit model uses exogenous variables to estimate choice of urban location, which we use in a regression model with natural log of per capita CO₂ emissions as the dependent variable.

The endogenous binary variable model requires a suitable set of exclusionary variables to identify the treatment mechanism, that is, variables affecting location choice, which have little or no impact on the dependent variable (automotive CO₂ emissions). We utilize the individual’s housing-type choice, assuming that housing type is a predominant factor in location choice or that households seek first to fulfill their housing needs. Importantly, this is constrained by the availability of preferred housing in local markets. Our decision to use this broad measure of the local housing market is heavily directed by the availability of suitable exclusionary variables in our data. However, the rich literature on the measurement of quality of life in cities (Blomquist, Berger, and Hoehn 1988; Roback 1982) highlights the pivotal role played by local housing characteristics in location decisions. For this purpose, we use variables from the NHTS that characterize housing type: whether the housing unit is single or multifamily and whether the unit is attached or detached. Our treatment model is thus as described above with complete specification of demographic and geographic variables including the two measures of housing type, and the outcome model contains the same covariates except for the housing-type measures.

Regression Analysis: Results

Table 4 presents the results of our endogenous treatment effects model. The three columns of results include (1) the baseline model without inclusion of agglomeration or accessibility variables, (2) the inclusion of urban population by urban area or urban cluster (Census definition) as a measure of agglomeration, and (3) the set of accessibility measures based on population density. Below we present ranges of estimates across the three specifications, finding broad consensus in the effects of control variables. Importantly, the validity of our approach is verified by the statistical significance of the correlation of the error term between the outcome and the treatment equation, as indicated by ρ (second to last row in Table 4). Further, the broad congruence of results between models 2 and 3, agglomeration versus accessibility, respectively, supports our approach of exploring a more nuanced understanding of the gravitational pull of urban amenities.

Table 4.

Endogenous Treatment-regression, MLE: Log, per Capita CO₂ Emissions; Treatment: Living in an Urban Location.

	Baseline	Agglomeration, ln(population)	Accessibility, ln(population)
Household size	−0.18**	−0.17**	−0.18**
Number of drivers in household	0.25**	0.23**	0.22**
Number of workers in household	0.063*	0.072*	0.087**
Mean age of drivers	0.021*	0.019*	0.019*
Mean age of drivers square	−0.00019^*	−0.00017*	−0.00017*
Life cycle: no children (omitted)
Life cycle: youngest child 0–5 years	−0.11	−0.10	−0.13
Life cycle: youngest child 6–15 years	−0.047	−0.026	−0.047
Life cycle: youngest child 16–21 years	−0.15^*	−0.14	−0.14
Life cycle: retired, no children	−0.072	−0.043	−0.071
Homeowner	0.54**	0.54**	0.50**
Climate region: northeast (omitted)
Climate region: southeast	0.50**	0.46**	0.44**
Climate region: central	0.34**	0.28**	0.27**
Climate region: south	0.46**	0.41**	0.38**
Climate region: east north central	0.52**	0.46**	0.42**
Climate region: west north central	0.55**	0.31**	0.25**
Climate region: southwest	0.50**	0.47**	0.41**
Climate region: west	0.79**	0.80**	0.82**
Climate region: northwest	0.42**	0.38**	0.37**
Household income: < US$10 k (omitted)
Household income: US$10 k–US$15 k	0.59**	0.58**	0.63**
Household income: US$15 k–US$25 k	0.97**	0.98**	1.02**
Household income: US$25 k–US$35 k	1.06**	1.09**	1.12**
Household income: US$35 k–US$50 k	1.19**	1.22**	1.26**
Household income: US$50 k–US$75 k	1.35**	1.41**	1.43**
Household income: US$75 k–US$100 k	1.30**	1.37**	1.40**
Household income > US$100 k	1.31**	1.41**	1.44**
Respondent race: white (omitted)
Respondent race: black	−0.50**	−0.45**	−0.43**
Respondent race: Asian	−0.20	−0.21	−0.21
Respondent race: Hispanic	0.045	0.061	0.069
Respondent race: other	−0.25	−0.26	−0.24
Household has immigrant	−0.10	−0.052	−0.031
Highest household degree: < high school (omitted)
Highest household degree: high school	0.16	0.25	0.20
Highest household degree: associate’s	0.19	0.28	0.25
Highest household degree: bachelor’s	0.15	0.26	0.24
Highest household degree: graduate	0.16	0.27	0.25
Mean household fuel price per gallon	−0.51	−0.54	−0.60*
Per capita internet deliveries	0.011*	0.014*	0.012*
Kilometers to work per capita	0.16**	0.17**	0.14**
Times worked from home per capita	−0.00026	0.00085	0.0016
Number bike trips per capita	−0.067**	−0.068**	−0.066**
Number walking trips per capita	−0.022**	−0.025**	−0.023**
Household has self-employed person	0.13**	0.10*	0.091*
Average treatment effect: urban	−1.24**	−1.07**	−0.91**
Agglomeration: ln (urban population, thousands)		−0.060**
Accessibility: population
ln (population in 3 km radius)			−0.11**
ln (population in 3–13 km annulus)			−0.042
ln (population in 13–26 km annulus)			−0.020
ln (population in 26–61 km annulus)			0.075**
ln(population in 61–106 km annulus)			−0.043*
Constant	6.78**	7.12**	8.41**
Observations	114994	104871	114586
ρ	0.12**	0.12**	0.11**
Log-likelihood	−188,429,261.4	−175,293,879.2	−186,206,683.6

Note: MLE = Maximum Likelihood Estimation; km = kilometer.

*p < .05.

**p < .01.

Across the three models, we find that a one-person increase in household size is associated with a 17–18 percent decrease in CO₂ emissions, reflecting the economies of scale as family size increases. An additional driver in the household increases CO₂ emissions by 22–25 percent, while an additional worker increases CO₂ emissions by 6–9 percent. We also find that increasing the mean age of drivers in the household by one year increases CO₂ emissions by 2 percent, but this effect begins to diminish after the mean age reaches approximately fifty-five years (as per the coefficient on the age square term).

Interestingly, we find that households at every life cycle stage do not have different emissions than adults with no children (the omitted category), except for having children in the age-group of sixteen to twenty-one years results in 14 percent⁴ fewer emissions than households without children, when we do not control for agglomeration or accessibility. Homeowners are found to have 50–54 percent greater automotive CO₂ emissions than nonowners.

For the regional binary variables, we use the Northeast as the reference region and find that households across the nation have significantly higher CO₂ emissions in all specifications. The West has the largest CO₂ emissions at 120–127 percent greater than households in the Northeast, with other regions exhibiting 31–73 percent more than households in the Northeast.

Categorical variables for household income show that greater incomes lead to greater automotive CO₂ emissions. Compared to households with incomes less than US$10,000, households earning US$10–15,000 emit 79–88 percent more CO₂ and those earning more than US$100,000 emit 271–322 percent more. This is an expected result, assuming that driving is an externality of consuming normal goods. This is also a direct reflection of the geographic colocation of households of different socioeconomic status: it is typical for the poorest households to be living in locations where they have ready access to public transportation (Glaeser, Kahn, and Rappaport 2008). With regard to CO₂ emissions and race, we find that blacks emit less CO₂ than whites, on the magnitude of 35–39 percent. This may be an additional reflection of the geographical locations of black households that is not captured by income measures.

The presence of one or more immigrants in a household has no significant effect on automotive CO₂ emissions. Education level also has little effect on automotive CO₂ emissions. In the model that includes our accessibility measure, the mean household fuel price per gallon is significant where a US$1 increase in fuel price results in a 60 percent drop in CO₂ emissions. We find that per capita internet deliveries increases CO₂ emissions by 1 percent. As expected, distance to work increases emissions by 14–16 percent per kilometer, cube root.⁵ Increasing the number of bike trips decreases emissions by 7 percent, while increasing walking trips decreases emissions 2 percent. Working from home has no apparent effect on CO₂ emissions. Being self-employed raises emissions by 10–14 percent.

Turning to our main variables of interest, greater accessibility decreases automotive CO₂ emissions, as we expected. At distances up to three kilometers, greater population density is associated with greater accessibility and with lower emissions. In this radius, a 1 percent increase in population results in a 0.11 percent decrease in CO₂ emissions. The effect of population at distances of three to twenty-six kilometers is not statistically different from zero, while population density at twenty-six to sixty-one kilometers creates a significant attractive force that is associated with greater automotive CO₂ emissions. A 1 percent increase in the population at a distance of twenty-six to sixty-one kilometers results in a 0.075 percent increase in CO₂ emissions. At the furthest distances we studied, 61–106 kilometers, we see a return to a reduction in CO₂ emission of 0.043 percent for every 1 percent increase in population. At such distances from population concentrations, households may meet all their travel needs within smaller areas. Our results for the accessibility measures also offer valuable insight into the nuances of the agglomeration effect rather than the aggregated effect of urban population, which reduces emissions by 6 percent (model 2, agglomeration variable).

We find that the exclusion restriction variables in the treatment equation—residential type—are consistently significant,⁶ providing one validation for our endogenous-treatment-effects modeling approach. We also find that across all our specifications, the outcome and treatment equations are not independent: the residuals of the two equations are correlated, as indicated by the p value for the test of whether ρ equals zero (Table 4, second row from bottom). Further, ρ is consistently positive in our specifications within a range of 0.11–0.12. This implies that unobservables that raise CO₂ emissions tend to occur with unobservables that raise the likelihood of living in an urban area. Rather than attracting individuals who would likely have low CO₂ emissions anyway, urban location apparently mitigates the emissions of people who would otherwise tend to have high automotive CO₂ emissions.

We find that the average treatment effect of residing in an urban location is significant at the 1 percent level, a finding that we believe to be a unique contribution to this literature. Living in an urban area, defined in the urban/rural continuum codes developed by Nielsen Claritas (NHTS 2011) significantly decreases automotive CO₂ emissions by 60 percent (model 3, using accessibility variables based on population density). That is, transplanting an average household in our sample from a nonurban area to an urban one would decrease the household’s automotive CO₂ emissions by approximately 60 percent, all else equal. This large effect stems in part from holding constant variables that would otherwise increase the emissions of individuals in urban areas (since ρ > 0, as shown above), and thus reduce observed emissions differences across the rural–urban gradient.

As an indirect measure of the sensitivity of our results to our specification, we examined whether owning a vehicle was a suitable treatment, as 4.8 percent of our sample do not have a personal vehicle. We found results largely consistent on our variables of interest (the accessibility measures). We saw very minor changes in magnitude and significance for some other controls. Further, we also examined the viability of a straight-up selection model (Heckman 1976, 1979) with selection on residential location, framing the outcome as missing data on CO₂ emissions. In our sample, 14.8 percent of urban dwellers do not have CO₂ emissions related to driving a personal vehicle, compared to only 4.63 percent for nonurban dwellers. Again, our results were fairly consistent. A notable divergence was the insignificance of the price of fuel in the Heckman selection model. Therefore, we strongly prefer the model that allows us to identify what may even be considered a behavioral truth: increases in fuel prices lead to reductions in driving. We also validated our approach by checking that results were consistent with a model that specified VMT as our dependent variable, instead of CO₂ emissions. As a final sensitivity test, we created accessibility variables using income density rather than population density, since either may proxy for availability of amenities. Again, results are largely consistent.

While our spatial variables capture effects of increasing distances from urban centers, note that these are straight-line distances that do not account for size or quality of road connections, geographic obstacles, and so on, which likely affect accessibility and could be captured in local or regional travel models. Further, our data set contains no variables to capture driving disamenities like traffic and parking cost, which likely vary between areas of similar density and affect observed travel behavior. It is also important to note that our econometric methodology hinges on the validity of the bivariate normality assumption of the errors of the treatment and outcome equation. While this is a standard assumption, work by Bhat and Eluru (2009) utilizes an alternative based on a multivariate functional form for the joint distribution. Even within their approach, however, different distributional assumptions are found to offer varying outcomes regarding the importance of residential self-selection. While our model passes tests of over identification, which is sufficient for the validity of our approach, the endogenous switching model does not allow for directly testing the identification of the treatment equation through our restriction. Addressing these econometric limitations is beyond the scope of this article but would offer valuable steps forward in expanding our understanding of travel behavior.

Discussion

After controlling for a number of demographic, regional, and behavioral differences, and addressing the underlying residential location selection mechanism, we find that in the United States in 2009, living in and near high-density urban areas was associated with lower CO₂ emissions from car transportation than living in rural and lower-density settings. The average treatment effect of living in an urban area is large and statistically significant, and higher population densities within three kilometers of a household tend to reduce automotive CO₂ emissions. While population and income densities at distances three to twenty-six kilometers do not have statistically significant effects, all else equal, population densities at twenty-six to sixty-one kilometers appear to exert a significant gravitational pull, increasing driving by those outside urban peripheries but within driving distance. The advantage, in terms of reduced emissions, of population density (and associated land uses) for those who live within urban areas is thus partially offset by the increased travel of others into those denser areas. Based on our model, it seems that the outer limit of gravitational pull is sixty-one kilometers. Beyond that distance, population and income concentrations may be so far away that households travel closer to home.

To illustrate the gravitational effect estimated in our model, consider population in twenty-six to sixty-one distance bands in Ohio (Figure 2). High values of population at this distance indicate outer suburb conditions, as illustrated by the light rings around major cities in Figure 2. In Ohio, the highest value for population in a twenty-six to sixty-one kilometer band is found in the northeastern part of the state, near Cleveland, while the minimum value of this variable is in the southeastern part of Ohio, where there are no major cities nearby (Figure 2). Using coefficients estimated in model 3 (Table 4) and Ohio median or modal values for all other variables, the minimum twenty-six to sixty-one kilometer population value results in a predicted value of 5.22 metric ton automotive CO₂ per capita, while the maximum value results in a prediction of 6.31 metric ton. This suggests that the largest difference in twenty-six to sixty-one kilometer population found in Ohio has a gravitational effect resulting in a 1.09 metric ton increase in CO₂, or a 21 percent increase from the minimum case. Since this is the largest possible difference in Ohio, gravitational effects in all other Ohio locations are below this bound.

There are at least three possible explanations for the reduced automotive CO₂ emissions of those in and near dense areas. The theoretical explanation posited above is that households in denser areas are able to obtain utility from consumption at closer proximities than households in less dense areas. Another plausible explanation is that our accessibility measures do not completely capture urban driving disamenities, such as reduced average driving speed and difficulty of parking, which may reduce urban driving and automotive CO₂ emissions. A third possibility is that vehicle fuel efficiency changes systematically with density, that is, living in denser areas may promote use of smaller and more fuel-efficient vehicles. However, our results from models of vehicle distance traveled are largely consistent with our models of CO₂ emissions, suggesting that systematic differences in vehicle efficiency are not an important factor.

It is also interesting to note that our results indicate no reduction in emissions due to telecommuting. Early research on this topic found substantial reductions in emissions due to telecommuting, though partially offset by noncommute trips (Koenig, Henderson, and Mokhtarian 1996). More recent work (Zhu and Mason 2014) has found that telecommuting does not have any benefits with regard to VMT and overall emissions. In light of these recent findings, including our own, this is a topic that deserves significant interest especially as telecommuting continues to be viewed as one avenue of reducing personal travel emissions.

Since the 2009 NHTS, electric and plug-in hybrids have become more available, using electric motors that are far more efficient than internal combustion engines, and using renewable energy to charge electric vehicles would avoid most CO₂ emissions from vehicle operation. Thus, a technological path exists for greatly reducing automotive CO₂ emissions and their differences across the rural–urban gradient.

Conclusions and Policy Implications

In this study, we find that greater population densities significantly decrease automotive environmental impact for those living in denser areas, though based on a gravity effect, denser population clusters also induce some additional travel from less-dense areas nearby. Using CO₂ emissions as a dependent variable captures this impact more accurately than VMT, since CO₂ emissions reflect both VMT and any spatial differences in vehicle fuel efficiency. While other studies have found positive environmental effects of population density in comparing different urban and suburban areas, we extend these findings to the entire rural–urban continuum. Our gravity model captures effects of densities at different proximities using new continuous measures of contextual density rather than simple categorical measures. We use an endogenous binary treatment model to account for self-selection of nondriving lifestyles as well as endogeneity of driving and housing decisions. These enhancements provide additional confidence in our results.

At −0.11, the population density elasticity in the most proximate three kilometer distance band is more than double the −0.04 elasticity value reported in a metastudy using nine estimates of this figure (Ewing and Cervero 2010). The endogeneity correction and other enhancements in our approach appear to make population density a more important factor in driving behavior than previously believed.

While spatial effects on automotive CO₂ emissions are consistent and significant, we also show that they may not be the most important variables from a policy perspective. Population density appears to create tendencies toward greater or lesser vehicular CO₂ emissions, but these tendencies are not deterministic. We observe many households in every settlement type with very low (and very high) automotive CO₂ emissions, and there is much overlap in interquartile ranges of emissions by urbanization extent (Figure 1).

While public policy may significantly affect future spatial arrangement of the population and associated CO₂ emissions, existing settlement patterns represent investments that society cannot easily forego. Even if a household should choose to move from a rural area to an urban one to reduce its CO₂ footprint, it is likely that another household would then move into the vacant rural home. Such financial realities ensure that spatial effects of existing settlement patterns will not change quickly. But planning can affect CO₂ emissions for decades and centuries hence, and the effects of population spatial patterns in already-developed areas can provide valuable lessons for future construction in less-developed areas.

Our study suggests that in a country like the United States, urbanization and population densification are likely to have positive effects with respect to CO₂ emissions from personal automobiles, though this conclusion might not hold in countries at all levels of industrialization. For example, we find greater incomes associated with greater emissions. If moving to an urban area means a substantial increase in income (perhaps enough to buy a car), urbanization in this case might be associated with greater rather than lesser CO₂ emissions. Our study also supports other studies finding large differences in automotive use between different urban contexts (Glaeser and Kahn 2010; Newman and Kenworthy 1999), providing examples of urban areas with lesser emissions. Of course, in most places, population spatial patterns have developed over very long periods of time. While urbanization clearly affects emissions of greenhouse gases like CO₂ in the long run, urbanization may not present the easiest policy lever to pull.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Amemiya

1985. Advanced Econometrics. Cambridge, MA: Harvard University Press.

Anderson

W. P.

Kanaroglou

P. S.

Miller

E. J.

. 1996. “Urban Form, Energy and the Environment: A Review of Issues, Evidence and Policy.” Urban Studies 33:7–35.

Bento

A. M.

Cropper

M. L.

Mobarak

A. M.

Vinha

. 2005. “The Effects of Urban Spatial Structure on Travel Demand in the United States.” Review of Economics and Statistics 87:466–78.

Bhat

C. R.

Eluru

. 2009. “A Copula-based Approach to Accommodate Residential Self-selection Effects in Travel Behavior Modeling.” Transportation Research Part B: Methodological 43:749–65.

Blomquist

G. C.

Berger

M. C.

Hoehn

J. P.

. 1988. “New Estimates of Quality of Life in Urban Areas.” The American Economic Review 78:89–107.

Boarnet

Crane

. 2001. “The Influence of Land Use on Travel Behavior: Specification and Estimation Strategies.” Transportation Research Part A: Policy and Practice 35:823–45.

Boarnet

M. G.

Sarmiento

. 1998. “Can Land-use Policy Really Affect Travel Behaviour? A Study of the Link between Non-work Travel and Land-use Characteristics.” Urban Studies 35:1155–69.

Brownstone

Golob

T. F.

. 2009. “The Impact of Residential Density on Vehicle Usage and Eenergy Consumption.” Journal of Urban Economics 65:91–98.

Cameron

A. C.

Trivedi

P. K.

. 2005. Microeconometrics: Methods and Applications. New York: Cambridge University Press.

10.

Cao

Mokhtarian

P. L.

Handy

S. L.

. 2009. “Examining the Impacts of Residential Self-selection on Travel Behaviour: A Focus on Empirical Findings.” Transport Reviews 29:359–95.

11.

Cervero

Kockelman

. 1997. “Travel Demand and the 3Ds: Density, Diversity, and Design.” Transportation Research Part D: Transport and Environment 2:199–219.

12.

Cox

N. J.

2011. “Stata Tip 96: Cube Roots.” Stata Journal 11:149–54.

13.

Crane

2000. “The Influence of Urban Form on Travel: An Interpretive Review.” Journal of Planning Literature 15:3–23.

14.

Crane

Crepeau

. 1998. “Does Neighborhood Design Influence Travel?: A Behavioral Analysis of Travel Diary and GIS Data.” Transportation Research Part D: Transport and Environment 3:225–38.

15.

Cullen

J. B.

Levitt

S. D.

. 1999. “Crime, Urban Flight, and the Consequences for Cities.” Review of Economics and Statistics 81:159–69.

16.

EPA. 2014a. Emission Factors for Greenhouse Gas Inventories. Accessed April 5, 2014. http://www2.epa.gov/sites/production/files/2015-07/documents/emission-factors_2014.pdf.

17.

EPA. 2014b. Inventory of U.S. Greenhouse Gas Emissions and Sinks: 1990-2012. Accessed July 1, 2014. http://www3.epa.gov/climatechange/Downloads/ghgemissions/US-GHG-Inventory-2014-Main-Text.pdf.

18.

Ewing

Cervero

. 2010. “Travel and the Built Environment--A Meta-analysis.” Journal of the American Planning Association 76:265–94.

19.

Frank

L. D.

Pivo

. 1994. “Impacts of Mixed Use and Density on Utilization of Three Modes of Travel: Single-occupant Vehicle, Transit, and Walking.” Transportation Research Record 1466:44–52.

20.

Frey

W. H.

2013. A Big City Growth Revival? Accessed February 1, 2014. http://www.brookings.edu/research/opinions/2013/05/28-city-growth-frey.

21.

Glaeser

E. L.

Kahn

M. E.

. 2001. “Decentralized Employment and the Transformation of the American City.” National Bureau of Economic Research w8117, Cambridge, MA.

22.

Glaeser

E. L.

Kahn

M. E.

. 2010. “The Greenness of Cities: Carbon Dioxide Emissions and Urban Development.” Journal of Urban Economics 67:404–18.

23.

Glaeser

E. L.

Kahn

M. E.

Rappaport

. 2008. “Why Do the Poor Live in Cities? The Role of Public Transportation.” Journal of Urban Economics 63:1–24.

24.

Glaeser

E. L.

Kolko

Saiz

. 2001. “Consumer City.” Journal of Economic Geography 1:27–50.

25.

Glaeser

E. L.

Gottlieb

J. D.

. 2006. “Urban Resurgence and the Consumer City.” Urban Studies 43:1275–99.

26.

Gordon

Richardson

H. W.

. 1989. “Gasoline Consumption and Cities: A Reply.” Journal of the American Planning Association 55:342–46.

27.

Handy

1993. “Regional Versus Local Accessibility: Implications for Nonwork Travel.” Transportation Research Record 1400:58–66.

28.

Heckman

1976. “The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models.” Annals of Economic and Social Measurement 5:475–92.

29.

Heckman

1978. “Dummy Endogenous Variables in a Simultaneous Equation System.” Econometrica: Journal of the Econometric Society 46:931–59.

30.

Heckman

1979. “Sample Selection as a Specification Error.” Econometrica: Journal of the Econometric Society 47:153–61.

31.

Heilig

G. K.

2012. World Urbanization Prospects the 2011 Revision. New York: United Nations, Department of Economic and Social Affairs (DESA), Population Division, Population Estimates and Projections Section.

32.

Kahn

M. E.

2000. “The Environmental Impact of Suburbanization.” Journal of Policy Analysis and Management 19:569–86.

33.

Kahn

M. E.

2006. Green Cities: Urban Growth and the Environment. Cambridge, UK: Cambridge University Press.

34.

Khattak

A. J.

Rodriguez

. 2005. “Travel Behavior in Neo-traditional Neighborhood Developments: A Case Study in USA.” Transportation Research Part A: Policy and Practice 39:481–500.

35.

Kim

Brownstone

. 2013. “The Impact of Residential Density on Vehicle Usage and Fuel Consumption: Evidence from National Samples.” Energy Economics 40:196–206.

36.

Koenig

B. E.

Henderson

D. K.

Mokhtarian

P. L.

. 1996. “The Travel and Emissions Impacts of Telecommuting for the State of California Telecommuting Pilot Project.” Transportation Research Part C: Emerging Technologies 4:13–32.

37.

Levinson

D. M.

Kumar

. 1997. “Density and the Journey to Work.” Growth and Change 28:147–72.

38.

Maddala

G. S.

1983. Limited-dependent and Qualitative Variables in Econometrics. Cambridge, UK: Cambridge University Press.

39.

Marshall

W. E.

Garrick

N. W.

. 2012. “Community Design and How Much We Drive.” Journal of Transport and Land Use 5:5–21.

40.

Mokhtarian

P. L.

Cao

. 2008. “Examining the Impacts of Residential Self-selection on Travel Behavior: A Focus on Methodologies.” Transportation Research Part B: Methodological 42:204–28.

41.

Mokhtarian

P. L.

Salomon

Redmond

L. S.

. 2001. “Understanding the Demand for Travel: It’s Not Purely Derived.” Innovation: The European Journal of Social Science Research 14:355–80.

42.

Newman

P. W. G.

Kenworthy

J. R.

. 1989. “Gasoline Consumption and Cities.” Journal of the American Planning Association 55:24–37.

43.

Newman

P. W. G.

Kenworthy

J. R.

. 1999. “Costs of Automobile Dependence: Global Survey of Cities.” Transportation Research Record: Journal of the Transportation Research Board 1670:17–26.

44.

NHTS (National Household Travel Survey). 2011. “2009 National Household Travel Survey User’s Guide. Version 2.” Accessed July 1, 2013. http://nhts.ornl.gov/2009/pub/UsersGuideV2.pdf.

45.

Puhani

2000. “The Heckman Correction for Sample Selection and Its Critique.” Journal of Economic Surveys 14:53–68.

46.

Raphael

Stoll

M. A.

. 2010. Job Sprawl and the Suburbanization of Poverty. Washington, DC: Metropolitan Policy Program at Brookings.

47.

Rees

Wackernagel

. 1996. “Urban Ecological Footprints: Why Cities Cannot be Sustainable- and Why They Are a Key to Sustainability.” Environmental Impact Assessment Review 16:223–48.

48.

Roback

1982. “Wages, Rents, and the Quality of Life.” Journal of Political Economy 90:1257–78.

49.

Roy

1951. “Some Thoughts on the Distribution of Earnings.” Oxford Economic Papers 3:135–46.

50.

Shen

2000. “Spatial and Social Dimensions of Commuting.” Journal of the American Planning Association 66:68–82.

51.

US Census Bureau.2010. Growth in Urban Population Outpaces Rest of Nation, Census Bureau Reports. Washington, DC: US Census Bureau.

52.

US Department of Transportation. 2009. Federal Highway Administration, National Household Travel Survey 2009. Accessed July 1, 2013. http://nhts.ornl.gov.

53.

Wooldridge

J. M.

2010. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT press.

54.

Zhu

Mason

S. G.

. 2014. “The Impact of Telecommuting on Personal Vehicle Usage and Environmental Sustainability.” International Journal of Environmental Science and Technology 11:2185–200.

Contextual Density and US Automotive CO 2 Emissions across the Rural–Urban Continuum

Abstract

Keywords

Introduction

Literature Review

Automotive CO2 Emissions: Spatial Variation across the Rural–Urban Continuum

Regression Analysis: Data and Methods

Regression Analysis: Results

Discussion

Conclusions and Policy Implications

Footnotes

Declaration of Conflicting Interests

Funding

Notes

References

Automotive CO₂ Emissions: Spatial Variation across the Rural–Urban Continuum