You are on page 1of 69

The McMansion Effect: Top House Size and

Positional Externalities in U.S. Suburbs


Clement S. Bellet*

July 18, 2023

Abstract

Combining multiple waves of survey data on house satisfaction with a ge-


olocalized data set of three million suburban houses, this paper provides direct
quasi-experimental evidence of positional externalities in the field. Homeown-
ers exposed to the construction of very large houses in their suburban county are
found to be less satisfied with their own houses, but not with their neighborhoods.
This effect is conditional on the visual salience of the large houses—captured by
their proximity to roads. Indirect evidence on home size improvements within
households supports the direct evidence of reduced experienced utility. (JEL D12,
R21, Z13, I31, E70)

Keywords: Positional Goods, Suburbanization, Relative Income Hypothesis,


Subjective Wellbeing, Inequality

* Erasmus University Rotterdam, bellet@ese.eur.nl. An early version of this paper circulated under
the title “The Paradox of the Joneses: Superstar Houses and Mortgage Frenzy in Suburban America” and
was released as a CEP Discussion Paper. I am extremely grateful to Samuel Bowles, Andrew Clark, Nico-
las Coeurdacier, Stefano Dellavigna, Paul Frijters, Pedro Gardete, Steve Gibbons, Sergei Guriev, Thomas
Piketty, and Alexandra Roulet for very helpful discussions and comments. I also thank the seminar and
conference participants at Sciences Po Department of Economics, INSEE-CREST, London School of Eco-
nomics, Paris School of Economics, INSEAD, Erasmus School of Economics, 2017 ECINEQ conference,
and 2018 Berkeley Haas SICS conference for useful comments. I acknowledge financial support from
CREST, Centre for Economic Performance (LSE) and INSEAD.

Electronic copy available at: https://ssrn.com/abstract=3378131


1 Introduction
Over recent decades, U.S. suburbs have experienced an impressive upscaling in the
size of new houses: since 1980, the median size of new-build single-family houses
increased by nearly 800 square feet (75 square meters), despite a decline in average
household size. The median new-build house is now twice as large as in 1945. Over
the same period, house size inequality has widened, driven by the construction of ever
bigger houses at the top of the size distribution. In 1980, only 5% of new-build sub-
urban houses were larger than 3,000 square feet (280 square meters); on the eve of
the 2008 financial crisis, these “McMansions” represented 15% of all new construc-
tions, and the largest 10% of all homes were some four times the size of below-median
houses, compared with a factor of three 30 years earlier.1 Surprisingly, these major
improvements in material comfort did not lead to a concomitant rise in experienced
satisfaction: suburban homeowners report similar levels of house satisfaction today as
they did in the early 1980s.
Does the consumption of richer others lower the utility that consumers derive from
their own purchases? In a well-known thought experiment, Frank (2005) describes a
situation where an individual has to choose between society A, in which he will live
in a 4,000-square-foot house while others live in 6,000-square-foot houses, and soci-
ety B, in which he will live in a 3,000-square-foot house while others live in 2,000-
square-foot houses. Most people favor society B over society A, namely, a situation
in which their absolute house size is smaller but their relative house size is higher.
Such goods—whose utility depends on the relative quantity consumed—have been de-
scribed as “positional,” along with the negative externalities generated by the higher
quantities consumed by richer others (Duesenberry 1949; Frank 1985, 1991, 2005;
Hirsch 1976).
In this article, I map consumers’ experienced satisfaction with a highly consequen-
tial purchase to unpredicted but observable shocks in the (larger) quantities consumed
by others, thereby providing unique quasi-experimental evidence of positional exter-
nalities on consumption utility in the field. I focus on housing purchases, which ac-
count for around one third of household expenditures each year, and more specifically
on the size of houses in U.S. suburbs, where more than half of the U.S. population now
live. Owning a big house is arguably the most visible signal of material success in U.S.
suburbs, and housing ranks among the most positional consumption categories in the
1 In this paper, I use the terms “McMansions” and “top house size” interchangeably to characterize the
significant increase in house size driven by the top of the house size distribution.

Electronic copy available at: https://ssrn.com/abstract=3378131


lab (Solnick and Hemenway 2005).2 I show that the construction of large and visually
salient houses over homeowners’ tenure period lowers the satisfaction that they derive
from their own house.
Strictly speaking, positional externalities imply that the utility (or satisfaction) an
individual experiences from the purchase of a positional product goes down when this
same individual sees others spending more on that same product. This hypothesis is as
old as political economy itself. As early as 1847, Marx argued that “a house may be large
or small; as long as the neighboring houses are likewise small, it satisfies all social require-
ment for a residence. But let there arise next to the little house a palace, and the little house
shrinks to a hut.” Dissatisfied consumers may, in turn, “keep up with the Joneses” to
maintain their relative position and recover material satisfaction (Bertrand and Morse
2016). Consumption inequality may therefore generate dissatisfaction for households
who cannot match the consumption levels of richer households (Frank 2013, 1985).
However, the difficulty of jointly mapping the satisfaction that consumers derive from
their purchases to the observable quantities they (and others) consume over time has
seriously hindered the ability of the literature to directly test this hypothesis in the
field.
Building on past work on the measurement of experienced utility (Kahneman and
Deaton 2010; Luttmer 2005; Perez-Truglia 2020), I use an evaluative measure of house
satisfaction capturing more than 184,000 subjective ratings of houses by their owners,
I first document that within any given year, new movers to larger houses systemati-
cally report higher levels of house satisfaction. Typically, living in a 500-square-foot
larger home improves new movers’ reported house satisfaction by around 0.25 points
(about 18% of a standard deviation). However, since 1980, and despite a major up-
scaling of average home size, average house satisfaction has remained steady.3 While
this pattern is consistent with the presence of positional externalities in house size,
identifying such effects in field settings poses a number of challenges. A first source of
endogeneity relates to the decision to build larger (versus smaller) houses within par-
ticular suburbs, as many of the local drivers for new constructions, such as land prices
or real estate taxes, may correlate with house satisfaction (Rossi-Hansberg et al. 2010).
Local income shocks may also be correlated to the construction of larger houses and to
2 In Building the Dream, Wright (1983) writes: “Allotments were based on a position and wealth of a family
within the community, so that the land divisions and size of houses varied considerably.”
3 This contradiction between a strong positive correlation in cross-sectional data and an absence of (or
much weaker) positive longitudinal correlation in the long run echoes the so-called Easterlin paradox
(Easterlin et al. 2010; Kahneman and Deaton 2010) within the domain of (housing) consumption, also
documented by Bao and Meng (2023) for the United Kingdom. Of note, Easterlin’s original explanation
for this paradox relied on Duesenberry (1949)’s theory of relative consumption.

Electronic copy available at: https://ssrn.com/abstract=3378131


the wealth of existing homeowners, who may then feel dissatisfied with their current
asset. Second, households may endogenously sort across suburbs based on their sen-
sitivity to positional externalities (Bottan and Perez-Truglia 2022). For instance, those
most negatively affected by positional externalities will choose to live in suburbs with
smaller houses, which would bias any estimate towards zero.4
My identification strategy relies on two key features of the suburban housing mar-
ket. First, while the construction of other houses may correlate with homeowners’
satisfaction in various ways, positional externalities are conditional on existing home-
owners being visually exposed to larger houses (Heffetz 2011; Perez-Truglia 2013).
Because American suburbs were designed for cars, within a suburb, visual exposure
to the size of other houses is most likely to occur while driving. In order to iden-
tify houses most (versus least) likely to be seen during commute, I match an original
data set of more than three million single-family houses to their suburban road net-
work. This allows me to construct reference size measures depending on whether new
houses are built in close proximity (high visual salience) versus in isolation (low visual
salience) from the road network. Second, while homeowners may endogenously sort
into certain suburbs based on the initial characteristics of those suburbs at the time of
purchase, suburban homeowners have high moving costs, which translates into high
length of tenure. This means that at the time of purchase, homeowners will have
little ability to simultaneously anticipate the size, timing, and location of future hous-
ing developments in their suburbs, and how they will affect their future satisfaction.
Exposure to the size of houses built after their purchase has been made (or personal
histories in the construction of new houses) should be orthogonal to the initial choice
of moving. I therefore relate households’ current house satisfaction to cross-sectional
differences in the visual salience of larger houses built in their suburb over their tenure
period.
This strategy, which conditions significant treatment effects to homeowners’ visual
exposure to treatment while driving, and to the unpredictability of the treatment as-
signment, allows me to mitigate concerns related to the endogenous construction of
new houses and endogenous sorting of homeowners within a suburb. For instance, lo-
cal income shocks leading higher income homeowners to build bigger houses are not
a threat to identification as long as this effect equally applies to all new houses within
a suburb, regardless of their proximity to roads. Besides tenure period fixed effects, I
4 In a prior study conducted in parallel to this research, Kuhlmann (2020) found a positive association
between house satisfaction and house size percentile using data from the 1993 wave of the American
Housing Survey (AHS). However, it remains unclear whether the observed correlation persists once the
identification challenges mentioned earlier are adequately addressed.

Electronic copy available at: https://ssrn.com/abstract=3378131


directly control for any time-varying suburb characteristics affecting households’ cur-
rent house satisfaction regardless of when they moved (i.e. suburb-year fixed effects),
from congestion costs to local amenities or land prices.5 I show that conditional on
suburb-year effects, the size of visually-salient houses built in homeowners’ suburbs
over tenure is not predictive of their current housing market conditions. How close
from roads future houses were built is also independent from the main characteristics
of existing homeowners.
I document significant positional externalities in the American housing market.
A 1% increase in the size of houses at the top of the distribution almost offsets the
satisfaction gains from a 1% increase in absolute size. Top size—or McMansions—is
defined as the 90th percentile of the local size distribution. I find smaller or non-
significant effects when the growth in size is driven by the middle or lower percentiles
of the size distribution, consistent with the “trickle-down consumption” effect previ-
ously documented (Bertrand and Morse 2016). Importantly, I find that the reference
size effect increases linearly with the visual salience of houses built over homeown-
ers’ tenure period, with no reference effect found when new constructions are isolated
from roads. The effect could capture more general neighborhood externalities such
as aversion to residential segregation or gentrification. However, while top house size
reduces homeowner’s satisfaction with their own house, no such effect is found on
neighborhood satisfaction. I also find that the McMansion externality is stronger for
households living in larger houses.
While the direct evidence is the main contribution of this paper, I add to the prior
literature on relative consumption and discuss indirect consequences of relative size
concerns on housing markets. I look at whether people react to top size growth by
increasing their own house size. Indeed, if the effect is driven by positional exter-
nalities in house size, homeowners wishing to maintain their house satisfaction over
time may upscale the size of their own house (or perform home improvements in gen-
eral) following the construction of large houses in their suburb. I find that existing
owners strive to “keep up” with the construction of visually salient mansions: a 10%
increase in the size of visually salient houses at the top of the distribution leads ex-
isting homeowners to increase their own house size by about 1.2%. For instance, the
suburban area of Queen-Anne county (Maryland) experienced a 20% rise in size of vi-
sually salient houses at the top of the distribution between 1998 and 2007. Assuming
5 Tenure period fixed effects are more restrictive than moving cohort fixed effects used in previous
research on the causal impact of personal experiences (Malmendier and Nagel 2016; Malmendier and
Shen 2018).

Electronic copy available at: https://ssrn.com/abstract=3378131


a causal interpretation, I estimate that this effect led existing homeowners to further
increase the size of their own house by an estimated 45 square feet over the period, on
average.
A long history of behavioral research has examined how variations in own and oth-
ers’ income affect overall subjective well-being as a proxy for experienced utility (see
Clark et al. 2008; Deaton and Stone 2013, for a review). While own income or wealth
increases well-being (Kahneman and Deaton 2010; Lindqvist et al. 2020) and shows, in
some instances, evidence of marginally decreasing returns (Layard et al. 2008; Steven-
son and Wolfers 2013), others’ income reduces it (Ferrer-i Carbonell 2005; Luttmer
2005). Despite the central role played by relative consumption as an explanation for
the “Easterlin paradox”,6 the relative income literature never directly tested the rela-
tive consumption hypothesis, and its implications in terms of positional consumption.
Indeed, it implicitly equates income to the sum of consumption utilities. However,
in the absence of income transparency (Cullen and Perez-Truglia 2022; Perez-Truglia
2020), only a fraction of income can be observed by others, namely, visible consump-
tion (Heffetz 2011). This issue becomes particularly salient when looking at top in-
come, as the fraction of unobserved savings tends to increase with income. Moreover,
a multiplicity of alternative explanations may explain the relative income effect, from
job satisfaction (Card et al. 2012) to envy or fairness (Breza et al. 2018), access to pub-
lic amenities (Brodeur and Flèche 2019), mental health (Pickett and Wilkinson 2015),
or self-esteem (Bursztyn et al. 2018). By effectively mapping own and others’ con-
sumption of a highly consequential product to the satisfaction derived from that same
product, the present paper provides unique field evidence on the relative consumption
hypothesis.
These findings also contribute to the literature on positional goods and relative
consumption. First, whereas existing field evidence on conspicuous consumption ex-
clusively relies on (indirect) revealed-preferences approaches (Bertrand and Morse
2016; Bursztyn et al. 2018; Charles et al. 2009), I provide direct evidence of posi-
tional externalities in consumption. For instance, Charles et al. (2009) find that racial
minorities spend more on visible goods in states where the average income of their
group is lower, while Bertrand and Morse (2016) show that when the rich spend more
within a state, poorer households increase their own expenditures. If these effects are
driven by positional externalities, we should expect the utility derived from house-
holds’ prior quantities of visible goods to fall when consumption by others increases.
6 Easterlin’s seminal paper “Does economic growth improve the human lot?” itself referred to Duesen-
berry’s demonstration effect in consumer behavior as a way to solve the paradox (Easterlin 1974).

Electronic copy available at: https://ssrn.com/abstract=3378131


Although past literature has looked at car or clothing consumption and overall well-
being (Kuhn et al. 2011; Perez-Truglia 2013), because consumers’ satisfaction from
prior purchases (e.g. their cars or clothing) is not measured, this literature is unable to
confirm the existence of positional externalities directly, and alternative explanations
such as social learning or positive aspirations cannot be ruled out. I show that the type
of trickle-down effects evidenced in Bertrand and Morse (2016) can be traced back
to the positional attributes of highly consequential purchases. Beyond the evidence
on relative consumption utility itself, I can also address the question of the reference
group. I am able to more fully appreciate the direction of the positional externalities
by looking at which part of the size distribution acts as a reference group.7
Lastly, this paper contributes to the economics literature on housing and neighbor-
hood externalities (Guerrieri et al. 2013; Ioannides and Zabel 2003; Rossi-Hansberg
et al. 2010). Using a model of housing demand, Ioannides and Zabel (2003) provide
empirical evidence of social interaction effects on home improvements. However, the
literature on neighborhood income effects tends to emphasize the contribution of pos-
itive externalities (Guerrieri et al. 2013; Rossi-Hansberg et al. 2010). The fact that I
identify negative positional externalities in housing consumption suggests a tradeoff
between the decision to live in richer neighborhoods and the social inclusion costs im-
plied by status competition. Prior literature has documented that internal reference
points (or habits) can have long-lasting effects on house prices (Bordalo et al. 2019;
Genesove and Mayer 2001), but also that exposure to (geographically-distant) external
reference points can affect housing decisions (Bailey et al. 2018). My findings add to
this literature and provide empirical support to Loewenstein et al. (2003), who argued
that the benefits of moving to better neighborhoods may be overestimated by home-
owners who do not anticipate the satisfaction loss resulting from the behavior of their
future neighbors.8
The rest of the paper proceeds as follows. Section 2 presents the two main data sets
and discusses important stylized facts on house size and house satisfaction. Section
3 provides direct evidence of positional externalities on house satisfaction. Section
4 documents complementary indirect evidence of positional externalities on house
market outcomes. Section 5 concludes.
7 Using a sample of 93 consumers in a field experiment, Bursztyn et al. (2018) identify the presence of
positional externalities in the demand for platinum credit cards from differences in takeup rates. How-
ever, the scope of their experiment is limited to a case in which middle-income consumers affect higher-
income consumers (“trickle-up”).
8 Odermatt and Stutzer (2022) provide empirical support to this type of projection bias, showing
homeowners systematically overestimate the positive consequences of home ownership on their future
life satisfaction.

Electronic copy available at: https://ssrn.com/abstract=3378131


2 Data and Stylized Facts
2.1 House Satisfaction and the American Housing Survey
The American Housing Survey (AHS) is a survey of U.S. houses, conducted by the
U.S. Census Bureau since 1973. It provides highly detailed representative repeated
cross-sectional data on urban, rural, and suburban homeowners and collects exten-
sive information at the national and metropolitan levels on house, neighborhood, and
household characteristics, including information on house satisfaction and house mar-
ket values.9
Starting in 1984, the U.S. Census Bureau added two subjective satisfaction ques-
tions about households’ home and neighborhood, as follows: “The next two questions
are about your home and your neighborhood and how you feel about each of them, consid-
ering everything that we have talked about during this interview. On a scale of 1 to 10,
how would you rate the house as a place to live? 10 is best, 1 is worst. How would you
rate the neighborhood on a scale of 1 to 10? 10 is best, 1 is worst.”10 Both questions are
close to evaluative (or cognitive) measures of experienced utility, like the life satisfac-
tion question, as opposed to more affective measures that do not require the cognitive
effort necessary to answer evaluative questions (Diener et al. 1999). Evaluative mea-
sures tend to increase more linearly with respect to log income than affective measures
(Kahneman and Deaton 2010; Stevenson and Wolfers 2013), and are more reactive to
long-run effects (Lindqvist et al. 2020).
The measurements of experienced utility using subjective scales have a long his-
tory in behavioral economics (Kahneman and Deaton 2010; Kahneman et al. 1997)
and consumer research (Johnson and Fornell 1991; Oliver 1980; Weaver et al. 2015).
In particular, subjective satisfaction has been shown to be a good predictor of economic
choices across house-related attributes (Benjamin et al. 2014). For instance, the link
between house satisfaction and house prices is strong (Figure S14), and conditional
on house market value, house amenities or defects affect house satisfaction in the ex-
pected way (Table S3). There is also consistent evidence that evaluative measures of
satisfaction can capture the influence of past experiences or habits through “remem-
bered utility” (Kahneman and Thaler 2006; Kimball and Willis 2006). In economics,
subjective well-being has been used to evaluate utility over wealth or income (Kahne-
9 Although a household identifier can be proxied from consistent answers across waves within the
same house (e.g., age, sex, race) on a sub-sample of the dataset, the AHS is not a household panel.
10 The introductory sentence preceding the two questions was introduced only after 1995. There is
no sign of discontinuity before and after 1995 in the way new movers answered the question. In the
empirical analysis, the inclusion of survey-year fixed effects accounts for any framing bias.

Electronic copy available at: https://ssrn.com/abstract=3378131


man and Deaton 2010; Lindqvist et al. 2020), but also as an alternative to revealed-
preferences methods when it comes to the evaluation of externalities or nonmarket
goods (Luechinger and Raschky 2009; Luttmer 2005; Van Praag and Baarsma 2005).
In particular, hedonic pricing methods rely heavily on market-clearing conditions and
assumes perfect information (Roback 1982; Rosen 1974). Such conditions are unlikely
to be met, given homeowners’ moving costs, which makes the satisfaction measure
valuable in my context. In Appendix D, I discuss some of the differences distinguish-
ing the hedonic pricing approach from the house satisfaction approach and why the
hedonic pricing approach is likely to underestimate the true externality.
On average, homeowners are satisfied with their house, and more so than renters,
with a mean house satisfaction slightly above 8.5 points out of 10 between 1985 and
2013 (versus 7.5 for renters). Suburban homeowners represent more than half of
homeowners (versus 25% for rural households and 20% for urban households). Around
7% report a value below or equal to six, 10% report seven, 26.5% eight, 18.5% nine,
and 38% ten (Figure S1). Compared with the average U.S. household, suburban home-
owners have higher incomes and are known to be less representative of racial minori-
ties (Boustan 2010; Boustan and Margo 2013). In 2001, their average household in-
come was $87,000, while the national average at that time was $58,000.11 The propor-
tion of African Americans and Hispanic Americans was 6.3% and 6.2% respectively,
compared with a national average of 12.3% and 12.5% (Table S1).
Do homeowners living in larger houses report higher house satisfaction than those
living in smaller houses? Since housing is a durable good, and the AHS includes both
new and old movers, I first restrict the analysis to the flow of new suburban movers,
defined as homeowners who bought their house within the last two years. Focusing
on new movers only also means abstracting from hedonic adaptation to absolute size.
In cross-sectional data, house size systematically increases house satisfaction, as can
be seen in Figure 1, which plots the regression coefficients of log house size on house
satisfaction for each wave of new suburban movers between 1985 and 2013. The coef-
ficient on own size is always positive and significant, with a point estimate of around
0.5. This coefficient is of similar magnitude to the regression coefficient of log income
on the Cantril ladder of life satisfaction (also measured on a 1 to 10 scale) found in the
subjective well-being literature (Deaton and Stone 2013; Stevenson and Wolfers 2008).
Despite a positive relationship between income and house satisfaction in cross-
sectional data, no such correlation can be found in longitudinal data. Figure 2 reveals
11 All statistics are weighted using the AHS-provided weights and deflated to 2000 using the CPI defla-
tor from the BLS.

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure 1: House Size and House Satisfaction, New Movers
Cross-Sections (1985-2013)

1.5
Impact of log house size on house satisfaction
0 .5 -.5 1
5

3
201
198

198

198

199

199

199

199

199

200

200

200

200

200

201
Notes: House satisfaction is regressed on the log house size of new-movers, defined as homeowners who
bought their house within the last two years (N = 22,772). Coefficients on log house size are plotted
by year. Each dot corresponds to a separate regression. 90% confidence intervals drawn using robust
standard errors. Sampling weights are included. Source: AHS national surveys (1985-2013).

an Easterlin paradox in housing consumption. Here, I plot the average house size
of new movers in suburban areas between 1985 and 2013 against their average house
satisfaction. Despite an increase in average house size of more than 600 square feet be-
tween 1985 and 2007, there is no long-term rise in house satisfaction over this period.
Since average household size fell from 3.0 to 2.8, this paradox cannot be explained by a
fall in average space per person. In Appendix A.2, I show that this puzzle is robust for
both new and old movers to the inclusion of controls related to house quality (house
price per square foot and age of house) and household characteristics (e.g., household
size, length of tenure or income).
Can positional externalities explain this paradox? To answer this question, I com-
bine 18 AHS waves of metropolitan files from 1984 to 2009. The metropolitan surveys
were conducted every year from 1984 to 1996 and around every two years thereafter.
They are representative at the level of metropolitan statistical area (MSA), which com-
prises an average of five metropolitan counties. Metropolitan statistical areas (MSA)
are delineated by the U.S. Office of Management and Budget (OMB). They consists in
a core urbanized area containing a substantial population together with adjacent com-

10

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure 2: Joneses’ Paradox, New Movers

2700 10
Average house size

Average house satisfaction (1-10), new movers


Average house satisfaction
2600

Average house size (sqft), new movers


9.5
2500
9
2400

2300 8.5

2200
8
2100
7.5
2000

1900 7
85

87

89

91

93

95

97

99

01

03

05

07

09

11
13
20
19

19

19

19

19

19

19

19

20

20

20

20

20

20
Notes: The dashed-line shows the average house satisfaction of new suburban movers on a 1 to 10 scale
(vertical-left axis). The solid line shows the average size of their house in square feet (vertical-right axis).
New movers are defined as homeowners who bought their house within the last 2 years before being
surveyed (N = 22,772). Sampling weights included. Source: AHS national surveys (1985-2013).

munities having a high degree of economic and social integration with that core. A
different set of MSAs are surveyed each year, so that on average, each MSA is surveyed
four times over the period from 1984 to 2009. Each county is further divided between
its central city area and its suburban area. A “suburb” therefore corresponds to the
suburban area of a metropolitan county. These suburbs represent around 54% of the
total U.S. population and are fully representative of the suburban population. After
dropping observations with missing values, I end up with a pooled sample of 182,583
observations of homeowners surveyed between 1984 and 2009 within 152 suburbs (see
Table S2 for detailed summary statistics). A map showing the spatial distribution of
these suburbs can be found in Figure S7.
Following Deaton and Stone (2013), a first test for positional externalities consists
of comparing the coefficient of own house size on individual satisfaction at the house-
hold level with the same coefficient at the suburban level. Conditional on suburb-year
fixed effects, and whether a household’s reference house is located within the same
suburb, a household-level regression of house satisfaction on own house size will re-
veal the individual-level coefficient. However, the aggregate suburb-level regression
will reveal the difference between the individual and the suburb coefficient. In the

11

Electronic copy available at: https://ssrn.com/abstract=3378131


presence of positional externalities, the effect of a suburb’s average house size on the
suburb’s average house satisfaction will be lower than the individual-level coefficient.
If the coefficient is higher, it indicates a positive externality (e.g., public goods). I first
regress individual house and neighborhood satisfaction on the log of own house size
and a suburb-year fixed effect.12 I then run similar regressions on average suburb-year
level data. Results are shown in Table 1.
Table 1: Self-Reported Satisfaction and Positional Externalities: A Simple Test

House Satisfaction Neighborhood Satisfaction


Household-level Suburb-year average Household-level Suburb-year average
Ln(own size) 0.387∗∗∗ -0.145∗∗∗ 0.218∗∗∗ 0.364∗∗∗
(0.010) (0.050) (0.010) (0.055)
Observations 182583 508 182583 508
R2 0.271 0.551 0.260 0.611
Adjusted R2 0.269 0.545 0.258 0.606
Controls Yes Yes Yes Yes
Suburb × Year FE Yes — Yes —
Notes: House satisfaction and neighborhood satisfaction, respectively, are used as dependent vari-
ables (measured on a 1 to 10 scale), and I report the regression coefficient on log house size. Fol-
lowing Deaton and Stone (2013), controls include age, sex, and race, either at the household level or
averaged within suburb-year cells. Regressions on house (neighborhood) satisfaction also control for
neighborhood (house) satisfaction. Sampling weights are included in all regressions. Source: AHS
Metropolitan Surveys (1984-2009)

House size is positively related to both house and neighborhood satisfaction within
suburbs at the household level. However, once aggregated, the effect of size on house
satisfaction becomes negative, suggesting that the positive effect of own house size is
more than offset by the negative externality of others’ house size. Conversely, average
house size affects neighborhood satisfaction positively: the suburb-year coefficient is
higher than the microdata coefficient.13

2.2 Web Scraping of Geolocalized Housing Data


Zillow.com, a leading online real estate company in the United States, makes publicly
available information on millions of houses for sale or rent, including houses already
sold. Using web-scraping techniques, I gathered an original geolocalized sample of
around 3.2 million single-family houses located in each of the 152 metropolitan coun-
ties surveyed in the AHS. This corresponds to an average of around 20,000 houses per
12 As in Deaton and Stone (2013), I control for age, sex, and race. The regression on house (neighbor-
hood) satisfaction also controls for neighborhood (house) satisfaction.
13 This suggests a possible tradeoff between a positive neighborhood externality and a negative house
size externality.

12

Electronic copy available at: https://ssrn.com/abstract=3378131


county. I restrict my web-scraping program to houses built between 1920 and 2009. I
collected information on the location of the house (latitude and longitude), the size of
the house, its lot size, and the year the house was built. To illustrate, Figure 3 maps the
location of the nearly 70,000 single-family houses in my sample that were built within
each of the seven suburbs of the Baltimore–Columbia–Towson MSA.
Figure 3: Mapping of Houses Within Baltimore–Columbia–Towson
MSA

Notes: The Figure maps all single-family houses web-scraped from Zillow that are located within one of
the seven counties of Baltimore–Columbia–Towson MSA.

The geolocalized data set provides highly detailed information about the evolution
and distribution of house size across suburbs. Figure S2 plots the median and 90th
percentile size of all houses by year of construction. Over the last 50 years, the median
newly built house doubled in size. Considering that average household size has de-
creased by around 20% since 1960, the amount of private space per person has risen
at an even higher rate. A strong rise in size also occurred at the top of the distribution.
Between 1980 and 2009, the 90th percentile house size grew 1.4 times faster than the
median house size.
This McMansion effect becomes even more salient when looking at how the con-
struction of new houses has altered the distribution of the housing stock since 1980.
Whether I look at the size of the 90th percentile house relative to the median house for
the stock of existing units each year, or the share of houses larger than 3,000 square
feet, a clear U-shape pattern emerges (Figure S3). This rise in top house size inequality

13

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure 4: The McMansion Effect, Baltimore MSA (1920-2009)

2.2 Baltimore–Columbia–Towson MSA


Anne Arundel Suburb Carroll Suburb
Baltimore Suburb Howard Suburb

90th percentile size relative to median


2.1

1.9

1.8

1.7

1.6

1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

Notes: The Figure plots the 90th percentile house size relative to the median house size for the stock
of existing houses each year within four suburbs of the Baltimore–Columbia–Towson metropolitan area
(Maryland). A value of 1.5 means the 90th percentile house is 50% bigger than the median house in the
suburb. Source: Author’s own calculation from Zillow.

echoes the evolution of top wealth inequality over the last century (Saez and Zucman
2016).14 From 1940 to 1960, the 90th percentile size of the housing stock declined
by 10 percentage points compared with the median house. The ratio remained stable
until 1980, when the reverse trend occurred. The 90th-percentile-size house then in-
creased by almost 15 percentage points relative to the median. On the eve of the 2008
financial crisis, the ratio had recovered to the level of the 1930s. Over the same pe-
riod, the share of houses larger than 3,000 square feet increased from 5% to 15% of all
houses built. Looking at the entire distribution of houses by construction decade, it is
clear that this increase in size inequality was driven by the top of the size distribution,
which became increasingly fat-tailed after 1980 (see Figure S4).
Importantly, the distribution of house size varies greatly between and within sub-
urbs, even for houses located within the same MSA. To illustrate this, I show the evolu-
tion of the ratio of 90th percentile size to median size in four suburbs of the Baltimore
MSA for the stock of houses since 1920 (Figure 4). The latter varies between 1.6 and
2.25 times the median house size. In 1960, the 90th percentile house was 70% larger
14 Using decennial Census data, Albouy and Zabek (2016) find a similar result for home values and rent
by looking at the variance of logs, the Gini coefficient, and the Theil entropy index.

14

Electronic copy available at: https://ssrn.com/abstract=3378131


than the median-size house in both Baltimore and Howard County suburbs. Twenty
years later, these houses were only 60% larger in the Howard County suburb, while in
the Baltimore County suburb, they rose to 80%.

2.3 Proximity to Suburban Road Networks


American suburbs were designed for driving rather than walking. In fact, the vast
majority of Americans living in suburbs use their car on a daily basis, whether for
commuting, to visit friends, or to go shopping. For instance, only 7% of households in
my sample of suburban households do not own a car (vs. 32% within central Ameri-
can cities), and about 52% own at least two cars (vs. 25% only for residents of central
cities). A suburb is defined by the American Housing Survey as the area of a metropoli-
tan county surrounding the primary (or secondary) central city of the MSA. The aver-
age suburb in my data is 30 miles long and 30 miles wide, which is about 20% smaller
than an average U.S. county. Within American suburbs, a daily 10-miles commute—a
15-minutes car drive at 36 miles per hour—is frequent. In fact, respondents in my
sample report a daily commuting distance of 15.75 miles and a commuting time of 26
minutes, on average (i.e. suggesting an average speed of 36 miles per hour).15 Recent
evidence using GPS data on individual trips also find that the average shopping or
recreational trip in the U.S. is close to 10 miles (Abbiasov et al. 2022). Finally, within
American suburbs, multiple similar houses are usually built at the same time on the
same area (or tract) of land, which is also known as “tract housing.”
These features of American suburbs have three major implications in my context.
First, within a suburb, visual exposure to large (or small) houses is most likely to oc-
cur while driving. The appropriate distance for positional externalities ultimately de-
pends on homeowners’ moving behavior. In that sense, it differs from supply-side
driven housing externalities identified in previous research (Campbell et al. 2011;
Rossi-Hansberg et al. 2010). In other words, suburban homeowners are more likely
to be visually exposed to houses built in close proximity to the suburb’s main road net-
work relative to houses isolated from that network, on average. Second, when think-
ing of what a short (vs. long) distance means within American suburbs, driving time
is more relevant than walking time. Geographically distant houses within a suburb
may still affect house satisfaction as long as they are built in areas in close proximity
to roads. Third, the size of any given house within a suburb is highly predictive of
the average size of other houses around it. In fact, half of the within-county variance
15 Since 2002, the American Housing Surveys ask homeowners about their daily commuting distance
and commuting time to work.

15

Electronic copy available at: https://ssrn.com/abstract=3378131


in individual house size in my sample of geolocalized houses can be explained by the
average house size of houses built in a given year within the same zipcode.
I match each geolocalized house to its main road network using data from the Na-
tional Highway Planning Network (NHPN). The NHPN is a comprehensive geospatial
network database of the main American road system: the National Highway System
(NHS). It consists of the nation’s national highway system routes, along with all ru-
ral and urban arterials and collector roads. The advantage of this dataset is twofold.
First, because it does not include local streets with limited mobility, it is highly rep-
resentative of the actual road usage in the United States. For instance, while the NHS
represented less than a third of the Nation’s total public road miles in 2000, it carried
over 86% of the total travel.16 Second, local streets are not systematically coded by
all counties, and some streets may be added or removed over the years. Instead, the
NHS consistently covers all suburbs in my dataset, and provides a spatial coverage of
suburban road networks that remained very stable over time.
Using GIS software, I construct two measures of visual salience captured by the
proximity of a geolocalized house to its suburban road network, namely the number
of roads around the house in a radius of 5 minutes driving time at 36 mph (i.e. a ra-
dius of 2 kilometers—or 1.27 miles—using Manhattan distance) and a measure of the
nearest distance to a road (in miles). Figure 5 describes the distribution of these two
newly constructed variables for all houses bigger than 3,000 square feet (versus other
houses).17 There is significant variation in both measures, which allows me to exploit
differences in the visual salience of large or small houses within a suburb depending
on whether they have little access to the suburban road network (low visual salience)
or get built in close proximity to this network (high visual salience).
Finally, I complement the above measure with predicted geodetic distances (and
driving time) between an AHS house and any other house within its suburb. Since
the AHS does not provide coordinates, I impute missing longitude and latitude infor-
mation using a random-forest algorithm, which is trained on the geolocalized hous-
ing dataset separately within each suburb, using overlapping variables between both
datastets (house size, lot size, and construction year) as predictors (Appendix A.4 pro-
vides details on how predicted driving time is estimated). While this procedure gives
16 US Department of Transportation, Federal Highway Administration: https://www.fhwa.dot.gov/
ohim/onh00/onh2p5.htm
17 I use NHPN road network data from 2005. I also compute the same measures using data from earlier
years (1992, 1998 and 2001) and find nearly 1-to-1 correlation (r=0.95 to r=0.99). Figure S8 shows that
the full distributions of these variables obtained using 2005 data overlap almost perfectly with the ones
obtained using 1992 data.

16

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure 5: Proximity of Houses to Suburban Road Networks

Large Houses (> 3000 sqft) Large Houses (> 3000 sqft)
1
.3
Density

Density
.2
.5
.1

0 0
0 5 10 15 20 25 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5
Number of roads around the house (1.27 miles radius) Distance to nearest road (miles)

Other Houses Other Houses

.2 1
Density

Density
.1 .5

0 0
0 5 10 15 20 25 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5
Number of roads around the house (1.27 miles radius) Distance to nearest road (miles)

Notes: The Figure shows the distribution of the number or roads around a house within a 1.27 miles (2
kilometers) radius, and the distance to the nearest road (in miles) separately for all houses bigger than
3,000 square feet (upper panels) and for all houses smaller than 3,000 square feet (lower panels). I use
geolocalized housing data collected on Zillow together with data from the National Highway Planning
Network (NHPN). Roads include national highway routes, but also rural and urban arterials, and rural
and urban collectors.

me a rough approximation of the area within the surburb where an AHS house is most
likely located (hence its distance to other houses), it is subjected to various forms of
measurement errors related to the limited number of predictors used in the prediction
task itself, but also the assumptions needed to convert a distance into a driving time.
For instance, the geodetic distance measure does not account for the available net-
work of streets, one-way streets, traffic congestion, speed limits, turning time, or stop
signs and traffic lights. Proximity to road networks therefore remains my preferred
measures of visual salience within a suburb.

17

Electronic copy available at: https://ssrn.com/abstract=3378131


3 Direct Evidence of Positional Externalities
3.1 Preference for Relative House Size
Assume a household i who lives in suburb s at time t in a house of size hist . Follow-
˜ = hist /H σ , where Hst captures the
ing Abel (1990), relative size can be defined as hist st
observable (external) reference house size at time t in suburb s. I define McMansions
by the 90th percentile size within suburbs, so that σ captures households’ sensitivity
to top house size locally. House satisfaction U H can be expressed as a Cobb–Douglas
utility function:

N
H
Y βj
Uist = (hist /Hstσ )α zjist
j=1

The use of Cobb–Douglas is common in empirical analyses of durable consump-


tion (particularly housing) in the United States.18 House size is a complement to the
other features of a house, and to nonhousing consumption more generally. Coefficient
α captures the importance of size for house utility, and βj the effect of any other house
or neighborhood characteristics zjist . If 0 < α ≤ 1, house satisfaction shows diminish-
ing returns to absolute size. If σ = 0, only absolute size matters. In the presence of
positional externalities, σ > 0, and a household’s satisfaction with their own house is
negatively impacted by top house size. Notably, σ may be higher than 1 if the relative
utility motive dominates absolute utility. A negative σ would indicate a positive size
externality, for instance, because the presence of large houses is associated with better
amenities locally. The case where σ < 0 leads to the specification of utility chosen by
Guerrieri et al. (2013) to model (positive) neighborhood externalities.
Figure 6 plots nonparametrically the measure of house satisfaction as a function of
own house size in suburbs characterized by a weak (versus strong) McMansion effect.
The latter are defined, respectively, as the bottom and top quartile suburbs of my data
set in terms of the 90th percentile size of the housing stock. First, the link between
house satisfaction and house size echoes the concave relationship found between well-
being and income (Stevenson and Wolfers 2013). Figure 6 also suggests a positive σ
coefficient: a householder from a top quartile suburb living in 2,000 square feet must
increase the size of their house by around 500 square feet to experience the same level
of satisfaction than a householder from a bottom quartile suburb living in a similar-
size house. Interestingly, the drop in house satisfaction is more pronounced for larger
18 See Davis and Ortalo-Magné (2011), Fernandez-Villaverde and Krueger (2011), or Guerrieri et al.
(2013).

18

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure 6: Nonparametric House Satisfaction Curves

9.25

House satisfaction (1-10)


8.75

8.5

8.25

7.75
750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250
Own house size (sqft)

(A) Top quartile suburbs (B) Bottom quartile suburbs

Notes: Nonparametric house satisfaction curves (measured on 1 to 10 scale) as a function of own house
size (in square feet) for households living in bottom quartile (A) and top quartile (B) suburbs in terms
of the 90th percentile size of the housing stock. A household from suburb B living in a 2000 square feet
house must increase the size of his house by 500 square feet to report the same level of house satisfaction
than a household from suburb A living in a house of similar size. Sampling weights are included. Sources:
AHS metropolitan files and author’s own calculation from Zillow.

houses; households with less than 1,250 square feet seem much less affected by the Mc-
Mansion externality. Formally, this implies that decreasing marginal returns to size are
more pronounced in the presence of positional exernalities (i.e., σ > 0). This further
supports a ratio specification of relative size, as the difference specification sometimes
˜ = hist − σ Hst ) would lead to the opposite prediction.19
used in the literature (hist
H
Knowing Uist , sensitivity to the size of other houses can be directly estimated from
a log linear regression:

N
X
H
ln(Uist ) = constant + αln(hist ) − ασ ln(Hst90 ) + βj ln(zjist ) + uist (1)
j

19 Since σ > 0 and since house satisfaction can be shown to be a concave function of relative size (0 <
∂U H /∂hist   
α ≤ 1), sign( ist
∂Hst
= sign − α 2 σ < 0. This specification of relative size implies that the interaction
term between own and reference house size should be negative, and diminishing marginal utility to size
should be more pronounced under higher levels of σ .

19

Electronic copy available at: https://ssrn.com/abstract=3378131


This specification is similar to the reduced-form regressions commonly estimated
in the literature on relative income and relative consumption.20

3.2 Identification Strategy


The identification of σ in equation (1) poses a number of challenges related to i) omit-
ted variables affecting house satisfaction and correlated to the decision to build larger
houses and ii) the endogenous sorting of households across suburbs.

Visual Salience. The decision to build large houses may be driven by unobserved
suburban characteristics that in turn correlate with house satisfaction. Some of those
characteristics likely affect all houses, i.e. regardless of where a house ranks in the local
size distribution. This provides a first placebo test of positional externalities: I expect
the estimated σ coefficient will be insignificant when house size growth is driven by
lower percentile of the distribution (e.g. the 10th house size percentile), rather than
the top of the distribution (i.e. the 90th size percentile).
However, this provides only a weak test as unobserved shocks affecting rich house-
holds more than poor households may still right skew the house size distribution, and
be endogenous to satisfaction. I therefore propose a stronger placebo test. Unlike most
local drivers of house construction, positional externalities are conditional on existing
homeowners being visually exposed to larger houses. To identify positional externali-
ties in house size, I rely on the fact that, on average within a suburb, the visual salience
of other houses is higher when these houses are built in close proximity to (versus in
isolation from) the suburban road network. Importantly, this should be true “on av-
erage”, i.e. regardless of where exactly within the suburb an existing homeowner is
located. Indeed, as long as those houses can be observed while driving, new construc-
tions do not need to occur within a household’s “local” neighborhood for positional
externalities to be identified.21 I define visual salience as the average number of roads
around a reference house in a radius of 5 minutes driving time (i.e. at 36 mph, a radius
of 2 kilometers—or 1.27 miles—using Manhattan distance). I also show robustness us-
ing a measure of the average distance to the nearest road (in miles).
20 See in particular Luttmer (2005), Charles et al. (2009), or Bertrand and Morse (2016). Notably, the
well-being literature usually reports level-log or ordered logit estimates. Results from these alternative
specifications are shown in Appendix C.5.
21 The main concern here relates to more general types of housing externalities that would also affect the
value of the house, for instance. However, unlike positional externalities (Frank et al. 2014), those type of
externalities are hyper-local: Rossi-Hansberg et al. (2010) show they decline by half approximately every
1,000 feet.

20

Electronic copy available at: https://ssrn.com/abstract=3378131


This allows me to exclude drivers of house size growth that would equally apply
to all new housing developments regardless of how close to the roads new houses get
built. Similarly, I can exclude factors of house constructions that correlate with their
proximity to roads as long as those factors equally apply to all houses regardless of
their rank in the house size distribution. In other words, a positive and significant σ
coefficient estimated on house size growth driven by the top of the distribution that
i) does not replicate for larger houses built in isolation from the road network and ii)
does not replicate for house size growth driven by the bottom of the size distribution,
is consistent with positional externalities in house size.

Pre- vs. Post-Purchase Constructions. Homeowners visually exposed to larger houses


may simply share stronger preferences for bigger houses. In this case, households who
are least negatively affected by positional externalities will endogenously sort into sub-
urbs with less equally sized distributions (or where large houses are more visible),
while those who are most negatively affected will live in suburbs with fewer (or more
isolated) McMansions (Bottan and Perez-Truglia 2022). Regressing house satisfaction
on the 90th percentile house size using the full distribution of houses at t, hence in-
cluding the stock of houses that were already built at the time the household decided
to buy, will result in σ biased toward zero. Furthermore, in the presence of social in-
teractions, the fact that a new mover and her neighbors may move to a new house at
the same time makes it difficult to know who influences whom (also known as Manski
(1993)’s reflection problem).
To depart from this endogenous sorting bias, I rely exclusively on houses con-
structed after the purchase has been made (vs. prior to purchase). Rather than taking
the entire stock of visually salient (versus isolated) McMansions at t in a given suburb,
i.e., Hst90 , I construct a household-level measure of visually salient houses (or Hist
90
).
This measure captures the 90th percentile size of visually salient houses built over
their tenure period (i.e. post-purchase). As Loewenstein et al. (2003) stress, “if a person
buys a small house in a wealthy neighborhood in part because it has a certain status
value in her apartment building, she may not fully appreciate that her frame of refer-
ences may quickly become the larger houses and bigger cars that her new neighbors
have.” Indeed, the extensive literature on mispredicted utility suggests that accurate
“affective forecasting” is very unlikely (see Kahneman and Thaler 2006, for a review).
The identification of σ hence relies on the unpredictability of the treatment assign-
ment for existing homeowners. More specifically, it makes the plausible assumption
that homeowners are unable to simultaneously anticipate (i) the timing and location

21

Electronic copy available at: https://ssrn.com/abstract=3378131


of future constructions, (ii) the future size of those constructions, and (iii) how their
future size will affect their house satisfaction.

Empirical Specification. Given that Cobb–Douglas utility seems to match the rela-
tionship between house satisfaction and (relative) house size well (section 3.1), I esti-
mate log-log regressions of the following form:

H 90
log(Uist ) = α+γ1 log(hist )+γ2 log(Hist )+βXist +SuburbY earst +T enureP eriodit +uist (2)

H
where Uist is the house satisfaction evaluated from 1 to 10 at t by i in s, hist captures
90
homeowners’ own house size, and Hist the reference house size. Reference size is de-
fined as the 90th percentile size of visually salient houses built after purchase. I use the
number of roads within a 5 minutes driving time around the house (2 kilometers ra-
dius) as my main measure of visual salience and, for each individual household, define
visually salient houses as houses built over the household’s tenure period that belong
to the top quartile of houses in terms of visual salience. I also construct a series of
placebo measures of reference house size by varying both the reference size threshold
(e.g. lower size percentile) and visibility of houses (e.g. lower visual salience quartile).
In this specification, σ can be interpreted as the ratio between two elasticities
(σ = −γ2 /γ1 ), where coefficients γ1 and γ2 correspond to the elasticity of house sat-
isfaction with respect to own and others’ house size, respectively. In other words, a σ
coefficient of 1 implies that a household visually exposed to house size growth at the
90th percentile must increase the size of their own house at the same rate to maintain
a constant house satisfaction level.
Xist is an extensive list of controls for household (and house) characteristics, be-
sides house size.22 Besides addressing potential sources of omitted variable bias, those
controls make sure major indirect effects of positional externalities on house market
outcomes (e.g. home prices, or home size improvements) that may also affect house
satisfaction are held constant. On top of this large set of controls, I add a set of
508 dummies for suburb-year effects (SuburbY earst ), which allows me to compare
households who share similar average suburban characteristics. This controls for any
suburb-specific characteristics affecting households’ current house satisfaction regard-
less of when they moved, such as the current characteristics of the housing stock, the
22 Household controls include age, race, sex, household size, education, annual income, number of cars,
tenure length, mortgage conditions (monthly mortgage payments, term and interest rate) and overall
neighborhood satisfaction evaluated by the household at t on a scale from 1 to 10. House controls include
house market value per square foot, local real estate taxes, age of house, quality features (balcony, porch,
extra bathrooms, heating equipment, recent remodeling), and defects (holes, water leak).

22

Electronic copy available at: https://ssrn.com/abstract=3378131


level of land prices, or any local differences in preferences for larger houses. Finally,
I add a full set of 1162 dummies for i’s tenure period.23 This is critical as households
who moved more recently may be currently more happy about their house, but at the
same time would have been exposed to larger houses, while those who purchased later
would have been exposed to a mix of larger and smaller houses, on average. Finally,
because the Metropolitan AHS data results from separate samples collected by suburb-
year, residuals are likely to be correlated within suburb-year cells. For sampling de-
sign reasons (Abadie et al. 2023), I hence adjust the standard errors for clustering on
suburb-year.24
Overall, the main identification compares households who reside within suburbs
that share similar characteristics at any point in time, but who experienced different
construction shocks of visually salient (versus not visually salient) houses after pur-
chase.

3.3 Main Results


Table 2 Panel A shows the result of the pooled ordinary least-squares (OLS) regression
(2). Column (1) shows the main effect of own house size and reference house size on
house satisfaction without any controls, then controlling for household characteristics
in column (2). As expected, the coefficient on own size is positive and significant, but
significantly negative for reference size. In columns (3) to (5), I add suburb-year fixed
effects, tenure period fixed effects, and finally the full list of house characteristics. The
magnitude and significance of the coefficient on own and others’ house size remain
quite stable across specifications. In terms of effect size, a 10% increase in the 90th
percentile house size after purchase has a negative effect on house satisfaction that is
as bad as reporting an inside water leak.25 Table 2 also reports the corresponding σ
coefficient resulting from those estimates. The estimated σ on top house size goes from
0.52 to 0.82.
In column (6), I control for alternative measures of reference house size. I con-
struct four such measures: the 50th and 10th percentile size of visually salient houses
built after purchase, the 90th percentile size of non-visually salient houses built after
23 Tenure period is the interaction between moving year and survey year.
24 Table S5 demonstrates the robustness of our results to alternative clustering of standard errors.
25 See Table S3 for the full table. The sign of each coefficient is consistent with what should be ex-
pected: home improvements or neighborhood satisfaction are positively related to house satisfaction,
while an aging house, higher real estate taxes, water leaks, or a dysfunctional heating system reduce
house satisfaction. The market value of the house per square foot is also positively associated with house
satisfaction.

23

Electronic copy available at: https://ssrn.com/abstract=3378131


Table 2: Evidence of Positional Externalities in House Size

Main Effect Placebo Tests


(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Ln(House Satisfaction)
Ln(own size) 0.080∗∗∗ 0.054∗∗∗ 0.054∗∗∗ 0.055∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗
Electronic copy available at: https://ssrn.com/abstract=3378131

(0.002) (0.002) (0.002) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Ln(reference size) -0.041∗∗∗ -0.033∗∗∗ -0.040∗∗∗ -0.045∗∗∗ -0.039∗∗∗ -0.046∗∗∗ -0.013 0.012 -0.003 -0.007
(0.006) (0.006) (0.009) (0.009) (0.009) (0.011) (0.009) (0.011) (0.008) (0.011)
σ 0.52∗∗∗ 0.62∗∗∗ 0.74∗∗∗ 0.82∗∗∗ 0.77∗∗∗ 0.90∗∗∗ 0.25 -0.24 0.06 0.13
Observations 182,570 182,570 182,570 182,570 182,570 182,009 182,570 182,570 182,425 182,180
R2 0.030 0.184 0.194 0.203 0.223 0.223 0.223 0.223 0.223 0.223
Panel B: Ln(Neighborhood Satisfaction)
Ln(reference size) -0.005 0.008 0.031∗∗ 0.030∗∗ 0.031∗∗ 0.006 0.047∗∗∗ 0.022 0.013∗ 0.012
(0.007) (0.007) (0.013) (0.013) (0.012) (0.015) (0.012) (0.016) (0.008) (0.017)
Observations 182,570 182,570 182,570 182,570 182,570 182,009 182,570 182,570 182,425 182,180
R2 0.019 0.172 0.183 0.192 0.196 0.196 0.196 0.196 0.196 0.196
Controls:
24

Household Controls — Yes Yes Yes Yes Yes Yes Yes Yes Yes
Suburb × Year FEs — — Yes Yes Yes Yes Yes Yes Yes Yes
Tenure Period FEs — — — Yes Yes Yes Yes Yes Yes Yes
Full House Controls — — — — Yes Yes Yes Yes Yes Yes
Placebo Reference Size Controls — — — — — Yes — — — —
Reference House Specification:
Size Percentile 90th 90th 90th 90th 90th 90th 50th 10th 90th 90th
Visual Salience High High High High High High High High Low High
Post-Purchase Yes Yes Yes Yes Yes Yes Yes Yes Yes No
Notes: OLS coefficients reported. Panel A: House satisfaction (logged) as the dependent variable. Coefficients on own house size and reference size (logged)
reported. Panel B: Neighorhood satisfaction (logged) as the dependent variable. All regressions include own house size (logged) and reference size. Household
controls include age, race, sex, household size, education, annual income (logged), number of cars, monthly mortgage payments, mortgage conditions (term
and interest rate) and overall neighborhood satisfaction. House controls include house market value per square foot (logged), local real estate taxes (logged),
age of house, quality features (balcony, porch, extra bathrooms, heating equipment, recent remodeling), and defects (holes, water leak). High (low) visual
salience corresponds to the top (bottom) quartile of houses within suburbs in terms of proximity to roads (i.e. number of roads within 5 minutes driving
time). Robust standard errors clustered at the suburb-year level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
purchase (i.e. the bottom quartile of houses in terms of visual salience within sub-
urb), and the 90th percentile size of visually salient houses built before purchase. In
this more restrictive specification, the estimated σ is now close to 0.9, which means
the average homeowner must upscale their house at almost the same rate as top house
size if they want to maintain their level of house satisfaction over time (the difference
between both coefficients is not statistically significant). This effect is similar in mag-
nitude to prior estimates on the relationship between self-reported happiness, own
income and neighbors’ income (see Perez-Truglia 2020, for a review).
In columns (7) to (10) of Table 2 Panel A, I run a series of placebo tests where I
replace my main measure of reference size by each one of the alternative measures de-
scribed above. This allows me to more directly test three key features of my identifica-
tion strategy. First, past literature defines positional externalities as upward-looking,
which means they can be traced back to house size growth driven by the top of the size
distribution (Bertrand and Morse 2016; Frank et al. 2014).26 Using the 50th or 10th
percentile of visually salient houses built after purchase leads to a non-significant σ
coefficient, as shown in columns (7)-(8). Figure S11 (left panel, high visual salience)
plots the estimated σ coefficients, replacing the 90th size percentile of visually salient
houses by its 70th, 50th, 30th and 10th size percentile. The results show a clear and
consistent decrease in the estimated σ coefficient when house size growth after pur-
chase is driven by lower percentiles of the size distribution. This suggests the effect
is not driven by general changes in the housing stock unrelated to house size, like the
number of new constructions. In Table S4, I look at the total number of new houses
built in the household’s suburb over tenure (logged) as an additional placebo test. I
find no significant effects on house satisfaction, but a negative and significant effect on
neighborhood satisfaction. Controlling for this measure does not affect my estimates
on reference house size.
Second, positional externalities in house size are conditional on existing homeown-
ers being visually exposed to larger houses. I therefore expect the effect of reference
house size to be smaller (or non-significant) for houses built in isolation from the sub-
urban road network. This is because those houses are less visually salient, on average.
Column (9) uses as reference size the 90th percentile size of non-visually salient houses
(bottom quartile) built after purchase. In stark contrast to column (5), the estimated
σ coefficient is close to 0. Figure 7 plots the estimated σ by quartile of visual salience.
It shows a clear linear decline of the effect as the visual salience of larger houses goes
26 See Frank et al. (2014) or Bertrand and Morse (2016) for further discussion and evidence on trickle-
down consumption and expenditure cascades.

25

Electronic copy available at: https://ssrn.com/abstract=3378131


down, while no such effect is found for house size growth at the bottom of the size
distribution (10th percentile). Similarly, the higher sensitivity to house size growth
driven by higher percentiles of the size distribution in the high visual salience con-
dition does not replicate for non-visually salient houses (Figure S11; right panel, low
visual salience).27

Figure 7: Positional Externalities, Visual Salience and Trickle-Down Effects

(a) By visual salience (b) By own house size


2
1
Sensitivity to Size of Reference Houses

Sensitivity to Size of Reference Houses


.75 1.5

.5 1
(σ coefficient)

(σ coefficient)
.25
.5

0
0
-.25

-.5
-.5

-.75 -1
4th (High) 3rd 2nd 1st (Low) Above median Below median
Visual Salience (Quartile) Own House Size

Size Percentile of Reference Houses:


10th 90th

Notes: Figure (a) plots the sensitivity coefficients (σ = −γ2 /γ1 ) to the 90th and 10th percentile house size
of houses built after purchase by quartile of visual salience. Visual salience is measured by the proximity
to roads of reference houses (i.e. number of roads within 2 kilometers radius). Figure (b) plots the
sensitivity coefficients to the 90th and 10th percentile house size of visually salient houses (top quartile)
for homeowners living in above (vs. below) median house size within suburbs. Sampling weights are
included in all regressions. The 95% confidence intervals are drawn using standard errors clustered at
suburb-year level.

Are positional externalities stronger or weaker for households already living in


large houses? Figure 7 plots the σ coefficient of equation (2), estimated on homeown-
ers who live in above (vs. below) median house size within their suburb. The results
show higher sensitivity to top house size for owners of larger houses. In fact, positional
externalities are twice as large for owners or large houses compared to owners of be-
low median size houses. For the former, satisfaction returns from house size is mostly
27 If I estimate reference house size regardless of their proximity to roads, the coefficient estimate is less
precisely estimated (p < 0.05). However, I still find a negative coefficient on reference house size.

26

Electronic copy available at: https://ssrn.com/abstract=3378131


relative (i.e. σ > 1). Moreoever, the construction of smaller houses does not signifi-
cantly affect households already living in equally small houses. This suggests that i)
positional externalities are stronger at the top of the size distribution and ii) upward-
looking “trickle-down” comparison effects seem more prevalent than peer effects (or
downward-looking comparisons), at least within American suburbs.
Overall, these results are consistent with positional externalities (Bertrand and
Morse 2016; Frank et al. 2014): any relative size effect comes down to the largest
houses built, those effects are conditional on constructions being visually salient (i.e.
observable) and made after-purchase (i.e. unanticipated), and households ranked down
the hierarchy are less affected by the largest houses than households who live in large
houses themselves. If affected customers react to relative size depreciation by moving
to larger houses, or increasing own house size, positional externalities in size may in
turn amplify the rise in top house size inequality previously described.

3.4 Robustness
Neighborhood Satisfaction. Rather than positional externalities, coefficient σ could
capture a more general form of neighborhood externality. For instance, the construc-
tion of McMansions may be associated with disutility from gentrification, or a signal
residential segregation, as it alters the composition of neighbors.28 The opposite effect
could also arise if gentrification or residential segregation lead to positive neighbor-
hood externalities. The inclusion of neighborhood satisfaction as a control variable al-
ready captured these effects. However, a more direct test can be performed through re-
placing house satisfaction by neighborhood satisfaction on the left-hand side of equa-
tion (2). Panel B of Table 2 reports the resulting γ2 coefficient on reference house size.
If anything, exposure to very large houses generate positive externalities on neighbor-
hood satisfaction. Moreover, contrary to house satisfaction, neighborhood satisfaction
is driven by size increases coming from the middle of the size distribution (50th per-
centile house size). Overall, these results confirm the evidence discussed in Table 1.
Positional externalities in size are distinct from general neighborhood satisfaction ef-
fects, as both measures behave in opposite ways with respect to others’ consumption
levels.29
28 The construction of McMansions could also be considered unaesthetic, which should mostly lower
neighborhood satisfaction.
29 Those results suggest households may face a tradeoff between the social benefits associated with
living in neighborhoods with larger houses and the social costs of maintaining their relative status, espe-
cially when growth in size is driven by the top of the distribution.

27

Electronic copy available at: https://ssrn.com/abstract=3378131


Visual Salience. I exploit two alternative measures capturing homeowners’ visual
exposure to reference houses. First, using the geolocalized dataset, I construct an
additional measure of visual salience using the average distance between reference
houses and the nearest road (in miles). I replicate the analysis on the 90th house size
of houses built after purchase by quartile of visual salience using this measure and
find similar effects (Table S6). Second, I exploit a more direct (although noisier) proxy
of household-level visual exposure using the predicted driving time (at 36 miles per
hour) between a household’s own house and newly constructed houses after purchase
(see Appendix A.4 for the measure’s construction). Column (1) of Table S7 shows that
the coefficient on predicted driving time is negative but not significantly related to
house satisfaction. However, the interaction between top house size and driving time
shown in column (2) is positive and significant.30 Figure S12 plots the estimated σ co-
efficients obtained on the size of visually-salient (vs. non-visually salient) houses as a
function of the predicted (driving) time to newly constructed units. It shows the mag-
nitude of positional externalities is highest within the first 5 minutes of driving time,
and quickly drops until 15 minutes (i.e. about 10 miles). Beyond that point, predicted
driving time does not seem to affect my estimates. I also find a clear drop of estimated
σ at any level of predicted driving time when looking at houses built in isolation from
the road network (low visual salience).

Forecasting Ability. Identification relies on homeowners’ inability to anticipate the


impact of post-purchase visual exposure to larger houses on their house satisfaction.
A few households may still be able to form rational beliefs on tear-down candidates
and large empty lots based on what they see of the original housing stock at the time
of purchase. In this less restrictive case, regression (2) underestimates the true exter-
nality. Formally, it only identifies the presence (or absence) of positional externalities
(i.e., σ > 0) rather than an unbiased estimate of the true parameter σ . As a corollary,
top house size at time of purchase should have no significant impact on their current
house satisfaction. For each homeowner, I construct the 90th percentile size of all vi-
sually salient houses built in their suburb prior to purchase. I then use this variable
as a placebo measure of reference size. Column (10) in Table 2 confirms pre-purchase
exposure has no discernible impact on their current house satisfaction.
30 The impact of within-suburb driving time persists in column (3) after excluding Californian counties,
which tend to be much larger than the U.S. average.

28

Electronic copy available at: https://ssrn.com/abstract=3378131


Supply-Side. Suburban supply-side factors (e.g. real estate taxes or land prices) may
drive the decision to build a house close or far from the suburban road network (Rossi-
Hansberg et al. 2010). However, conditional on suburb-year fixed effects, the drivers
of new constructions are likely exogenous to the local characteristics of existing home-
owners.31 My data suggest that this is indeed the case. First, I test if within a suburb-
year, drivers of new constructions are systematically related to the local housing mar-
ket conditions of existing homeowners. I regress reference house size over tenure on i)
the log of homeowners’ house market values32 and ii) the log of real estate taxes paid,
conditional on suburb-year fixed effects. Table S8 confirms the absence of correlation
between the size of visually salient houses built over existing homeowners’ tenure and
these two variables. Second, I test whether the visual salience of new houses itself is
correlated to their house and household characteristics. For each household, I compute
the total number of roads (within a radius of 1.27 miles, or 2 kilometers) surrounding
the largest 10% houses built in their suburb over tenure and regress this variable on
house and household characteristics. Households more likely to be visually exposed to
new houses do not significantly differ in terms of house satisfaction, own house size,
or house market values. Conditional on suburb-year and tenure period fixed effects,
how close from the road network houses get built is not systematically related to any
of the main characteristics of existing homeowners (Table S9, column (2)).33

Attrition Bias. The use of Zillow to recover historical variation in reference house
size may be problematic if attrition or home-remodeling effects alter the measurement
of past house size in a systematic way. For instance, assuming that large mansions
are more likely to be demolished than smaller houses, the average historical size of
older houses will be underestimated. Conversely, assuming that improvements are
made to the size of old houses, their average historical size will be overestimated. I
address this concern in Appendix C.4 by comparing the mean to median house size
ratio in Zillow with the equivalent measure from the U.S. Census Bureau’s Survey of
Construction (SOC) across regions. Looking at a period of nearly 40 years, there is a
perfect correlation between Zillow and SOC data, and no evidence of a diverging gap
31 As long as it is visually salient (i.e. can be observed while driving), a reference house does not have
to be built in the homeowner’s specific neighborhood to generate positional externalities. And unlike
positional externalities (Frank et al. 2014), these supply-side effects are hyper-local (Campbell et al. 2011;
Rossi-Hansberg et al. 2010).
32 Market values in the AHS are higher in levels than transaction prices, but have similar time-series
patterns (DiPasquale and Somerville 1995; Kiel and Zabel 1999).
33 Local real estate taxes is the only predictor that remains weakly significant, but the effect size is very
small.

29

Electronic copy available at: https://ssrn.com/abstract=3378131


in size distribution over time between the two data sets (Table S10).

Functional Forms. The preferred log-log regressions in Table 2 directly result from
the Cobb–Douglas specification of house satisfaction in Section 3.1. They echo the
relative consumption literature and lead to a straightforward interpretation of σ . Al-
though the use of OLS remains widespread, the literature typically also estimates or-
dered probit regressions. However, recent work on the cardinality of subjective scales
show that when respondents use response scales in an approximately linear manner,
OLS estimates are valid (see Kaiser and Vendrik 2020, for a review). I follow a similar
approach as Kristoffersen (2017) and regress the log of house market values on dum-
mies for each of the response categories of the house satisfaction question.34 Figure
S14 confirms the validity of the linearity of scale use assumption. Table S11 also re-
produces the main estimates from Table 2 using level-log and ordered probit models.
The results still hold: the coefficients on own and reference house size, along with their
interactions, remain significant and of the expected signs and relative magnitude.

Heterogeneity. Identification relies on post-purchase constructions, which should


reduce concerns related to the endogenous sorting of homeowners across suburbs.
However, I test whether the effects are driven by specific groups of the population,
or specific suburbs. I look for heterogeneous effects on major household character-
istics such as household size (one or two members versus three or more), household
income (above versus below median) or household education (lower versus higher ed-
ucation) but find little evidence for it (Table S12). I then test whether households who
spend more time commuting are also more sensitive to the size of houses built in prox-
imity to the suburban road network. While this test only relies on more recent waves
of the AHS (starting in 2002), the estimated σ coefficient goes from about 1 for home-
owners with shorter commutes (12 minutes, 7.5 miles on average) to about 1.5 for
homeowners with longer commutes (40 minutes, 25 miles on average). Finally, I split
the sample by suburb size (in square miles) and suburban population density (above
vs. below median). While the suburb size does not seem to drive any heterogeneous
effects, more densely populated counties show somewhat higher sensitivity to visually
salient houses.
34 Given the lack of responses below 5, I regroup the latest as a ”smaller than 5 category” (0.9% of
observations).

30

Electronic copy available at: https://ssrn.com/abstract=3378131


Moving Costs. In the presence of positional externalities, negatively affected home-
owners may ultimately decide to sell their house and move out of the suburb. This cre-
ates an additional attrition issue as the selected sample of those who stayed translates
into a measured impact of the McMansion effect that underestimates the true exter-
nality. However, a bias may only arise if the cost of staying (i.e., the satisfaction loss)
outweighs the cost of moving.35 I look at whether the construction of visually salient
McMansions after purchase explain homeowners’ decision to leave a house. In any
survey year, I identify as “movers” households whose house becomes occupied by new
owners in the next survey year (versus those who remain in the same housing unit).36
I run a linear probability model using this indicator as my new dependent variable
in regression (2).37 Table S13 shows there is little evidence that the construction of
visually salient McMansions in homeowner’s suburbs significantly predicts moving
decisions. If anything, homeowners living in above median size houses are less likely
to move after experiencing larger increase in top house size. This suggests high mov-
ing costs. Under those conditions, existing owners are more likely to react through
home improvements, a possibility explored in the next Section.

4 Direct vs. Indirect Evidence of Positional Externalities


The identification of positional externalities, i.e. that the utility (or satisfaction) con-
sumers experience from their own purchase goes down when these same individuals
see others spending more on these same products, represents the primary contribu-
tion of this paper. Indeed, building on relative utility theory (and the relative income
hypothesis), I presented direct field evidence of significant positional externalities in
highly consequential purchases. Specifically, post-purchase visual exposure to larger
houses nearly offsets homeowners’ satisfaction with their own house size, while no
such effect is found on neighborhood satisfaction. Those results are consistent with
the “Joneses’ Paradox” documented in Figure 2, i.e. that fact that since 1980, and de-
spite a major upscaling of average home size, average house satisfaction has remained
steady in U.S. suburbs.
In this final section, I complement the direct evidence with indirect evidence of po-
35 Prior evidence suggests that moving costs are particularly high for homeowners relative to renters
and are an increasing function of the length of tenure (DiPasquale and Glaeser 1999).
36 Although the data is not structured as a household panel, I identify the same household from con-
sistent answers within the same housing unit to questions about householder’s sex, age, and race, along
with the year in which the household moved to the unit and unit purchase price.
37 By design, the sample size is smaller as I cannot recover information on the decision to move for the
last tenants of any given house.

31

Electronic copy available at: https://ssrn.com/abstract=3378131


sitional externalities, echoing past work on relative consumption and relative income
(Bertrand and Morse 2016; De Giorgi et al. 2020). This literature adopts a revealed
preferences approach where own consumption is regressed on other’s consumption
(or income). Here, I examine whether existing homeowners react to the construction
of larger houses through home size improvements. I discuss the magnitude of these in-
direct effects in comparison to the existing literature, and to direct evidence on house
satisfaction, highlighting the importance of considering both types of evidence in un-
derstanding the full impact of positional externalities in consumer markets.

4.1 Empirical Approach


So far, house market outcomes that may be indirectly affected by the presence (or ab-
sence) of positional externalities, like own house size, were held constant. In the pres-
ence of positional externalities in size, however, homeowners who wish to maintain
their level of house satisfaction may keep up with the Joneses by moving to an even
larger home, or by adding extra space (or rooms) to their current home. Home size im-
provements is a likely behavioral outcome if, as documented in Section 3.4, suburban
homeowners have high moving costs.
A relevant parameter to estimate is the elasticity of homeowners’ house size to
changes in top house size. Such a parameter, however, cannot be measured using cross-
sectional data on current house size. Indeed, it requires within-household changes in
own house size. Therefore, a more promising approach is to look at the decision of
existing homeowners to add to the size of their own house over time, as their visual
exposure to top house size increases.
To do so, I rely on a particular feature of the AHS, which is that the same housing
units (not homeowners) within an MSA may be repeatedly surveyed from one wave to
the next, on average every five years. I then assume households who provide consis-
tent answers within each housing unit to questions about householder’s sex, age, race,
moving year, and unit purchase price, should be considered as the same households.
After removing missing values, I end up with a short, unbalanced panel of 27,461
homeowners surveyed two to four times between 1984 and 2009. Using this much
smaller sample, I estimate the following OLS-FE model:

log(hist ) = ξlog(Hst90 ) + βXist + µi + Γt + uist (3)

where regression (3) takes (logged) house size as the dependent variable. Each regres-
sion includes household/house fixed effects µi and year fixed effects Γt . I use Hst90 as the

32

Electronic copy available at: https://ssrn.com/abstract=3378131


measure of reference size, namely the 90th percentile size of visually salient houses in
suburb s during year t. Like before, visually salient houses belong to the top quartile
of houses within suburb in their proximity to roads. In this within-household specifi-
cation, ξ captures the elasticity on own house size to an increase in house size coming
from the construction of very large houses in proximity to roads. On top of the log
of own house size, I also use an alternative measure of remodelling which is the total
number of bedrooms and bathrooms in the house. Xist is a vector of major time varying
variables that may explain home improvements aside from visual exposure to larger
houses. This includes changes in the house market value per square foot (logged),
local real estate taxes (logged), household income (logged), education dummies, mort-
gage conditions (term and interest rates), neighborhood satisfaction, number of cars
and household size. Like before, I vary the measure of reference house size looking at
lower percentiles of the size distribution, and more or less visually salient houses. In
all regressions, robust standard errors are clustered at the suburb-year level.

4.2 Indirect Evidence on Home Improvements


Table 3 (Panel A) shows the main results from the within-household regressions. Panel
A uses the log of own house size as the dependent variable. Conditional on household
and year fixed effects, the coefficient on reference house size is positive and signifi-
cant with an estimated elasticity of home size improvements to top house size of 0.12
(column (1)). Every 10% increase in the size of visually salient houses at the top of
the distribution leads existing homeowners to increase their own house size by about
1.2%. Those results replicate when looking at the total number of bedrooms and bath-
rooms in the house (Table 3; Panel B).
The estimated coefficients remain unchanged controlling for the full list of time-
varying controls (column (2)). When using the 90th percentile size of houses built
in isolation from roads (bottom quartile) as a placebo measure of reference size, the
estimates are not significantly different from zero (column (3)). Keeping both measures
in the same regression does not affect the main coefficient estimates on reference house
size (column (4)). Columns (5) and (6) also test whether the effect of reference size
is robust to controlling for changes in suburban house size driven by the bottom of
the size distribution. Again, the 10th percentile size of visually salient houses does
not significantly explain homeowners’ choice to remodel their house by adding extra
space or rooms. When controlling for both placebo measures (column (7)), only new
and visually-salient constructions at the top of the distribution significantly explain
home size improvements.

33

Electronic copy available at: https://ssrn.com/abstract=3378131


Table 3: Indirect Evidence of Positional Externalities (Within-Households)

(1) (2) (3) (4) (5) (6) (7)


Panel A: Ln(Own size)
Ln(reference size) 0.121∗∗ 0.132∗∗∗ 0.125∗∗∗ 0.129∗∗∗ 0.124∗∗∗
(0.047) (0.047) (0.043) (0.050) (0.046)
Ln(90th, low salience) 0.070 0.033 0.033
(0.065) (0.058) (0.058)
Ln(10th, high salience) 0.135 0.017 0.010
(0.096) (0.095) (0.094)
Observations 61,631 61,631 61,631 61,631 61,631 61,631 61,631
R2 0.974 0.975 0.975 0.975 0.975 0.975 0.975
Panel B: Bedrooms/Bathrooms
Ln(reference size) 0.352∗∗ 0.338∗∗ 0.378∗∗ 0.411∗∗∗ 0.441∗∗∗
(0.141) (0.141) (0.148) (0.152) (0.157)
Ln(90th, low salience) -0.089 -0.199 -0.179
(0.200) (0.192) (0.196)
Ln(10th, high salience) -0.049 -0.423 -0.385
(0.314) (0.306) (0.313)
Observations 61,631 61,631 61,631 61,631 61,631 61,631 61,631
R2 0.906 0.907 0.907 0.907 0.907 0.907 0.907
Household FEs Yes Yes Yes Yes Yes Yes Yes
Year FEs Yes Yes Yes Yes Yes Yes Yes
Controls — Yes Yes Yes Yes Yes Yes
Notes: OLS-FE models reported. Log of own house size used as dependent variable. Controls include
the house market value per square foot (logged), local real estate taxes (logged), household income
(logged), education dummies, mortgage conditions (term and interest rates), neighborhood satisfac-
tion, number of cars and household size. High (low) visual salience corresponds to the top (bottom)
quartile of houses within suburbs in terms of proximity to roads (i.e. number of roads within 1.27
miles radius). Robust standard errors clustered at the suburb-year level reported in parentheses.
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

4.3 Effect Size and Generalizability


What do these results mean for our understanding of the average house size trends
prior to the Great Recession of 2008? In 1998, single-family houses in the suburban
area of Queen-Anne county (Maryland) had an average size of 2,000 square feet. The
county experienced a 20% rise in the 90th percentile house size of visually salient
houses between 1998 and 2007. As a result and based on the estimates from Table 3,
existing homeowners added an extra 45 square feet (about 4 square meters) to their
houses, on average. In sum, those results suggest positional externalities may lead
existing owners to upscale or remodel their homes.
More generally, is an elasticity coefficient of 0.12 high or low relative to prior re-
duced form estimates on relative consumption? Using cross-sectional variation on top
income across U.S. states over time, Bertrand and Morse (2016) document an elastic-
ity of non-rich consumption (i.e. households ranked below the 80th percentile of the
income distribution) to the income of the rich (the 90th percentile of the income distri-

34

Electronic copy available at: https://ssrn.com/abstract=3378131


bution) of 0.214. Consistent with our results, they find that trickle-down effects better
explain relative consumption patterns than peer consumption effects, at least in the
U.S. context. They also find stronger effects when conditioning on the mean visibility
of consumer expenditures, as measured in surveys. Besides methodical differences,
effect size heterogeneity may therefore relate to the ability of the existing literature
to fully capture the visibility of consumption. Indeed, the moderating role of con-
sumption visibility distinguishes status consumption and positional externalities from
more general forms of consumption externalities, like peer effects. For instance, us-
ing a measure of consumer spending from administrative data of Danish households,
De Giorgi et al. (2020) estimate an elasticity of own consumption to peer consump-
tion of 0.3. However, they find no evidence that peer effects differ between visible and
non-visible goods.
Perhaps more importantly, however, while these estimates have strong economic
relevance, they remain suggestive of relative consumption utility and may seriously
underestimate the full extent of positional externalities. As documented in this paper,
consumers exposed to a rise in other’s consumption but constrained in their ability
to keep up with others’ consumption levels may still feel dissatisfied with their own
purchases. In that sense, the indirect evidence can limit one’s ability to capture the
full extent of positional externalities in contexts where consumers are most likely to
be constrained, like housing. In this case, using a revealed preferences approach based
on moving decisions would wrongly conclude for the absence of positional external-
ities in house size. The direct evidence on house satisfaction documented earlier is
less sensitive to those constraints, and can help identify latent behavioral responses to
positional externalities.
In fact, a significant amount of work in economics has relied on subjective mea-
sures of well-being (happiness, life satisfaction or job satisfaction) to test for the pres-
ence or absence of relative income effects. Direct tests of the relative income hypothesis
typically regress affective or cognitive measures of well-being on own and other’s in-
come. Prior research has estimated a wide range of coefficients for the relative income
effect, from 0.22 (Perez-Truglia 2020) to 0.82 (Luttmer 2005). My main estimate of a
0.82 coefficient for relative consumption utility falls within this range and suggests
that positional externalities on consumption may be driving the relative income effect.
While recent field studies have investigated the relative income hypothesis in settings
where income can be directly observed (Perez-Truglia 2020), a complete understand-
ing of the underlying mechanisms remains challenging. This paper provides novel
field evidence of the direct impact of other’s observable consumption levels on con-

35

Electronic copy available at: https://ssrn.com/abstract=3378131


sumption utility itself, contributing to a deeper understanding of the relative income
effect.

5 Conclusion
This paper provided direct field-level evidence of positional externalities in the Amer-
ican housing market. I identify positional externalities from cross-sectional variation
in house satisfaction and the size of visually-salient houses built within a household’s
suburb over tenure. I show that a rise in size of visually-salient McMansions—defined
as the 90th percentile size of houses built in close proximity to the suburban road
network—nearly offsets the satisfaction gains from living in a larger house. The ev-
idence is consistent with the puzzling fact that despite a major upscaling in size of
single-family houses in U.S. suburbs since 1980, households did not experience a sig-
nificant increase in house satisfaction. I complement the cross-sectional analysis with
indirect evidence of positional externalities on house market outcomes. I show that
within-households, changes in visual exposure to McMansions also affects homeown-
ers’ decision to upscale the size of their own house.
These results have major implications for our understanding of consumer markets
in unequal economies. They imply consumption inequality driven by the top of the
distribution can lower the satisfaction consumers derive from their own purchases.
Competing for relative size becomes a zero sum game (Frank 2013). For those who de-
cide to upscale, once the satisfaction gains of living in a relatively large house vanish,
households may be left with the negative long-run impact of their decisions in terms
of extra costs (from extra energy costs to re-furnishing costs). While I cannot directly
test whether relative visible consumption (or consumption inequality) better explains
life satisfaction than relative income (or income inequality), this paper also calls for
a reassessment of the relative income hypothesis in terms of positional consumption
choices.
The documented magnitude of positional externalities also suggest regulations
aimed at restricting the size of houses at the top of the distribution, from zoning
laws to housing permits, may be welfare improving (Frank 2005). Such regulations
already exist in a number of U.S. counties, but their overall welfare impact is yet to be
evaluated. In most cases, they take the form of maximum lot size ratios that restrict
the size of houses. However, suburban communities tend to favor the extensive use
of minimum lot size requirements, which may have amplified positional externalities
and increased financial distress, with no long-term improvements in house satisfaction

36

Electronic copy available at: https://ssrn.com/abstract=3378131


overall.

References
Abadie, A., S. Athey, G. W. Imbens, and J. M. Wooldridge (2023). When should you adjust
standard errors for clustering? The Quarterly Journal of Economics 138(1), 1–35.
Abbiasov, T., C. Heine, E. L. Glaeser, C. Ratti, S. Sabouri, A. S. Miranda, and P. Santi (2022).
The 15-minute city quantified using mobility data. Technical report, National Bureau of
Economic Research.
Abel, A. B. (1990). Asset prices under habit formation and catching up with the joneses. The
American Economic Review, 38–42.
Albouy, D. and M. Zabek (2016). Housing inequality. Technical report, National Bureau of
Economic Research.
Bailey, M., R. Cao, T. Kuchler, and J. Stroebel (2018). The economic effects of social networks:
Evidence from the housing market. Journal of Political Economy 126(6), 2224–2276.
Bao, H. X. and C. C. Meng (2023). Housing wealth distribution, inequality and residential
satisfaction. Regional Studies, 1–14.
Benjamin, D. J., O. Heffetz, M. S. Kimball, A. Rees-Jones, et al. (2014). Can marginal rates of
substitution be inferred from happiness data? evidence from residency choices. American
Economic Review 104(11), 3498–3528.
Bertrand, M. and A. Morse (2016). Trickle-down consumption. Review of Economics and Statis-
tics 98(5), 863–879.
Bordalo, P., N. Gennaioli, and A. Shleifer (2019). Memory and reference prices: an application
to rental choice. In AEA Papers and Proceedings, Volume 109, pp. 572–76.
Bottan, N. L. and R. Perez-Truglia (2022). Choosing your pond: location choices and relative
income. Review of Economics and Statistics 104(5), 1010–1027.
Boustan, L. P. (2010). Was postwar suburbanization “white flight”? evidence from the black
migration. The Quarterly Journal of Economics 125(1), 417–443.
Boustan, L. P. and R. A. Margo (2013). A silver lining to white flight? white suburbanization
and african–american homeownership, 1940–1980. Journal of Urban Economics 78, 71–80.
Breza, E., S. Kaur, and Y. Shamdasani (2018). The morale effects of pay inequality. The Quar-
terly Journal of Economics 133(2), 611–663.
Brodeur, A. and S. Flèche (2019). Neighbors’ income, public goods, and well-being. Review of
Income and Wealth 65(2), 217–238.
Bursztyn, L., B. Ferman, S. Fiorin, M. Kanz, G. Rao, et al. (2018). Status goods: Experimental
evidence from platinum credit cards. The Quarterly Journal of Economics 133(3), 1561–1595.
Campbell, J. Y., S. Giglio, and P. Pathak (2011). Forced sales and house prices. American
Economic Review 101(5), 2108–31.
Card, D., A. Mas, E. Moretti, and E. Saez (2012, October). Inequality at Work: The Effect of
Peer Salaries on Job Satisfaction. American Economic Review 102(6), 2981–3003.

37

Electronic copy available at: https://ssrn.com/abstract=3378131


Charles, K. K., E. Hurst, and N. Roussanov (2009, May). Conspicuous Consumption and Race.
The Quarterly Journal of Economics 124(2), 425–467.
Clark, A. E., P. Frijters, and M. A. Shields (2008). Relative income, happiness, and utility: An
explanation for the easterlin paradox and other puzzles. Journal of Economic Literature 46(1),
95–144.
Cullen, Z. and R. Perez-Truglia (2022). How much does your boss make? the effects of salary
comparisons. Journal of Political Economy 130(3), 766–822.
Davis, M. A. and F. Ortalo-Magné (2011). Household expenditures, wages, rents. Review of
Economic Dynamics 14(2), 248–261.
De Giorgi, G., A. Frederiksen, and L. Pistaferri (2020). Consumption network effects. The
Review of Economic Studies 87(1), 130–163.
Deaton, A. and A. A. Stone (2013). Two happiness puzzles. The American Economic Re-
view 103(3), 591.
Diener, E., E. Suh, R. E. Lucas, and H. L. Smith (1999). Subjective well-being: Three decades
of progress. Psychological Bulletin 125(2), 276–302.
DiPasquale, D. and E. L. Glaeser (1999). Incentives and social capital: Are homeowners better
citizens? Journal of Urban Economics 2(45), 354–384.
DiPasquale, D. and C. T. Somerville (1995). Do house price indices based on transacting units
represent the entire stock? evidence from the american housing survey. Journal of Housing
Economics 4(3), 195–229.
Duesenberry, J. (1949). Income, saving, and the theory of consumer behavior.
Easterlin, R. A. (1974). Does economic growth improve the human lot? some empirical evi-
dence. In Nations and households in economic growth, pp. 89–125. Elsevier.
Easterlin, R. A., L. A. McVey, M. Switek, O. Sawangfa, and J. S. Zweig (2010). The happiness–
income paradox revisited. Proceedings of the National Academy of Sciences 107(52), 22463–
22468.
Fernandez-Villaverde, J. and D. Krueger (2011). Consumption and saving over the life cycle:
How important are consumer durables? Macroeconomic Dynamics 15(05), 725–770.
Ferrer-i Carbonell, A. (2005). Income and well-being: an empirical analysis of the comparison
income effect. Journal of public economics 89(5-6), 997–1019.
Frank, R. (2013). Falling Behind: How Rising Inequality Harms the Middle Class, Volume 4. Univ
of California Press.
Frank, R. H. (1985). The demand for unobservable and other nonpositional goods. The Ameri-
can Economic Review, 101–116.
Frank, R. H. (1991). Positional externalities. Strategy and Choice. Cambridge, Mass.: The MIT
Press.
Frank, R. H. (2005). Positional externalities cause large and preventable welfare losses. Amer-
ican economic review 95(2), 137–141.
Frank, R. H., A. S. Levine, and O. Dijk (2014). Expenditure cascades. Review of Behavioral
Economics 1(1–2), 55–73.
Genesove, D. and C. Mayer (2001). Loss aversion and seller behavior: Evidence from the

38

Electronic copy available at: https://ssrn.com/abstract=3378131


housing market. The Quarterly Journal of Economics 116(4), 1233–1260.
Guerrieri, V., D. Hartley, and E. Hurst (2013). Endogenous gentrification and housing price
dynamics. Journal of Public Economics 100, 45–60.
Heffetz, O. (2011). A test of conspicuous consumption: Visibility and income elasticities. Re-
view of Economics and Statistics 93(4), 1101–1117.
Hirsch, F. (1976). Social Limits to Growth. Cambridge, Massachusetts: Havard University Press.
Ioannides, Y. M. and J. E. Zabel (2003). Neighbourhood effects and housing demand. Journal
of Applied Econometrics 18(5), 563–584.
Johnson, M. D. and C. Fornell (1991). A framework for comparing customer satisfaction across
individuals and product categories. Journal of Economic Psychology 12(2), 267–286.
Kahneman, D. and A. Deaton (2010). High income improves evaluation of life but not emo-
tional well-being. Proceedings of the National Academy of Sciences 107(38), 16489–16493.
Kahneman, D. and R. H. Thaler (2006). Anomalies: Utility maximization and experienced
utility. Journal of economic perspectives 20(1), 221–234.
Kahneman, D., P. P. Wakker, and R. Sarin (1997). Back to bentham? explorations of experienced
utility. The quarterly journal of economics 112(2), 375–406.
Kaiser, C. and M. Vendrik (2020). How threatening are transformations of happiness scales to
subjective wellbeing research? Available at SSRN 3743129.
Kiel, K. A. and J. E. Zabel (1999). The accuracy of owner-provided house values: The 1978–
1991 american housing survey. Real Estate Economics 27(2), 263–298.
Kimball, M. and R. Willis (2006). Utility and happiness. University of Michigan, 1–67.
Kristoffersen, I. (2017). The metrics of subjective wellbeing data: An empirical evaluation
of the ordinal and cardinal comparability of life satisfaction scores. Social Indicators Re-
search 130(2), 845–865.
Kuhlmann, D. (2020). Coveting your neighbour’s house: understanding the positional nature
of residential satisfaction. Housing Studies 35(6), 1142–1162.
Kuhn, P., P. Kooreman, A. Soetevent, and A. Kapteyn (2011). The effects of lottery prizes on
winners and their neighbors: Evidence from the dutch postcode lottery. American Economic
Review 101(5), 2226–47.
Layard, R., G. Mayraz, and S. Nickell (2008). The marginal utility of income. Journal of Public
Economics 92(8-9), 1846–1857.
Lindqvist, E., R. Östling, and D. Cesarini (2020). Long-run effects of lottery wealth on psycho-
logical well-being. The Review of Economic Studies 87(6), 2703–2726.
Loewenstein, G., T. O’Donoghue, and M. Rabin (2003). Projection bias in predicting future
utility. The Quarterly Journal of Economics 118(4), 1209–1248.
Luechinger, S. and P. A. Raschky (2009). Valuing flood disasters using the life satisfaction
approach. Journal of Public Economics 93(3), 620–633.
Luttmer, E. F. (2005). Neighbors as negatives: Relative earnings and well-being. The Quarterly
Journal of Economics 120(3), 963–1002.
Malmendier, U. and S. Nagel (2016). Learning from inflation experiences. The Quarterly Journal
of Economics 131(1), 53–87.

39

Electronic copy available at: https://ssrn.com/abstract=3378131


Malmendier, U. and L. S. Shen (2018). Scarred consumption. Technical report, National Bureau
of Economic Research.
Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem. The
Review of Economic Studies 60(3), 531–542.
Odermatt, R. and A. Stutzer (2022). Does the dream of home ownership rest upon biased
beliefs? a test based on predicted and realized life satisfaction. Journal of Happiness Studies,
1–33.
Oliver, R. L. (1980). A cognitive model of the antecedents and consequences of satisfaction
decisions. Journal of Marketing Research 17(4), 460–469.
Perez-Truglia, R. (2013). A test of the conspicuous–consumption model using subjective well-
being data. The Journal of Socio-Economics 45, 146–154.
Perez-Truglia, R. (2020). The effects of income transparency on well-being: Evidence from a
natural experiment. American Economic Review 110(4), 1019–1054.
Pickett, K. E. and R. G. Wilkinson (2015). Income inequality and health: a causal review. Social
science & medicine 128, 316–326.
Roback, J. (1982). Wages, rents, and the quality of life. Journal of Political Economy, 1257–1278.
Rosen, S. (1974). Hedonic prices and implicit markets: Product differentiation in pure compe-
tition. Journal of Political Economy 82(1), 34–55.
Rossi-Hansberg, E., P.-D. Sarte, and R. Owens III (2010). Housing externalities. Journal of
Political Economy 118(3), 485–535.
Saez, E. and G. Zucman (2016). Wealth inequality in the united states since 1913: Evidence
from capitalized income tax data. The Quarterly Journal of Economics 131(2), 519–578.
Solnick, S. J. and D. Hemenway (2005). Are positional concerns stronger in some domains than
in others? The American Economic Review 95(2), 147–151.
Stekhoven, D. J. and P. Bühlmann (2011). Missforest—non-parametric missing value imputa-
tion for mixed-type data. Bioinformatics 28(1), 112–118.
Stevenson, B. and J. Wolfers (2008). Economic growth and subjective well-being: Reassessing
the easterlin paradox. Technical report, National Bureau of Economic Research.
Stevenson, B. and J. Wolfers (2013). Subjective well-being and income: Is there any evidence
of satiation? Technical report, National Bureau of Economic Research.
Van Praag, B. and B. E. Baarsma (2005). Using happiness surveys to value intangibles: The
case of airport noise*. The Economic Journal 115(500), 224–246.
Weaver, K., K. Daniloski, N. Schwarz, and K. Cottone (2015). The role of social comparison for
maximizers and satisficers: Wanting the best or wanting to be the best? Journal of Consumer
Psychology 25(3), 372–388.
Wright, G. (1983). Building the Dream: A Social History of Housing in America. MIT press.

40

Electronic copy available at: https://ssrn.com/abstract=3378131


Supplementary Online Materials
The McMansion Effect: Top House Size and Positional Exter-
nalities in U.S. Suburbs
Clement Bellet

Appendix A Data and Stylized Facts


Appendix A.1 Descriptive Statistics

Table S1: Summary Statistics (National Sample, Selected Waves)

1985 1993 2001 2009


Annual income (2000 thousands dollars) 59.0 61.4 87.0 72.3
House market value (2000 thousands dollars) 136.4 157.2 204.8 231.2
Length of tenure (years) 13.6 13.9 13.9 14.6
Monthly mortgage payments (2000 thousands dollars) 0.4 0.5 0.7 0.8
Age of household’s head (years) 50.2 51.0 51.3 52.2
Household size (number) 3.0 2.9 2.9 2.8
Bachelor degree (%) 15.2 17.4 20.9 23.8
Graduate degree (%) 12.2 14.5 12.4 14.6
Latinos (%) 3.3 4.2 6.2 8.1
Blacks (%) 4.5 5.1 6.3 6.9
Observations 10151 11433 12826 14335
Notes: Summary statistics for the main characteristics of suburban homeowners across four selected
waves of the AHS National Surveys. All statistics are weighted using the AHS-provided weights and
deflated to 2000 using the CPI deflator from the BLS.

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S2: Summary Statistics (Metropolitan Sample)

Variable Obs Mean Std. Dev.


Household characteristics
Housing satisfaction 182583 8.67 1.45
Neighborhood satisfaction 182583 8.45 1.66
Annual household income 182583 73620.32 58749.72
Total mortgage monthly payment 182583 670.52 699.54
Mortgage’s interest rate 182583 5.09 4.41
Mortgage term (years) 182583 14.91 12.13
Number of cars 182583 1.63 .95
Household size 182583 2.96 1.42
Respondent’s age 182583 50.06 14.89
12th grade or below 182583 .32 .47
High School Graduate 182583 .08 .28
Some College 182583 .24 .43
Undergraduate degree 182583 .21 .41
Graduate degree 182583 .15 .36
Black respondent 182583 .05 .22
Hispanic respondent 182583 .06 .23
Male respondent 182583 .72 .45
Lenght of tenure 182583 12.77 11.72
House characteristics
House size (sqft) 182583 2163.89 1227.77
House market value per square foot 182583 97.48 107.67
Annual real estate tax payment 182583 766.48 1906.25
Size change last 12 months 182583 .02 .15
Number of bathrooms 182583 1.77 .73
Age of house 182583 30.23 19.7
Inside water leaks 182583 .09 .29
Heating equipment broken 182583 .02 .13
Unit has porch/balcony 182583 .9 .31
Reference size over tenure (visually salient)
90th percentile size 182570 3306.08 767.1
50th percentile size 182570 1970.94 402.28
10th percentile size 182570 1249.68 215.51
Reference size over tenure (non-visually salient)
90th percentile size 182425 3871.38 1873.96
50th percentile size 182425 2259.48 614.2
10th percentile size 182425 1405.65 458.22
Notes: Summary statistics for the main characteristics of suburban homeowners, AHS Metropolitan Sur-
veys (1984-2009). All statistics are weighted using the AHS-provided weights and deflated to 2000 using
the CPI deflator from the BLS.

ii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S1: Distribution of Self-Reported House Satisfaction

25 40
35
30
Percent
20 15
10
5
0

1 2 3 4 5 6 7 8 9 10
House satisfaction (1-10)

Note: Histogram of house satisfaction reported on a 1 to 10 scale by AHS suburban homeowners between
1985 and 2013 (pooled sample). Statistics are weighted using the AHS-provided weights. Source: AHS
National Surveys (1985-2013)

Figure S2: General Upscaling (1920-2009), New Houses Built

5000
90th percentile size 50th percentile size

4000
Size of new houses (sqft)

3000

2000

1000
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

Notes: The Figure plots the median house size (solid line) and 90th percentile size (dashed line) of new-
build houses by year of construction from 1920 to 2009. Source: Author’s own calculation from Zillow.

iii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S3: The McMansion Effect (1920-2009), U.S. Suburbs

1.9

15
% Houses Bigger than 3000 sqft
1.85
P90/P50 house size ratio

10
1.8

5
1.75 1.7

0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

P90/P50 house size ratio % Houses > 3000 sqft

Notes: Figure S3 takes the stock of houses each year and plots the 90th percentile house size relative to
the median house size between 1920 and 2009 (left axis), along with the share of houses bigger than 3000
square feet (right axis). Source: Author’s own calculation from Zillow.

Figure S4: Distribution of House Size by Construction Decade


.0008

1960s 1970s 1980s


1990s 2000s
.0006
Density
.0004 .0002
0

0 2000 4000 6000 8000 10000


Housing size (sqft)
kernel = epanechnikov, bandwidth = 44.0575

Notes: Kernel distribution of house size by decade (1960-2010). Source: Author’s own calculation from
Zillow.

iv

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix A.2 Joneses’ Paradox: Robustness

Figure S5: House Size and House Satisfaction (With Controls)

1.5
Impact of log house size on house satisfaction
0 .5 -.5 1
5

3
201
198

198

198

199

199

199

199

199

200

200

200

200

200

201
House satisfaction regressed on log house size of homeowners each year. Controls: price per square
foot, age of the house, household income, monthly mortgage payments, household size, householder’s
age, race, sex, education, length of tenure and number of cars. Coefficients on log house size are plotted
by year. Each dot corresponds to a separate regression. 90% confidence intervals drawn using robust
standard errors. Sampling weights included. Source: AHS national surveys (1985-2013).

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S6: Joneses’ Paradox, All Homeowners (With Controls)

2500 .2
Average house size
Residual house satisfaction

Residual house satisfaction (1-10)


2400
.1
Average house size (sqft)

2300

0
2200

-.1
2100

2000
-.2
85

87

89

91

93

95

97

99

01

03

05

07

09

11
13
20
19

19

19

19

19

19

19

19

20

20

20

20

20

20

Notes: Dashed-line: average residual house satisfaction of homeowners each year (vertical-left axis) after
controlling for the price per square foot, age of the house and household characteristics (income, monthly
mortgage payments, household size, householder’s age, race, sex, education, length of tenure and number
of cars). Vertical-right axis: average house size. Sampling weights included. Source: AHS national
surveys (1985-2013).

vi

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix A.3 Mapping of Suburban Houses

Figure S7: Mapping of Metropolitan Counties

Notes: The figure maps the 152 counties and pseudo-counties located within the 45 Metropolitan Statis-
tics Areas (MSA) surveyed in the American Housing Survey. These locations cover 55% of the total US
population and almost the entire suburban population.

vii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S8: Proximity of Houses to Suburban Road Networks: 1992
vs. 2005

Large Houses (> 3000 sqft) Large Houses (> 3000 sqft)
.3 1
Density

Density
.2
.5
.1

0 0
0 5 10 15 20 25 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5
Number of roads around the house (1.5 miles radius) Distance to nearest road (miles)

Other Houses Other Houses

.2 1
Density

Density
.1 .5

0 0
0 5 10 15 20 25 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5
Number of roads around the house (1.5 miles radius) Distance to nearest road (miles)

1992 2005

Notes: Distribution of the proximity to the suburban road network measures for the full sample of geolo-
calized houses. We compare the visual salience measures obtained using NHPN data from either 2005 or
1992 (i.e. 13 years gap). Roads include all national highway routes, rural and urban arterials, and rural
and urban collector roads.

viii

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix A.4 Predicted Driving Time Measure
This appendix describes the construction of the predicted driving time measure. I do this in
three steps.

Step 1. First, I impute missing longitude and latitude information for each house in the AHS
sample based on overlapping variables with the geolocalized sample.
To impute longitude and latitude to each AHS house, I follow Stekhoven and Bühlmann
(2011) and train a random forest algorithm within each suburb separately. Multiple imputa-
tion methods based on linear regressions do not fully consider the dependence between each
of the imputed variables, which in the case of spatial data is particularly important. Random
forest allows for interactions between continuous and discrete variables within suburbs and
simultaneously predicts both latitude and longitude. I use longitude and latitude as target
variables and unit size, lot size, year of construction and any possible interactions between
these variables as predictors.

Figure S9: Cumulated Distribution of Correlation Coefficients Between True and


Simulated Latitudes/Longitudes
1 .75
Cumulative distribution
.25 .5 0

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
Correlation coefficient between true and simulated values (within suburbs)

Longitude Latitude

To test the robustness of the imputation method, I simulate random missing values for
each suburbs in Zillow in the same proportions as the corresponding missing values in the
AHS. I then look at the correlation coefficient between imputed and true values within suburbs.
Figure S9 plots the distribution of the correlation coefficients for all suburbs. More than half of
the suburbs have a correlation higher than 0.3, and 25% show a correlation higher than 0.5.38
38 Due to the much larger size of Californian suburbs, removing the state of California greatly improves
the average suburb prediction.

ix

Electronic copy available at: https://ssrn.com/abstract=3378131


Step 2. Second, using the imputed longitude and latitude data together with the geolocalized
dataset, I compute the predicted geodetic distance (in miles) between an AHS house and each
other geolocalized house located in the same suburb as the AHS house, i.e. the length of the
shortest curve between two points along the surface of a mathematical model of the earth. I
use the geodist command available on Stata. For each AHS homeowner in the sample, I then
compute the average distance between the homeowner’s own house and all other houses built
over homeowner’s tenure period in her suburb.
Figure S10 shows the distribution of the estimated average distance between a household’s
house and all other houses built during his tenure period. The average household is located
about 10 miles away from these other houses. Considering these are suburban homeowners,
such a distance is easily reachable by car, which is the principal commuting mode used by
American suburban households.

Figure S10: Distribution of Predicted Distance Between AHS House and New-Build
Houses (Sources: author’s own computation from AHS and Zillow)
10
8
6
Percent
4
2
0

0 5 10 15 20 25 30 35 40 45 50
Average estimated distance from new-build houses (miles)

Step 3. Finally, I convert the predicted distance measure into an average predicted driving
time (in minutes) using the following formula:

(Distance × 60)
DrivingT ime = (4)
36
I assume an average speed of 36 miles per hour, which corresponds to the average speed
reported by respondents in my sample. Indeed, since 2002, the American Housing Surveys ask
homeowners about their daily commuting distance and commuting time to work. Respondents
in my sample report a daily commuting distance of 15.75 miles and a commuting time of 26

Electronic copy available at: https://ssrn.com/abstract=3378131


minutes, on average.

xi

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix B Main Results

Table S3: Positional Externalities in House Size, Full Table

Ln(house satisfaction)
Coeff. S.E.
Ln(own size) 0.051∗∗∗ (0.003)
Ln(reference size) -0.039∗∗∗ (0.009)
Household Controls
Ln(income) 0.004∗∗∗ (0.001)
Total mortgage monthly payment 0.000 (0.000)
Mortgage’s interest rate -0.001∗∗∗ (0.000)
Mortgage term (years) 0.000 (0.000)
Number of cars 0.001 (0.001)
Household size -0.005∗∗∗ (0.001)
Respondent’s age 0.001∗∗∗ (0.000)
Respondent’s education -0.005∗∗∗ (0.000)
Black respondent 0.017∗∗∗ (0.003)
Hispanic respondent 0.021∗∗∗ (0.004)
Male respondent -0.004∗∗∗ (0.001)
Ln(neighborhood satisfaction) 0.278∗∗∗ (0.006)
Other House Controls
Ln(house value per sqft) 0.020∗∗∗ (0.002)
Ln(real estate tax payment) 0.001 (0.001)
Size change last 12 months 0.017∗∗∗ (0.005)
Number of bathrooms 0.014∗∗∗ (0.001)
Age of house -0.001∗∗∗ (0.000)
Inside water leaks -0.031∗∗∗ (0.002)
Heating equipment broken -0.033∗∗∗ (0.006)
Unit has porch/balcony 0.020∗∗∗ (0.003)
Observations 182570
R2 0.223
Adjusted R2 0.216
Tenure Period FE Yes
Suburb × Year FE Yes
Notes: OLS coefficients reported. Reference size is defined as the 90th percentile size of visually
salient houses built over households’ tenure period. Sampling weights are included in all regressions.
Robust standard errors clustered at the suburb-year level reported in parentheses. ∗ p < 0.10, ∗∗ p <
0.05, ∗∗∗ p < 0.01

xii

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S4: Reference House Size vs. New Constructions Over Tenure Period

Ln(House Satisfaction) Ln(Neighborhood Satisfaction)


(1) (2) (3) (4) (5) (6)
Ln(reference size) -0.039∗∗∗ -0.038∗∗∗ 0.031∗∗ 0.026∗∗
(0.009) (0.009) (0.012) (0.013)
Ln(new constructions) 0.006∗ 0.002 -0.013∗∗∗ -0.010∗∗
(0.004) (0.004) (0.005) (0.005)
Observations 182,570 182,583 182,570 182,570 182,583 182,570
R2 0.223 0.223 0.223 0.196 0.196 0.196
Controls:
Household Controls Yes Yes Yes Yes Yes Yes
Suburb × Year FEs Yes Yes Yes Yes Yes Yes
Tenure Period FEs Yes Yes Yes Yes Yes Yes
Full House Controls Yes Yes Yes Yes Yes Yes
Reference House Specification:
Size Percentile 90th — 90th 90th — 90th
Visual Salience High — High High — High
Post-Purchase Yes — Yes Yes — Yes
Notes: OLS coefficients reported. House satisfaction (logged) used as the dependent variable in
columns (1), (2) and (3). Neighorhood satisfaction (logged) used as the dependent variable in columns
(4), (5) and (6). Column (1) and (4) include the 90th percentile size of visually-salient houses built
over homeowners’ tenure period in their suburb as a dependent variable. Columns (2)-(3) and (5)-(6)
add the total number of new-build houses in their suburb over the same period (logged). Household
controls include age, race, sex, household size, education, annual income (logged), number of cars,
monthly mortgage payments, mortgage conditions (term and interest rate) and overall neighborhood
satisfaction. House controls include own house size (logged), house market value per square foot
(logged), local real estate taxes (logged), age of house, quality features (balcony, porch, extra bath-
rooms, heating equipment, recent remodeling), and defects (holes, water leak). High (low) visual
salience corresponds to the top (bottom) quartile of houses within suburbs in terms of proximity to
roads (i.e. number of roads within 5 minutes driving time). Robust standard errors clustered at the
suburb-year level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

xiii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S11: Estimation of Positional Externalities Across Reference House Size
Distribution for High vs. Low Visual Salience of Reference Houses

(a) High Visual Salience (b) Low Visual Salience


1
1

.75 .75
Homeowners' Sensitivity to Size of Reference

Homeowners' Sensitivity to Size of Reference


.5 .5
Houses (σ coefficient)

Houses (σ coefficient)
.25 .25

0 0

-.25 -.25

-.5 -.5

-.75 -.75

-1 -1
10th 30th 50th 70th 90th 10th 30th 50th 70th 90th
pth percentile size of reference houses after purchase pth percentile size of reference houses after purchase

Notes: The Figure plots the sensitivity coefficient (σ = −γ2 /γ1 ) to the size of reference houses built over
households’ tenure in their suburbs estimated from regression (2) for visually salient houses (top quartile
in terms of proximity to roads) and non-visually salient houses (bottom quartile). Each dot represents a
separate regression where the pth percentile size of newly built houses is used as a measure of reference
house size. Sampling weights included. 90% confidence intervals drawn using standard errors clustered
at suburb-year level.

xiv

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C Robustness
Appendix C.1 Alternative Clustering

Table S5: Alternative Clustering of Standard Errors

(1) (2) (3) (4)


Ln(own size) 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗
(0.003) (0.003) (0.006) (0.003)
Ln(reference size) -0.039∗∗∗ -0.039∗∗∗ -0.039∗∗∗ -0.039∗∗∗
(0.009) (0.009) (0.007) (0.010)
σ
Observations 182,570 182,570 182,570 182,570
R2 0.223 0.223 0.223 0.223
Controls:
Household Controls Yes Yes Yes Yes
Suburb × Year FEs Yes Yes Yes Yes
Tenure Period FEs Yes Yes Yes Yes
Full House Controls Yes Yes Yes Yes
S.E. Clustering:
Suburb-Year Yes — — —
Suburb — Yes Yes Yes
Year — — Yes —
Tenure Period — — — Yes
Notes: OLS regressions reported. Reference size is defined as the 90th percentile size of houses
built over households’ tenure in their suburbs. Each regression varies the degree of visual salience of
reference houses. Controls include householder’s age, race, sex, education, household size, annual in-
come (logged), number of cars, monthly mortgage payments, mortgage conditions (term and interest
rate), neighborhood satisfaction, house market value per square foot (logged), local real estate taxes
(logged), age of house, quality features (balcony, porch, extra bathrooms, heating equipment, recent
remodeling), and defects (holes, water leak). Sampling weights included. Cluster-robust standard
errors reported in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

xv

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C.2 Visual Salience

Table S6: Visual Salience (Robustness)

(1) (2) (3) (4) (5) (6) (7) (8)


Ln(own size) 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗ 0.051∗∗∗
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Ln(reference size) -0.003 -0.013∗∗ -0.032∗∗∗ -0.039∗∗∗ 0.000 -0.022∗∗ -0.032∗∗∗ -0.035∗∗∗
(0.008) (0.006) (0.009) (0.009) (0.009) (0.010) (0.010) (0.010)
σ 0.05 0.25∗∗ 0.63∗∗∗ 0.76∗∗∗ -0.00 0.44∗∗ 0.63∗∗∗ 0.69∗∗∗
Observations 182,425 182,430 182,571 182,570 182,407 182,541 182,577 182,583
R2 0.223 0.223 0.223 0.223 0.223 0.223 0.223 0.223
Controls Yes Yes Yes Yes Yes Yes Yes Yes
Suburb × Year FEs Yes Yes Yes Yes Yes Yes Yes Yes
Tenure Period FEs Yes Yes Yes Yes Yes Yes Yes Yes
Visual Salience:
Measure # Roads # Roads # Roads # Roads Dist. (m) Dist. (m) Dist. (m) Dist. (m)
Quartile Q1 Q2 Q3 Q4 Q4 Q3 Q2 Q1

Notes: OLS regressions reported. Reference size is defined as the 90th percentile size of houses built
over households’ tenure in their suburbs. Each regression varies the degree of visual salience of ref-
erence houses. Columns (1) to (4) use the total number of roads around a reference house within a
2 kilometers (1.27 miles) radius as a measure of visual salience (higher quartile means higher visual
salience). Columns (5) to (8) use the distance to the nearest road around a reference house (in miles)
as a measure of visual salience (lower quartile means higher visual salience). Controls include house-
holder’s age, race, sex, education, household size, annual income (logged), number of cars, monthly
mortgage payments, mortgage conditions (term and interest rate), neighborhood satisfaction, house
market value per square foot (logged), local real estate taxes (logged), age of house, quality features
(balcony, porch, extra bathrooms, heating equipment, recent remodeling), and defects (holes, water
leak). Sampling weights included. Robust standard errors clustered at the suburb-year level are
reported in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xvi

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S7: McMansion Effect and Within-Suburb Distance to Post-Purchase Construc-
tions

Ln(house satisfaction)
(1) (2) (3)
All Counties All Counties Excl. California

Ln(own size) 0.050∗∗∗ 0.050∗∗∗ 0.049∗∗∗


(0.003) (0.003) (0.003)
Ln(reference size) -0.037∗∗∗ -0.114∗∗∗ -0.136∗∗∗
(0.009) (0.027) (0.028)
Ln(predicted driving time) -0.001 -0.203∗∗∗ -0.256∗∗∗
(0.002) (0.069) (0.072)
Ln(reference size) × Ln(predicted driving time) 0.025∗∗∗ 0.032∗∗∗
(0.009) (0.009)
Observations 158920 158920 137300
R2 0.229 0.229 0.230
Adjusted R2 0.221 0.221 0.221
Suburb x Year FEs Yes Yes Yes
Tenure period FEs Yes Yes Yes
Controls Yes Yes Yes
Notes: OLS models reported. Reference size is defined as the 90th percentile size of visually salient
houses built over households’ tenure in their suburbs. The log of predicted driving time (in min-
utes) is interacted with the log of reference house size (see Appendix Appendix A.4 for a detailed
description of the predicted driving time measure). Controls include householder’s age, race, sex,
education, household size, annual income (logged), number of cars, monthly mortgage payments,
mortgage conditions (term and interest rate), neighborhood satisfaction, house market value per
square foot (logged), local real estate taxes (logged), age of house, quality features (balcony, porch,
extra bathrooms, heating equipment, recent remodeling), and defects (holes, water leak). Sampling
weights included. Robust standard errors clustered at the suburb-year level are reported in paren-
theses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xvii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S12: McMansion Effect as a Function of Predicted Driving Time to
Post-Purchase Constructions

4
Sensitivity to 90th Percentile Size

-1

-2
0

5
10

15

20

25

30

35

40

45

50

55

60

Average Driving Time (Minutes)

High Visual Salience Low Visual Salience

Notes: The Figure plots the estimated σ as a function of predicted driving time from new-build houses
(measured in minutes at 36 miles per hour), excluding the State of California, separately for high visual
salience and low visual salience houses. Coefficients are estimated using specification (2), where own
house size and the 90th percentile size of new houses built in proximity (vs. isolation) to road networks
(top vs. bottom quartile) are interacted with 30 equally-spaced bins of two minutes driving time. Each
dummy captures a given average driving time bracket (shown in minutes) separating each house from
each other new-build houses in their suburbs since they moved in. The dashed line represents the non-
parametric fit using a locally weighted regression.

xviii

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C.3 Supply-Side

Table S8: Reference House Size and Local Housing Market Conditions

Ln(House market value) Ln(Real estate taxes)


(1) (2) (3) (4)
Ln(own size) 0.559∗∗∗ 0.250∗∗∗ 0.513∗∗∗ 0.206∗∗∗
(0.014) (0.010) (0.017) (0.011)
Ln(reference size) 0.000 0.019 -0.064 -0.059
(0.027) (0.028) (0.057) (0.056)
Observations 182570 182570 182570 182570
R2 0.419 0.506 0.869 0.881
Adjusted R2 0.413 0.501 0.868 0.880
Suburb × Year FE Yes Yes Yes Yes
Tenure period FE Yes Yes Yes Yes
Controls — Yes — Yes
Notes: OLS coefficients reported. Columns (1)-(2) have house market value (logged) as the dependent
variable and columns (1)-(2) have annual real estate taxes (logged) as the dependent variable. I report
coefficients on the logged value of own house size and the 90th percentile size of visually-salient
houses built in the household’s suburb over tenure. Controls include the age of the house, amenities
(balcony, porch, extra bathrooms, heating equipment, recent remodeling), defects (holes, water leak),
and satisfaction with neighborhood (1-10), householder’s age, race, sex and education, household
annual income, household size, number of cars, mortgage payments and mortgage characteristics.
Sampling weights are included in all regressions. Robust standard errors clustered at the suburb-
year level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xix

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S9: Predictors of Visual Salience of Reference Houses Built Over Tenure

# Roads Around McMansions (1.5 Miles Radius)


(1) (2)
Coeff. S.E. Coeff. S.E.
Housing satisfaction -0.001 (0.002) -0.000 (0.000)
Neighborhood satisfaction -0.009∗∗∗ (0.002) 0.000 (0.000)
Ln(own size) 0.023 (0.021) 0.001 (0.002)
Ln(income) -0.014∗ (0.007) -0.001 (0.001)
Number of cars 0.014∗∗∗ (0.005) -0.000 (0.000)
Respondent’s age 0.003∗∗∗ (0.000) 0.000 (0.000)
Household size 0.009∗∗∗ (0.003) 0.000 (0.000)
Respondent’s education -0.001 (0.003) 0.000 (0.000)
Male respondent 0.013 (0.009) 0.001 (0.001)
Total mortgage monthly payment -0.000 (0.000) 0.000 (0.000)
Mortgage’s interest rate 0.001 (0.002) 0.000 (0.000)
Mortgage term (years) 0.001∗ (0.001) -0.000 (0.000)
Ln(real estate tax payment) -0.037∗∗∗ (0.012) -0.002∗ (0.001)
Age of house 0.002∗∗∗ (0.001) 0.000 (0.000)
Ln(house value per sqft) 0.070∗∗ (0.028) 0.000 (0.001)
Observations 182,583 182,583
Pseudo-R2 0.008 0.145
Suburb x Year FEs No Yes
Tenure Period FEs No Yes
Notes: Poisson regression models reported. The dependent variable is the number of roads sur-
rounding the biggest 10% houses built over tenure within 2 kilometers (1.27 miles) radius. Sampling
weights are included in all regressions. Robust standard errors clustered at the suburb-year level are
reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xx

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C.4 Zillow vs. Survey of Construction (SOC)
To check whether Zillow measures of house size are accurate, I compare them to the US Census
Survey of Construction (SOC). The Survey of Construction (SOC) provides measures for the
mean and median size of new single-family housing units constructed each year since 1971.
Figure S13 plots the house size of the average new-build house from Zillow and SOC datasets
over the period 1971-2009. The two measures are perfectly correlated over this forty years
period.

Figure S13: Average Size of New-Build Single Family Houses (1971-2009)

3000
Mean size of new-build houses (sqft)

2750

2500

2250

2000

1750

1500
1970 1975 1980 1985 1990 1995 2000 2005 2010
year

Zillow sample Zillow sample: top-coded


SOC sample SOC sample: MSA only

Notes: Figure S13 plots the mean size of new-build single family houses from 1971 to 2009 using data
from Zillow (dark lines) and from the Survey of Construction (grey lines).
Sources: Survey of Construction (SOC) and author’s own calculations from Zillow.

However, Zillow captures on average bigger houses than the SOC. Two important reasons
can explain this gap in levels. First, the SOC historical data regroups both MSA and non-MSA
single-family houses, while the Zillow sample is restricted to metropolitan suburbs, where
houses are on average bigger. A better comparison is to restrict the SOC to houses built within
MSA. This only partly addresses the problem as suburban and central city houses cannot be
distinguished in the SOC. However, the Figure shows it reduces half of the gap.39 Second, the
SOC is top-coded for the biggest 1% houses, which means Zillow does a better job at measuring
the true size of the biggest houses built. If I truncate the Zillow sample to exclude the top
percentile, the remaining gap is negligible.
Lastly, attrition bias may affect the distribution of houses over time if big houses are more
39 The Census Bureau does not compute averages at the MSA level for the period 1971-2009, and access
to the micro data of the SOC is restricted to the 1999-2009 period.

xxi

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S10: Testing for Attrition

Attrition Index
(1) (2)
Coeff. S.E. Coeff. S.E.
Time since the house was built (years) -0.007 (0.020) -0.027 (0.027)

Northeast 4.618∗∗∗ (0.626) 6.144∗∗∗ (1.281)


South 2.627∗∗∗ (0.430) 0.825 (0.865)
West 2.915∗∗∗ (0.503) 1.649∗ (0.927)

Northeast × Time since the house was built (years) -0.080 (0.062)
South × Time since the house was built (years) 0.095∗∗ (0.038)
West × Time since the house was built (years) 0.067 (0.045)
Observations 156 156
R2 0.311 0.377
Adjusted R2 0.293 0.348
Notes: In column (1), I regress the measure of attrition defined in equation (5) against the number
of years since the house was built, controlling for Census region dummies. In column (2), I interact
the number of years since the house was built with the Census region dummies. Sources: Survey of
Construction (SOC) and author’s own calculations from Zillow.

likely to be demolished than small houses. House remodeling and upscaling should also bias
the Zillow measures. To check whether this is the case, I take the ratio of mean to median size
in each census region for each year t as a first approximation of the size distribution for both
datasets. I then compute the difference between the two measures and construct the following
index:
h Mean Mean i
Attrition measuret = 100 × ( )Zillow,t − ( )SOC,t (5)
Median Median
To test whether the distribution of house size between the two datasets varies over time,
I regress this index on the number of years since houses were built and region fixed effects.
Table S10 shows the results of the regression. The coefficient on time is not significant. I can
also interact time with region fixed effects, as attrition and remodeling could play differently
across regions. Except in the South, I find no significant differences across regions. This further
reduces the attrition concern.

xxii

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C.5 Functional Forms

Table S11: Estimation of Positional Externalities: Alternative Model Specifications

Log-Log OLS Level-Log OLS Ordered Probit


(1) (2) (3)
Ln(own size) 0.051∗∗∗ 0.384∗∗∗ 0.334∗∗∗
(0.003) (0.018) (0.015)
Ln(reference size) -0.039∗∗∗ -0.257∗∗∗ -0.173∗∗∗
(0.009) (0.059) (0.055)
Observations 182570 182570 182570
R2 0.223 0.252 —
Adjusted R2 0.216 0.245 —
Pseudo-R2 — — 0.089
Suburb x Year FEs Yes Yes Yes
Tenure Period FEs Yes Yes Yes
Controls Yes Yes Yes
Notes: OLS level-log and ordered probit models reported. Regressions have house satisfaction as the
dependent variable and I report coefficients on own house size and reference size. Reference size
is the 90th percentile size of visually salient houses built over households’ tenure in their suburb.
Controls include householder’s age, race, sex, education, household size, annual income (logged),
number of cars, monthly mortgage payments, mortgage conditions (term and interest rate), neigh-
borhood satisfaction, house market value per square foot (logged), local real estate taxes (logged), age
of house, quality features (balcony, porch, extra bathrooms, heating equipment, recent remodeling),
and defects (holes, water leak). Sampling weights included. Robust standard errors clustered at the
suburb-year level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xxiii

Electronic copy available at: https://ssrn.com/abstract=3378131


Figure S14: Testing for Linearity of House Satisfaction Scale

.6
Impact on Ln(house market value)

.4

.2

0
5 6 7 8 9 10
House Satisfaction (vs. Below 5)

Without Controls With Controls

Notes: The Figure plots the estimates of an OLS regression where the log of house market value is re-
gressed on dummies for each of the response categories of the house satisfaction question conditional on
suburb-year and period fixed effects. Due to low sample size, I regroup answers below 5 as a single cate-
gory (0.9% of answers) and use it as the excluded category. Controls include householder’s age, race, sex,
education, household size, annual income (logged), number of cars, monthly mortgage payments, mort-
gage conditions (term and interest rate), neighborhood satisfaction, house market value per square foot
(logged), local real estate taxes (logged), age of house, quality features (balcony, porch, extra bathrooms,
heating equipment, recent remodeling), and defects (holes, water leak). Sampling weights included. 95%
confidence intervals drawn using robust standard errors clustered at the suburb-year level.

xxiv

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix C.6 Heterogeneous Effects

xxv

Electronic copy available at: https://ssrn.com/abstract=3378131


Table S12: Heterogeneous Effects

HH Size HH Income HH Education Commute Time Suburb Size Pop. Density


(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
1 or 2 3+ Low High Low High Low High Low High Low High
Ln(own size) 0.048∗∗∗ 0.054∗∗∗ 0.050∗∗∗ 0.054∗∗∗ 0.054∗∗∗ 0.044∗∗∗ 0.045∗∗∗ 0.038∗∗∗ 0.046∗∗∗ 0.059∗∗∗ 0.054∗∗∗ 0.049∗∗∗
Electronic copy available at: https://ssrn.com/abstract=3378131

(0.003) (0.004) (0.004) (0.003) (0.003) (0.003) (0.005) (0.006) (0.004) (0.003) (0.004) (0.004)
Ln(reference size) -0.033∗∗∗ -0.041∗∗∗ -0.040∗∗∗ -0.036∗∗∗ -0.037∗∗∗ -0.039∗∗∗ -0.046∗∗ -0.056∗∗ -0.037∗∗∗ -0.041∗∗∗ -0.031∗∗∗ -0.047∗∗∗
(0.011) (0.011) (0.012) (0.012) (0.010) (0.013) (0.020) (0.025) (0.012) (0.014) (0.010) (0.014)
σ 0.70 0.76 0.80 0.68 0.68 0.87 1.03 1.48 0.81 0.70 0.57 0.97
Observations 84,545 97,884 91,879 90,579 118,112 64,340 23,693 19,337 91,311 91,051 92,472 89,507
R2 0.225 0.237 0.224 0.240 0.215 0.278 0.255 0.278 0.230 0.227 0.231 0.228
Controls:
Household Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Suburb × Year FEs Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Tenure Period FEs Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Full House Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
xxvi

Notes: OLS coefficients reported. House satisfaction (logged) as the dependent variable. Coefficients on own house size and reference size (logged) reported.
Each column reports heterogeneous effects splitting the full sample between low (less than 2 persons) and high (more than 2) household size (columns
1-2), low (below median) and high (above median) household income (columns 3-4), lower (no college) or higher (undergraduate or graduate) education
(columns 5-6), low (below median) and high (above median) suburb size in square miles (columns 7-8), population density (columns 9-10) and commuting
time (columns 11-12). Data on commuting time is only available since 2002. Robust standard errors clustered at the suburb-year level are reported in
parentheses. Household controls include age, race, sex, household size, education, annual income (logged), number of cars, monthly mortgage payments,
mortgage conditions (term and interest rate) and overall neighborhood satisfaction. House controls include house market value per square foot (logged),
local real estate taxes (logged), age of house, quality features (balcony, porch, extra bathrooms, heating equipment, recent remodeling), and defects (holes,
water leak). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Appendix C.7 Moving Costs

Table S13: Moving Decisions and McMansion Effect

Decision to Move = 1
(1) (2)
Ln(own size) -0.032∗∗∗ -0.014
(0.007) (0.011)
Ln(reference size) -0.038∗ -0.023
(0.023) (0.025)
Ln(own size) × Above median size 0.003
(0.015)
Ln(reference size) × Above median size -0.031∗
(0.019)
Above median size 0.206
(0.177)
Observations 73666 73666
R2 0.136 0.136
Adjusted R2 0.124 0.124
Controls Yes Yes
Tenure Period FEs Yes Yes
Suburb × Year FEs Yes Yes
Notes: OLS coefficients reported. Reference size is defined as the 90th percentile size of visually
salient houses built over households’ tenure period. In any survey year, I identify as “movers” house-
holds whose house becomes occupied by new owners in the next survey year (versus those who re-
main in the same housing unit). I use this indicator as a dependent variable in regression (2), instead
of house satisfaction. Sampling weights included in all regressions. Robust standard errors clustered
at the suburb-year level reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

xxvii

Electronic copy available at: https://ssrn.com/abstract=3378131


Appendix D House Satisfaction vs. Hedonic Pricing Approach
Assume a household with income y has the choice between two similar houses in suburbs s1
and s2 at time τ 0 . The only difference between the two suburbs is the size of the other houses
at that time Hτ10 > Hτ20 (hereafter called H 1 and H 2 ). The household chooses h to maximize

max U (x, h, H s ) such that y = x + ph

with x a composite commodity, h the size of the house, H s the McMansion externality in
suburb s and p the housing price per square foot. The marginal utility is positive in own
house size Uh > 0 and negative in reference house size UH s < 0. In a perfectly competitive
economy, the housing market internalizes the externality so p and y adjust to variations in H s .
In equilibrium, utility is equalized across the two suburbs so that the household is equally
happy in both places, with no incentive to move. The problem can be rephrased from the
indirect utility function V as
 
V y(H s ), p(H s ), H s = k ∀ s (6)

where k is a constant. This market equilibrium condition is the starting point of the hedo-
nic pricing (HP) approach introduced by Rosen (1974) or Roback (1982). The indirect utility
of housing is an increasing function of income (Vy > 0) and a decreasing function of house
prices for new movers (Vp < 0)40 . The marginal impact of a change in the housing size stock
depends on whether the externality is positive (VH s > 0) or negative (VH s < 0). The implicit
cost of positional externalities C experienced by an existing home owner can be defined as the
increase in income required to make new movers indifferent net of the variation in the market
value of houses:

C = dy/dH s − h(dp/dH s ) with h = −Vp /Vy (Roy’s identity) (7)

Taking the total derivative of equation (6) gives

dV /dH s = Vy (dy/dH s ) + Vp (dp/dH s ) + VH s = 0 (8)

And combining equation (8) and (7), the implicit hedonic cost of the McMansion external-
ity equals
 
C = dy/dH s − Vp /Vy (dp/dH s ) = −VH s /Vy > 0 (9)

When the labor and housing markets are in equilibrium, the implicit cost of positional
externalities exactly equals the marginal willingness to pay (MWTP) to avoid relative down-
40 The fact that higher income allows for better house quality logically leads to a positive marginal
utility of income. The estimation of the later is therefore very sensitive to the inclusion of dwelling
specific controls for quality.

xxviii

Electronic copy available at: https://ssrn.com/abstract=3378131


scaling. Therefore, by regressing house prices and households’ income on the experienced
variation in reference house size, one can recover the MWTP to avoid positional externalities.
However, if a direct proxy of house utility is available, the right hand side of equation (9)
can be estimated directly. This method is known as the life satisfaction (LS) approach41 . In the
context of this article, it consists in regressing a subjective measure of house satisfaction on
income and the externality, holding house prices and income constant, to recover respectively
Vy and VH s . In the case presented above, it requires that the subjective measure of house
satisfaction at time τ 0 be a function of the cumulative instantaneous utility flows over the T
periods since the person moved in42 . If the two methods give similar estimates, one can claim
the market perfectly internalizes the externality through higher price differentials between
relatively small and relatively big houses.
There exist various reasons why the market equilibrium condition is unlikely to hold. A
classical issue is the presence of moving costs. This generates a downward bias in the cost of
the positional externality, as households who would like to move to a relatively bigger house
must also pay an extra moving cost and may be unable to do so. A similar bias may arise in the
presence of loss aversion, which is typically associated with reference dependent preferences
(Genesove and Mayer 2001). Loss aversion can be experienced by existing home owners but
not by potential buyers.
Formally, if condition (6) does not hold, house satisfaction is not equalized across all coun-
ties, so that dV /dH s < 0. It follows that the new implicit cost of positional externalities esti-
mated through the HP approach C̃ is in fact lower than the true MWTP as estimated by the LS
approach:
 
C̃ = dy/dH s + Vp /Vy (dp/dH s ) = −VH s /Vy + (dV /dH s )/Vy < −VH s /Vy (10)

The hedonic cost of positional externalities computed from the wage and price gradients
would therefore give a downward biased estimate of the true cost, as it neglects the residual
effect (dV /dH s )/Vy not capitalized in private markets.

41 For a discussion of the LS approach, see Luechinger and Raschky (2009); Van Praag and Baarsma
(2005).
42 Evidence on “remembered utility” and on the fact that subjective well-being may differ from flow
utility is reviewed by Kahneman and Thaler (2006); Kimball and Willis (2006).

xxix

Electronic copy available at: https://ssrn.com/abstract=3378131

You might also like