American Economic Review 2025, 115(5): 1439–1484

https://doi.org/10.1257/aer.20230346

The Role of People versus Places

in Individual Carbon Emissions†
By Eva Lyubich*

There is substantial spatial heterogeneity in household carbon emis-

sions. I leverage movers in two decades of administrative Decennial
Census and American Community Survey data to estimate place
effects—the amount by which carbon emissions change for the same
household living in different places—for almost 1,000 cities and
roughly 61,500 neighborhoods across the United States. I estimate
that place effects account for 1­ 4–23 percent of overall heterogeneity.
A change in n­ eighborhood-level place effects from 1 standard devi-
ation above the mean to 1 below would reduce household carbon
emissions from residential energy and commuting by about 40 per-
cent. (JEL D12, D15, L94, Q41, Q54, R11, R23)

Increased carbon and other greenhouse gas emissions since the onset of the
industrial revolution have caused global average temperatures to rise by over 1°C
(1.8°F) relative to preindustrial levels (NASA 2020). In 2015, the United States
signed the Paris Accord, a global agreement aimed at mitigating the potential dam-
ages from climate change by limiting overall warming to below 2°C. In search of
opportunities for decarbonization, researchers and policymakers have pointed to
substantial spatial heterogeneity in household carbon emissions, suggesting that
perhaps ­higher-emissions places could adopt features of l­ower-emissions places,
such as density and h­ igh-quality public transportation infrastructure, in order to
lower household carbon emissions (e.g., Jones and Kammen 2014; International
Energy Agency 2021; Pomponi et al. 2021; Wagner 2021).

*US Census Bureau (email: [email protected]). Rema Hanna was the coeditor for this article. Any opin-
ions and conclusions expressed herein are those of the author and do not represent the views of the US Census
Bureau. The Census Bureau’s Disclosure Review Board and Disclosure Avoidance Officers have reviewed this data
product for unauthorized disclosure of confidential information and have approved the disclosure avoidance practices
applied to this release (Release authorization numbers: ­CBDRB-FY24-CES014-016, ­CBDRB-FY24-CES014-017,
and ­CBDRB-FY24-CES005-017). I am grateful to Reed Walker, Pat Kline, Emmanuel Saez, and Joe Shapiro for
their guidance and support throughout this work. I also thank the editor and three anonymous referees for their con-
structive feedback. This article has benefited tremendously from comments and suggestions from Alan Auerbach,
Matilde Bombardini, Stephanie Bonds, Severin Borenstein, David Card, Lucas Davis, Karl Dunkle Werner, Ben
Faber, Meredith Fowlie, Cecile Gaubert, Kendall Houghton, Hilary Hoynes, Jon Kolstad, Enrico Moretti, Andrés
­Rodríguez-Clare, Raffaele Saggio, Jim Sallee, Elif Tasar, Danny Yagan, Katherine Wagner, Randall Walsh, Chris
Walters, and seminar participants at the University of California, Berkeley. I am grateful to the National Science
Foundation Graduate Research Fellowship Program (DGE 1752814), the Berkeley Opportunity Lab and Smith
Richardson Foundation, and Washington Center for Equitable Growth for financial support. The collection of data
used in this study was partly supported by the National Institutes of Health under grant number R01 HD069609
and R01 AG040213, and the National Science Foundation under awards SES 1157698 and 1623684. I thank Matt
Mullins for editing assistance.
†
Go to https://doi.org/10.1257/aer.20230346 to visit the article page for additional materials and author
disclosure statement(s).

1439
1440 THE AMERICAN ECONOMIC REVIEW MAY 2025

However, differences in mean carbon emissions across places reflect a combina-

tion of local amenities, household characteristics, and t­aste-based sorting. The rel-
ative contributions of these pieces is a central determinant of whether ­place-based
interventions that change urban form would lead to significant reductions in carbon
emissions. For instance, if places with large s­ingle-family homes and ­car-oriented
transportation infrastructure are h­ igh-emissions because the people who live there dis-
like m ­ ultifamily homes and public transit, then deregulating zoning or building new
rail lines would have little impact on household emissions. Conversely, if the lack of
denser housing and transit options is a constraint on household choices, rather than a
reflection of their preferences, then interventions that change these local public ame-
nities have the potential to decrease carbon emissions for many households at once.
In this paper, I decompose variation in household carbon emissions into a com-
ponent driven by household characteristics and a component driven by place effects,
that is, the amount by which the same household’s carbon emissions would differ
from place to place due to differences in the underlying features of those places.
To do so, I construct a longitudinal panel of residential and transportation energy
use for over 1 million individuals from 20 years of ­restricted-access Decennial
Census and American Community Survey (ACS) (US Census Bureau 2000, 2010,
2001–2019). The longitudinal nature of these data allows me to link individual
survey respondents over time and across places. I use a mover design, examining
changes to household carbon emissions for over 250,000 movers across roughly
1,000 cities and 61,500 neighborhoods, to estimate place effects and their contribu-
tion to heterogeneity in carbon emissions.
I begin my analysis by documenting observational patterns of city- and
­neighborhood-level variation in household carbon emissions in my sample. While
previous work has shown evidence of substantial variation in household carbon
emissions (e.g., Jones and Kammen 2014; Ummel 2014; Green and Knittel 2020),
the level of geographic granularity in publicly available data has limited researchers
to predicting ­neighborhood-level carbon emissions from national data projected onto
local place and household characteristics. In contrast, detailed geographic identifiers
in the administrative Census Bureau microdata make it possible to directly estimate
­neighborhood-level means. I estimate that on average, households living in cities
with emissions 1 standard deviation above the mean emit 50 percent more than those
living in cities 1 standard deviation below the mean. Similarly, households in neigh-
borhoods 1 standard deviation above the mean emit, on average, just under twice as
much as those in neighborhoods 1 standard deviation below the mean. Accounting
for variation driven by observed household characteristics such as household size
and income decreases the dispersion across place estimates by less than 10 percent.
The heterogeneity that remains after accounting for observable household char-
acteristics reflects some combination of unobserved household characteristics and
causal place effects. Unobserved household characteristics might include prefer-
ences for spending time in a private yard versus a public park, risk tolerances for
biking versus driving, sensitivities to hot or cold temperatures, or simply environ-
mental consciousness. Place effects could stem from a variety of local amenities
and ­supply-side factors that determine patterns of household energy use. They
could reflect aspects of urban form such as public transportation, bike and pedes-
trian infrastructure, highway networks, density, or zoning regulations. They could
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1441

also be driven by natural amenities such as climate. Lastly, they could arise from
­supply-side factors that determine fuel shares and electricity emissions factors, both
of which shift the amount of carbon emitted for a given level of energy use.1 I show
how place effects can be interpreted through the lens of a consumer energy demand
model in which average energy demand, energy demand elasticities, energy prices,
and average emissions factors vary across places.
My empirical strategy uses movers to estimate the contributions of place effects
and household characteristics to heterogeneity in household carbon emissions. The
mover design accounts for unobserved differences between households by compar-
ing carbon emissions for the same household living in different places. Consider the
following thought experiment. Imagine two households identical in every observ-
able way—same household size, same income, same education levels, etc. One lives
in a dense, urban neighborhood ­well-served by public transit, while the other lives
in a ­car-dependent suburb with large homes. The urban household likely generates
lower carbon emissions. But is that because the place itself determines emissions or
because the urban household has different unobservable characteristics, like stronger
environmental preferences, that led it to select into the lower emissions neighbor-
hood? If I observe the same household move between neighborhoods, any resulting
change in that household’s emissions can no longer be attributed to fixed unobserved
household preferences or characteristics, and I can use those changes to infer differ-
ences in place effects. In order for estimates from the mover design to be unbiased,
the central assumption is that mover destinations are uncorrelated with changes to
unobserved determinants of household carbon emissions. A crucial advantage of
undertaking this analysis with Census Bureau microdata is that I observe, and can
control for, many t­ime-varying household characteristics that could correlate with
both potential emissions and destination choices and thereby confound estimates of
place effects. In other words, the identifying strategy does not allow for households to
move in response to a sudden shift in environmental preferences, but it does allow for
households to move in response to changes in income, the number of children, and
­age-based or life cycle preferences, because I observe these characteristics.
For the first set of results, I use an event study to estimate how much carbon emis-
sions change after households move, as a share of the mean difference between their
origin and destination. Intuitively, if spatial heterogeneity is driven in part by under-
lying differences between places, when a household moves, I should see its emissions
shift toward the mean of its new location. The larger the shift, the more important
the role of place. I find that, on average, when households move to a new city, their
carbon emissions change by about 85 percent of the mean difference between origin
and destination cities. Sorting plays a larger role in n­ eighborhood-level variation
than in variation across cities, but the role of place remains meaningful; when house-
holds move to a new neighborhood, their carbon emissions change by ­53–60 percent
of the mean difference between their origin and destination neighborhoods.

1
For instance, a household in an area with predominantly ­coal-fired power plants could emit significantly more
carbon than a household with identical electricity consumption in an area powered mainly by natural gas or renew-
able sources. Similarly, many households in the Northeast use heating oil, in part due to legacy equipment choices,
which produces substantially more carbon dioxide for the same amount of heating than natural gas.
1442 THE AMERICAN ECONOMIC REVIEW MAY 2025

I explore several dimensions of heterogeneity, finding that my estimates remain

stable when restricting the sample to households without significant changes in
observable characteristics, when splitting the sample by duration between obser-
vations, and when grouping movers based on the magnitude of o­ rigin-destination
differences in observational mean emissions. The consistency of estimates across
these analyses suggests limited systematic sorting of households to places, as such
sorting would manifest in a n­ onlinear relationship between observational means and
household emissions. In the absence of systematic sorting, the event study estimates
can be interpreted as causal estimates of the effect on emissions of any household
moving between any pair of places. While the heterogeneity analysis lends support
to this interpretation, the additional assumption of no systematic sorting is quite
strong, particularly at the neighborhood level. Under the weaker baseline assump-
tions, the event study estimates yield unbiased predictions about how household
carbon emissions change for any set of observed moves. This is valuable, as it makes
it possible to calculate the carbon emissions externality of policies or regulatory
restrictions that drive existing patterns of household migration, for instance, policies
that restrict housing supply in o­ n average lower-emissions cities.
In the second set of results, I estimate the full ­nonparametric distribution of
household and place effects using a t­wo-way fixed effects model and then do a
variance decomposition to estimate the share of overall heterogeneity explained by
each component. This approach allows for unrestricted patterns of sorting, but this
weaker assumption comes at the cost of limited mobility bias (Andrews et al. 2008):
Estimates of place effects are noisy because they can be derived from a small num-
ber of movers to and from each place. This imprecision creates an upward bias in
the naïve ­plug-in variance estimate relative to the true variance of place effects,
even if estimates of place effects themselves are unbiased. I account for this upward
bias using the h­ eteroskedasticity-robust “­leave-out” estimator proposed by Kline,
Saggio, and Sølvsten (2020). I find low correlations between unobserved house-
hold and place effects, even at the neighborhood level. This suggests that sorting
on unobserved household characteristics contributes to differences between places
through “segregation” of households, but not in a way that is systematically cor-
related with unobserved neighborhood attributes. City effects explain ­14–16 percent
of overall heterogeneity, while neighborhood effects explain roughly ­22–23 percent
of overall heterogeneity. Climate, electricity emissions factors, and energy prices
together account for over half of the core-based statistical area (CBSA) variance
component. At the neighborhood level, controlling for these factors decreases the
place share by less than half, to about 15 percent. While this leaves the majority of
overall heterogeneity to other factors, my estimates nevertheless imply the potential
for considerable reductions to household carbon emissions from interventions that
decrease place effects: I estimate that if a neighborhood went from having a place
effect 1 standard deviation above the national mean to having a place effect 1 stan-
dard deviation below the national mean, household emissions for residents of that
neighborhood would decrease by about 40 percent.
I characterize low- and h­ igh-emissions neighborhoods by presenting correlations
between estimated tract effects and observable t­ract-level characteristics. I char-
acterize local amenities using observable characteristics from within the Census
Bureau microdata; p­ ublic-use data on climate and electricity; and commercial data
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1443

from Walk Score, a private company that generates estimates of the ­walkability,
­transitability, and b­ ikeability of every address in the United States (Walk Score
2021). These data lend unique insight into highly granular variation in neighbor-
hood characteristics. I find that the correlations between amenities and neighbor-
hood place effects for the most part mirror the relationships in the observational
data. ­Low-emissions places have clean electricity and mild climates, and they have
amenities characteristic of urban areas. Among the local amenities, density and
average home size, proximity to principal cities, and the quality of local bike infra-
structure appear to have the most explanatory power.
I conclude my analysis by examining how household carbon emissions would
differ if some households were exposed to the place effects of neighborhoods more
urban than the one they currently live in. If suburban and rural households lived in
a place with the average place effect of the principal city closest to them—a sce-
nario that captures in spirit how households’ exposure to place effects might shift
in response to regulations that limit urban sprawl and encourage ­upzoning and infill
development—I estimate that their emissions from residential energy use and com-
muting would decrease by about 15 percent. To put this estimate into context, the
Inflation Reduction Act (IRA), which was signed into law in August 2022 and is
the largest federal effort to address climate change to date, is projected to decrease
­economy-wide emissions in 2030 by an additional 15 percent relative to projected
reductions from 2005 levels under business as usual.2
In many ways, the basic physical design and urban fabric of cities, suburbs, and
towns creates the foundational patterns of transportation and residential energy use.
Decades of housing, transportation, and land use policies shape these features. The
wide distribution of place effects estimated in this paper implies that there may be
potential for “­place-based climate policies”—policies that aim to reduce household
carbon emissions from residential and transportation energy by changing the under-
lying characteristics of the places people live in—to lead to meaningful reductions
in carbon emissions. While this paper does not estimate the causal drivers of place
effects, the correlational analysis presented here and many observational and mod-
el-based studies (e.g., Shammin et al. 2010; Timmons, Zirogiannis, and Lutz 2016;
Ribeiro, Rybski, and Kropp 2019; Pomponi et al. 2021; Ko 2013) provide some
hypotheses. Estimating causal relationships between specific amenities and place
effects using credible exogenous variation remains an important direction for future
research. One might have worried about whether the settings in which such stud-
ies can be done are too selected; what if the households that reside in places that
make changes to local infrastructure or regulatory restrictions are different from
households in places that do not make those changes? The core results of this paper
suggest that variation in place effects drives a meaningful share of differences in
emissions between places, mitigating some of these concerns about external validity.

2
Three separate efforts to model IRA reductions have been commonly cited by advocates and lawmakers. The
Rhodium Group estimates that under the IRA, emissions will fall 3­ 1–44 percent from 2005 levels by 2030, with
­24–34 percent reductions under business as usual (Larsen et al. 2022). For the same time frame, Energy Innovation
estimates 3­ 7–41 percent reductions under the IRA and 24 percent under business as usual (Mahajan et al. 2022),
and the REPEAT project estimates 42 percent reductions under the IRA and 17 percent under business as usual
(Jenkins et al. 2022).
1444 THE AMERICAN ECONOMIC REVIEW MAY 2025

This paper makes several contributions. A large body of work in labor and
urban economics finds significant wage, employment, and productivity benefits
from density and integrated land use and transit policies (Tsivanidis 2022; Allen
and Arkolakis 2022; Duranton and Puga 2020). These studies suggest that spatial
equilibria are inefficient due to agglomeration economies and other externalities but
have largely not considered the carbon emissions externality in their analysis. The
theoretical justification for using ­place-based policies in cases where agglomeration
economies and other local externalities exist is w ­ ell-established. Federal interven-
tion can correct inefficient market equilibria and improve welfare by supplement-
ing local government provision of ­underprovided amenities, offering a “big push”
toward an optimal equilibrium when several exist, fostering the growth of produc-
tive areas and agglomeration externalities or insuring residents against ­place-based
shocks (Glaeser and Gottlieb 2008; Kline 2010; Kline and Moretti 2014; Glaeser
2013; Austin, Glaeser, and Summers 2018). However, the empirical evidence on the
efficacy of ­place-based policies is mixed. While some studies find that tax incen-
tives targeting areas with lower employment can improve welfare (Busso, Gregory,
and Kline 2013; Austin, Glaeser, and Summers 2018; Bilal 2023), others suggest
that spatial policies promoting growth in less developed areas may have negligible
or even negative aggregate effects on productivity and welfare (Kline and Moretti
2013; Gaubert 2018). Duranton and Venebles (2021) highlight the challenges in
evaluating ­place-based policies, including in the context of urban transport, hous-
ing, and infrastructure. The welfare implications of ­place-based climate policies
would crucially depend on their design, implementation details, costs, and house-
hold preferences for local amenities; a welfare analysis is beyond the scope of this
paper. Carbon emissions are a canonical example of a global externality, but many
of the theoretical justifications for p­ lace-based policy outlined above could be rele-
vant, given the observed relationship between carbon emissions and factors of urban
form, such as density and urban transport. This, together with evidence from this
paper that places play an important role in driving household carbon emissions,
suggests that further research is warranted.
Methodologically, I build on a large literature in labor examining wage inequal-
ity across firms and a growing literature that uses mover designs to estimate place
effects on other individual outcomes, e.g., nutritional choices (Allcott et al. 2019),
health outcomes and health care utilization (Eid et al. 2008; Finkelstein, Gentzkow,
and Williams 2016, 2021), intergenerational mobility (Chetty and Hendren 2018),
and wages (De la Roca and Puga 2017; Card, Rothstein, and Yi 2025). This paper is
the first to use a mover design to study household energy use and carbon emissions,
yielding new insights into spatial heterogeneity in these outcomes. Previous work
has highlighted the consequences of spatial heterogeneity in carbon emissions for
allocative efficiency (Glaeser and Kahn 2010; Colas and Morehouse 2022) as well
as for distributional impacts and the political economy of hypothetical climate pol-
icies (Cronin, Fullerton, and Sexton 2019; Sallee 2019; Green and Knittel 2020),
but these papers did not examine the causal role of places in their findings. Several
papers in the literature have generated estimates of heterogeneous energy demand
parameters, but they have necessarily done so in spatially limited and s­ ector-specific
contexts (Auffhammer and Rubin 2024; Gillingham 2014; Nowak and Savage 2013;
Spiller et al. 2014). The estimates generated in this paper on the relative roles of
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1445

place effects versus household sorting could assist in resolving some of the chal-
lenges identified in the aforementioned literature around distributional impacts
and political economy. For instance, if l­ock-in of urban form limits the share of
household carbon emissions that could be targeted by pricing instruments in the
­short-to-medium term, estimates of the share of spatial heterogeneity driven by
place effects could inform decisions around how much to redistribute carbon div-
idends when using geography as a tag. And as highlighted previously, they raise
an important question as to whether p­ lace-based climate policies could serve as a
­welfare-improving complement to traditional instruments for addressing this global
externality.

I. Data and Stylized Facts about Carbon Emissions in the United States

Seventy-five percent of US greenhouse gas emissions are from burning fossil

fuels. Of these, 20 percent are from residential energy use (including electricity),
and another 20 percent are from light-duty (i.e., passenger) vehicles (US Energy
Information Administration 2020b). The focus of this analysis is on carbon emis-
sions from these two sectors; they are a meaningful portion of overall emissions, and
they are the forms of emissions that are most directly related to ­locked-in charac-
teristics of urban form. In the remainder of this section, I first discuss my data and
the construction of relevant analysis variables and then provide some descriptive
evidence on heterogeneity in carbon emissions in my sample.

A. Data and Key Variables

I build a 2­ 0-year panel of individual and h­ ousehold-level data using the 2000
restricted-access Decennial Census long form and the ­ 2001–2019 American
Community Survey (US Census Bureau 2000, 2010; 2001–2019). The 2000
Decennial Census long form consists of a stratified random sample covering one
in six households in the United States. After 2000, the ACS replaced the Decennial
Census long form in order to gather detailed information on individuals and house-
holds more regularly. The ACS is a stratified random sample covering roughly
0.4 percent of households in ­2001–2005 and roughly 1 percent of households in
each year after 2005 (US Census Bureau 2014). I link individuals across surveys
using protected identification keys, which are unique person identifiers assigned by
the US Census Bureau based on names, addresses, dates of birth, other household
members, and social security numbers (when available).3
For every individual in the panel, I observe measures of residential and trans-
portation energy use and a rich set of demographic, household, workplace, and
home characteristics, including detailed geographic identifiers. I supplement the

3
Neither the Decennial Census nor the ACS asks respondents for their social security numbers. Wagner and
Layne (2014) use data with social security numbers to show that the error rate in assigning protected identification
keys without social security numbers is below 1 percent. See Bond et al. (2014) for detailed discussion of the
assignment algorithm used by the US Census Bureau. Assignment success rates vary across demographic groups—
in particular, White and higher-income individuals are more likely to be successfully assigned a protected identifi-
cation key—but for all demographic subgroups, the success rate is greater than 85 percent. See Bond et al. (2014)
for additional discussion of the variation in assignment rates across population subgroups.
1446 THE AMERICAN ECONOMIC REVIEW MAY 2025

Decennial Census and ACS with several external sources of data in order to convert
energy expenditures to energy services and emissions, and to characterize places.

Geographic Units of Analysis.—Throughout the study, I analyze spatial hetero-

geneity at two levels of geographic granularity that roughly represent a city or labor
market and a neighborhood.
My first geographic unit of analysis is a core-based statistical area. CBSAs are
designated by the Office of Management and Budget and cover the population of
metropolitan and micropolitan areas in the United States. Each CBSA is a set of
contiguous counties with strong commuting ties and at least one urban core area of
at least 10,000 people. In addition to formally designated CBSAs, I define ­state-level
residual CBSAs from unassigned rural areas. My second geographic unit of anal-
ysis is a census tract. Census tracts are county subdivisions that typically cover
contiguous areas, have populations of 1,­200–8,000 people (4,000 on average), and
are delineated with boundaries that follow identifiable physical features. They are
designed to be relatively stable but are split or merged every 10 years if populations
exceed or fall below the 1,­200–8,000 window.4

Carbon Emissions.—My primary outcome is metric tons of carbon emissions

from residential energy and passenger vehicle use, which together account for
roughly one-third of US greenhouse gas emissions. I implement the main analysis
at the household level: Carbon emissions are given by household residential emis-
sions plus the sum of individual commuting emissions over all individuals in the
household.
I estimate carbon emissions from residential energy use from ­household-reported
expenditures on electricity, natural gas, and other home heating fuels in the last
year, combined with external data on local annual retail prices and fuel emissions
factors. For electricity, I calculate county-level average prices using data from the
Annual Electric Power Industry Report (US EIA 2020a). This report contains sales,
revenues, and total customers for every major utility in the United States, by sector
and state. It also delineates counties contained in each utility’s service territory.
I calculate ­county-level retail electricity prices using ­customer-weighted average
prices (revenue divided by sales) across all utilities with service territories con-
taining the county, and I compute household electricity consumption by dividing
reported expenditures by my price estimates. I then assign households to one of 12
National Electric Reliability Council (NERC) regions using a t­ract-level crosswalk
from the the Infrastructure Foundation-Level Database (US DHS 2021), and com-
pute emissions using the average annual emissions rates assigned to each region
by the Emissions and Generation Resource Integrated Database (US EPA 2021).
For natural gas and other home heating fuels, I obtain average retail prices at the
state level from the State Energy Data System (US EIA 1960-2019). If a household
reports ­nonzero expenditures on “other home heating fuels,” I impute the fuel used

4
Census geographic definitions vary over time to account for changes in administrative boundaries and pop-
ulations. To ensure that I don’t erroneously identify people who live in places where the designation changed as
movers, I use the 2­ 000–2010 census block concordance to assign 2010 geographic definitions to all years in the
data, combining blocks in cases where they correspond to a single 2000 block.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1447

from its answer to the question, “What was the primary fuel used for home heating?”
Finally, I obtain fuel emissions factors from the Emission Factors for Greenhouse
Gas Inventories (US EPA 2018).
I estimate carbon emissions from transportation energy from the sum of
­individually reported commuting behavior within a household.5 I estimate com-
mute distance using the geodesic distance between home and place of work census
blocks, and I estimated commute speed from estimated mileage and reported t­ime
length of commute. I estimate gasoline usage using annual national average fuel
economy from the Bureau of Transportation Statistics (2000–2019) accounting for
the fact that in general, fuel economy is roughly 30 percent higher on highways than
in cities. Finally, I estimate the number of annual commutes using reported weeks
worked last year and hours worked last week and convert annual gallons of gaso-
line to carbon emissions using the motor gasoline emissions factor from the State
Energy Data System (US EIA 1960–2019). Individuals who commute by rail, sub-
way, streetcar, bus, bike, or walk and individuals who work from home are assigned
an emissions factor based on their mode of transit and data from the National Transit
Database (Federal Transit Administration 2002–2019) (see Supplemental Appendix
for more details). Altogether, this portion of the outcome captures variation in car-
bon emissions driven by commute lengths, number of commutes, and mode of tran-
sit. I examine the sensitivity of my results to using the National Household Travel
Survey (NHTS, Federal Highway Administration 2022) to predict heterogeneous
fuel economy and n­ oncommute miles from household and geographic character-
istics available in both the census and NHTS. This is not my baseline approach,
as it infers how much of variation in vehicle fleets and fuel economy observed
in the NHTS is driven by individual preferences versus p­ lace-based factors from
­cross-sectional variation.6

Individual and Household Characteristics.—Throughout the analysis, I use

demographic and household characteristics to control for variation driven by
observable characteristics. Specifically, I control for age, education (completion of
a bachelor’s degree), sex, race and ethnicity, household income (from salaries and
wages, interest, social security, supplemental security, public assistance, retirement,
and self-employment), household size, number of children, and homeowner status.
I aggregate ­individual-level demographics to the household level by taking the mean
across individuals within a household.
I also observe whether a household lives in a detached ­single-family home, the
number of rooms in a home, and the number of vehicles in the household. These char-
acteristics are intermediate outcomes, which directly affect carbon emissions. They
also very likely reflect a combination of household preferences and place character-
istics; many places impose restrictions on m
­ ultiunit homes and/or minimum lot sizes
and lack transportation options for households without a car, so household choices

5
Commuting accounts for about 28 percent of all ­vehicle-miles traveled and 39 percent of ­person-miles traveled
on transit systems (US Department of Transportation 2015), which means I underestimate carbon emissions from
overall personal vehicle use for most people in my sample.
6 ­
Place-based factors that contribute to variation in vehicle fleets could include social norms, perceptions of
safety (e.g., if everyone around you is driving a big car, it is safer for you to drive a big car; certain types of cars
may be able to handle adverse weather better), road widths, ease of parking, etc.
1448 THE AMERICAN ECONOMIC REVIEW MAY 2025

are likely to differ from place to place depending on these constraints. Therefore, I
do not treat these variables as observable household characteristics when estimating
place and household effects, but I do use them later to explore correlates of unob-
served place and household heterogeneity.
It is not obvious whether homeownership should be considered an observable
household characteristic or part of a place effect. Homeownership rates vary dra-
matically across CBSAs (Raetz 2021; Mateyka and Mazur 2021), and housing
regulations can price people out of homeownership. In these cases, treating home-
ownership as an intermediate outcome seems appropriate. On the other hand, the
choice to become a homeowner simultaneously impacts where people live and fac-
tors related to their carbon emissions. For instance, homeowners may want extra
space for potential family expansion or ­space-intensive leisure activities or may
choose homeownership in order to be able to install solar panels and have a place
to charge their electric vehicle. To be conservative and err on the side of finding a
smaller role of place effects, I treat homeowner status as an observable household
characteristic in my baseline analysis.

Place Characteristics and Amenities.—I supplement Census Bureau m ­ icrodata

with several external sources of data to characterize amenities at the block, tract,
city, and regional level. I focus on amenities that are directly relevant to energy con-
sumption and carbon emissions in the residential and transportation sectors.
To capture variation in climate, I use data on annual heating degree days
(HDDs) and cooling degree days (CDDs) at the CONUS Climate Division level
(National Oceanic and Atmospheric Administration 2020). The National Oceanic
and Atmospheric Administration (NOAA) divides the contiguous states into a total
of 344 climate divisions based on regional differences in climates within states.
Degree days represent the annual sum of the daily difference between that day’s
temperature and 65°F and are meant to quantify the heating and cooling require-
ments of a place.
To account for ­neighborhood-level variation in transportation and leisure ameni-
ties, I use data from Walk Score, a private company that generates estimates of the
­walkability, t­ ransitability, and ­bikeability of every address in the United States (Walk
Score 2021).7 Walk Score rankings capture proximity to different commercial ame-
nities, such as grocery stores, as well as street characteristics, such as block lengths
and intersection widths. Bike Score indexes capture characteristics that make biking
more or less accessible, such as the existence of bike lanes, road connectivity, and
hilliness. Transit Score ratings capture proximity to different types of transit and
the frequency and connectivity of nearby options. For transit, I also observe the
number of bus routes and rail routes within one-­half mile. Other than route counts,
each score is an index from ­0–100. I assign over 6 million unique walk score points
reflecting data from late 2020 through through early 2021, one to every populated
census block in the United States, by matching census block centroids to the nearest
walk score l­atitude-longitude coordinate.

7
Data can be viewed at www.walkscore.com and were provided by Redfin Real Estate (www.redfin.com).
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1449

Finally, I estimate density at the ­tract level using 2010 census ­block-level infor-
mation on area and population. I define urban tracts as those that are characterized
as urban by the US Census Bureau and surpass the density threshold set for urban
centers by the ­EU-OECD definition of a functional urban area (Dijkstra, Poelman,
and Veneri 2019).8 I define suburban tracts as those contained within a CBSA but
not designated as urban. Tracts outside of CBSAs are classified rural.

Sample Restrictions.—I restrict the analysis to individuals who are at least 18

years of age, who are not identified as the householder’s child or grandchild, and
who are not missing any of the outcome variables or key explanatory or control
variables described above. I also impose several additional restrictions related to
energy variables. I exclude from the sample households for which residential energy
costs are included in rent or gas costs are included in electricity bills because I
don’t observe expenditures in those cases. I discuss this s­ ubsample of households
and the potential impact of its exclusion in detail in the Supplemental Appendix. I
also exclude individuals in households in which residential energy use is top-coded
or whose commute time is top-coded, as the t­op-coding will obfuscate changes in
individual consumption for the highest-demand individuals. Lastly, I exclude indi-
viduals if the sum of their household residential energy expenditures is 0, if they
are in the bottom 1 percent of ­nonzero residential energy cost observations, or if
they are in the top 1 percent of commute distance observations, as these outliers
more likely reflect survey misreporting. My full sample consists of all individu-
als who meet these restrictions across the 48 continental states and the District of
Columbia—almost 17 million people across over 12 million households (Table 1,
column 1). I use the full sample to estimate observational geographic and household
heterogeneity.
I construct a panel sample by restricting the full sample to individuals for whom
I have at least two observations in which they did not indicate that they had moved
within the last year. This restriction on very recent migration ensures that I am esti-
mating household carbon emissions at the correct location, as households report
their residential energy expenditures over the past year.9 The panel sample consists
of 1,097,000 people across 916,000 households (Table 1, column 2).
Finally, I impose two additional sample restrictions, which are necessary for the
implementation of my empirical strategy. ­First, because residential energy is deter-
mined at the household level and place effects are identified from the variation in
outcomes of movers between places, I restrict the panel sample to only households
consisting of the same full-sample individuals across observations. This ensures that
changes in emissions across observations are not driven by changes in the com-
position of adults in the household. Importantly, this approach does retain house-
holds where children are born or grow up and move out between observations, as
individuals under 18 are not part of the full sample. Second, I restrict CBSAs and
tracts to the “­leave-out connected set”—the network of CBSAs or tracts that remain
connected to each other by at least one mover after I drop all the observations of

8
This threshold is 1,500 people per square kilometer.
9
This restriction applies to all ACS years. The 2000 Decennial Census asked whether respondents had moved
within the last five years. Since this is significantly more restrictive, I don’t drop these individuals.
1450 THE AMERICAN ECONOMIC REVIEW MAY 2025

Table 1—Sample Statistics

Panel sample Mover sample

Full All CBSA Tract CBSA Tract
(1) (2) (3) (4) (5) (6)
Panel A. Demographics
College 0.25 0.25 0.25 0.25 0.35 0.31
Age 44 46 46 46 43 43
White 0.82 0.89 0.89 0.90 0.89 0.88
Female 0.48 0.48 0.48 0.48 0.45 0.47
Household income 103,400 114,100 114,200 114,900 116,100 115,400
Household kids 1.0 1.0 1.0 1.0 1.0 1.0
Household size 2.8 2.9 2.9 2.9 2.8 2.9
Homeowner 0.75 0.85 0.85 0.85 0.72 0.73

Panel B. Outcomes
Tons CO​​2​ ​​​ 15.1 16.1 16.0 16.1 15.0 15.0
Tons CO​​2​ ​​​—residential 11.9 12.8 12.8 12.8 11.8 11.9
Tons CO​​2​ ​​​—commute 3.3 3.3 3.3 3.3 3.3 3.1

Panel C. Intermediate outcomes

Detached home 0.72 0.81 0.81 0.81 0.73 0.73
Use electricity only 0.30 0.24 0.24 0.24 0.30 0.28
Commute by car 0.94 0.96 0.96 0.96 0.95 0.96
Commute minutes 25.1 24.9 24.9 24.8 26.0 25.6

Panel D. Place characteristics

Urban 0.29 0.22 0.23 0.22 0.20 0.25
Suburban 0.62 0.62 0.63 0.63 0.68 0.67
Rural 0.09 0.15 0.15 0.15 0.11 0.08
Walk score 27.2 23.0 23.1 22.4 23.0 25.2
Bike score 35.9 33.4 33.4 33.1 34.1 35.4
Transit score 9.0 6.7 6.7 6.3 6.7 7.9
N bus routes 1.6 1.1 1.1 1.0 1.2 1.3
N rail routes 0.15 0.09 0.09 0.07 0.10 0.10
Cooling degree days 1,359 1,213 1,215 1,205 1,356 1,335
Heating degree days 4,376 4,824 4,815 4,851 4,494 4,518

N people 16,900,000 1,097,000 1,073,000 1,042,000 107,000 290,000

N households 12,600,000 916,000 860,000 833,000 93,000 248,000
CBSAs 1,000 1,000 1,000 1,000 1,000 1,000
Tracts 71,500 70,000 70,000 61,500 54,500 61,500

Notes: Column 1 shows statistics for the full sample. Column 2 shows statistics for the panel sample, with no
restrictions that individuals be in the same household or live in a connected geography. Columns 3 and 4 show the
panel samples restricted to individuals in a consistent household over time and the CBSA and tract l­eave-one-out
connected sets, respectively. Columns 5 and 6 show statistics for the CBSA and tract mover samples. All sample
statistics are weighted using Census Bureau sample weights. Sample counts are unweighted and rounded according
to Census Bureau disclosure rules.

any given household (see the Supplemental Appendix for an illustration). I do this
after dropping tracts with fewer than 10 full-sample household observations. The
­leave-out connected sets are constructed separately at the CBSA and tract levels. This
means it is possible for a household to be in the CBSA panel but not the tract panel.
The ­leave-out restriction drops a negligible share of (residual) CBSAs and roughly
12 percent of (disproportionately rural) tracts, yielding approximately a 5 percent
reduction in the number of households in the sample (Table 1, columns 3 and 4).
CBSA movers are households in the CBSA panel that live in different CBSAs
across observations (93,000 households, column 5), and similarly, tract movers are
households in the tract panel that live in different tracts (within or across CBSAs)
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1451

across observations (248,000 households, column 6). The CBSA panel, tract panel,
CBSA movers, and tract movers make up my four primary analysis samples.

B. Sample Statistics

Table 1 shows sample statistics for the full sample, unrestricted panel sample, the
two geographically restricted panel samples, and the two mover samples.
A comparison across the samples yields three main t­akeaways. First, households
in the panel are on average more likely to be White, have higher income, and are
more likely to be homeowners than households in the full sample (columns 1 and 2).
This reflects known heterogeneity in protected identification key assignment rates
within the Census Bureau (Bond et al. 2014). The panel sample is also 7 percentage
points less likely to live in an urban tract, 9 percentage points more likely to live
in a detached home, and 2 percentage points more likely to commute by car. The
appreciable drop in Transit Score when going from the full sample to the panel is
consistent with these differences in urbanity and is likely in part driven by dispro-
portionately dropping households in the densest areas when dropping households
whose electricity and/or heating is included in rent, as discussed earlier and also
in the Supplemental Appendix. Second, further restricting the baseline panel to the
CBSA and tract panels (columns 3 and 4) does not meaningfully change the distri-
bution of demographics, (intermediate) outcomes, or place characteristics. Finally,
movers (columns 5 and 6) tend to be younger, more college educated, and have
higher income than both stayers and the full sample. Movers also are more likely
than stayers to live in urban tracts, less likely than stayers to live in detached homes,
and they have higher rates of electric heating and lower emissions from residential
energy, making them more comparable to the full sample on all of these dimensions.
Overall, a little under 80 percent of household carbon emissions in my sample are
from residential energy, and a little over 20 percent are from commuting.10 Close to
three-quarters of the sample live in a detached, single-family home, a vast majority
of the sample commutes by car, and on average, households live within half a mile
of only 1 bus route and only 0.1 rail routes.
Table 2 shows additional statistics for the panel samples. I observe the vast major-
ity of households in the panel sample exactly twice, with on average ­8–10 years in
between observations. Movers tend to be younger than stayers the first time I observe
them and are much more likely to have had a child, experience a large change in real
household income11, or transition from renting to owning their home. Households
tend to move to places with higher shares of detached s­ ingle-family homes and worse
­non-car transportation amenities. The majority of moves are from urban to urban
tracts, urban to suburban tracts, or suburban to suburban tracts. Finally, consistent
with secular trends of mobility in the United States, households generally move to

10
Household carbon emissions from residential and transportation energy are roughly evenly split (US Energy
Information Administration 2020b). Given that commuting makes up about 30 percent of transportation energy
emissions, we would expect a slightly higher than 3:1 ratio of residential to commuting energy. My estimates appear
to overstate residential energy use relative to commuting by a few percentage points at most.
11
I define a large change in income as a greater than 0.5 (in absolute value) change in log income. This approxi-
mates a 50 percent increase/decrease in income and corresponds to about the top quarter in absolute income change
in my panel sample.
1452 THE AMERICAN ECONOMIC REVIEW MAY 2025

Table 2—Panel Statistics

Panel sample Mover sample

CBSA Tract CBSA Tract
Panel A. Sample characteristics
First observed in 2000 0.10 0.10 0.16 0.14
Years between observations 7.9 7.9 10.3 9.8

Panel B. Demographic characteristics

Age first observed 42.0 42.0 37.1 37.2
Share with large change in income 0.28 0.28 0.45 0.41
Share with change in N kids 0.45 0.45 0.55 0.55
Change in N kids −0.12 −0.12 0.08 0.08
Share rent to own 0.11 0.11 0.26 0.26

Panel C. Mover place changes

Δ walk score −7.0 −7.1
Δ bike score −4.2 −4.0
Δ transit score −2.3 −2.7
Δ N bus routes −0.57 −0.56
Δ N rail routes −0.04 −0.04
Δ tract share detached home 0.05 0.05
% moves ­urban to urban 0.08 0.14
% moves ­urban to suburban 0.17 0.14
% moves ­suburban to suburban 0.44 0.46
Δ cooling degree days 214 136
Δ heating degree days −331 −188

N people 1,073,000 1,042,000 107,000 290,000

N households 860,000 833,000 93,000 248,000
CBSAs 1,000 1,000 1,000 1,000
Tracts 70,000 61,500 54,500 61,500

Notes: Columns 1 and 2 show panel statistics for the CBSA and tract panel samples. Columns 3 and 4 show
panel statistics as well as summary measures of mobility patterns for the CBSA and tract mover samples.
All sample statistics are weighted using census sample weights. Sample counts are unweighted and rounded
according to Census Bureau disclosure rules.

warmer places. For additional comparisons of movers versus stayers, estimates of

the likelihood of moving given shocks to household income, number of children, or
changes in homeownership, and the full set of transition probabilities across urban,
suburban, and rural places, see the Supplemental Appendix.

C. Observational Heterogeneity

Carbon emissions from residential energy and commuting vary immensely

across individuals in the full sample (see Supplemental Appendix). This varia-
tion is strongly correlated with both geographic and household attributes. At the
regional level, carbon emissions vary with climate and with the emissions inten-
sity of local fuel sources. At the more local level, many have observed a relation-
ship between emissions and local amenities that characterize urban form, such as
local public transit, bike infrastructure, green space, and density (Ou et al. 2013;
Philips, Anable, and Chatterton 2022).
Figure 1 presents average household carbon emissions across urban, suburban,
and rural neighborhoods. Households residing in suburban and rural areas have sub-
stantially higher emissions than those living in urban areas. While those differences
could, in theory, result from sorting based on observable household characteristics
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1453

20

Mean household CO2 (tons)

Without controls
With controls
15

10

0
Urban Suburban Rural

Figure 1. Household Carbon Emissions in Urban, Suburban, and Rural Places

Notes: This figure shows estimates of household carbon emissions by urbanity. Estimates are derived from regres-
sions of log carbon emissions (in metric tons) on indicators for urban, suburban, and rural tracts, with year fixed
effects. The specification with controls additionally includes age, gender, race, education, household size, number
of children, and homeowner status. I define urban tracts according to Dijkstra, Poelman, and Veneri (2019). I define
suburban tracts as n­ onurban tracts within a ­CBSA. I define rural tracts as those outside CBSAs. Mean household
carbon emissions are precisely estimated and statistically different from one another. The unconditional regression
has an R​​​​​2​​of 0.11; the conditional regression has an R​​​​​2​​ of 0.29.

such as college education, race and ethnicity, household income, or number of

children, among others (see Supplemental Appendix), controlling for observable
household characteristics only marginally reduces the differences between rural,
suburban, and urban areas. There is some sorting of ­higher-emissions households to
suburban and rural areas and l­ower-emissions households to urban areas, but even
after accounting for these differences, households in suburban tracts still emit over
20 percent more per year than observationally similar households in urban tracts,
while households in rural tracts emit about 50 percent more.
To examine spatial heterogeneity in more detail, I estimate unconditional and
​ j​​​​, using an ordinary least squares regression of log
conditional t­ract-level means, ​μ
of household carbon emissions onto place fixed effects, year fixed effects ​τ​ t​​​​, and in
the conditional regression, individual and household observable characteristics ​X ​ i​t​​​.

(1) ​ln C​O​2it​​  = ​μ​j​(i,t)​​ ​ + ​X​'it​ ​  β + ​τt​​​  + ​εit​ ​​​.

I use an empirical Bayes “shrinkage estimator” to adjust the estimates for statisti-
cal noise (see Supplemental Appendix for details). In practice, the distributions and
relevant moments are almost identical for the adjusted and unadjusted estimates.
Figure 2 presents the adjusted conditional and unconditional distributions of ​​μ​ˆ ​j​​.​​  ­I
estimate that households living in neighborhoods 1 standard deviation above the
mean emit on average approximately 1.9 times more than those living in neighbor-
hoods 1 standard deviation below the mean, or 1.8 times more after accounting for
differences in observable characteristics. For the remainder of this paper, I refer to
means conditional on observable characteristics as “observational means,” follow-
ing the terminology used by Abaluck et al. (2021).
1454 THE AMERICAN ECONOMIC REVIEW MAY 2025

Without controls
With controls

Density

−1 −0.5 0 0.5
Observational tract mean log CO2

Figure 2. Heterogeneity in ­Tract-Level Carbon Emissions

Notes: This figure shows kernel density estimates, using a Gaussian kernel function, of t­ract-level observational
mean emissions. The distributions are censored at the top and bottom 1 percent of observations in order to abide
by Census Bureau Disclosure Avoidance rules. The dotted line labeled “Without controls” corresponds to the dis-
tribution of log CO​​2​ ​​​conditional on year fixed effects only and has a standard deviation of 0.33, while the solid line
labeled “With controls” conditions on observable household characteristics and has a standard deviation of 0.30.
Both distributions are ­de-meaned to match the model with controls. Observable characteristics include age, gender,
race, ethnicity, education, homeowner status, household income, household size, and number of children.

The remaining heterogeneity in observational means reflects some combination

of ­place-based characteristics and unobserved household characteristics. To illus-
trate this, I rewrite equation (1) as a ­two-way fixed effects model:

(2) ​ln C​O​2it​​  = ​α​i​​  + ​ψj​​​  + ​X​'it​ ​  β + ​τt​​​  + ​εi​t​​​,

where ​​ψj​ ​​​represents the ­place-based characteristics and ​α ​ i​ ​​​represents the unob-
served, fixed, household characteristics. Comparing equation (1) and equation (2)
highlights the bias that can arise when inferring place effects from observational
means, as ​​μ​j​​  = ​ψ​j​​  + E​[​αi​ ​​  | i ∈ j]​​. In other words, observational means reflect a
combination of place effects and an average over the unobserved characteristics of
the residents of a place.
The primary objective of this paper is to disentangle these components and quan-
tify the extent to which heterogeneity in household carbon emissions is driven by
unobserved household characteristics and preferences, and how much is driven by
causal place effects, that is, the amount by which the same household’s carbon emis-
sions would differ from place to place due to the underlying features of each place,
holding household characteristics fixed. This motivates the mover design, which
leverages this exact variation.

II. Model

In the previous section, I demonstrated how a ­two-way fixed effects model can
be used to mitigate biases that arise when making inferences based on observational
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1455

means. To provide further insight into the model, I now illustrate how it relates to a
standard way of modeling consumer energy demand and discuss the interpretation
of place and household effects.
Consider a household, ​i​, living in place, ​j​, that consumes quantity ​Q​of energy in
the form of four categories of fuels, ​f​. In the residential sector, it can consume elec-
tricity (​e​), natural gas (​n​), and other heating fuels (​o​). In the transportation sector, it
can consume motor gasoline (​m​).12
Average demand ​​aj​​​​, price elasticities of demand ​​ρ​jf​  ​​, and prices ​​P​fj​​​  are permitted
to vary by place. P ­ lace-based differences in average energy demand and in price
elasticities of demand could stem from a range of fixed and malleable characteristics
of places. These characteristics include climate, local public goods and urban form
(e.g., density, public transit, pedestrian and bike infrastructure, proximity to high-
ways and availability of parking, and proximity to leisure and commercial ameni-
ties), and regulatory characteristics (e.g., zoning restrictions that change the size and
density of homes and building codes that change energy efficiency requirements or
eliminate natural gas hookups). All of these features could potentially shift energy
demand, both on average and in slope.
In addition to ­place-based characteristics, demand also depends on observable
fixed and time-varying household characteristics (such as age, household size, and
income) ​​Xi​t​​​, individual fixed unobserved determinants of demand (such as a person’s
intrinsic risk tolerance for biking on shared roads, aversion to public transit versus
traffic, or relative enjoyment of spending time in their own backyard versus a public
park) ​​αi​​​​, individual ­time-varying unobserved determinants of demand (perhaps an
increased willingness to bike after reading the latest Intergovernmental Panel on
Climate Change report) ​​εi​t​​​, and national annual variation ​​τt​​​​. Taken together, house-
hold demand for residential and transportation energy is given by

(3) ​ln ​Q​it​​  = ​a​j​​  + ​∑ ​ ​​ ​ρ​jf​  ​⋅ ln ​P​fj​  + ​X​'it​ ​  β + ​τt​​​  + ​αi​​​  + ​εi​t​​​.
f∈

If all households within a place used the same proportion of fuel types, it would
be simple to express the above equation, with log carbon emissions as the outcome,
–
in terms of a ­place-based average emissions factor ​​​ϕ​j​​,​​  and to in turn rewrite that
expression as the ­two-way fixed effects model defined in the previous section:

ln C​O​​2it​​​​​  = ln​(​​ϕj​​ ​​​  Qit​ ​​)​​

–
(4)   ​

∑
–
= ​​ln​ 
​     ​ϕj​​ ​​  + ​aj​​​  + ​
   ​ ​​ ​ρ​jf​  ​⋅ ln ​P​fj​ ​​​+ ​X​'it​ ​  β + ​τt​​​  + ​αi​​​  + ​εi​t​​​.
​  

f∈
​ψj​​​

In this simplified setting, place effects ​​ψj​ ​​​capture the combination of ­place-based
variation in five key factors: average energy demand, fuel price elasticities, fuel

12
Electric vehicles are a negligible share of driving in my sample time frame. If someone has an electric vehi-
cle, I o­ verestimate their emissions because the electricity they use to charge their vehicle is included in residential
energy (if they charge at home), but I also assign them gasoline emissions. As electric vehicles (EVs) become a
larger share of the market, observing household vehicle choices, and the p­ lace-based role of EV charging networks
on these choices, will be critical for future research.
1456 THE AMERICAN ECONOMIC REVIEW MAY 2025

prices, electricity emissions factors, and fuel shares. In reality, fuel shares vary not
only across places but also across households. For instance, some states and munici-
palities are attempting to ban natural gas connections for new construction (O’Brien
2023; Cornfield 2023), but even if natural gas is available as an option, some house-
holds may still opt to fully electrify their homes, while others may have a strong
preference for cooking on gas stoves.
In the Supplemental Appendix, I examine how the interpretation of the ­two-way
fixed effects model is affected when fuel shares are allowed to vary across individu-
als. The more complex model includes a term that captures the interaction between
fuel emissions factors and a household’s fuel shares relative to the p­ lace-based
average. Because electricity emissions factors vary across places, a household that
disproportionately uses electricity will experience a larger drop in emissions when
moving from a place with relatively high-emissions electricity to a place with rela-
tively clean electricity, compared to the average household. This lack of separability
implies that there is some inherent m ­ isspecification in the ­two-way fixed effects
model in this setting.
The model makes two additional simplifications. First, place effects are assumed
to be fixed, implying that any time variation, including changes in prices, is absorbed
in the place effects, which reflect average differences between places over the sam-
ple time frame. Second, while the model allows price elasticities of demand to vary
across places, it does not allow for them to systematically vary across households.
For example, the model does not allow for different elasticities between high- and
­low-income households. Allowing for heterogeneity in demand elasticities across
household characteristics would introduce additional interaction terms in the error,
as elasticities would be interacted with ­place-specific prices. Together, this inter-
action along with the interaction between household fuel shares and ­place-specific
electricity emissions factors motivate treating the errors as heteroskedastic.

III. Empirical Strategy

My empirical strategy uses movers to estimate place effects and their contribu-
tion to spatial heterogeneity in carbon emissions. The intuition behind the mover
design is the following: Suppose ­high-emissions places are ­high-emissions because
of causal place effects—for example, because there are no alternatives to commut-
ing other than by car or because zoning regulations impose constraints on minimum
home sizes and density. If so, there would be households that live in ­high-emissions
places for work or to be near family or near other amenities they enjoy, which may
otherwise make ­ lower-emissions residential and transportation choices but are
unable to. If those households move from an o­ n average h­ igh-emissions place to an
­on average ­low-emissions place where l­ower-emissions alternatives become avail-
able to them, their carbon emissions should decrease. Conversely, if spatial hetero-
geneity is driven by strong preferences, then households currently living in detached
­single-family homes and commuting by car would continue to do so even given
alternate options, and moving from on average high- to ­low-emissions places should
have little effect on household carbon emissions.
I use the mover design to estimate two versions of a heterogeneity decompo-
sition of household carbon emissions. I begin with an event study, which, under
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1457

strong assumptions, characterizes the share of differences between places attrib-

utable to place effects, inferred from movers’ changes in emissions relative to
­origin-destination mean differences. Under weaker assumptions, the event study
serves as additional descriptive evidence and model validation for the second decom-
position, as well as unbiased prediction of how carbon emissions will change for the
set of observed moves. I then estimate the ­nonparametric distribution of household
and place effects and decompose overall heterogeneity into variance components. In
the following subsections, I first discuss modeling and identifying assumptions, and
then I describe each of these decompositions in more detail.

A. Main Assumptions

My empirical strategy at its foundation pairs a two-way fixed effect model with
a mover design. In order for estimates from this approach to be unbiased, three
assumptions need to hold: (i) additive separability of place effects, or constant
effects; (ii) n­ onpersistence of outcomes; and (iii) exogenous mobility, or conditional
orthogonality. I discuss each below.

ASSUMPTION 1: Additive Separability of Place Effects, or Constant Effects.

A core modeling assumption of the t­wo-way fixed effect design is that the out-
come—log carbon emissions—is additively separable in household and place
effects. The log specification is statistically appealing, as it reflects the approximate
­log-normality of the household carbon emissions distribution (see Supplemental
Appendix). It is also conceptually appealing, as it implies that place effects increase
and decrease carbon emissions proportionally by the same amount for everyone,
which aligns with many potential mechanisms through which place effects could
arise. For example, it is natural to model climate as scaling residential heating or
cooling needs up or down by the same factor for all households, regardless of their
baseline energy consumption. If density drives place effects, it is reasonable to
expect denser places to decrease the size of homes (and therefore residential energy
requirements) or the length of commutes (and therefore transportation energy
requirements) by the same factor for all households. Similarly, an increase in trans-
portation alternatives to cars might decrease the share of trips taken by car for all
households proportionally.
Nevertheless, the t­ wo-way fixed effects model imposes a substantial restriction: It
does not allow for heterogeneous treatment effects or match effects. Heterogeneous
place effects could arise in various scenarios. For instance, a new public transit
option might significantly reduce emissions for relatively low-emissions house-
holds, while barely affecting ­high-emissions households’ behavior. Paired with
changes in density, this could even produce divergent effects if high-emissions
individuals find themselves idling in traffic more due to increased congestion.
Alternatively, heterogeneous place effects might arise if low-emissions households
make the same consumption choices wherever they live, while high-emissions
households respond strongly to changes in amenities. Moreover, Section II reveals
that there is an interaction between households’ fuel shares and local emissions
factors; a household that prefers to use electricity for heating will see a larger drop
1458 THE AMERICAN ECONOMIC REVIEW MAY 2025

in emissions when moving to an area with cleaner electricity than a household that
prefers natural gas heating, all else equal. This interaction introduces a degree of
misspecification into the ­two-way fixed effect model, rendering it an approximation.
The critical question is whether households systematically sort into locations based
on this (or any other) interaction. As long as they do not, the mover design will con-
tinue to yield unbiased estimates of average place effects.
To rule out selection on heterogeneous effects, I follow Card, Heining, and Kline
(2013) and test whether moves from a ­low-emissions place to a ­high-emissions
place and moves from a­ high-emissions place to a l­ow-emissions place are associ-
ated with equal and opposite changes in household carbon emissions. To see why
symmetry rules out selection on heterogeneous effects, consider differences in
potential outcomes across an origin o​ ​and destination d​ ​, allowing now for there to
be an interaction ​η(​ ​αi​​​, ​ψj​​​)​​between person and place types:

E[​ C​O2​it​​​(d)​]​  − E​[C​O2​it​​​(o)​]​ = ​(​ψd​​​  − ​ψ​o​​)​  + η​(​αi​​​, ​ψd​​​)​  − η​(​α​i​​, ​ψo​​​)​​.

​

As long as

(5) ​η(​ ​αh​​​, ​ψd​​​)​  − η​(​α​h​​, ​ψo​​​)​  ≠ η​(​α​l​​, ​ψd​​​)​  − η​(​αl​​​, ​ψo​​​)​​,

this type of interaction, paired with selection, would lead to asymmetries between
changes in household carbon emissions for moves to higher ­on average places ver-
sus moves to lower ­on average places. Note that condition (5) holds for a broad class
of functions, including the simplest interaction, ​η = ​ α​i​​  ⋅ ​ψ​j​​​, since ​α
​ ​i​​  ⋅ ​(​ψd​ ​​  − ​ψ​o)​​ ​ 
≠ ​α​j​​  ⋅ ​(​ψd​ ​​  − ​ψ​o​​)​ ∀ i ≠ j, o ≠ d​.13 Returning to an earlier example, suppose
place effects are due to a public transit option that only l­ow-emissions households
use. Suppose also that households that wish to use public transit disproportionately
move to places where it is available, whereas households that don’t want to use
public transit disproportionately move to places where it is not available; the subse-
quent decline in emissions for households moving from a h­ igh-emissions place to
a ­low-emissions place will be larger than the increase in emissions for households
moving in the opposite direction.
I group places into four quartiles based on observational averages of carbon emis-
sions, and I estimate household carbon emissions for each ­origin-destination quar-
tile pair, adjusting for annual trends and controlling for demographic and household
characteristics. Figure 3 displays the results. For parsimony, the figure shows only
moves from the lowest-quartile emissions places to all four quartiles and vice versa,
as well as moves within fi ­ rst-quartile places and moves within f­ ourth-quartile places
as bounds in gray.
First and foremost, this figure shows that moves across quartiles lead to equal
and opposite changes in household carbon emissions. This symmetry suggests that
the ­log-linear t­wo-way fixed effect model of household carbon emissions is a good
approximation for the role of place effects and alleviates concern about selection on
heterogeneous treatment effects. Second, this figure provides evidence of selection

∂η ​ ​2​η
∂
13
More generally, equation (5) implies that ​_​∂ α ​(​ψd​​− ​ψo​​)​ ≠ 0​, and in turn, ​_
​∂ α∂ ψ ​ ≠ 0​. In words, the sym-
metry check rules out any interaction in which the change between places grows or shrinks with household type.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1459

Panel A. CBSA Panel B. Tract

3 3
Mean household log CO2

Mean household log CO2

2.75 2.75

2.5 2.5

2.25 2.25

2 2

Before move After move Before move After move

1 to 1 1 to 2 1 to 3 1 to 4
4 to 4 2 to 1 3 to 1 4 to 1

Figure 3. Changes in Household CO​​2​ ​​​When Moving across Quartiles of Mean CO​​2​ ​​​

Notes: This figure shows average household carbon emissions for movers across places classified into quartiles
based on their observational mean carbon emissions in the full sample. Symmetric responses to moves in opposite
directions provide support for the ­log-linear ­two-way fixed effect model of household carbon emissions because
they are consistent with place effects entering log carbon emissions additively. This figure shows only the subset of
moves to and from the ­lowest-emissions places (quartile 1), as well as moves within the ­highest-emissions places
(quartile 4). Estimates are conditional on year fixed effects and the standard set of household characteristics used
throughout this analysis. Estimates are weighted by Census Bureau sample weights.

at the tails, particularly for t­ ract-level moves; households that move from the lowest
quartile to a different place in the lowest quartile have lower emissions on average
than households that move between the lowest quartile and any of the three higher
quartiles. Similarly, households that move between places in the fourth quartile have
higher emissions than those that move between the fourth and the first, though the
difference at the top is less pronounced. Finally, note that households that move
between tracts within the same quartile experience a small increase in emissions,
consistent with the general trend of households moving to o­ n average higher-emis-
sions places.

ASSUMPTION 2: Nonpersistent Outcomes.

As highlighted above, I identify relative place effects from pairwise comparisons

of household carbon emissions between origin and destination,

E[​ C​O2it
​ ​ ​​​(d)​ | ​αi​​​, ​Xit​ ​​, ​τ​t​​]​  − E​[C​O2it
​ ​​​(o)​ | ​αi​​​, ​Xit​ ​​, ​τt​​​]​ = ​ψd​​​  − ​ψ​o​​​.

This expression holds for any two households moving between ​o​and ​d​, which
means that it cannot differ from household to household as a result of differences in
their residential histories. In other words, the expected change for two households
moving between the same origin and destination should be the same even if one of
the households previously lived in Houston, while the other previously lived in New
1460 THE AMERICAN ECONOMIC REVIEW MAY 2025

York. Note, however, that the nonpersistent outcomes assumption does not rule out
that the place someone was born and raised may have a persistent effect on their
preferences and carbon emissions. Because I include household effects in the model
and only include individuals over the age of 18 in the sample, any persistent effect of
place of birth and upbringing on carbon emissions will be captured by the household
fixed effect.

ASSUMPTION 3: Exogenous Mobility or Conditional Orthogonality.

When the first two assumptions hold, the model serves as a reasonable approxi-
mation to the real world, and consequently, random variation in exposure to place
identifies place effects. Thus the final, identifying, assumption necessary for this
empirical approach to be unbiased is that moves are conditionally exogenous; in
other words, household destination choices are not related to changes in unobserved
determinants of carbon emissions:

E​[​ϵ​it​​  | ​αi​​​, ​ψj​​(i,t)​​,​ ​Xi​t​​, ​τt​​​]​  = 0​.

(6) ​

It is important to emphasize here that the t­ wo-way fixed effects model allows for a
broad set of sorting behaviors. First, it allows for unrestricted sorting of households
on fixed or t­ime-varying observable characteristics. A key advantage of the census
microdata is the ability to observe numerous ­time-varying household characteris-
tics that could potentially bias the estimates if unobserved. Factors such as enter-
ing middle age, having children, experiencing a change in income, or becoming a
homeowner are all associated with an increase in energy consumption generally (see
Supplemental Appendix), and the latter three also significantly increase the proba-
bility of moving (see Supplemental Appendix). However, this endogeneity does not
bias the estimates, as I observe age, household size, number of children, household
income, and ­homeowner status and am therefore able to separate the effect of these
characteristics on household emissions from their influence on a household’s choice
of new city or neighborhood.
Second, and crucially, the t­wo-way fixed effect model allows for unrestricted
sorting on fixed unobservable characteristics. In other words, if households have
heterogeneous, but fixed, preferences for neighborhood amenities—for instance,
if someone has a particular distaste for public transit, a strong preference for large
homes, or a particular love for walking or biking—and their choice of what neigh-
borhood to live in reflects those preferences, estimates of place effects are unbi-
ased by this selection because household fixed effects capture these unobserved
but fixed determinants of carbon emissions. The ability to account for, and allow
unrestricted sorting on, these time-invariant unobserved preferences is a critical
benefit of the pairing of the two-way fixed effect model and mover estimation
strategy.
Note also that the fact that I might never observe a very high-emissions household
move to a very low-emissions place is not a problem if you accept Assumption 1,
that places do not have heterogeneous effects on different types of households.
It is sufficient to have a connected network of pairwise moves between places in
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1461

order to identify relative place effects, and household effects are identified relative
to other households within the same place. Consequently, even if I observe only
­lower-emissions households moving from Pittsburgh to New York, and only higher
­emissions-households moving from Pittsburgh to Houston, a comparison of house-
hold emissions within Pittsburgh combined with the observed changes to household
emissions at their destinations identifies the relevant household and place effects. I
have already provided some evidence that place effects do not appear to be heteroge-
neous in the symmetry check shown in Figure 3. I will provide additional evidence
of this when I examine heterogeneity in the event study results, in Section IVA. If
this model assumption is violated and there is selection on heterogeneous treatment
effects, then the place effects I estimate should be thought of as local average treat-
ment effects for the population moving to those places.
Thus, the main threat to identification stems from the possibility that moves cor-
respond to changes in unobserved preferences—either through an idiosyncratic
shock or via preference “drift,” that is, a gradual evolution in preferences that is not
captured by aging. Revisiting the example in the previous paragraph, this would
reflect a scenario in which the two households in Pittsburgh were initially simi-
lar, but then one household becomes increasingly concerned about climate change
and makes lifestyle choices to reduce its carbon emissions, subsequently relocating
to New York to facilitate these choices, while the other household’s preferences
remain unchanged. In this example, the first household’s carbon emissions would
have changed to some extent even if it had stayed in Pittsburgh, and that portion of
the change would be incorrectly attributed to the New York place effect.
A standard approach for ruling out endogenous moves is to test for parallel trends
between movers and stayers prior to the move. Unfortunately, I observe the majority
of my sample only twice, which makes this impossible. Instead, I use data from
the Panel Study of Income Dynamics (PSID; Lyubich 2024), over the same sample
period, and assess whether movers in the PSID exhibit any changes to energy expen-
ditures (defined as monthly spending on electricity, home heating, and gasoline)
prior to their move. I do not know where households move from or to, so I look sep-
arately at households whose energy expenditures are higher after moving and those
whose energy expenditures are lower after moving, controlling for household size,
income, householder age, and year fixed effects. I find no evidence of p­ re-trends for
either group (see Supplemental Appendix).
In Section IVA, I will also show evidence on preference “drift.” If moves were
endogenous to evolving preferences, the selection component, and consequently, the
parameter estimate, would likely increase with the duration since the move. Some
evidence of drift emerges in the baseline analysis, but the magnitude of drift appears
to be quite small and becomes insignificant when looking at a restricted sample with
no observable major life events.

B. Event Study Decomposition

The first decomposition I estimate is an event study, following the approach used
by Finkelstein, Gentzkow, and Williams (2016). The event study begins with the
­two-way fixed effect model defined in equation (2) but summarizes heterogene-
ity with a single parameter rather than the full distribution of J place effects. For
1462 THE AMERICAN ECONOMIC REVIEW MAY 2025

household ​i​moving from origin o​ ​to destination d​ ​, the expected change in carbon
emissions is given by

E​[ln C​O2it
​ ​ ​​​(o)​ | ​αi​​​, ​Xit​ ​​, ​τt​​​]​ = ​ψd​​​  − ​ψ​o​​​.
​ (​​​ d)​  − ln C​O2it

I then express the change in place effects as a share of the differences between
observational means:

​ψd​​​  − ​ψ​o​​
​ψd​​​  − ​ψ​o​​  = ​______ ​⋅ ​ ​​y–​​​​  − ​​y–​​o​​)​​
​​y–d​​​​  − ​​y–o​​​​ ( d
​

≡ ​θ​o,d​​ ⋅ (​ ​​y–​​d​​  − ​​y–​​o​​)​​.

​

Incorporating this expression into the ­two-way fixed effect model yields the event
study equation, which I use to estimate θ​ ​, the share of b­ etween-place differences
attributable to place effects:

ln C​O​2it​​  = ​α​i​​  + ​ψj​​​  + ​τt​​​  + ​X​'it​ ​  β + ​εi​t​​​

(7)   ​

= ​α​i​​  + ​ψo​​​  + ​1​{​ moved}​​​ ⋅ (​ ​ψd​​​  − ​ψ​o​​)​  + ​τt​​​  + ​X​'it​ ​  β + ​εi​t​​​

​

= ​​α̃ i​​ ​​  + ​1​{​ moved}​​​ ⋅ θ ⋅ (​ ​​y–​d​ ​​  − ​​y–​​o​​)​  + ​τt​​​  + ​X​'it​ ​  β + ​εi​t​​​.
​

By characterizing the place share of heterogeneity with a single parameter, ​θ​, the
event study approach vastly reduces the dimensionality of the estimation problem.
This efficiency comes at the cost of an additional assumption: In order for ​​θˆ ​​to reflect
an unbiased, causal parameter, it cannot be correlated with other parameters in the
model. In other words, unlike the ­two-way fixed effect model, this approach does
not permit systematic sorting of certain types of households, either based on observ-
able characteristics or unobservable types, to certain types of places. This is because
it infers place types from observational means; if there were sorting of, for example,
high-type households to high-type places, y​​​–​​d​​  − ​​y–​​o​​​would grow faster than ​ψ
​ ​d​​  − ​ψ​o​​​
as places grow more different, leading to biased estimates.
The stronger restriction on sorting is more plausible at the CBSA level, in
which people are more likely to move for job opportunities or to be close to fam-
ily, than at the neighborhood level, where choices are more likely to be driven by
local amenities. In Section IVA I will provide evidence that estimates of the share
parameter are not heterogeneous along several dimensions, suggesting little bias
from this stronger assumption. However, even under weaker baseline assump-
tions, under which event study results can’t be interpreted as causal, the results
remain informative for two reasons. First, they serve as useful descriptive evidence
and additional model intuition and validation for the Kline, Saggio, and Sølvsten
(2020)—henceforth, KSS—analysis to come. Second, they yield unbiased predic-
tions about how household carbon emissions will change for any set of observed
moves. This is particularly useful, as researchers have highlighted that restrictive
zoning and high cost of living in productive areas drives people to higher-emissions
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1463

locations (Glaeser and Kahn 2010). The labor and urban literatures have identified
that such moves lead to decreased welfare due to loss of agglomeration external-
ities. The event study specification makes it possible to additionally estimate the
carbon emissions externality of these regulatory restrictions.

C. Variance Decomposition

The second decomposition is based on the full set of nonparametric fixed effect
estimates from the model. Specifically, heterogeneity in household carbon emis-
sions can be decomposed as below, lumping ​​τt​​​​with ​​Xit​ ​​​for brevity:

(8) ​var​[​yi​j​​]​  = var​[​ψj​​​]​  + 2 ⋅ cov​[​αi​​​, ​ψj​​​]​  + var​[​αi​​​]​​

+ var​[​X​'it​ ​  β]​+ 2 ⋅ cov​[​αi​​​, ​X​'it​ ​  β]​+ 2 ⋅ cov​[​ψj​​​, ​X​'it​ ​  β]​+ var​[​εi​t​​]​​.

​

This analysis focuses on the first three terms: the variance component of place
effects, the variance component of unobserved person effects, and their covariance,
which captures the spatial heterogeneity that results from systematic sorting on
unobserved preferences. Each variance component describes the share of overall
heterogeneity attributable to the relevant component.
In contrast to the event study decomposition, the ­two-way fixed effects decom-
position allows unrestricted sorting of households across places; it imposes no lim-
itations on the magnitude or sign of the covariance terms. This flexibility comes
at an econometric cost. A w ­ ell-documented challenge to estimating variance com-
ponents in ­two-way fixed effect models is limited mobility bias (Andrews et al.
2008). ­Limited mobility bias arises because place effects are estimated based on
the outcomes of people who move between different locations; however, for any
given place, there might only be a small number of people moving in or out. This
can lead to imprecise estimates, which creates an upward bias in the naïve ­plug-in
variance estimate relative to the true variance of place effects, even if estimates of
place effects themselves are unbiased. To address this, I estimate variance compo-
nents using the ­heteroskedasticity-unbiased ­leave-out estimator proposed by KSS.
The KSS estimator uses a ­leave-out estimate of standard errors to correct estimates
of the variance components for sampling variability.
I implement the ­leave-out estimator at the household level, leaving out all obser-
vations corresponding to a household match. In the mover sample, the KSS esti-
mator is robust to unrestricted heteroskedasticity and serial correlation within each
match. Because it is not possible to leave out matches for stayers without dropping
all their observations, if there is serial correlation in the error term, KSS estimates
of the person variance component in the panel sample are an upper bound on the
true value. See the Supplemental Appendix for additional computational details, and
KSS for a complete discussion of the ­leave-out estimator.

IV. Results

This section presents the core results of my paper: estimates of the share of het-
erogeneity in household carbon emissions attributable to place effects. I begin by
1464 THE AMERICAN ECONOMIC REVIEW MAY 2025

Table 3—Share of Spatial Variation in Mean CO​​2​ ​​​ Attributable to Place Effects

CBSA Tract
(1) (2) (3) (4) (5) (6)
Panel A. Panel sample
Place share of mean difs. 0.91 0.86 0.86 0.78 0.60 0.55
(0.008) (0.007) (0.016) (0.003) (0.003) (0.007)

Observations 1,764,000 1,764,000 633,000 1,710,000 1,710,000 613,000

R2 (adj.) 0.73 0.75 0.77 0.74 0.76 0.77

Panel B. Mover sample

Place share of mean difs. 0.89 0.85 0.84 0.76 0.57 0.53
(0.009) (0.009) (0.020) (0.004) (0.004) (0.009)

Observations 191,000 191,000 36,000 508,000 508,000 102,000

R2 (adj.) 0.64 0.70 0.70 0.69 0.73 0.73

Household controls X X X X
No big life events X X

Notes: This table reports event study estimates of the place share of spatial heterogeneity in household carbon emis-
sions. The place share estimate (​​θˆ ​​) represents the proportion of differences in average carbon emissions ( ​​y–​  ) between
a mover’s origin and destination attributable to place effects. Panel A reports estimates from the panel sample,
while panel B restricts the sample to movers only, allowing for systematic differences between movers and stayers.
Columns 1 and 4 show estimates from CBSA and tract moves, respectively, with only year fixed effects. Columns 2
and 5 add controls for the standard set of household characteristics. Columns 3 and 6 restrict the sample to house-
holds without changes in number of children, income changes exceeding 50 percent, or changes in homeownership.
All estimates use Census Bureau sample weights.

showing results from the event study specification, which—even if the stronger
assumptions on selection are violated—serve as additional descriptive evidence and
can be used to predict how household carbon emissions will change for movers
under existing patterns of mobility. I then present results from the variance decom-
position of the unrestricted t­wo-way fixed effect model. I conclude the section with
a discussion on interpreting the two versions of the analysis.

A. Event Study Decomposition

This section presents estimates from the event study derived in Section IIIB:

​ln C​O​​2it​​​​​  = ​​α̃ ​​i​​  + ​1​{​ moved}​​​ ⋅ θ ⋅ (​ ​​y–​​​d​−i​​​​  − ​​y–​​​o−​ i​​​​)​  + ​τt​​​  + ​X​'it​ ​  β + ​εit​ ​​​,

where ​​​y–​​​j−​ i​​​​​ are sample means estimated from the full sample, leaving out the house-
hold observation.14
Table 3 presents event study estimates of ​​θˆ ​​, which captures the share of differ-
ences between observational place means attributable to differences between place
effects. Columns 1–3 show results from regressions examining CBSAs, while col-
umns 4–6 show results from regressions examining tracts. Panel A shows results

14
To the extent that there is sampling variability in the distribution of observational means, my estimate of the
relationship between o­ rigin-destination mean changes and individual changes in ​log C​O2​​​may be biased. In practice,
using a linear empirical Bayes estimator to adjust means for sampling variability does not materially change the
results.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1465

estimated from the panel sample, which consists of both movers and stayers; panel
B shows results using just the mover sample. Both samples use mover variation to
identify place effects, but estimates may differ if the relationship between household
characteristics and emissions varies systematically between movers and stayers. For
instance, if having children induces divergent preference shocks—prompting mov-
ers to seek larger, potentially ­higher-emission homes, while reinforcing stayers’
preference to be within walking distance of friends and local amenities—including
stayers in the analysis could underestimate the direct impact of children on mov-
ers’ emissions, thereby overstating the place effect. Across specifications, the mover
sample yields place share estimates that are at most 4 percentage points lower than
the panel sample estimates. This implies that movers and stayers have only margin-
ally different responses to changes in observables, and including stayers introduces
a small upward bias in the place effect estimates.
Columns 1 and 4 present estimates from a model with just year fixed effects.
These estimates show that when a household moves, its emissions change by
­89–91 percent of o­ rigin-destination CBSA mean differences and 7­ 6–78 percent
of ­origin-destination tract mean differences. However, these estimates are con-
founded by the correlation between household characteristics and location choices.
Columns 2 and 5 address this by controlling for observable household characteris-
tics (age, household income, household size, number of children, and homeowner
status) that likely influence both emissions and neighborhood choice. This adjust-
ment reduces CBSA share estimates by up to 5 percentage points (to ­85–86 percent)
and tract share estimates by nearly 20 percentage points (to ­57–60 percent). The
larger impact on t­ract-level estimates suggests that, as you would expect, house-
hold sorting plays a more significant role in ­neighborhood-level emissions variation,
while C­ BSA-level moves may be driven more by factors like job opportunities or
family considerations. In the Supplemental Appendix, I ­reestimate columns 2 and
5 controlling for average heating degree days, cooling degree days, and electricity
emissions factors. Doing so decreases the C ­ BSA share to 6­ 8–70 percent and the
tract share to 5­ 1–54 percent. While these estimates lend insight into the mechanisms
underlying place effects, I focus my main event study analysis on total place effects,
which are more relevant for predicting emissions changes from observed migration
patterns.
Even after accounting for the rich set of observable characteristics available in the
data, household destination choices may still be influenced by unobserved heteroge-
neous preference shocks. While I cannot rule this out entirely, I explore the poten-
tial bias from such selection by restricting the sample to households with minimal
observable shocks, motivated by the premise that these households are less likely to
have experienced large concurrent unobservable changes. Columns 3 and 6 present
results from this “no big life events” subsample, which excludes households that
experienced a change in the number of children, a greater than 0.5 log point change
in income, or a change in homeownership status between observations. This sub-
sample, comprising approximately 20 percent of the original observations, yields
place share estimates at most 5 percentage points lower than the baseline estimates.
The fact that households least likely to have experienced unobservable shocks yield
qualitatively similar place share estimates to the baseline sample is reassuring and
suggests limited bias from shocks to unobserved preference heterogeneity.
 1466 THE AMERICAN ECONOMIC REVIEW MAY 2025

1
Movers
Movers, no big life changes
Place share

0.5

0
0 5 10 15 20
Years between observations

Figure 4. Tract Share of Spatial Variation, by Duration between Observations

Notes: This figure shows event study estimates of the share of spatial variation in mean carbon emissions that can
be explained by place effects, by duration between mover observations. In other words, each coefficient is the esti-
mate for place effects generated from the ­subsample of households that I observe x​ ​years apart. Coefficients plot-
ted in light gray are estimated from the model using the full panel of stayers and movers. Coefficients plotted in the
dark blue are estimated from the model using the s­ ubsample of stayers and movers with no changes in the number
of children, less than 50 percent change in household income, and no change in homeownership status between
observations. All estimates are weighted using census sample weights.

In addition to unobserved shocks, another potential source of bias arises from

“preference drift”—gradual changes in households’ preferences over time that
are not fully captured by changes in age or other observable characteristics (Card,
Heining, and Kline 2013). If such drift exists and influences relocation decisions,
place effect estimates would conflate true causal effects with selection, particularly
for moves observed after long time intervals. This would manifest as place effect
estimates that trend upward with the duration between observations. To explore this
possibility, Figure 4 presents t­ract-level place share estimates by duration between
observations, controlling for demographic and household characteristics. The light
gray points depict estimates for the baseline mover sample, while the dark blue bars
correspond to the restricted subsample of movers with no big life changes between
observations.
For the full mover sample, a slight upward trend emerges, with place share esti-
mates for durations in the ­15- to 19-year range exceeding the pooled estimate. This
pattern dissipates when considering the restricted subsample of movers without
observable big life changes; the pooled estimate for this group is not contained within
the 95 percent confidence interval for 4 out of the 19 ­duration-specific estimates, but
with no clear trend across durations. The Supplemental Appendix presents analogous
CBSA estimates, which exhibit a similar pattern, except the upward trend observed in
the l­onger-duration coefficients becomes insignificant in the baseline mover sample
as well.
While these results suggest some preference drift, potentially leading to
upward-biased place share estimates, the magnitude appears small relative to
­
observed changes in emissions between places, and this bias appears to be negligible
in the sample restricted to households without big life events. One additional result
 VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1467

Panel A. CBSA Panel B. Tract

0.5 0.5
Change household log (CO2)

Change household log (CO2)

0.25 0.25

0 0

−0.25 −0.25

−0.5 −0.5

−0.5 −0.25 0 0.25 0.5 −0.5 −0.25 0 0.25 0.5

Change mean log (CO2) Change mean log (CO2)

Figure 5. Place Share of Spatial Variation, by Move Type

Notes: This figure shows mover changes in household carbon emissions, by size of o­ rigin-destination differences
in mean carbon emissions. All estimates are from a model that conditions on changes in observable household
characteristics and year fixed effects. I split movers into 10 deciles, according to the size of the (conditional) gap
in mean carbon emissions across their origin and destination, and estimate standard errors using bootstrap. Point
estimates are shown with 95 percent confidence intervals. The solid lines show the regression estimates from the
pooled model, and the dotted line denotes 45 degrees, that is, the scenario in which moving to on average higher- or
lower-emissions places leads to a ­one-for-one increase in own carbon emissions. All estimates are weighted using
Census Bureau sample weights.

that emerges from this analysis is that household carbon emissions appear to change
instantaneously, suggesting place effects stem from attributes that directly impact
emissions rather than gradual influences like peer effects or habit formation.15
As a final specification check, I explore heterogeneity across moves of different
magnitudes in terms of place mean changes. A potential concern with the results
presented thus far is that households choosing to relocate from ­low-emissions places
to h­ igh-emissions places (or vice versa) may systematically differ from those mov-
ing between areas with similar emissions levels. If a household’s choice of destina-
tion reveals newly observable information about unobserved preference shocks, the
estimates would be biased, mistakenly attributing these shocks to place effects. To
explore this possibility, Figure 5 depicts changes in mover households’ carbon emis-
sions across deciles of ­origin-destination mean differences, controlling for demo-
graphic and household characteristics. The x­ -axis represents conditional changes
in mean emissions between origin and destination, while the ­y-axis shows to con-
ditional changes in household emissions. The 4­ 5-degree line (in gray) represents a
scenario in which place effects fully explain variation in carbon emissions across
locations. The solid line’s slope corresponds to the pooled estimate of the rela-
tionship between ­tract-level mean changes and ­household-level changes. Standard
errors are estimated using bootstrap.

15
While I don’t observe how long ago a household moved, the expected time since moving increases with the
duration between observations.
1468 THE AMERICAN ECONOMIC REVIEW MAY 2025

I find that, for both moves across CBSAs and moves across neighborhoods,
changes in household carbon emissions as a share of changes in o­ rigin-destination
means are symmetric and linear across move types. In other words, the share of
spatial heterogeneity attributable to p­ lace-based differences is consistent for moves
from low- to h­ igh-emissions places (rightmost points), moves from high- to low-
emissions places (leftmost points), and moves between places with similar aver-
age emissions (central points). The symmetry and stability of share estimates have
several implications. First, this result extends the symmetry check presented in
Figure 3, further validating the l­og-linear model specification and suggesting that
the t­wo-way fixed effect model reasonably approximates household carbon emis-
sions. Recall that this implies place effect estimates are unbiased even if I never
observe some household types moving to some place types. Second, this result indi-
cates that the event study estimates presented in Table 3 are not driven by a subset of
movers or mover destinations, alleviating concerns that estimates primarily reflect
changes in the emissions of households who moved to a vastly different destina-
tion in response to a large unobserved preference shock. Lastly, this result provides
another dimension along which heterogeneity in the estimated share parameter
appears to be limited.
In summary, the event study estimates indicate that movers’ carbon emissions
change by over half of ­origin-destination differences in neighborhood means (and
about 85 percent of o­ rigin-destination differences in CBSA means) when they move.
Under strong assumptions on sorting, this can be interpreted as a causal share parame-
ter, which would determine the change in carbon emissions resulting from any house-
hold moving between any pair of places. I estimate that heterogeneity in the share
parameter is limited along several dimensions explored throughout this section—esti-
mates are within a 5 percentage point range of each other across subsamples with
vastly different changes to observed characteristics, across different observation time
horizons, and across varying move types (between similar versus dissimilar places)—
suggesting that perhaps bias from violations of this assumption may also be minimal.
Under weaker assumptions, the event study estimates serve as a useful specifica-
tion check for the ­two-way fixed effect model, and perhaps more interestingly, they
provide unbiased prediction of how household carbon emissions will change for any
observed move. Recent research and media coverage have highlighted a pattern of
increasing migration from more expensive but on average lower-emissions places to
less expensive but on average higher-emissions places (Kolko 2021; Eisen 2019). For
instance, in 2019, there was a net migration of about 45,00 people from California to
Texas. While not all these moves were from San Francisco to Houston, using these
two cities as an illustrative example, carbon emissions in Houston are about 21 percent
higher than they are in San Francisco (Jones and Kammen 2014). Combined with
the event study findings, this implies that regulatory constraints that restrict housing,
increase the cost of living, and drive this pattern of migration increase household car-
bon emissions by 12 percent or more, imposing a sizable carbon externality.16

16
The 12 percent figure would correspond to treating San Francisco and Houston as neighborhoods. If we were
to treat them more like CBSAs, the predicted increase in emissions would be closer to 19 percent.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1469

Table 4—Unobserved Heterogeneity in CO2: Variance Decomposition

CBSA Tract
(1) (2) (3) (4) (5) (6) (7)
Panel sample
Variance of log(CO​​​2​​​) 0.31 0.31 0.31 0.31 0.31 0.31 0.31
Share attributable to places 0.16 0.07 0.08 0.17 0.23 0.15 0.15
Share attributable to hhs 0.50 0.50 0.50 0.30 0.36 0.36 0.36
Corr. of place and hh effects 0.01 0.03 0.03 0.02 0.02 0.08 0.11
SD of place effects 0.23 0.15 0.16 0.23 0.26 0.22 0.22

Mover sample
Variance of log(CO​​​2​​​) 0.35 0.35 0.35 0.33 0.33 0.33
Share attributable to places 0.14 0.04 0.04 0.22 0.16 0.16
Share attributable to hhs 0.14 0.16 0.17 0.10 0.10 0.10
Corr. of place and hh effects 0.07 0.08 0.08 0.08 0.16 0.18
SD of place effects 0.22 0.12 0.12 0.27 0.23 0.23

Climate + electricity CO​​2​ ​​​ X X X X

Price index X X
­Time-varying fixed effects X

Notes: This table reports results from the ­heteroskedasticity-robust KSS estimation of variance components. For
each sample and specification, the table reports the overall outcome variance, the share of variance attributable
to place effects (​​ψ​j​​​), the share of variance attributable to household effects (​​αi​​​​), the correlation between place
effects and household effects (which captures how much households sort on unobserved characteristics), and the
­bias-corrected standard deviation of place effects. All specifications include year fixed effects and the standard set of
household controls used throughout the paper. Columns 1 and 5 report the baseline variance decompositions at the
CBSA and tract levels. Columns 2 and 6 add controls for local mean heating degree days, cooling degree days, and
log electricity emissions factors. Columns 3 and 7 additionally control for a price index, constructed from lagged
fuel shares interacted with national retail prices. Finally, column 4 computes ­time-varying CBSA place effects using
five-year periods (­2000–2004, ­2005–2009, ­2010–2014, and ­2015–2019), using stayer observations across periods
to identify time variation in place effects, while movers, as before, identify ­cross-sectional variation.

B. Variance Decomposition

In this section, I weaken the restriction on sorting imposed by the event study and
present estimates from the variance decomposition shown in equation (8):
 ​var​[​yi​j​​]​  = var​[​ψj​​​]​  + 2 ⋅ cov​[​αi​​​, ​ψj​​​]​  + var​[​αi​​​]​​
+ var​[​X​it' ​ ​  β]​+ 2 ⋅ cov​[​α​i​​, ​X​'it​ ​  β]​+ 2 ⋅ cov​[​ψj​​​, ​X​'it​ ​  β]​+ var​[​εi​t​​]​​.
​

For each sample and specification, Table 4 presents the overall variance of the
outcome, log(CO​​2​​​); the share of variance attributable to each of the unobserved het-
erogeneity components (place effects, ​​ψj​​​​, and household effects, ​​ψ​i​​​); the correlation
between the unobserved heterogeneity components; and the ­bias-corrected standard
deviation of place effects. The top panel presents estimates from the entire panel of
movers and stayers, while the bottom panel presents estimates from the ­mover-only
sample.
Columns 1 and 5 present the baseline analysis, which includes year fixed effects
and the standard vector of household controls. I estimate that CBSA effects account
for 1­ 4–16 percent of overall heterogeneity (column 1), and tract effects account for
­22–23 percent of overall heterogeneity (column 5). Columns 2 and 6 introduce
1470 THE AMERICAN ECONOMIC REVIEW MAY 2025

controls for mean heating degree days, cooling degree days, and log electricity emis-
sions factors, excluding their impact on household emissions from estimated place
effects. This specification not only provides insight into the mechanisms driving
place effects, as in the event study analysis, but also allows for more precise coun-
terfactual analysis within the KSS framework. When considering interventions that
impact household carbon emissions through changes to local amenities and the built
environment, isolating the place effect component not driven by climate or electric-
ity-generating sources may be more relevant, though the appropriate specification
remains ­context dependent. For instance, a ­city-wide initiative to install solar panels
on parking lots, potentially coupled with EV charging stations, could decrease the
average emissions intensity of electricity while promoting EV adoption. Moreover,
while regional climate is, to first order, exogenous to local actions, there are many
interventions cities and neighborhoods could enact to decrease heat island effects
and in turn reduce local heating degree days (Druckenmiller 2023).
I find that controlling for climate and electric grid intensity together decreases
the CBSA share of spatial heterogeneity by roughly 10 percentage points, or by
more than half, to ­4–7 percent of overall heterogeneity. At the neighborhood level,
controlling for climate and electric grid intensity decreases the place share of het-
erogeneity by roughly 6­ –7 percentage points (column 6), by less than half, leaving
the remaining neighborhood attributes explaining approximately 15 percent. In col-
umns 3 and 7, I additionally partial out variation driven by prices, using a price index
constructed from interacting local lagged fuel shares with national retail prices. This
does not further change the results at either the CBSA or tract level.
In the Supplemental Appendix, I control separately for electricity emissions factors
and climate, and I find that the majority of the decline in place share observed in col-
umns 2 and 6 comes from controlling for electricity emissions factors. It is well under-
stood that climate has a robust effect on energy demand (e.g., Goldstein, Gounaridis,
and Newell 2020; Levinson 2016), so why the disparity? Many places that are other-
wise ­low-emissions are in regions that have made efforts to decarbonize their electric-
ity (Murray and Maniloff 2015; Petek 2020), so there is a positive correlation between
electricity emissions factors and high other emissions attributes. In contrast to this, I
estimate a negative correlation between having a­ bove-average heating degree days and
low other carbon emissions attributes; many colder cities are in the Northeast, a region
that disproportionately contains older, denser cities and neighborhoods (Tomer et al.
2021). These correlation terms enter the place effect variance component, amplifying
it when electricity emissions factors are included in the place effect and diminishing it
when heating and cooling degree days are included.
Across all three specifications discussed so far, estimates of the place share of
heterogeneity are quite comparable between the panel and mover sample. This is not
true of the estimates of the household components; a comparison between the panel
sample and the mover sample reveals that the contribution of unobserved house-
hold characteristics to overall heterogeneity is highly sensitive to which sample the
model is estimated on. In the panel sample, unobserved household heterogeneity
accounts for 50 percent of overall heterogeneity when defining place at the CBSA
level, and 36 percent when measuring place at the neighborhood level. Using the
­mover-only sample substantially decreases the unobserved household contribution
across specifications, to ­14–17 percent in the CBSA specification and 10 percent
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1471

in the tract specification. The correlation estimates between unobserved place and
household characteristics are also sensitive to sample choice, increasing across all
specifications by up to 8 percentage points when switching from the panel to the
mover sample. This suggests some assortative matching of household types to place
types (especially neighborhoods), but in the specification where the correlation coef-
ficient is largest, it remains relatively low, at under 20 percent. The share of overall
heterogeneity attributable to matching implied by the covariance term (which is not
shown in Table 4 but can be computed from the correlation coefficient and variance
components) is less than 5 percent across all specifications.
The panel and mover samples may yield different estimates due to inherent differ-
ences between stayers and movers, or KSS’s inability to correct for serial correlation in
stayers’ error terms. To distinguish between these causes, I compare the KSS decom-
position estimates with an uncorrected Abowd, Kramarz, and Margolis (1999)—
henceforth, AKM—decomposition, presented in the Supplemental Appendix. If
differences between stayers and movers drive the discrepancy, both KSS and AKM
estimates should reflect this pattern. Conversely, if serial correlation in stayers’ error
terms is the main factor, AKM should show similar relative contributions of unob-
served household heterogeneity across samples (both inflated due to limited mobility
bias), with significant differences emerging only after the KSS correction. The AKM
estimates reveal relatively stable place and household shares across panel and mover
samples, with only minor reductions in the household component’s relative size in
some specifications. This contrasts with the more dramatic shifts observed in the KSS
analysis, suggesting that KSS estimates of household variance components are likely
more reliable in the mover sample. Panel sample estimates likely represent an upper
bound, with upward bias primarily driven by serial correlation in stayer error terms.
The Supplemental Appendix discusses potential sources of this serial correlation,
many arising from the survey nature of the data, and their implications for result
interpretation. Additionally, the Supplemental Appendix evaluates the sensitivity of
results to alternative outcome definitions in the KSS decomposition.
Up until now, I have been estimating place effects pooling over the entire sam-
ple time frame; however, place effects may evolve over time in ways that differ
from national average trends in carbon emissions. Local or state governments par-
ticularly concerned about climate change may enact regulatory changes or make
­place-based investments aimed at reducing emissions for their residents. Changes to
place effects could also arise from local or regional planning initiatives motivated by
factors completely unrelated to decision-makers’ climate objectives. For instance,
the Phoenix metropolitan area—one of the fastest growing metropolitan areas in the
United States—grew by nearly 1.6 million residents between 2000 and 2020. This
period of growth has been accompanied by a mix of suburban expansion, urban
development, the opening of a new light rail system, and several highway expan-
sions (Maricopa Association of Governments 2020). To allow for such ­place-specific
changes, I follow Lachowska et al. (2023) and estimate ­time-varying fixed effects​​
ψ​jt​​​at the CBSA level, using stayers to identify variation across time within place.17

17
Because tracts by definition consist of many fewer observations than CBSAs, including a time-varying compo-
nent introduces either substantial noise, or a substantial geographic restriction to only the most populated tracts, so I do
not estimate t­ ime-varying tract effects. Studying n­ eighborhood-level changes is an important direction for future study.
1472 THE AMERICAN ECONOMIC REVIEW MAY 2025

To maintain connectivity in my set of places, and because for the most part places
evolve slowly, I define the ­time-varying place effects using ­five-year intervals.
Thus, there’s a different ­time-varying place effect for each period: ­2000–2004,
­2005–2009, 2­ 010–2014, and 2­ 015–2019. Results are shown in column 4 of Table 4;
allowing CBSA effects to evolve increases their variance share by only 1 percent-
age point relative to the baseline specification. This implies that changes in places
over my sample period are either mostly captured by national secular trends or are
unrelated to household carbon emissions. In the Supplemental Appendix, I pro-
vide some descriptive results on the nature of changing t­ime-varying place effects
in my sample. I also discuss some of the ways in which place effects may have
evolved following the C ­ OVID-19 pandemic and the ensuing transition to remote
work, which was accompanied by a steep decline in commuting and a shift toward
larger homes to accommodate home offices (Van Nieuwerburgh 2023; D’Lima,
Lopez, and Pradhan 2022).
As a final specification test, in the Supplemental Appendix, I show binned scat-
terplots similar to those presented in Section IVA, but now with deciles of changes
in estimated place effects rather than observational means, on the x-axis. I plot
these against two sets of changes in household mean outcomes: changes for the
full mover sample and changes in the sample restricted to only households with
no big life events. In a correctly specified model, changes in place effects should
lead to o­ ne-for-one changes in household carbon emissions, though attenuation bias
from noisily estimated place effects is expected to decrease the slope by some. This
is roughly what I observe. Moreover, I find no discernible difference between the
primary sample and the subsample of households experiencing no big life events,
which provides additional reassurance that selection on heterogeneous preference
shocks isn’t a first-order threat to identification in my analysis.

C. Interpreting Decomposition Results

How do the event study and KSS results inform one another? First, recall that
event study estimates are unbiased only if heterogeneity in the share parameter
is uncorrelated with observed and unobserved household characteristics. I have
shown evidence that the event study estimates are fairly stable across several
observable dimensions of heterogeneity in the data. I also showed in the KSS
decomposition that the covariance between unobserved components of heteroge-
neity is relatively small; the largest correlation coefficient across the four base-
line estimates corresponding to the specifications examined in the event study is
0.08. Together, this evidence suggests that bias from this assumption on selection
should be minimal.
Second, even when unbiased, the event study yields estimates of shares of
mean differences between places attributable to place effects, while the KSS esti-
mates yield a variance decomposition of overall variation, and this can lead to
meaningful discrepancies in magnitudes. To help illustrate this, consider a sim-
plified version of the ­two-way fixed effects model, where ​​yi​t​​  = ​α​i​​  + ​ψj​​​  + ​εit​ ​​​.
Song et al. (2019) show that unobserved heterogeneity can be decomposed fur-
ther into a ­between-place component v​a​r​j​​​[​​y–​​j​​]​​
, which captures the variation in
mean household carbon emissions across places, and a w ­ ithin-place component
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1473

v​ a​ri​​​​[​yit​ ​​  − ​​y–​​j​​  | i ∈ j]​, which captures the heterogeneity in carbon emissions of house-
holds living in the same place:

(9) ​var​[​yit​ ​​]​  = va​rj​​​​[​​y–​​j​​]​  + va​ri​​​​[​yit​ ​​  − ​​y–​​j​​  | i ∈ j]​

= ​​var​
​     [​ψj​​​]​  + 2 ⋅ cov​
   [​​α​​j​  
–
​​, ​ψj​​​]​  + var​[​​α
–
​​j​​]​ ​​​+ ​​var​
   [​αi​​​  − ​​α
  –
​​j​​]​  +
​  
var​[​εij​ ​​]​​.​​
 
Between Within

Equation (9) highlights that heterogeneity between places reflects variation in

place effects, sorting of certain types of households to certain types of places (the
covariance term), and “segregation” of households—the extent to which households
of different types segregate across places, whether or not this pattern reflects system-
atic sorting on place types.18 In addition to the b­ etween-place heterogeneity, overall
heterogeneity reflects heterogeneity in household carbon emissions within places, as
well as heterogeneity that cannot be explained by the ­two-way fixed effects model.
For intuition, imagine two places, one ​ψ ​ ​low​​​and one with ​ψ
​ ​high​​​, with identical
populations. If there is high variation in carbon emissions within populations and a
small difference between ​ψ ​ ​low​​​and ​ψ
​ ​high​​​, the event study would yield a share coeffi-
cient of one (since populations are identical across places, all between differences
are driven by place effects), but the KSS decomposition would yield a place vari-
ance component of close to zero (because of a large within component to the vari-
ance). In practice, this is very close to what happens at the CBSA level; the vast
majority (85 percent) of differences between CBSAs can be attributable to variation
in place effects and not household attributes, but there is much more variation in
household carbon emissions within CBSAs than there is across, leading to a vari-
ance component of approximately 15 percent in the KSS estimation (more than
half of which is attributable to climate and electric grid intensity). At the neighbor-
hood level, household sorting contributes more to variation between places, drop-
ping event study estimates of the place share to roughly 5­ 5–60 percent. Accounting
for variation within neighborhoods in the denominator decreases the neighborhood
share to 2­ 2–23 percent of overall heterogeneity, or about 15 percent when excluding
climate and electric grid intensity from place effects.

V. The Characteristics of Low- and ­High-Emissions Places

With estimates of place effects in hand, I proceed to characterizing high- and

l­ow-emissions places. As highlighted in the conceptual model, place effects reflect
a mix of differences in demand for energy services, energy prices, energy demand
elasticities, fuel mixes, and emissions factors. The urban planning literature has
identified many local amenities that could contribute to differences in average
household energy demand and energy demand elasticities. In the residential energy
sector, larger homes tend to use more energy, as do s­ ingle-family homes; there is a
strong relationship between carbon emissions and density, though it is potentially

​​α​​j​​ ≡ E​[​α​i​| i ∈ j]​​.

18 ​–
1474 THE AMERICAN ECONOMIC REVIEW MAY 2025

Panel A. Bivariate regressions Panel B. Single regression

Urbanity Urbanity
Suburban Suburban
Rural Rural
Density Density

Local amenities Local amenities

Walk score Walk score
Bike score Bike score
Transit score Transit score
Rail routes Rail routes
Bus routes Bus routes
Dist. to closest city Dist. to closest city
Dist. to largest city Dist. to largest city

Capital stock Capital stock

Detached home Detached home
Number rooms Number rooms
Number cars Number cars

Exogenous amenities Exogenous amenities

Heating degree days Heating degree days
Cooling degree days Cooling degree days
Electricity emissions Electricity emissions

Demographics Demographics
Age 40+ Age 40+
Non-White Non-White
Hispanic Hispanic
College College
Household income Household income
Has children Has children
Homeowner Homeowner
2

3
0.

0.

0.

0.

0.

0.

0.

0.

0.

0.
−

−
Change in tract effects Change in tract effects

Figure 6. Correlates of Tract Effects

Notes: This figure presents estimates from OLS regressions of estimated tract effects on a set of observable
­place-based and household characteristics. Panel A shows results from separate bivariate regressions, while panel
B shows results from a single regression on all covariates. All amenity variables are tract-level means, normalized
to have mean 0 and standard deviation 1, except the rural and suburban indicators, which are retained as indicators.
Regressions are weighted using ACS sample weights. Thus, estimated coefficients reflect the average change in
place effect—and in turn the percent change in carbon emissions, given the log specification of the model—associ-
ated with a 1 standard deviation change in observable characteristics.

n­ onmonotonic due to the effect of density on ­microclimates; and parks, plants,

and green surface coverage are all negatively correlated with energy use (see, e.g.,
Ko 2013 for a review). In the transportation sector, car use is lower in places with
better public transportation, less parking, and more direct road connectivity (e.g.,
Transportation Research Board 2009; B ­ arrington-Leigh and M ­ illard-Ball 2017),
and mechanically, people drive fewer miles when they live closer to where they
work, shop, and spend their leisure time. Many of these amenities are intertwined
and simultaneously relate to residential and transportation emissions, underscoring
the value of studying these sectors together. For instance, higher density not only
implies smaller homes but also enables more efficient public transportation net-
works by reducing distances between transit stops and destinations.
In Figure 6, I show the results from regressions of estimated tract effects onto a
set of observational tract-level characteristics. In panel A, I show coefficients from
bivariate regressions of tract effects onto each of the shown ­tract-level ­characteristics
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1475

independently (except for suburban and rural, which are estimated in a single regres-
sion), while in panel B, I show coefficients from a single regression onto all of the
characteristics.19 Observable characteristics are all measured at the tract level and
are normalized across tracts to have mean 0 and a standard deviation of 1 (except
for suburban and rural, which are retained as indicators). Thus, coefficients should
be interpreted as showing approximately the average percent change in carbon emis-
sions associated with a 1 standard deviation change in n­ eighborhood-level observ-
able characteristics.
Observable characteristics fall into five categories. Both “urbanity” and “local
amenities” cover aspects of urban form that households effectively taken as given
but that vary neighborhood to neighborhood. These include whether a tract is
urban, suburban, or rural, its density, measures of local public transportation ame-
nities (walk scores, bike scores, transit scores, number of nearby rail routes and
bus routes), and measures of sprawl (captured by geodesic distance between tract
centroids and the centroid of the closest city and the largest city within the CBSA).
“Capital stock” characteristics reflect household choices among the set of options
available in a place, which are determined at least in part by local zoning regulations
or other policies. These include the share of ­single-family detached homes, average
house sizes, and the number of cars in a household. “Exogenous amenities” are pre-
dominantly shaped by factors beyond local control, and households also take these
as given. These include annual heating degree days, annual cooling degree days, and
electric grid intensity.20 Finally, the “demographics” category captures tract-level
variation in household observable characteristics.
The results of the bivariate regressions (panel A) are consistent with obser-
vational data and conclusions in the urban planning literature. H ­ igher-emissions
neighborhoods tend to be less urban and dense, with poorer walkability, bikabil-
ity, and public transit options, and are typically situated farther from city centers
within their CBSA. H ­ igher-emissions neighborhoods feature a higher proportion
of detached and larger homes, with households owning more cars on average. They
experience less mild climates and have higher electricity emissions factors. Finally,
looking at the projection of tract effects onto demographics, I find that ­non-White,
Hispanic, c­ ollege-educated, and higher-income households are more likely to live in
­low-emissions tracts, while older households, households with children, and home-
owners are more likely to live in ­high-emissions tracts.
When including all of these characteristics in a single regression (panel B),
directionally, the relationships remain the same, but many of the coefficient mag-
nitudes decrease substantially. Electricity emissions factors, heating degree days,
and cooling degree days emerge as three of the four strongest correlates of place
effects. The second-largest correlational coefficient is on the number of rooms in
a house; a standard deviation change in tract mean number of rooms is associated

19
I used a LASSO regression with 10x crossfold validation to select characteristics, and it retained all of the
variables.
20
As mentioned earlier, neither climate nor electricity emissions factors are impervious to local influence.
Urban form can quite meaningfully impact local microclimates through urban heat island effects, but in this paper
my measure of climate is at the NOAA climate division level, a geographically coarser definition. And municipal
utilities can shape local electricity emissions factors through energy procurement choices but currently make up a
small share of overall electricity supply.
1476 THE AMERICAN ECONOMIC REVIEW MAY 2025

with a roughly 7 percent change in tract-level carbon emissions. The remainder of

tract effects appear to load primarily onto measures of sprawl, the quality of local
bike infrastructure, rurality, density, and household income, with standard deviation
changes in these characteristics being associated with a 2­ –4 percent change in car-
bon emissions each. It is notable that measures of walkability, transit quality, share
of detached homes, and number of cars appear to become less important after con-
ditioning on density, sprawl, and home size. It is possible that this is because these
characteristics are very c­ olinear in the observational data; as highlighted earlier,
many of these relationships are interconnected.
The Supplemental Appendix shows analogous projections of household fixed
effects onto observable ­neighborhood-level characteristics. I find that correlations
between household effects and observable characteristics are about an order of mag-
nitude weaker than those between tract effects and observable characteristics in both
the bivariate regressions and the full regression. This is consistent with minimal
sorting on unobserved household characteristics that I estimate in KSS. The larg-
est coefficients imply that households with high unobservable potential for carbon
emissions sort to tracts where people are on average higher income and have more
children, where it is warmer, where electricity is higher-emissions, and where houses
are bigger. The Supplemental Appendix presents additional results on the correlates
between observable household characteristics and observable place characteristics.
These regressions do not elucidate causal relationships. As is the case with
inference about place effects as a bundle, inference about the role of specific ame-
nities from observational data alone is likely to be biased by household sorting.
Nevertheless, the observational data suggest features of urban form whose effect
on household carbon emissions should be studied further with credible exogenous
variation in the amenity itself. Previously, one might worry about whether a setting
for such a case study is selected. Can we learn about the effect of lifting zoning
restrictions and allowing for smaller lot sizes if people who live in denser cities have
significantly different preferences than those who don’t? Are improvements to bike
infrastructure only effective in cities where a lot of people already bike? The core
results of my paper suggest that a meaningful share of variation between places is
driven by variation in place effects, with evidence that place effects are constant
across heterogeneous populations, mitigating some of these concerns about external
validity.

VI. Implications for Aggregate Carbon Emissions

The wide distribution of place effects suggests that there may be an opportunity
to substantially reduce household carbon emissions from residential and transpor-
tation energy through what I refer to as p­ lace-based climate policies. P
­ lace-based
climate policies—distinct from regional energy sector regulations or ­market-based
mechanisms designed to encourage decarbonization—aim to reduce household car-
bon emissions by altering local characteristics that shape household energy choic-
es.21 They could be implemented at federal or local levels and, building on insights

21
As transportation and residential sectors electrify and the power grid decarbonizes, the impact of p­ lace-based
drivers of household emissions will diminish. However, given current challenges in clean electrification, particularly
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1477

from labor and urban economics, could be designed to simultaneously reduce car-
bon emissions and address economic development goals, housing shortages, or local
externalities like pollution and congestion.
Infrastructure and local public goods investments make up one category of poten-
tial ­place-based climate policies. The federal government has a history of invest-
ing into transportation infrastructure, with the Interstate Highway System being
its most notable example (US Department of Transportation 1977). More recently,
the Bipartisan Infrastructure Law provided a significant increase to transit funds
across the country; the Federal Transit Administration announced in early 2024 that
it would be investing $9.9 billion to support local transit systems in urban areas
across the United States (US Department of Transportation 2024). While adding and
expanding public transit networks is perhaps the most obvious example of an inter-
vention that might decrease place effects, many cities are implementing less costly
initiatives that repurpose existing urban space for community use. Paris’s Plan Velo
has converted over 200 miles of roads into (often protected) bike lanes as of 2024,
with further expansions planned through 2026 (Ville de Paris 2021). Barcelona’s
Superillas program aims to transform more than half of its ­car-dominated streets into
­mixed-use public spaces (Roberts 2019), while New York City’s High Line proj-
ect has repurposed an abandoned elevated railway into a popular pedestrian walk-
way and park. The ­COVID-19 pandemic accelerated this trend, sparking “slow” or
“open” streets programs that temporarily restricted car traffic to benefit pedestrians
and cyclists in many cities of the United States, with some cities now working to
make these changes permanent (New York City Department of Transportation 2023;
Combs 2020). Expanding the network of electric vehicle chargers is another exam-
ple of a lower-cost ­place-based infrastructure investment that can reduce carbon
emissions.
A second category of potential ­place-based policies includes regulatory changes
intended to encourage—or remove barriers previously preventing—sustainable
urban development. Zoning deregulation, land use reform, and t­ransit-oriented
development have all gained traction in the United States in recent years. In 2018,
Minneapolis became the first city in the United States to enact a ­city-wide ban on
exclusionary zoning, a common practice across the United States that restricts land
to be used for s­ ingle-family homes only (Mervosh 2018). In 2021, the California
State Assembly passed Bills 9 and 10, which streamlined the process of “­upzoning”
residential land, allowing for the development of up to four units on land previously
zoned for ­single-family homes only and facilitating h­ igher-density construction near
transit corridors. This was followed by AB 2097 in 2022, which eliminated most
parking minimum requirements for new development (Fulton et al. 2023). At the
federal level, President Biden’s original infrastructure bill proposal in March 2021
included grants to cities that eliminated exclusionary zoning. While this portion of
the bill did not get passed, and many states and municipalities are still debating
but not implementing zoning reforms, these examples illustrate the relevance of
such approaches in the current policy debate. At the same time, some states and

transmission capacity constraints and aging infrastructure (e.g., US Department of Energy 2015), ­place-based
policies that reduce energy demand and alleviate pressure on the electricity grid could be complementary to more
traditional decarbonization instruments.
1478 THE AMERICAN ECONOMIC REVIEW MAY 2025

municipalities are making regulatory efforts to make building codes more stringent
by, for example, enacting energy efficiency minimum requirements (Levinson 2016)
or mandating electrification through natural gas hookup bans (Payne 2021).
This paper does not identify a causal relationship between any specific amenity
and place effects, or any specific intervention and place effects, but it shows evi-
dence that l­ow-emissions places are ­low-emissions not simply due to sorting of
­low-emissions people to those places, and that low emissions places tend to have
amenities that are characteristic of more urban, less sprawling neighborhoods. Using
this as motivation, I examine how carbon emissions would change if the national dis-
tribution of place effects was more urban than it currently is. Specifically, to approx-
imate a scenario in which the United States limits suburban sprawl and increases the
share of households within a region that live in an urban neighborhood, I estimate
the effect on emissions if households currently living within a suburban or rural area
lived instead in a place with the average place effect of the nearest principal city. A
naïve comparison of household emissions shows that, on average, households living
in principal cities emit 20 percent less than households living in surrounding areas.
After accounting for sorting of households between suburban and urban neighbor-
hoods, my estimates suggest that if every tract had the place effect of the nearest
principal city, the emissions of suburban and rural households would decrease by
approximately 15 percent.
The majority of principal cities in the United States would not be considered par-
ticularly urban on a global scale. To consider the potential effect of deeper urbaniza-
tion, I examine how household carbon emissions would change if more people lived
in places like Manhattan, which is uniquely dense, walkable, and transit oriented
within the US context. Specifically, I examine how the emissions of households
in the principal cities of the nine largest CBSAs after the New York Metropolitan
Area in the United States (US Census Bureau 2024) would change if those cities
developed into a place with the average place effect of Manhattan. The observa-
tional gap is enormous: Even using large cities as a comparison group, households
in Manhattan emit about 73 percent less than observably comparable households in
the other 9 large cities. Accounting for unobserved fixed differences between house-
holds reduces the gap to 60 percent; people who choose to live in Manhattan are a
selected s­ ubsample, but their emissions are nevertheless much lower as a result of
living in Manhattan than they would be if they lived elsewhere.
These exercises lend insight into how development that shifts population shares
across place types by “expanding” places with lower place effects—either by mak-
ing their neighbors look more like them or by allowing more people to live in such
places without changing their fundamentals—could affect emissions in the future.
My estimates yield only a ­first-order, partial equilibrium approximation to the effect
of such interventions, as in practice, there would be some ­re-sorting of populations
in response to ­place-based changes, which would change the distribution of house-
hold types living in each place and thereby change aggregate carbon emissions.
More importantly, my estimates don’t lend insight into the specific interventions
that would result in the largest changes to place effects; this is a critical direction
for future research. My results highlight that the potential for ­urbanization-induced
reductions to carbon emissions is overstated when inferred from observational
means across places because people choose where to live based in part on their
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1479

potential carbon emissions. Nevertheless, because there is substantial variation in

household carbon emissions and neighborhood effects explain 1­ 5–23 percent of this
variation—depending on whether climate and electricity emissions are included in
the place effects—my results also imply that even relatively small shifts in the distri-
bution of place effects that households are exposed to could meaningfully decrease
aggregate carbon emissions.

VII. Discussion

This paper is the first to estimate the causal effects of places on household carbon
emissions and decompose spatial heterogeneity in carbon emissions into a compo-
nent driven by these place effects and a component driven by household character-
istics and sorting. I find that up to 23 percent of overall heterogeneity in household
carbon emissions from residential energy use and commuting across the United
States can be explained by neighborhood effects, or roughly 15 percent can be
explained by neighborhood effects after accounting for variation driven by climate
and electric grid intensity. Paired with high overall heterogeneity, these estimates
imply that interventions that change the population-weighted distribution of place
effects across the United States, either through direct changes to places or through
regulatory changes that make it possible for more people to live in low-emissions
places, could result in meaningful reductions in household carbon emissions.
There are several limitations of my empirical analysis that should be taken into
consideration while interpreting my results. The first is that due to the survey nature
of my data, carbon emissions are noisily measured. This leads to lower explanatory
power of the model than is standard in papers using these methods with adminis-
trative data to estimate firm wage premia. The relatively low explanatory power of
the model could also reflect model ­misspecification, but with only two observations
per household for the majority of my sample, the number of specification tests I can
do is limited. Second, there is relatively little variation in urban form and transpor-
tation options across the United States; 75 percent of residential land in the United
States is zoned for single-family homes only (Badger and Bui 2019), 95 percent
of commuters in my sample commute by car, and there is only 1 high-speed rail
line in the entire country, which operates at high speed over only roughly 50 miles
of track. This is in stark contrast with other parts of the world, where many cities
are denser and substantially less car oriented. Moreover, place effects are identified
from movers, who differ from the general US population in meaningful ways. Thus,
the external validity of my results relies upon estimates being stable to widening the
distributions of place and household types.
Establishing that place matters—and how much it matters—for household car-
bon emissions takes a critical first step toward investigating the welfare impacts of
specific ­place-based climate policies. The welfare effects of a given intervention
depend on several parameters whose estimation is outside of the scope of this paper.
First, they depend on the causal relationships between local amenities and place
effects, and on household preferences for local amenities. While Tiebout (1956)
posits that residential sorting allows for efficient provision of local public goods,
his framework only applies to amenities without scale economies. Moreover, there
is reason to believe that residential sorting is not efficient due to frictions and
1480 THE AMERICAN ECONOMIC REVIEW MAY 2025

exclusionary policies (e.g., Rothstein 2017; Hausman and Stolper 2021; Christensen
and Timmins 2023; Avenancio-León and Howard 2022). Estimating causal relation-
ships between local public amenities and household carbon emissions and quanti-
fying whether ­emissions-relevant local public amenities are at an efficient level are
important directions for future work. Second, the welfare impacts of ­place-based
interventions would depend on the costs of implementing them relative to the cost
of business as usual or other contender climate policies. Costs can vary dramatically
for the same intervention in different settings (Goldwyn et al. 2022), making this
estimation difficult, but incorporating cost estimates for a marginal value of public
funds analysis (Hendren and Sprung-Keyser 2020) is another important avenue for
future research. It is worth emphasizing that because built environment is sticky,
infrastructure investments and regulatory choices made in the present day have the
potential to increase or decrease the costs of carbon mitigation efforts in the future.
Finally, welfare impacts would depend on other externalities or agglomeration ben-
efits of the intervention. For example, the types of interventions highlighted in this
paper could also impact local air pollution, congestion, traffic fatalities, and labor
market productivity. These ­co-benefits and harms have been extensively studied in
the environmental and urban economics literatures, and estimates could be incorpo-
rated into an aggregate welfare effect.

REFERENCES

Abaluck, Jason, Mauricio Caceres Bravo, Peter Hull, and Amanda Starc. 2021. “Mortality Effects
and Choice across Private Health Insurance Plans.” Quarterly Journal of Economics 136 (3):
­1557–1610.
Abowd, John M., Francis Kramarz, and David N. Margolis. 1999. “High Wage Workers and High
Wage Firms.” Econometrica 67 (2): 251–333.
Allcott, Hunt, Rebecca Diamond, Jean-Pierre Dubé, Jessie Handbury, Ilya Rahkovsky, and Molly
Schnell. 2019. “Food Deserts and the Causes of Nutritional Inequality.” Quarterly Journal of Eco-
nomics 134 (4): 1793–1844.
Allen, Treb, and Costas Arkolakis. 2022. “The Welfare Effects of Transportation Infrastructure
Improvements.” Review of Economic Studies 89 (6): 2911–57.
Andrews, M. J., L. Gill, T. Schank, and R. Upward. 2008. “High Wage Workers and Low Wage Firms:
Negative Assortative Matching or Limited Mobility Bias?” Journal of the Royal Statistical Society
Series A: Statistics in Society 171 (3): 673–97.
Auffhammer, Maximilian, and Edward Rubin. 2024. “Quantifying Heterogeneity in the Price Elastic-
ity of Residential Natural Gas.” Journal of the Association of Environmental and Resource Econo-
mists 11 (2): 319–57.
Austin, Benjamin, Edward Glaeser, and Lawrence H. Summers. 2018. “Saving the Heartland: Place-
Based Policies in 21st Century America.” Brookings Papers on Economic Activity 2018 (Spring):
­151–255.
Avenancio-León, Carlos F., and Troup Howard. 2022. “The Assessment Gap: Racial Inequalities in
Property Taxation.” Quarterly Journal of Economics 137 (3): 1383–1434.
Badger, Emily, and Quoctrung Bui. 2019. “Cities Start to Question an American Ideal: A House with
a Yard on Every Lot.” New York Times, June 2019.
Barrington-Leigh, Chris, and Adam Millard-Ball. 2017. “More Connected Urban Roads Reduce US
GHG Emissions.” Environmental Research Letters 12 (4).
Bilal, Adrien. 2023. “The Geography of Unemployment.” Quarterly Journal of Economics 138 (3):
­1507–76.
Bond, Brittany, J. David Brown, Adela Luque, and Amy O’Hara. 2014. “The Nature of the Bias
When Studying Only Linkable Person Records: Evidence from the American Community Survey.”
CARRA Working Paper 2014-08.
Bureau of Transportation Statistics. 2000–2019. [Table 4-23] Average Fuel Efficiency of U.S. Light
Duty Vehicles. US Department of Transportation. https://www.bts.gov/content/average-fuel-
efficiency-us-light-duty-vehicles.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1481

Busso, Matias, Jesse Gregory, and Patrick Kline. 2013. “Assessing the Incidence and Efficiency of a
Prominent Place Based Policy.” American Economic Review 103 (2): 897–947.
Card, David, Jörg Heining, and Patrick Kline. 2013. “Workplace Heterogeneity and the Rise of West
German Wage Inequality.” Quarterly Journal of Economics 128 (3): 967–1015.
Card, David, Jesse Rothstein, and Moises Yi. 2025. “Location, Location, Location.” American Eco-
nomic Journal: Applied Economics 17 (1): 297–336.
Chetty, Raj, and Nathaniel Hendren. 2018. “The Impacts of Neighborhoods on Intergenerational
Mobility I: Childhood Exposure Effects.” Quarterly Journal of Economics 133 (3): 1107–62.
Christensen, Peter, and Christopher Timmins. 2023. “The Damages and Distortions from Discrimina-
tion in the Rental Housing Market.” Quarterly Journal of Economics 138 (4): 2505–57.
Colas, Mark, and John M. Morehouse. 2022. “The Environmental Cost of Land-Use Restrictions.”
Quantitative Economics 13 (1): 179–223.
Combs, Tabitha. 2020. Local Actions to Support Walking and Cycling during Social Distancing Data-
set. Pedestrian and Bicycle Information Center.
Cornfield, Jerry. 2023. “Washington Makes Another Run at Heat Pump Rules.” Washington State Stan-
dard, November 28.
Cronin, Julie Ann, Don Fullerton, and Steven Sexton. 2019. “Vertical and Horizontal Redistribution
from a Carbon Tax and Rebate.” Journal of the Association of Environmental and Resource Econ-
omists 6 (S1): S169–S208.
De La Roca, Jorge, and Diego Puga. 2017. “Learning by Working in Big Cities.” Review of Economic
Studies 84 (1): 106–42.
Dijkstra, Lewis, Hugo Poelman, and Paolo Veneri. 2019. “The EU-OECD Definition of a Functional
Urban Area.” Unpublished.
D’Lima, Walter, Luis Arturo Lopez, and Archana Pradhan. 2022. “COVID-19 and Housing Market
Effects: Evidence from US Shutdown Orders.” Real Estate Economics 50 (2): 303–39.
Druckenmiller, Hannah. 2023. Urban Heat Islands 101. Resources for the Future.
Duranton, Gilles, and Diego Puga. 2020. “The Economics of Urban Density.” Journal of Economic
Perspectives 34 (3): 3–26.
Duranton, Gilles, and Anthony J. Venebles. 2021. “Place-Based Policies: Principles and Developing
Country Applications.” In Handbook of Regional Science, edited by Manfred M. Fischer and Peter
Nijkamp, 1009–30. Springer.
Eid, Jean, Henry G. Overman, Diego Puga, and Matthew A. Turner. 2008. “Fat City: Questioning the
Relationship between Urban Sprawl and Obesity.” Journal of Urban Economics 63 (2): 385–404.
Eisen, Ben. 2019. “Workers Are Fleeing Big Cities for Smaller Ones—and Taking Their Jobs With
Them.” Wall Street Journal, September 7.
Federal Highway Administration. 2019. National Household Travel Survey.
Federal Transit Administration (FTA). 2002–2019. National Transit Database. US Department of
Transportation.
Finkelstein, Amy, Matthew Gentzkow, and Heidi Williams. 2021. “Place-Based Drivers of Mortality:
Evidence from Migration.” American Economic Review 111 (8): 2697–2735.
Finkelstein, Amy, Matthew Gentzkow, and Heidi Williams. 2016. “Sources of Geographic Varia-
tion in Health Care: Evidence from Patient Migration.” Quarterly Journal of Economics 131 (4):
­1681–1726.
Fulton, William, David Garcia, Ben Metcalf, Carolina Reid, and Truman Braslaw. 2023. New Path-
ways to Encourage Housing Production:A Review of California’s Recent Housing Legislation.
Terner Center for Housing Innovation, UC Berkeley.
Gaubert, Cecile. 2018. “Firm Sorting and Agglomeration.” American Economic Review 108 (11):
­3117–53.
Geospatial Management Office. 2021. Homeland Infrastructure Foundation-Level Data. US Depart-
ment of Homeland Security.
Gillingham, Kenneth. 2014. “Identifying the Elasticity of Driving: Evidence from a Gasoline Price
Shock in California.” Regional Science and Urban Economics 47: 13–24.
Glaeser, Edward L., and Matthew E. Kahn. 2010. “The Greenness of Cities: Carbon Dioxide Emis-
sions and Urban Development.” Journal of Urban Economics 67 (3): 404–18.
Glaeser, Edward. 2013. “Urban Public Finance.” In Handbook of Public Economics, Vol. 5, edited by
Alan J. Auerbach, Raj Chetty, Martin Feldstein, and Emmanuel Saez, 195–256. Elsevier.
Glaeser, Edward L., and Joshua D. Gottlieb. 2008. “The Economics of Place-Making Policies.” Brook-
ings Papers on Economic Activity 2008 (Spring): 155–239.
Goldstein, Benjamin, Dimitrios Gounaridis, and Joshua P. Newell. 2020. “The Carbon Footprint of
Household Energy Use in the United States.” PNAS 117 (32): 19122–30.
1482 THE AMERICAN ECONOMIC REVIEW MAY 2025

Goldwyn, Eric, Alon Levy, Elif Ensari, Marco Chitti, Abdirashid Dahir, Rachel Daniell, Yinan Yao,
and Anan Maalouf. 2022. The Transit Costs Project.
Green, Tomas W., and Christopher R. Knittel. 2020. “Distributed Effects of Climate Policy: A Machine
Learning Approach.” Unpublished.
Hausman, Catherine, and Samuel Stolper. 2021. “Inequality, Information Failures, and Air Pollution.”
Journal of Environmental Economics and Management 110: 102552.
Hendren, Nathaniel, and Ben Sprung-Keyser. 2020. “A Unified Welfare Analysis of Government Pol-
icies.” Quarterly Journal of Economics 135 (3): 1209–1318.
International Energy Agency. 2021. Net Zero by 2050: A Roadmap for the Global Energy Sector.
Jenkins, Jesse D., Erin N. Mayfield, Jamil Farbes, Ryan Jones, Neha Patankar, Qingyu Xu, and Greg
Schivley. 2022. Preliminary Report: The Climate and Energy Impacts of the Inflation Reduction Act
of 2022. REPEAT Project.
Jones, Christopher, and Daniel M. Kammen. 2014. “Spatial Distribution of US Household Carbon
Footprints Reveals Suburbanization Undermines Greenhouse Gas Benefits of Urban Population
Density.” Environmental Science and Technology 48 (2): 895–902.
Kline, Patrick. 2010. “Place Based Policies, Heterogeneity, and Agglomeration.” American Economic
Review 100 (2): 383–87.
Kline, Patrick, and Enrico Moretti. 2013. “Local Economic Development, Agglomeration Economies,
and the Big Push: 100 Years of Evidence from the Tennessee Valley Authority.” Quarterly Journal
of Economics 129 (1): 275–331.
Kline, Patrick, and Enrico Moretti. 2014. “People, Places and Public Policy: Some Simple Welfare
Economics of Local Economic Development Programs.” Annual Review of Economics 6: 629–62.
Kline, Patrick, Raffaele Saggio, and Mikkel Sølvsten. 2020. “Leave-Out Estimation of Variance Com-
ponents.” Econometrica 88 (5): 1859–98.
Ko, Yekang. 2013. “Urban Form and Residential Energy Use: A Review of Design Principles and
Research Findings.” Journal of Planning Literature 28 (4): 327–51.
Kolko, Jed. 2021. “The Most Urban Counties in the US Are Shrinking.” New York Times, August 12.
Lachowska, Marta, Alexandre Mas, Raffaele Saggio, and Stephen A. Woodbury. 2023. “Do Firm Effects
Drift? Evidence from Washington Administrative Data.” Journal of Econometrics 233 (2): ­375–95.
Larsen, John, Ben King, Hannah Kolus, Naveen Dasari, Galen Bower, and Whitney Jones. 2022. A
Turning Point for US Climate Progress: Assessing the Climate and Clean Energy Provisions in the
Inflation Reduction Act. Rhodium Group.
Levinson, Arik. 2016. “How Much Energy Do Building Energy Codes Save? Evidence from California
Houses.” American Economic Review 106 (10): 2867–94.
Lyubich, Eva. 2025. Data and Code for: “The Role of People versus Places in Individual Carbon
Emissions.” Nashville, TN: American Economic Association; distributed by Inter-university Con-
sortium for Political and Social Research, Ann Arbor, MI. https://doi.org/10.3886/E211141V1.
Mahajan, Megan, Olivia Ashmoore, Jeffrey Rissman, Robbie Orvis, and Anand Gopal. 2022. Model-
ing the Inflation Reduction Act Using the Energy Policy Simulator. Energy Innovation.
Maricopa Association of Governments. 2020. 2040 Regional Transportation Plan (RTP) Update Exec-
utive Summary.
Mateyka, Peter, and Christopher Mazur. 2021. Homeownership in the United States: 2005 to 2019.
US Census Bureau.
Mervosh, Sarah. 2018. “Minneapolis, Tackling Housing Crisis and Inequity, Votes to End Single-Fam-
ily Zoning.” New York Times, December 13.
Murray, Brian C., and Peter T. Maniloff. 2015. “Why Have Greenhouse Emissions in RGGI States
Declined? An Econometric Attribution to Economic, Energy Market, and Policy Factors.” Energy
Economics 51: 581–89.
NASA. 2020. Vital Signs.
National Oceanic and Atmospheric Administration (NOAA). 2020. CONUS Climate Division Data-
base. National Centers for Environmental Information. https://www.ncei.noaa.gov/access/
monitoring/dyk/us-climate-divisions.
New York City Department of Transportation. 2023. NYC DOT Taking New Steps to Expand Bike
Infrastructure and Encourage Safe Operation of E-bikes as Overall Bike Ridership Reaches All-
time High.
Nowak, William P., and Ian Savage. 2013. “The Cross Elasticity between Gasoline Prices and Transit
Use: Evidence from Chicago.” Transport Policy 29: 38–45.
O’Brien, Brendan. 2023. “New York State Bans Natural Gas in Some New Construction.” Reuters,
May 3.
Ou, Jinpei, Xiaoping Liu, Xia Li, and Yimin Chen. 2013. “Quantifying the Relationship between Urban
Forms and Carbon Emissions Using Panel Data Analysis.” Landscape Ecology 28: 1889–1907.
VOL. 115 NO. 5 LYUBICH: THE ROLE OF PEOPLE VERSUS PLACES 1483

Payne, Heather. 2021. “The Natural Gas Paradox: Shutting Down a System Designed to Operate For-
ever.” Maryland Law Review 80 (3): 693–762.
Petek, Gabriel. 2020. Assessing California’s Climate Policies—Electricity Generation. California Leg-
islative Analyst’s Office.
Philips, Ian, Jillian Anable, and Tim Chatterton. 2022. “E-bikes and Their Capability to Reduce Car
CO2 Emissions.” Transport Policy 116: 11–23.
Pomponi, Franceso, Ruth Saint, Jay H. Arehart, Niaz Gharavi, and Bernardino D’Amico. 2021.
“Decoupling Density from Tallness in Analysing the Life Cycle Greenhouse Gas Emissions of Cit-
ies.” npj Urban Sustainability 1: 33.
Raetz, Hayley. 2021. “Housing Characteristics of Small and Mid-Sized Cities.” The Stoop, NYU Fur-
man Center Blog, January 27.
Ribeiro, Haroldo V., Diego Rybski, and Jürgen P. Kropp. 2019. “Effects of Changing Population or
Density on Urban Carbon Dioxide Emissions.” Nature Communications 10: 3204.
Roberts, David. 2019. “Cars Dominate Cities Today. Barcelona Has Set Out to Change That.” Vox,
September 11.
Rothstein, Richard. 2017. The Color of Law. Liveright Publishing Corporation.
Sallee, James M. 2019. “Pigou Creates Losers: On the Implausibility of Achieving Pareto Improve-
ments from Efficiency-Enhancing Policies.” NBER Working Paper 25831.
Shammin, Md. R., Robert A. Herendeen, Michelle J. Hanson, and Eric J. H. Wilson. 2010. “A Multi-
variate Analysis of the Energy Intensity of Sprawl versus Compact Living in the US for 2003.” Eco-
logical Economics 69 (12): 2363–73.
Song, Jae, David J. Price, Faith Guvenen, Nicholas Bloom, and Till von Wachter. 2019. “Firming up
Inequality.” Quarterly Journal of Economics 134 (1): 1–50.
Spiller, Elisheba, Heather Stephens, Christopher Timmins, and Allison Smith. 2014. “The Effect of
Gasoline Taxes and Public Transit Investments on Driving Patterns.” Environmental and Resource
Economics 59: 633–57.
Tiebout, Charles M. 1956. “A Pure Theory of Local Expenditures.” Journal of Political Economy
65 (5): 416–24.
Timmons, David, Nikolaos Zirogiannis, and Manuel Lutz. 2016. “Location Matters: Population Den-
sity and Carbon Emissions from Residential Building Energy Use in the United States.” Energy
Research and Social Science 22: 137–46.
Tomer, Adie, Joseph W. Kane, Jenny Schuetz, and Caroline George. 2021. We Can’t Beat the Climate
Crisis without Rethinking Land Use. Brookings Institution.
Transportation Research Board. 2009. Driving and the Built Environment: The Effects of Compact
Development on Motorized Travel, Energy Use, and CO2 Emissions. National Research Council of
the National Academies.
Tsivanidis, Nick. 2022. “Evaluating the Impact of Urban Transit Infrastructure: Evidence from Bogo-
ta’s TransMilenio.” Unpublished.
Ummel, Kevin. 2014. “Who Pollutes? A Household-Level Database of America’s Greenhouse Gas
Footprint.” Center for Global Development Working Paper 381.
US Census Bureau. 2001–2019. American Community Survey [dataset].
US Census Bureau. 2024. Annual Estimates of the Resident Population for Metropolitan Statistical
Areas in the United States and Puerto Rico: April 1, 2020 to July 1, 2023 (CBSA-MET-EST2023-
POP) [dataset]. https://www.census.gov/data/tables/time-series/demo/popest/2020s-total-metro-
and-micro-statistical-areas.html
US Census Bureau. 2000, 2010. Decennial Census [dataset].
US Census Bureau. 2014. American Community Survey Design and Methodology. Technical Report.
US Department of Energy. 2015. Quadrennial Energy Review: Energy Transmission, Storage, and Dis-
tribution Infrastructure. US Department of Energy.
US Department of Homeland Security (US DHS). 2021. Homeland Infrastructure Foundation-Level
Data.
US Department of Transportation. 1977. America’s Highways 1776–1976: A History of the Feder-
al-Aid Program. US Government Printing Office.
US Department of Transportation. 2024. Biden-Harris Administration Sends Another $9.9 Billion
from the Bipartisan Infrastructure Law to Transit Agencies across the Country. US Department of
Transportation.
US Department of Transportation, Federal Highway Administration. 2015. Status of the Nation’s
Highways, Bridges, and Transit: Conditions and Performance. US Department of Transportation.
US Energy Information Administration (US EIA). 1960-2019. State Energy Data System. https://
www.eia.gov/state/seds/seds-data-complete.php?sid=US.
1484 THE AMERICAN ECONOMIC REVIEW MAY 2025

US Energy Information Administration (US EIA). 2020a. Annual Electric Power Industry Report,
Form EIA-861, detailed data files. https://www.eia.gov/electricity/data/eia861.
US Energy Information Administration (US EIA). 2020b. US Energy-Related Carbon Dioxide Emis-
sions Fell in 2019, Mainly in Electric Generation. US Energy Information Administration.
US Environmental Protection Agency (US EPA). 2018. Emission Factors for Greenhouse Gas Invento-
ries. US Environmental Protection Agency.
US Environmental Protection Agency (US EPA). 2021. Emissions and Generation Resource Integrated
Database. https://www.epa.gov/egrid.
Van Nieuwerburgh, Stijn. 2023. “The Remote Work Revolution: Impact on Real Estate Values and the
Urban Environment: 2023 AREUEA Presidential Address.” Real Estate Economics 51 (1): 7–48.
Ville de Paris. 2021. “Un Nouveau Plan Vélo pour une Ville 100% Cyclable.”
Wagner, Deborah, and Mary Layne. 2014. “The Person Identification Validation System (PVS):
Applying the Center for Administrative Records Research and Applications’ (CARRA) Record
Linkage Software.” CARRA Working Paper 2014-01.
Wagner, Gernot. 2021. “Sometimes a High Rise Is Better Than a Park.” Bloomberg, July 23.
Walk Score. 2021. "Walk Score APIs." https://www.walkscore.com/professional/walk-score-apis.php.