University of Southern California Dissertations and Theses

The impact of digital transformation on urban communities, welfare, and distributional outcomes

(USC Thesis Other)

The impact of digital transformation on urban communities, welfare, and distributional outcomes

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content THE IMPACT OF DIGITAL TRANSFORMATION ON
URBAN COMMUNITIES, WELFARE, AND
DISTRIBUTIONAL OUTCOMES
by
N. M. Tri Hoang
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2024
Copyright 2024 N. M. Tri Hoang

Acknowledgements
A Vietnamese proverb says “Ăn quả nhớ kẻ trồng cây,” or to eat fruit and remember the person
that planted the tree. Many people planted the tree so that I could eat its fruit today.
I am grateful to Matthew E. Kahn, my main advisor and committee chair, for his mentoring and
support. This dissertation would not be possible without his valuable inputs, as well as those of my
committee members, Paulina Oliva and Richard K. Green. I thank Jorge De la Roca and Simon
Quach for serving on my qualifying exam committee and providing constructive feedback in the
early stage of my work.
I would like to acknowledge the faculty and staff at the economics department for the incredible academic and administrative support throughout the program. Perhaps the best thing about
studying economics at USC is being surrounded by the talented and aspiring fellow Ph.D. students,
many of whom have greatly inspired me both in life and intellectually and with several of whom I
am lucky to forge great friendships.
My heartfelt gratitude extends to the many people at my high school in Thailand who took a
chance on me over a decade ago. Their kindness and generosity allowed me to pursue an education
in Phuket, which later led me to many exciting opportunities in life.
ii

My family – ba, mẹ, bà-nội, and em-gái – I owe much to you for your patience, unconditional love,
and unceasing sacrifice. Nothing is comparable to years of ba mẹ not having breakfast so that I
and my sister could have ours and that I could attend my English lessons. Ba mẹ: thank you for
being the best examples of hard work, resilience, and humility, and for all the sacrifices you quietly
made so I could reach today. A deepest tribute to bà-ngoại who was so excited that I started this
pursuit in America and but never got to see me completing it.
To Katsumi who has been by my side in this once-so-lonely jouney: thank you for believing and
enabling the best in me. Thank you for the countless times you spent caringly listening to and
commenting on my research. I am blessed to have you in my life; sharing our good and bad times
together – and with Papaya, too! To your family, thank you for always making me feel welcomed
and loved, and for the many fond family memories that we spent together.
iii

Table of Contents
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 2: Do the Poor Get Priced out of Popular Airbnb Communities? . . . . . . . . . . . 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Airbnb listings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 School enrollment and test scores . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Neighborhood-level characteristics . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3.1 Designated historical monuments (HMs) . . . . . . . . . . . . . . . . 17
2.3.3.2 Measures of neighborhood quality . . . . . . . . . . . . . . . . . . . 17
2.3.3.3 Other neighborhood characteristics . . . . . . . . . . . . . . . . . . . 19
2.4 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.1 School-level regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2 ZIP code-level regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Reduced-form results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5.1 Impact on student enrollment outcomes . . . . . . . . . . . . . . . . . . . . . 23
2.5.2 Impact on rents and housing reallocation . . . . . . . . . . . . . . . . . . . . 24
2.5.3 Impact on select measures of neighborhood quality . . . . . . . . . . . . . . . 27
2.6 Robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.1 Validity of the instrumental variable . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.2 Alternative measure of Airbnb activity . . . . . . . . . . . . . . . . . . . . . . 34
2.6.3 Further evidence from HSO enactment . . . . . . . . . . . . . . . . . . . . . . 35
2.7 Assessing welfare implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.7.1 A model for the long-term rental market . . . . . . . . . . . . . . . . . . . . . 37
2.7.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7.3 Structural model results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
iv

2.7.4 Counterfactual analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.4.1 Metric for welfare evaluations . . . . . . . . . . . . . . . . . . . . . . 45
2.7.4.2 Welfare and distributional impact of Airbnb . . . . . . . . . . . . . 46
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Chapter 3: Buy Now with 1-Click: Spatial Impacts of E-commerce . . . . . . . . . . . . . . . 73
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3.1 E-commerce facilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3.2 Measuring retail accessibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3.3 Other data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.4 Empirical strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.5.1 Impact on housing and local economic activity . . . . . . . . . . . . . . . . . 91
3.5.2 Does e-commerce cause gentrification? . . . . . . . . . . . . . . . . . . . . . . 91
3.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.6 Robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.6.1 Alternative baseline group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.6.2 Alternative measures of accessibility . . . . . . . . . . . . . . . . . . . . . . . 94
3.6.3 Other robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.7 Welfare implications of e-commerce . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.7.1 Resident’s choice problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.7.2 Model solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.7.3 Parameterization and estimation . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.7.4 Welfare metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.7.5 Counterfactual exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Appendix to “Buy Now with 1-Click: Spatial Impacts of E-commerce” . . . . . . . . . . . 138
A.1 Figures for additional robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . 138
A.2 Background on Amazon fulfillment centers . . . . . . . . . . . . . . . . . . . . . . . . 144
A.3 Handling routes with null travel times . . . . . . . . . . . . . . . . . . . . . . . . . . 145
A.4 Proof for Proposition 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
A.5 Imputing user cost for homeowners . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
v

List of Tables
2.1 Impact of STR on student enrollment . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.2 Impact of STR on rents and property values . . . . . . . . . . . . . . . . . . . . . . . 54
2.3 Impact of STR on Ellis Act withdrawals and housing supply . . . . . . . . . . . . . . 55
2.4 Impact of STR on student performance . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.5 Impact of STR on crimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.6 Impact of STR on local amenities and employment count . . . . . . . . . . . . . . . 58
2.7 Impact of STR on restaurant inspection score . . . . . . . . . . . . . . . . . . . . . . 59
2.8 Impact of STR on street cleanliness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.9 First-stage regression result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.10 Robustness check: Number of HMs and pre-2008 enrollment trend . . . . . . . . . . 62
2.11 Robustness check: Number of HMs and pre-2008 eviction and ZHVI trends . . . . . 63
2.12 Robustness check: Impact of STR on student enrollment, alternative metric . . . . . 64
2.13 Estimated linear parameters for LTR demand . . . . . . . . . . . . . . . . . . . . . . 67
2.14 Estimated non-linear parameters for LTR demand . . . . . . . . . . . . . . . . . . . 68
2.15 Count and share of reallocated housing by number of bedrooms . . . . . . . . . . . . 69
3.1 Summary for model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.2 Estimated travel time elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
vi

3.3 Estimated and recovered costs of online shopping before and after FC entry . . . . . 127
3.4 Welfare impact of e-commerce (2017 dollars) . . . . . . . . . . . . . . . . . . . . . . . 128
3.5 Welfare impact by pre-FC retail access time (2017 dollars) . . . . . . . . . . . . . . . 129
3.6 Welfare impact by CBSA (2017 dollars) . . . . . . . . . . . . . . . . . . . . . . . . . 129
vii

List of Figures
2.1 Global search volume for Airbnb on Google, 2004-2019 . . . . . . . . . . . . . . . . . 52
2.2 Locations of pre-2000 designated historical monuments . . . . . . . . . . . . . . . . . 52
2.3 Robustness check: Dist. of t-statistics from placebo test for spurious trend . . . . . . 65
2.4 Impact of home-sharing ordinances of student enrollment . . . . . . . . . . . . . . . . 66
2.5 Distribution of home size in STR and LTR markets . . . . . . . . . . . . . . . . . . . 69
2.6 Share of reallocated housing by neighborhood . . . . . . . . . . . . . . . . . . . . . . 70
2.7 Average counterfactual price changes by neighborhood . . . . . . . . . . . . . . . . . 70
2.8 Welfare impact by household size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.9 Welfare impact by race and ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.10 Welfare impact by education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.11 Welfare impact by household income . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.1 Number of daily physical trips per person, 1995-2017 . . . . . . . . . . . . . . . . . . 114
3.2 Map of the Amazon FC network over time . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3 Distribution of retail access in driving time (minutes) . . . . . . . . . . . . . . . . . . 116
3.4 Retail accessibility measure, Los Angeles & San Francisco MSAs . . . . . . . . . . . 117
3.5 Impact of e-commerce on home values and rents . . . . . . . . . . . . . . . . . . . . . 118
3.6 Impact of e-commerce on local economic activity . . . . . . . . . . . . . . . . . . . . 119
viii

3.7 Impact of e-commerce on local prices . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.8 Impact of e-commerce on gentrification measures . . . . . . . . . . . . . . . . . . . . 121
3.9 Robustness check: Alternative baseline group . . . . . . . . . . . . . . . . . . . . . . 122
3.10 Distribution of online shopping cost before and after FC entry . . . . . . . . . . . . . 123
3.11 Model prediction for share of online shopping . . . . . . . . . . . . . . . . . . . . . . 124
3.12 Welfare impact for homeowners in select cities . . . . . . . . . . . . . . . . . . . . . . 125
3.13 Welfare impact for renters in select cities . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.14 Welfare impact by CBSA population . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
A.1 Robustness check: Alternative retail accessibility measure . . . . . . . . . . . . . . . 139
A.2 Robustness check: Placebo test with physician accessibility measure . . . . . . . . . 140
A.3 Robustness check: MSA fixed effects . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
A.4 Robustness check: MSA linear time trends . . . . . . . . . . . . . . . . . . . . . . . . 142
A.5 Robustness check: ZIP code linear time trends . . . . . . . . . . . . . . . . . . . . . 143
ix

Abstract
This dissertation studies the effects of digital transformation on cities and neighborhoods, as well
as their welfare and distributional implications. The author focuses on two of the fastest growing
technologies in the past decade: short-term rentals and e-commerce, which have prompted many
critical questions of both academic and policy interest. The first chapter motivates the research
questions underlying the author’s work and summarizes the main findings. The second chapter
examines the impacts of short-term rentals in Los Angeles County, while the last chapter investigates
the spatial implications of e-commerce expansion across cities in the United States.
x

Chapter 1
Introduction
An estimated 70% of global economic growth in the next decade will originate from digitally enabled
business activity (World Economic Forum 2023). More than 80% of the global GDP is generated
in cities (World Bank 2023). Despite the importance of understanding the intersection of urban
economics and the digital transformation, the literature on this domain remains sparse. Economists
have long established compelling evidence that digital technology has substantial positive impacts
for both firms and consumers, ranging from increased productivity and efficiency gains to greater
consumer choice and convenience (Draca, Sadun and Van Reenen 2009, Brynjolfsson and Saunders
2010, Bloom, Sadun and Reenen 2012, Akerman, Gaarder and Mogstad 2015, Goolsbee and Klenow
2006, Brynjolfsson, Hu and Smith 2003, Cohen et al. 2016, Dolfen et al. 2023). However, the specific
implications of such impacts for cities and their economies, including the effects on urban structure, housing affordability, and quality of life, as well as the implications for overall welfare and
inequality, remain relatively underexplored. In their comprehensive review, Goldfarb and Tucker
(2019) discuss the regional effects of digital technology, citing several studies on this topic: from the
1

early speculation of Gaspar and Glaeser (1998), who suggest that the Internet could be a complement to cities, to more recent research affirming agglomeration effects as the primary explanation
why cities benefit the most from the Internet (Savage and Waldman 2009, Forman, Goldfarb and
Greenstein 2012, Eichengreen, Lafarguette and Mehl 2016). Many of these studies focus largely on
the digital impact along the urban-versus-rural line. The rapid growth of the platform economy
and the pandemic-induced uptick in remote work have recently inspired scholars to research how
the Internet economy has reshaped our cities and how that could have an impact on overall welfare
and within-city redistribution (Brueckner, Kahn and Lin 2021, Delventhal and Parkhomenko 2020,
Delventhal, Kwon and Parkhomenko 2022, Gorback 2020, Calder-Wang 2019). This dissertation
humbly joins this emerging literature in the intersection of digital transformation and urban economics, focusing on two of the fastest growing technologies: short-term rentals and e-commerce,
which have prompted many critical questions of both academic and policy interest in the United
States.
In the next chapter titled Do the Poor Get Priced Out of Popular Airbnb Communities?, I study
the relationship between short-term rentals and multiple neighborhood outcomes. Specifically, I
address two questions that are important to urban policy: (1) does Airbnb gentrify neighborhoods
or does it simply displace local, particularly low-income, residents?, and (2) what are the welfare and
distributional consequences of Airbnb for renters? By exploiting a shift-share instrumental variable
and the staggered adoption of home-sharing regulations across cities in Los Angeles County, I find
that an increase in Airbnb activity leads to displacement of low-income residents, with minimal
evidence of gentrification. I provide evidence suggesting that the displacement is primarily driven
by Airbnb-induced rent increase as units in the supply-inelastic housing stock are shifted from the
long-term to the short-term rental markets. Motivated by the reduced-form results, I employ a
random-coefficient logit framework to assess the welfare and distributional consequences due to
2

housing reallocation. The quantification suggests the average renting household incurs a loss of
$184 per year, amounting to a county-wide aggregate annual welfare impact of $330m or an average
cost of 33¢ on renters for every dollar spent on lodging by Airbnb guests. While high-income,
college-educated, and white renters face higher loss as their homes are more likely to be reallocated,
low-income counterparts are increasingly rent-burdened as the welfare impact represents a larger
share of their income. Taken together, these results suggest nuanced distributional implications of
short-term rentals and local regulations on the market.
Over the past decade, e-commerce has significantly transformed how we access consumption: with
just a few clicks, consumers in many urban areas can purchase and have their orders delivered to
their doorsteps within a day or two, sometimes even within just a few hours, without having to visit
the stores. How does this experience shape our cities and what are the welfare implications of ecommerce? The last chapter Buy Now with 1-Click: Spatial Impacts of E-commerce seeks to answer
this question. By exploiting the staggered rollout of Amazon fulfillment centers across metropolitan
areas in the United States as a proxy for local e-commerce expansion, I find that neighborhoods
that were previously less accessible to retail amenities experience an increase in home values, rents,
and service amenities in the years following e-commerce expansion. This effect is not driven by
gentrification or sorting of residents into these neighborhoods, nor is it explained by the spillovers
from increased employment due to large warehouse openings. Building upon the reduced-form
evidence, I employ a quantitative spatial model to assess the welfare and distributional consequences
of e-commerce. The analysis shows that FC entry reduces the average cost of online shopping by
29% and lowers the local retail price index by 0.6 points. Taken together, my quantification suggests
that the average renter gains $113 per month, while the average homeowner benefits twice as much
at $200 per month from e-commerce, thanks to capital appreciation translating into lower user cost.
Residents in historically less accessible locations enjoy higher surplus than those in more accessible
3

locations, and residents in larger cities see the highest welfare gains. This result suggests that
while e-commerce expansion induces a decentralization effect within cities, it amplifies the existing
agglomeration benefits observed across cities.
4

Chapter 2
Do the Poor Get Priced out of Popular Airbnb
Communities?
2.1 Introduction
The rise of short-term rentals driven by Airbnb in the past decade has attracted the attention of
many economists. Research on this topic mostly considers the impact on the housing and rental
markets with consistent evidence that Airbnb causes property values and rents in the long-term
rental market to increase (Barron, Kung and Proserpio 2021, Garcia-López et al. 2020, Koster, van
Ommeren and Volkhausen 2021, Duso et al. 2020, Calder-Wang 2019). While heightened property values and rents are generally associated with neighborhood phenomena such as gentrification
(Guerrieri, Hartley and Hurst 2013, Glaeser, Luca and Moszkowski 2020), establishing causality
between Airbnb and neighborhood change often faces two main challenges. The first is the classic
endogeneity issue, where the researcher does not fully observe neighborhood characteristics that
correlate with both local Airbnb growth and relavant outcomes of interest. As a result, estimates
5

of the treatment effects are often biased because Airbnb tends to enter neighborhoods that are
already gentrifying, or conversely neighborhoods where local residents do not want to live long-term
(Fontana 2020). The second challenge comes from the lack of precise estimates for demographic
and socio-economic characteristics at the neighborhood level. Such data is only available from the
Census Bureau every 5 and 10 years and thus falls short in capturing the rapid transformation
induced by the fast growth of the short-term rental market (Glaeser, Luca and Moszkowski 2020).
In this paper, I study the causal relationship between Airbnb and neighborhood change, as well
as the distributional impact for long-term renters. Specifically, I ask: (1) Does Airbnb gentrify or
revitalize neighborhoods, or does it mainly displace local residents? and (2) How are different renters
impacted differentially by short-term rentals? Unless noted otherwise, I refer to gentrification as the
process through which low-income residents are replaced by higher-income counterparts (Brummet
and Reed 2019, Couture et al. 2019, Ding, Hwang and Divringi 2016, Ellen, Horn and Reed 2019,
Guerrieri, Hartley and Hurst 2013, Pennington 2021). It is worth noting that, in some contexts,
gentrification may refer to improvements in overall neighborhood quality without much replacement
of incumbent residents. Following Pennington (2021), I refer to this phenomenon as “neighborhood
revitalization.” On the other hand, displacement can take place in the absence of improvements
in neighborhood quality as local residents, particularly low-income renters, are priced out due to
increased rent burden.
The context for my study is Los Angeles County, one of the most popular and the second largest
Airbnb market - by the number of listed rentals - in the U.S., after New York City. It is worth
noting L.A. renters are usually not allowed to sublease or list a vacancy on Airbnb without landlords’
written consent.1 This restriction on renters to participate in homesharing suggests that the impact
1Most rental leases in LA explicitly state that tenants are not allowed to sublease the units (Lipton (2014), cited in
Koster, van Ommeren and Volkhausen (2021)). Even with landlord’s consent, California Civil Code 1995.240 declares
6

of Airbnb may largely stem from absentee landlords shifting long-term rental units to the shortterm rental market, rather than from incumbent residents engaging in homesharing themselves. As
housing supply barely catches up with demand, the impact of reallocation could be exacerbated.
This motivates my paper to focus on Los Angeles, where one of the nation’s most serious housing
crises takes place. Another motivation to focus on Los Angeles is that, as of 2019, only 18 out of
88 cities in Los Angeles County have implemented Home Sharing Ordinances (HSOs) to regulate or
ban short-term rentals. The City of Los Angeles, the largest and most populous in the County, does
not have an HSO in place until July 2019 (Koster, van Ommeren and Volkhausen 2021). From an
inference perspective, this lack of government regulations compared to other major Airbnb markets,
in addition to the early penetration of Airbnb in 2008, allows for plausibly uncontaminated estimates
of the short-term rental effects.
To address endogeneity issues, I employ a shift-share-styled instrumental variable (IV) strategy that
is widely used in labor and migration literature, as well as in the recent literature that study the
effects of short-term rentals (Barron, Kung and Proserpio 2021, Garcia-López et al. 2020, Fontana
2020). Following the literature, the “shift” component is the global Google Trends index for the
keyword “Airbnb.” The “share” component is the number of designated historical monuments (HMs)
located in a neighborhood, as used in Fontana (2020) and Almagro and Domínguez-Iino (2022).
Intuitively, having more HMs would attract more tourists to stay in a neighborhood and thus the
IV should correlate with the number of Airbnb listings. This is the relevance condition. For the
estimates to be valid, exogeneity condition needs to be satisfied, which is that, conditional on the
observables, the excluded variable is uncorrelated with unobserved neighborhood characteristics
that would also impact outcomes. As global interest in Airbnb does not vary across neighborhoods,
that “landlord is entitled to some or all of any consideration the tenant receives from a transferee in excess of the rent
under the lease.”
7

this condition essentially requires count of HMs to be exogenous to factors other than Airbnb
at the neighborhood level. Equivalently, neighborhoods with more (or fewer) HMs should not
be on differential trends in the absence of Airbnb. This is usually known as the parallel trend
assumption often employed in difference-in-differences settings. I address endogeneity concerns with
neighborhood share of HMs in three steps. First, I restrict to the set of HMs that were designated
such status prior to 2000 to rule out reverse causality and under the assumption that designations
prior to 2000 (and the existence of HMs from many decades ago) would not place neighborhoods
on significantly differential trends in the 2010s. Following Fontana (2020), I then include a rich
set of 2005-2009 neighborhood characteristics interacted with time, which controls for potential
differential trends based on baseline outcomes. Lastly, I find that both short- and long-run changes
in neighborhood outcomes prior to 2008 is not correlated with the number of monuments, suggesting
that HMs only affect outcomes through the Airbnb channel.
In the reduced-form analysis, I establish two main results. The first result is that short-term rentals
cause displacement of students who are primarily from economically disadvantaged backgrounds,
which is indicative of low-income residents being displaced from popular Airbnb neighborhoods.
Specifically, enrollment of FRPM-eligible, Hispanic/Latino, and black students decreases by 0.25%,
0.10%, and 0.24% respectively when the rate of Airbnb penetration – defined as number of active
listings per 1,000 housing units – increases by 1%. I do not find statistically significant effects
on enrollment outcomes for white, and Asian and Pacific Islander students (henceforth Asian),
suggesting that the marginal effect of 0.14% decline in overall enrollment is driven primarily by
students from low-income backgrounds. I discuss several mechanisms that potentially explain these
estimates, and provide three evidence suggestive of rent increase and housing reallocation as the
primary displacement channel. I first show that 1% increase in Airbnb penetration would cause
rents in the long-term market to increase by 0.03%. The effect is stronger in neighborhoods with
8

lower owner-occupancy rate, suggesting that homes are more likely reallocated by absentee landlords.
These estimates essentially corroborate previous research, corresponding to an approximate marginal
increase of $10 to $15 in annual rents. Second, consistent with anecdotal evidence and existing
research hypothesis on landlords converting long-term rental housing to Airbnb units, I find that a
1% increase in active Airbnb listings per thousand housing units would result in a 0.06% increase
in the number of rental units withdrawn under the Ellis Act, a California state law that allows
landlords to evict incumbent tenants in their properties, mostly subject to rent control, to exit the
(long-term) rental businesses. While I do not provide direct parcel-level evidence that “ellisised”
properties are converted to Airbnb units, this result indicates that landlords do reallocate housing
away from the long-term rental market in more popular Airbnb communities. My estimate at the
neighborhood level suggests that, on average, approximately 1 out of every 13 Airbnb rentals is an
“ellisised” property. Lastly, I show that Airbnb does not have a statistically significant effect on the
supply of housing stock, at least in the short run. Hence the impact of Airbnb on reallocation and
rents are unlikely mitigated by housing supply forces, which again is suggestive of long-term rental
housing being converted to Airbnb units.
My results on displacement are not driven by pre-trends or spurious time trends, and are robust
to using an alternative definition of Airbnb activity. As additional robustness exercises, I exploit
the timing and spatial variation in the adoption of home-sharing ordinances (HSOs) across cities in
LA County in an event-study design, similar to Koster, van Ommeren and Volkhausen (2021). The
event-study results indicate that HSOs increase overall enrollment, particularly that of students
identified as Hispanic/Latino and of students who are FRPM-eligible. Taken together, these findings suggest landlords in HSO jurisdictions allocate housing back to the long-term rental market,
reducing rents and thus reversing the displacement impact. I find that this “recovery” effect is rather
short-term, only persisting for about 3 years following an HSO implementation. This is consistent
9

with existing anecdotal evidence and concerns regarding the effectiveness of HSOs and challenges
associated with enforcing such regulations in the long run.2
The second result is that Airbnb is unlikely to cause neighborhood gentrification or revitalization.
My estimates from the IV specification indicate that Airbnb has no positive effects on enrollment
of white students, nor does Airbnb have a positive effect on the number and quality of amenities
available in the neighborhood. As local incumbents are displaced and Airbnb guests are usually
young and budget travelers, these effects may explain why there is no additional demand for more
and better neighborhood amenities. In terms of neighborhood safety, I observe a long-run increase
in non-violent crimes as a result of Airbnb penetrating neighborhoods. That Airbnb only reduces
neighborhood safety in the long run is consistent with previous literature that high turnover rate can
gradually deteriorate neighborhood dynamics, such as trust and cooperation, which could otherwise
help them prevent crimes (Ke, T. O’Brien and Heydari 2021, Filippas and Horton 2020). On the
other hand, I find statistically significant improvements in public school quality, which is driven
entirely by the higher test scores of economically disadvantaged students. This result may be
explained by the plausibly higher teacher-to-student ratio due to enrollment decline, which tends to
benefit low-income students more, although this hypothesis does not rule out the possiblity of test
scores increasing simply due to the selection of more affluent students, even within the low-income
group, that remain. Taken together, these findings suggest no minimal evidence of gentrification
or revitalization, and therefore the displaced low-income population is unlikely to be replaced by
higher-income counterparts.
Having established evidence that Airbnb causes displacement and is unlikely to improve neighborhoods’ amenities, I use individual choice data to estimate a random-coefficient logit demand system,
2For example, Zahniser (2022) from the Los Angeles Times reports at least 1 out of 5 Airbnb listings in Los
Angeles is illegal. Wachsmuth (2021) estimates 35% of Airbnb listings in Los Angeles are illegal.
10

following Bayer, McMillan and Rueben (2004), Bayer, Ferreira and McMillan (2007) and CalderWang (2019), which allows me to quantify the welfare and distributional impact on renters. In the
model, supply of the housing stock is fixed and does not respond to opportunities in the short-term
rental market. Long-term renters, who have heterogeneous preferences over prices, neighborhood
and building characteristics, choose their preferred housing type via utility maximization. In equilibrium, aggregate demand for each housing type must equal to its supply net the number of reallocations to the short-term rental market. I estimate the model in two steps: the first step pins
down mean utlities and heterogeneous coefficients via maximum likelihood, and the second recovers
linear coefficients within the mean utilities via instrumental variable.
My counterfactual analysis, where all short-term rental units are returned to the long-term market,
suggests a substantial welfare impact. I find that rents decrease by an average of 0.66%, which
corresponds to an impact of approximately $15 per month or $184 per year for the average renting
household. Aggregating across all 1.8m renting households in LA County, I calculate an aggregate
annual loss of $331m. Breaking down the welfare impact, I find $66m is due to choice set reduction,
as 0.69% of the housing stock was converted to Airbnb. Rent increases account for the remaining $265m. Among this, the deadweight loss due displaced households is $3.6m per year whereas
$261.4m was transfered from incumbent, non-displaced renters to landlords. Given Airbnb’s reported host revenue within the City of Los Angeles of $279m net of transient occupancy tax around
the same period (Mitra, De Anda and Ritter-Martinez 2017), my analysis suggests a 33-cent impact
due to reallocation for every dollar spent on lodging by guests through Airbnb.
As to distributional consequences, I observe significant differential impact along observable demographics. Since Airbnb tends to penetrate neighborhoods that are high-income, college-educated,
and white, renters in these locations experience higher loss from reallocation. While the impact
11

on low-income, black, and Hispanic households is mitigated by a lower reallocation rate, their loss
takes a larger share of the household income. In fact, while the welfare impact represents only 0.1%
of the average household income for those in the top decile of the income distribution, the burden
is as much as 3.8% for those in the bottom decile. These findings, coupled with low-income households having a more elastic demand, are consistent with my reduced-form results that low-income
residents are increasingly rent-burdered and more likely to be displaced.
The rest of the paper is organized as follows. Section 2.2 discusses related literature. Section 3.3
outlines the data used in the paper, followed by a discussion of identification strategy in Section
2.4. Results are presented in Section 2.5. Section 2.6 reports robustness results. Section 2.7 studies
the welfare and distributional impact. And lastly, section 2.8 concludes.
2.2 Related literature
This paper relates and seeks to contribute to several strands of literature. It is closely related to
the literature on home-sharing platforms. There is a wealth of empirical evidence suggesting that
Airbnb activity leads to increases in both property values and long-term rental prices. Barron,
Kung and Proserpio (2021) use a shift-share variable to instrument for number of Airbnb listings
in ZIP codes across the U.S. to estimate the causal impact of Airbnb on housing and rental prices.
Their results suggest a sizeable impact due to Airbnb, which accounts for as much as one-fifth and
one-seventh of the observed growth in rents and home values, respectively. Garcia-López et al.
(2020) find similar results in Barcelona. Koster, van Ommeren and Volkhausen (2021) and Duso
et al. (2020) exploit temporal and spatial variation of local home-sharing regulations in Los Angeles
County and Berlin, Germany to provide quasi-experimental evidence on property values in these
two markets. I find limited research studying how the Airbnb-induced rent increase may affect local
12

residents. Calder-Wang (2019) is the only paper on this topic and focuses on the Airbnb market in
New York City. However, this paper does not provide direct evidence of displacement, and instead
assumes that as a natural consequence of the structural model. To the best of my knowledge, my
research is the first to provide empirical estimates suggestive of displacement due to Airbnb.
There is also a growing literature studying the effects of Airbnb on several measures of neighborhood
quality, such as crimes and local economic activity (Ke, T. O’Brien and Heydari 2021, Fontana 2020,
Basuroy, Kim and Proserpio 2020, Garate, Pennington-Cross and Zhao 2020), as well as overall
gentrification. Jain et al. (2021) and Wachsmuth and Weisler (2018) find a positive relationship
between short-term rentals and neighborhood gentrification. However, their analyses are largely
descriptive and correlational, thus do not overcome endogeneity issues. Results on other measures
of neighborhood quality are mixed, suggesting that the effects of Airbnb on neighborhood quality
may vary depending on contexts. For instance, Garate, Pennington-Cross and Zhao (2020) find
that Airbnb has a negative effect on crime rates in Milwaukee, Wisconsin, whereas Ke, T. O’Brien
and Heydari (2021) find that crime rates increase in the long run as a result of Airbnb penetration
into Boston neighborhoods. My paper adds Los Angeles, one of the largest Airbnb markets in the
world, to the pool of empirical evidence on the topic. In addition, I explore the impact of Airbnb
on school enrollment and student performance, which are outcome variables of both academic and
policy interests not yet explored in the literature.
Lastly, my paper also relates to literature on housing demand and sorting models employing a
BLP-styled random coefficient framework (Berry 1994, Bayer, McMillan and Rueben 2004, Bayer,
Ferreira and McMillan 2007, Calder-Wang 2019). Closest to my work is that of Calder-Wang (2019),
who studies the distributional impact of Airbnb in New York City. Using this framework, I provide
13

the first welfare results for the County of Los Angeles, which is another epicenter of the national
housing crisis.
2.3 Data
To study the effects of Airbnb on neighborhood change, I combine multiple data sources. I first
obtain Airbnb listings data from InsideAirbnb.com, a non-commercial website that regularly scrapes
listing information from Airbnb for many major cities in the world. Second, to proxy for neighborhood demographics, I obtain school-level enrollment and student poverty data for LA County
public schools from the California Department of Education. I also collect data on school quality
to investigate how Airbnb affects school quality, which is an important measure of neighborhood
quality that has yet to be explored in the literature. I collect neighborhood level data, including
number of HMs, crimes, street cleanliness, and local economic patterns, which are all publicly available from local authorities’ websites and the Open Data portals. ZIP code-level characteristics are
from the American Community Surveys (ACS), and ZIP code-level housing and rental indices are
from Zillow. In what follows, I will describe the data and variables used in the analysis in detail.
2.3.1 Airbnb listings
I obtain data on Airbnb listings from InsideAirbnb.com. The dataset contains monthly snapshots
of listings available on Airbnb.com between May 2015 and December 2019. Each observation has
information at the scraping date about each listing, including price per night, number of bedrooms
and bathrooms, location, amenities, listing type (shared room, private room, or entire house or
apartment), information about the host, number of reviews, and all past reviews associated with
the listing. Given the geographical coordinates of each listing, I reverse-geocode to locate the listing’s
neighborhood. The dataset contains 142,351 unique listings in Los Angeles County that were ever
14

listed on the platform. As discussed below, I restrict to the 100,212 listings (59,511 of which are
entire homes or apartments) discovered in the snapshots that received at least one review in their
lifetime when constructing a measure for Airbnb penetration over the 12-year period between 2008
and 2019.
A challenge that remains, even for Airbnb, is to determine the number of active listings at a
particular time using historical snapshots (Barron, Kung and Proserpio 2021). In fact, Fradkin
(2015) finds as many as 1 out of 3 Airbnb bookings is rejected due to hosts failing to deactivate
non-active listings from the platform. For this reason, I focus on entire houses and apartments,
which are less likely to be dead vacancies compared other types of listings (e.g. bed space, or
private room in a house/condo) - although my results remain robust when including all listing
categories. Furthermore, given that 69% of Airbnb guests leave a review after their stay (Fradkin,
Grewal and Holtz 2021) and that review data is more predictive of Airbnb activity (Jain et al.
2021), I consider a listing is ever active if it has received at least one review. Specifically, a listing
j is considered active in month m of year t if mt - the snapshot date - is between 90 days prior to
and after the dates of the first and last reviews, respectively:3,4
activejmt = I{
90 days prior to
first reviewj ≤ mt ≤
90 days after
last reviewj
} (2.1)
The number of active listings in neighborhood i in month m of year t is simply the sum of the
activejmt dummies over all listings in i. When aggregating to the neighborhood-year level, I consider
3Barron, Kung and Proserpio (2021) assume a listing is active from the date the host joins and considers multiple
end date measures. The start date estimate ignores the fact that the join date indicates when the account was created,
regardless as a host or a guest (InsideAirbnb.com 2021). Per my calculation for LA County, the average number of
days between host join date and the date of first review is 666 days, with the median being 464 days.
4This measure assumes the host does not remove the listing between the first and last reviews.
15

a listing is active in year t if it is discovered active for at least 6 months in year t, following CalderWang (2019). Formally, the measures of active listings are as follows:
numActiveListingsimt =
X
j∈i
activejmt (2.2)
numActiveListingsit =
X
j∈i
I{
X
m∈t
activejmt ≥ 6} (2.3)
Finally, the rate of Airbnb penetration into neighborhood i over time is computed by dividing the
number of active listings in time period (m)t by i’s housing stock in year t:
airbnbP enetrationi(m)t =
numActiveListingsi(m)t × 1000
totalHousingUnitsit
(2.4)
which gives the interpretation of number of active listings per 1,000 housing units.
2.3.2 School enrollment and test scores
I obtain school-level Census Day (first Wednesday of October) annual primary enrollment count for
all academic years between 2000 and 2018 for all public elementary schools in Los Angeles County
from the California Department of Education. The school-level data provides enrollment count by
race/ethnic group, gender, and grade. In my analysis, I aggregate enrollment counts by race/ethnic
group across genders and grades for each school. The Department of Education also provides data
on student poverty, which includes school-level annual counts of students eligible for FRPM starting
from academic year 2011-12.
Data on public school quality is also available publicly from the Department of Education for
the five academic years between 2014 and 2019. The data contains school-level performance by
grade and economic background on the California Assessment of Student Performance and Progress
16

(CAASPP) Smarter Balanced Assessments, which consists of an English language component and
a mathematics component. I aggregate test scores at the school-year level. For interpretational
purposes, I standardize the test score variables so that each has mean zero and variance one.
2.3.3 Neighborhood-level characteristics
2.3.3.1 Designated historical monuments (HMs)
I collect data data on designated historical monuments (HMs) from the National Register of Historic
Places (NRHP). The NRHP is the federal goverment’s official list, established by the National
Historic Preservation Act of 1966, to allow designation of historical monument status to landmarks
that have significant importance to local and national history. The National Park Service establishes
criteria and procedure for nomination, designation, preservation, and tax incentives associated with
preservation expenses. There are 462 HMs in Los Angeles County. To rule out potential reverse
causality, I exclude HMs that were designated after 2000 and assume that the effects of landmark
designation, if any, are not persistent after many years such that they would place neighborhoods
on permanent differential trends. This leaves me with 319 monuments. Figure 2.2 shows the spatial
distribution of monuments in Los Angeles County.
2.3.3.2 Measures of neighborhood quality
In addition to school quality, I exploit four additional measures of neighborhood quality, namely
street cleanliness, crimes, number of business establishments, and restaurant inspection scores. For
some measures such as street cleanliness, I am only able to gather data for the City of Los Angeles,
as other cities do not have such data available. In these cases, I restrict the analysis to only
neighborhoods located within the City of Los Angeles.
17

Crime data is obtained from the Los Angeles County Sheriff as well as the Los Angeles Police
Department. The Sheriff data contains over 2.7 million reported crime incidents across cities in the
Los Angeles County between 2005 and 2019, and the LAPD data contains over 2.1 million reported
incidents between 2010 and 2019. Taking advantage of GIS software, I identify the ZIP code where
each crime incident took place, for approximately 2 million incidents with missing such information,
based on reported coordinates. Based on each incident’s crime code, I follow the Universal Crime
Reporting (UCR) program’s classification of criminal activities to label whether the incident is a
violence crime, property crime, or neither. The LAPD data also provides Modus Operandi (MO)
codes for all reported incidents, allowing me to construct counts of crime incidents involving tourists
and short-term rentals (MO codes 1205, 2037, and 2042-2046). In the analysis, I aggregate crime
incidents to the ZIP code-month level.
Local economic patterns: I explore the impact of Airbnb on neighborhood economic activity
by using establishment counts reported in the Census ZIP Code Business Patterns and workplace
area employment counts from the Longitudinal Employer-Household Dynamics (LEHD). I focus on
NAICS industries that are indicative of neighborhood amenities, namely restaurants and accommodation services, retail, and healthcare services. Since the LEHD data is available at the block
level, I assign blocks to ZIP codes based on the census-defined internal points to obtain employment
count at the ZIP code level.
Service quality: Since establishment and employment counts are not indicative of business improvements on the intensive margin, I gather the Environmental Health Inspection results for restaurants and markets from the Los Angeles County’s Department of Public Health. This dataset provides inspection results of all active and closed restaurants and markets located in Los Angeles
18

County starting April 2016.5 Each observation represents an inspection result, with a raw score
between 0 and 100. Since a facility may be inspected few times a year, I average all scores in a
given year for each facility. For the analysis, I create a dummy variable equal 1 if the average score
is at least 90, which corresponds to grade A or “superior food handling practices and overall food
facility maintenance.”
Street cleanliness data is available from the City of Los Angeles Bureau of Sanitation through
the Open Data portal. This index is constructed for 22,775 miles of street segments within the City
of Los Angeles and based on four factors: loose litter, weeds, bulky items and illegal dumping. The
composite score ranges from 1 to 3, with 1 being coded as “clean,” 2 “somewhat clean,” and 3 “not
clean.” The assessment is conducted by the Bureau of Sanitation on a quarterly basis, starting from
the first quarter of 2016.
2.3.3.3 Other neighborhood characteristics
Housing and rental markets: I gather data on housing and long-term rental prices at ZIP codelevel from Zillow, namely the Zillow Home Value Index (ZHVI) and the Zillow Rent Index (ZRI).
These indices are available for 260 LA County ZIP codes. For each ZIP codes, Zillow provides the
median listed rent and listed sales price at monthly frequency. I obtain the ZHVI and ZRI for
all homes, including single-family and multi-family homes. Additionally, I proxy for new housing
supply using data on construction permits issued by the LA Department of Building and Safety
from the city’s Open Data portal.
5The dataset does not include businesses in Long Beach, Pasadena, and Vernon. Each of these cities has their
own public health agency responsible for inspecting restaurants and markets and does not make such data publicly
available.
19

Ellis Act eviction data: Data on Ellis Act-related evictions between 2008 and 2019 is publicly
available from the Los Angeles Housing & Community Investment Department (HCIDLA). Additionally, I made a public record request to the Department for pre-2007 Ellis Act data and am able
to obtain a list of “ellisised” properties between between 2001 and 2006. As many observations in
the pre-2007 data do not have a complete address, I look up each property’s address and ZIP code
using its parcel number through the LA County Assessor’s portal.
Other neighborhood-level characteristics data is obtained on the American Community Survey
5-year estimates and the 2000 Census, downloaded from the IPUMS National Historical Geographic
Information System (NHGIS).
2.4 Identification
2.4.1 School-level regression
Using public school enrollment data, I first study the Airbnb impact on student enrollment by
estimating the following model:
yst = β0 + β1 log airbnbP enetrationn(s)t + ϵst (2.5)
where yst is the outcome of interest of school s in year t, namely log count of students by race and
ethnic group, FRPM-eligible students, and total enrollment. airbnbP enetrationn(s)t
is number of
active Airbnb listings per 1,000 housing units in neighborhood n(s) in year t, where n(s) denotes
the approximate school neighborhood containing the area within a 3-mile radius from school s.
Since Airbnb is likely to penetrate neighborhoods that are already gentrifying, or alternatively
neighborhoods where locals do not desire to live, a simple OLS estimate will result in a biased
20

estimate of β1, the coefficient of interest. I address this endogeneity challenge in three steps. First,
I introduce a rich set of neighborhood controls Xn(s)t
similar to Fontana (2020), which is a vector of
interaction terms between pre-Airbnb neighborhood characteristics with time trend. The baseline
characteristics consist of total population (log), median household income (log), median rent to
median income ratio, median home value to median rent ratio (log), share of population over 25
with a college degree, employment rate, employment density in service and accommodation sector
(log), as well as shares of white and hispanic population.6 Second, I introduce school fixed effects
as well as school district-by-year fixed effects, denoted λs and λdt respectively. Here the school FEs
λs absorb school time-invariant characteristics that confound Airbnb penetration with enrollment
outcomes (e.g. a school being located in a popular tourist neighborhood), and district-by-year FEs
λdt isolate time-varying forces common to schools within the same school district (e.g. district-level
policies in response to enrollment outcomes). Lastly, I employ a shift-share-styled IV strategy in
the spirit of Goldsmith-Pinkham, Sorkin and Swift (2020). The preferred specification is as follows:
ysdt = β0 + β1airbnbP enetration \ n(s)dt + Xn(s)tβ3 + λs + λdt + ϵsdt (2.6)
airbnbP enetration \ n(s)dt = GoogleT rendst × numMonumentsn(s)
(2.7)
where ysdt indicates the outcome variable for school s in school district d in year t. airbnbP enetration \ n(s)dt
is the instrumented Airbnb penetration into neighborhood n(s) in year t, numMonumentsz is the
number of pre-2000 designated HMs located in neighborhood n(s), and GoogleT rendst
is average
global search volume for “Airbnb” in year t. The IV is numMonumentsn(s) × GoogleT rendst
.
Standard errors are clustered at the school district level, allowing for the outcomes to be correlated
across periods and among schools within the same school district.
6Neighborhood characteristics are constructed by aggregating the 2005-2009 5-year ACS estimates at the block
group level. A block group is assigned to a neighborhood if its census-defined internal point is within the neighborhood’s boundaries.
21

In order for the IV-2SLS estimate of β1 to be unbiased, the spatial variation of designated HMs must
not impact changes in the outcomes of interest through channels other than Airbnb penetration.
Furthermore, growth in global interest in Airbnb must also be exogenous to school-specific unobserved factors that impact change in the dependent variables. Intuitively, “randomness” in Airbnb
exposure comes from some neighborhoods having more designated HMs by chance, once controlled
for observables and fixed effects. To validate the identification assumption, I perform a series of
pre-trend tests if the number of HMs correlates with changes in outcomes prior to 2008. In fact,
the estimates presented in Section 2.6 suggest no statistically significant evidence suggesting the
identification assumption is violated.
2.4.2 ZIP code-level regression
To analyze the Airbnb effects on rents, home values, and neighborhood quality, I take the analysis
to the neighborhood-month level:
yzcmt = β0 + β1airbnbP enetration \ zmct + Xzmtβ3 + λzm + λcmt + ϵzmct (2.8)
airbnbP enetration \ zcmt = GoogleT rendsmt × numMonumentsz (2.9)
where zcmt refers to ZIP code, city, month, and year. Aside from subscript changes, all variable
names carry the same interpretations as in the school-level regresssion. Here the ZIP code-by-month
FE λzm absorbs ZIP code time-invariant and within-year seasonal confouding unobservables, while
city-by-month-by-year FE λcmt differences out time-varying forces common all neighborhoods within
the same cities (e.g. city-wide policies in response to increased Airbnb presence). Standard errors
are clustered at the city level. In some cases, when ZIP code data is only available for the city of
22

Los Angeles (e.g., Ellis Act evictions), I follow the literature in clustering standard errors at the
ZIP code level.
2.5 Reduced-form results
In Section 2.5.1, I provide estimates on student enrollments by poverty status as well as by race and
ethnicity. Section 2.5.2 discusses potential causal channels and provide three evidence suggestive of
rent increase and housing reallocation as the primary displacement mechanism. In Section 2.5.3, I
investigate Airbnb’s effects on five measures of neighborhood quality, namely public school quality,
crimes, street cleanliness, local economic patterns, and restaurant quality. Results across these
neighborhood quality measures are consistent with the hypothesis that Airbnb entrance does not
gentrify or revitalize neighborhoods.
2.5.1 Impact on student enrollment outcomes
Table 2.1 presents OLS and IV estimates for the impact of Airbnb on student enrollment. The
estimates represent the elasticity of enrollment with respect to Airbnb penetration rate. Comparing
the two sets of results suggests that the OLS estimates are consistently biased towards zero and
thus overlook the actual impact of short-term rentals. Specifically, I estimate that 1% increase
in Airbnb listings per 1,000 housing units, on average, would reduce overall enrollment by 0.14%
(Panel B, column 1), as opposed to the OLS estimate of just 0.04% (Panel A, column 1). Panel B
suggests that the decline in overall enrollment is driven by students from economically-disadvantaged
background, as implied by the negative elasticities of 0.25%, 0.1%, and 0.24% for FRPM-eligible,
Hispanic, and black students, respectively (Panel B, columns 2, 3, and 6). The fact that there is
no statistically significant effects on white and Asian students (Panel B, columns 4 and 5), who
are likely more economically advantaged, suggests that Airbnb-penetrated neighborhoods may not
23

undergo gentrification, but rather displacement of low-income incumbents. This observation is
consistent with my results in Section 2.5.3, where I find that Airbnb does not cause neighborhood
amenities to improve and thus is unlikely to attract higher-income residents.
It is also worth discussing the OLS bias in detail. Given the fixed effects, the most likely source
of bias must come from neighborhood time-varying unobservables that correlate with both Airbnb
penetration and school outcomes. For instance, schools in low-income neighborhoods may experience an increase in enrollment as a consequence of the Expo line introduction in the 2010s, which
might have simultenously resulted in more Airbnb listings in these neighborhoods in response to
the increased tourist demand driven by improved access to public transit. As a result, the OLS
specification underestimates and overestimates the Airbnb effects on enrollment of students from
low- and high-income backgrounds, respectively. One the other hand, while the spatial distribution of designated HMs prior to 2000 correlates with Airbnb penetration in the 2010s, it needs not
correlate with unobserved trends across neighborhoods in this period once controlled for observable
characteristics. This crucial assumption, which I show to be highly plausible in Section 2.6, allows
the IV estimates to correct for the bias due to unobserved confoundedness.
2.5.2 Impact on rents and housing reallocation
What explains the demographic change in enrollment outcomes? The most likely explanation is
rent increase and housing reallocation. Specifically, long-term rental properties are being converted
to Airbnb units, resulting in a shortage in the long-term market and thus driving up rents (Barron,
Kung and Proserpio 2021, Calder-Wang 2019). I present three evidence supporting this displacement
mechanism as an explanation for the observed outcomes.
24

First, I provide direct evidence for the Airbnb effects on long-term rents. Additionally, I show that
rent increases are concentrated in neighborhoods with historically lower owner-occupancy rates,
suggesting that long-term renters are at risk of displacement and that absentee landlords are more
likely to rellocate than owner-occupiers. These findings essentially corroborate prior research results.
Second, I provide additional reallocation evidence by showing that a higher rate Airbnb activity
causes an increase in the number rental units withdrawn under the Ellis Act. Lastly, I suggest that
the rent and reallocation effects are not mitigated due to the inleastic housing supply not responding
to Airbnb.
Table 2.2 summarizes the effects of Airbnb on rents and property values, measured by the Zillow
Rental and Home Value Indices, respectively. Column 1 shows that a 1% increase in Airbnb listings
per thousand units causes a 0.03% increase, statistically significant at 1%, in ZIP code market rents.
However, this estimate masks heterogeneous effects among neighborhoods. Column 2 suggests that
the effect on rents is smaller by 34%, or an absolute difference of 0.0162 percentage points, in
neighborhoods with historically higher share of owner-occupiers, which I define as ZIP codes with
2005-2009 owner-occupier rates exceeding the national average at 66%, i.e. at least 2 out of 3
occupied housing units are occupied by owners. Everything else equal, in dollar terms, this marginal
effect is equivalent to an annual increase of $15 and $10 in monthly rent for the average renters
in low and high owner-occupied neighborhoods, respectively. These results are in line with that of
Barron, Kung and Proserpio (2021), who employ similar identification strategy and find an 0.018%
or $9 annual increase in monthly rent. Columns 3 and 4 explore the effect of Airbnb on property
values. I find a statistically significant marginal effect of 0.1% increase when Airbnb penetration
increases by 1% and that such effect does not vary across neighborhoods by tenure status.
25

The Ellis Act is a 1985 California state law that allows landlords to withdraw long-term rental
properties, which are mostly subject to rent control, in order to “go out of [their] rental businesses.”
As a consequence of such exit, tenants may be evicted. Given California has some of the nation’s
strongest renter protection legislations, landlords who seek to convert long-term rental units may
take advantage of the Ellis Act to evict incumbent renters. In fact, many examples where landlords
“ellisise” their rental properties and convert them into Airbnb units have been documented anedotally
(Lee 2016). Columns 1 and 2 of Table 2.3 summarize the impact of Airbnb on Ellis Act-related
evictions within the City of Los Angeles. I find that 1% increase in Airbnb penetration causes
the rate of units withdrawn under the Ellis Act per 1,000 units to increase by 0.06%. Given that
the monthly average Ellis Act eviction rate is 0.03 per 1000 housing units, this marginal effect
corresponds to an additional 0.024 per 1,000 rental units being “elisised” when Airbnb penetration
increases by 10% or equivalently by 0.32 per 1,000 housing units. In other words, approximately 1
out of every 13 Airbnb rentals would be an “ellisised” property.
Lastly, columns 2-4 in Table 2.3 present estimate of Airbnb on housing supply. I detect no statistically significant effects of Airbnb on the number of permits issued for new buildings as well
as additional dwellings to existing structures. One the other hand, I find a statistically significant
0.18% decline in the number of permits issued for building repairs, marginally significant at 10%.
This result could be driven by that units in the short-term rental market may have lower utilization
rates relative to their counterparts in the long-term rental market.
Alternative channels
It is worth looking into other alternative channels that may explain the results. One channel would
be that incumbent low-income owner-occupiers, who may have lower cost to reallocate, are more
likely to move away to rent their homes through Airbnb, which in turn causes the decline in school
26

enrollment. While this remains a possibility, I argue that there is little reason to maintain that it
would be primary channel that driving the observed results. For example, having a school-age child
is known significantly increases the cost to participate in homesharing on Airbnb (Calder-Wang
2019), and likely also to reallocate. The returns to Airbnb for low-income owner-occupiers with
school-age children must be higher than those without, which I take to be unlikely the case. In
fact, as shown in Section 2.5.3, low-income students benefit academically from the decline in the
enrollment, and to the extent that low-income households tend to value public school quality more,
it is arguable that this may increase reallocation costs and rule out this alternative story.
On the other hand, local residents might voluntarily exit their neighborhoods if the arrival of
Airbnb guests leads to substantial decline in neighborhood quality. This may explain the observed
demographic change if these nuisance externalities disproportionately affect low-income residents in
a way that makes their continued residence difficult. In Section 2.5.3, I discuss this mechanism in
detail.
2.5.3 Impact on select measures of neighborhood quality
So far, I have established that Airbnb penetration causes displacement of low-income residents due to
reallocation of housing units away from the long-term rental market. This section explores whether
such displacement takes place in parallel or without gentrification, namely whether the displaced
incumbents are replaced by richer high-income newcomers. I hypothesize that new high-income
residents would not be attracted to Airbnb-penetrated neighborhoods if short-term rentals do not
cause local amenities and quality to improve. My analysis reveals that gentrification is unlikely the
case. While public school performance increases, this improvement is driven entirely by economically
disadvantaged students. I find little short-run effects of Airbnb on crimes, with suggestive evidence of
increasing crimes in the long run in more popular Airbnb communities. Furthermore, neighborhoods
27

do not become cleaner or attract new amenities. Taken together, these results imply little evidence
that Airbnb gentrifies or revitalizes neighborhoods.
School quality: The negative impact on total enrollment and enrollment of low- income students
poses an interesting question on whether school quality improves. On the one hand, lower studentto teacher-ratio may improve teaching effectiveness and student performance, particularly that of
disadvantaged students if they are allocated more resources. Conversely, average test scores may
increase simply to due the positive selection among students who remain. On the other hand, higher
housing costs may cause low-income parents who remain in the neighborhood to spend less on their
children’s education, which could lead to lower academic performance (Newman and Holupka 2014,
2015). Increased displacement risk as well as loss in socio-economic diversity at school may also
adversely affect test scores (Orfield 2001, Figlio et al. 2021). Table 2.4 shows how Airbnb affects
student perform in the two components, English and Math, of the CAASPP Smarter Balanced Assessments. Averaging among all students, 1% increase in Airbnb penetration rate leads to an score
increase of 0.046 and 0.041 standard deviations (columns 1 and 4) in English and Math assessments, respectively. Columns 2-3 and 5-6 report estimates of the same regression using school-level
average scores of students from economically disadvantaged and non-economically disadvantaged
backgrounds, respectively. I observe that the improved school quality is entirely driven by students
of economically-disadvantaged backgrounds, who heightened their English and Math performance
by a statistically significance of 0.052 and 0.056 standard deviations, compared to the statistically
insignificant increases of 0.0147 and 0.008 among the non-economically disadvantaged counterparts.
Crimes: The classic crime model by Becker (1968) suggests that an individual would commit
a crime if the expected net benefit of committing the crime is positive, and that the expected
net benefit is a function decreasing in the probability P of getting caught, search cost C for a
28

potential target, and legal punishment threshold L. Based on this model, Airbnb rentals may
impact neighborhood crime by affecting C and P. While the inflow of Airbnb guests likely reduces
criminals’ search cost C (and thus increases crimes) due to, for instance, tourists’ unfamiliarity with
local areas (Ryan 1993), how Airbnb affects the probability of getting caught P is less intuitive. On
the one hand, neighborhoods and local authorities may respond to the increased influx of Airbnb
guests through crime prevention initiatives, such as policing and patroling (McDonald 1986, Autor,
Palmer and Pathak 2014), thus resulting in higher P. On the other hand, P may decrease if
the presence of highly transient Airbnb guests (and the displacement of local long-term residents)
has a deteriorating effect on neighborhood social capital, such as trust, reciprocity, and mutual
cooperation, that may help the local community deter crimes (Ke, T. O’Brien and Heydari 2021,
Sampson, Raudenbush and Earls 1997).
Table 2.5 presents results on crimes. To separate the short-run and long-run effects, I report estimates corresponding to contemporaneous Airbnb listing count as well as to its interaction with average Airbnb penetration rate over the preceeding 12 months. This rolling average is then demeaned
so that interpretation for short-run effect is applicable to ZIP codes with the average long-run penetration rate. Under this specification, I detect no statistically significant contemporaneous effects
of Airbnb on overall crimes, whereas coefficients on the interaction terms suggest a positive effect
in the long run. Separated by crime categories, Panel A, which reports estimates using LA County
Sheriff data, indicates a 0.145% short-run decrease in property crimes (incidents reported per 1,000
residents), statistically significant at 5%. This crime reduction effect is weakened in the long run,
particularly in more popular Airbnb communities, when considering cofficient on the interaction
term. In fact, neighborhoods with 8% long-run Airbnb penetration rate higher (or more) than the
average are predicted to experience an increase in property crimes. No effects are observed for crimes
in other categories. To the extent that the county sheriff may disproportionately respond to certain
29

types of crime incidents – perhaps more serious offenses since less serious ones may fall under local
police departments, I report estimates using LAPD data in Panel B. The estimates remain largely
similar, both in their signs and magnitudes. In terms of statistical significance, estimates on the
interaction terms become statistically significantly in Panel B for all crime categories other violent
crimes. These results are indicative that Airbnb increases crime in the long run, which appears to
be driven mostly by an increase in crime in the non-violent categories.
What drives the observed crime results? In the short run, more tourist presence may reduce crimes
as increased traffic can result in more surveillence and an increase the probability of caught, which
could help deter crimes. Holding everything else equal, the high transient traffic in the long-run
may erode social capital that would otherwise strengthen neighborhood capacity to prevent crimes
(Ke, T. O’Brien and Heydari 2021). In column 5 of Panel B, I investigate whether Airbnb increases
reported incidents related to tourists, short-term rentals and Airbnb itself. The estimates exhibit
similar patterns, where I observe popular Airbnb communities experience an increase in touristrelated incidents. This result is in line with the hypothesis that tourists are more likely crime
targets (Ryan 1993) and with Fontana (2020), who finds Airbnb guests may engage in anti-social
crime activities themselves due to the lack of surveillence when staying at short-term rentals.
Local amenities: The arrival of Airbnb may also attract new high-end businesses to come to
neighborhoods. Using Census ZBP and LEHD data, I examine the effects of Airbnb on the establishment count and employment in sectors that are indicative of local amenities. Table 2.6 reports
no statistically significant impact of Airbnb on local business activity, although the sign of the estimates suggest that there could be some increase in more tourist-oriented establishments, such as
restaurants and retail outlets, and a decrease in amenities generally catering to long-term residents,
such as healthcare and social assistance.
30

Service quality: Since the ZBP data aggregate number of establishments by industry in each ZIP
code, these results may overlook movements within each industry. For instance, low-end eateries
may be replaced by high-end restaurants, leading to no significant change in the overall count
of local restaurants. I turn to restaurant score provided by the County Environmental Health
Inspection as a proxy for service quality. To account for unobserved entry and exit, I aggregate
establishment-level to the ZIP code level and examine two main outcomes: (1) share of active
restaurants receiving Grade A, the highest standard, and (2) ZIP code average score (log). Table 2.7
reports no improvements of local restaurants on the intensive margin. In fact, I detect a statistically
significant and negative effect on the share of restaurants meeting grade A (0.57 points) and overall
score (by 0.076%). This result runs against the general intuition that that the arrival of tourists
may improve service. A possible explanation is that restaurants may have less incentive to uphold
superior practices as transient Airbnb guests are not familiar with local inspection standards, less
likely to be returning customers, and are more likely to be budget travelers.
Street cleanliness: Another good proxy for neighborhood quality is how clean its streets are
because gentrified or gentrifying communities may maintain their public space better. Table 2.8
summarizes regression results on this measure. Column 1 presents result from a simple linear
probability model where the dependent variable takes value 1 when a street segment is rated “clean,”
and equals 0 otherwise. Overall, I do not find statistically significant estimate for the effect of Airbnb
on neighborhood cleanliness rating. Columns 2 to 6 break down to the individual components the
composite score, namely illegal dumping, loose litter, weed, and bulky item, of all which show no
correlation with Airbnb penetration.
To summarize, I show that the entrance of Airbnb does not cause consistent improvements in
neighborhood quality across different measures. These results conclusively suggest that Airbnb
31

does not gentrify or revitalize neighborhoods, but mainly displaces low-income residents due to rent
increase and housing reallocation.
2.6 Robustness checks
2.6.1 Validity of the instrumental variable
Table 2.9 presents results from several first-stage least square specifications, where I progressively
include location and time fixed effects as well as neighborhood controls. Overall, the first-stage
results are strong and consistent with the relevance assumption that neighborhoods with more HMs
would have a higher Airbnb penetration rates. Panels A and B report the first stages for schoollevel and ZIP code-level regression specifications described by equations (2.6)-(2.7) and (2.8)-(2.9),
respectively. Note that the first-stage coefficients are similar between Panels A and B as well as
across specifications. The Kleinbergen-Paap F-statistics are significantly above conventional critical
values, ruling out weak identification concerns.
A potential threat to my identification strategy is that if neighborhood share of pre-2000 designated
HMs correlates with unobserved factors that may impact displacement, then my estimates could
have simply picked up this confounding relationship. For instance, the number of HMs could
correlate with traditional hotel penetration into a neighborhood. I argue that this is unlikely for
two reasons. Firstly, official tourism statistics indicate that total room nights sold remains relatively
constant over the years with little to zero growth between 2013 and 2017 (LA Department of
Convention & Tourism Development 2017). Secondly, if the displacement effect is due to hotel
penetration or other pre-trends, we should also observe such effect prior to the entrance of Airbnb.
I test for this hypothesis by checking if number of HMs are correlated with both short-run and
32

long-run changes in student enrollment, Ellis Act evictions, and home values using the following
specification:
∆ysdt = α0 + α1 log numMonumentsn(s) + Xn(s)tα3 + γz + γdt + vsdt (2.10)
∆yzmt = α0 + α1 log numMonumentsz + Xztα3 + γcmt + vzmt (2.11)
where ∆ysdt and ∆yzmt denote change in the outcome variables month-over-month, year-over-year,
and between 2000 and 2007. X consists of year 2000 baseline characteristics as previously specified.
I report regression results for equations (2.10) and (2.11) in Tables 2.10 and 2.11. None of the
estimates are statistically significant at conventional levels, lending support to the parallel trends
assumption discussed in Section 2.4.
As to the shift component, a potential concern raised in the literature is that some neighborhoods
may exhibit strong correlation as a result of the time series being introduced to the model, causing
standard errors to be underestimated and thus an over-rejection of the null hypothesis (Barron, Kung
and Proserpio 2021, Christian and Barrett 2017). To address this, I implement the placebo test
suggested by Christian and Barrett (2017), which is also performed in Barron, Kung and Proserpio
(2021). In the placebo, all factors other than the endogenous variable, airbnbP enetrationit, are kept
constant. For each year, I identify neighborhoods with actual active Airbnb listings and randomly
assign without replacement the penetration rates among the identified locations. I obtain the IV
estimates for 1,000 draws of such randominzed assignments on the sample of neighborhoods with
positive Airbnb penetration rates. Note that the randomized treatments preserve aggregate trends
over time but attenuate the effect on the intensive margin of the penetration rate. In other words,
if my results are indeed driven by spurious trends, then we should continue observe statistical
significant estimates among the placebo tests. I do not observe such evidence. Figure 2.3 plots
33

the distributions of t-statistics of the placebo regressions for several school attendance outcome
variables. The dashed-lines plot the t-statitics from the actual regressions using non-randomized
data. Note among the dependent variables for which I observe statistically significant results (e.g. all
enrollment, FRPM-eligible, Hispanic, and black), the actual t-statistics lie beyond the conventional
critical values, whereas this is not the case for statistically insignificant outcomes, e.g. enrollments of
white and Asian students. This observation suggests that my estimates are unlikely to be driven by
spurious aggregate time trends, but rather due to the exogenous instrumented variation in Airbnb
penetration at the neighborhood level.
2.6.2 Alternative measure of Airbnb activity
To check whether the results are sensitive to the constructed measure of Airbnb activity, I employ
an alternative calculation of penetration rate following Fontana (2020):
airbnbGuestP enetrationit =
airbnbGuestN ightsit × 1000
residentN ightsi2008
(2.12)
which computes the average number of Airbnb tourists per 1,000 local residents7
in ZIP code i in
year t. Specifically, the guest flow variable in the numerator is computed by the formula
airbnbGuestN ightsit =
X
j
numReviewsjit
0.69
× numGuestsj × numN ightsj (2.13)
where numReviewsjit is the number of guest reviews that listing j in i received in year t, 0.69
is an estimate for how often an Airbnb guest leaves a review after their stay, taken from Fradkin, Grewal and Holtz (2021), numGuestsj is the number of guests that listing j can accommodate, and numN ightsj is the minimum number nights required by listing j. The denominator
7More precisely, number of Airbnb tourist nights per 1,000 resident nights.
34

residentN ightsi2008 is defined as population of neighborhood i in year 2008 multiplied by 350,
where it is assumed that an average local resident is not present in their neighborhood for 15 days
each year.
Table 2.12 reports results on school enrollment using the guest traffic metric as the independent
variable of interest. The estimates remain robust under the alternative measure. In terms of
magnitude, a 1% increase in Airbnb guests per 1,000 residents would have an impact of 0.17%
decline in overall enrollment, which is primarily driven by the decrease of 0.3%, 0.11%, and 0.32%
in the enrollment of FRPM-eligible, Hispanic, and black students, respectively.
2.6.3 Further evidence from HSO enactment
To provide further evidence, I employ an event study design exploiting the gradual enactments of
Airbnb regulations across LA County cities, collected from Koster, van Ommeren and Volkhausen
(2021). Note that this identification strategy does not depend on the researcher’s decision regarding
how to measure Airbnb activity, and a result, the estimates do not reflect the same marginal effect
of Airbnb activity as in the IV specification.
In their paper, Koster, van Ommeren and Volkhausen (2021) show that implementation of an HSO
leads a decrease in long-term rental prices because landlords, now facing regulations or restrictions
to provide accommodation to short-term guests, re-allocate housing back to the long-term rental
market. Because I find little evidence of gentrification, an implication for my paper is that the
displacement effect would be mitigated or reversed in neighborhoods that have an HSO in place. In
other words, we should observe a higher enrollment of low-income students in schools within cities
35

that implemented an HSO compared to those in cities without an HSO. I test this hypothesis using
the following event-study specification:
ysdt = α0 +
X
4
l=−3,l̸=−1
αl1{t − yearEnactedc(s) = l} + Xn(s)tβ + γs + γt + ϵsdt (2.14)
where yearEnactedc is the year when city c adopts an HSO, {αl} for l < 0 corresponds to pre-trends,
and {αl} for l > 0 to the dynamic impact l years after implementation of an HSO. I exclude non-HSO
schools located more than 3 miles away from an HSO border from the “never-treated” group, under
the assumption that schools closer an HSO border would share similar unobservable characteristics
to treated schools nearby. I estimate equation (2.14) using several heterogeneity-robust treatment
estimators, including De Chaisemartin and d’Haultfoeuille (2020), Callaway and Sant’Anna (2021),
Wooldridge (2021), and Dube et al. (2023). Essentially, these estimators prevent using outcomes
of the already-treated schools as controls for later-treated schools, which would generate negative
weights in the event that HSOs had heterogeneous effects across treatment groups, thus causing
the OLS estimates to be misleading. De Chaisemartin and d’Haultfoeuille (2023) and Roth et al.
(2023) provide detailed surveys of recent advancements in DiD estimators.
Figure 2.4 presents event-study results. Consistent within tuition, I find statistically significant postHSO increase in overall enrollment, as well as enrollment of of students from low-income background,
namely FRPM-eligible and Hispanic students. Furthermore, there are no clear pre-trends before
HSO enactment, suggesting that, conditional on observables, the decline in Airbnb penetration
across neighborhoods is almost as good as random. Schools located in HSO cities experience on
average a 3.1%, 3.5%, and 9% increases in overall enrollments, enrollmments of FRPM-eligible,
and Hispanic students, respectively, relative to nearby schools located in an non-HSO cities in the
36

year following an HSO implementation. I find that these reversing effects on low-income students’
enrollment persists for only three years after HSO implementation.
2.7 Assessing welfare implications
2.7.1 A model for the long-term rental market
In this section, I describe the model used for assessing the welfare and distributional implications
of short-term rentals. I employ the random coefficient logit demand framework to model long-term
rental demand as in Bayer, McMillan and Rueben (2004), Bayer, Ferreira and McMillan (2007)
and Calder-Wang (2019), which share many features with the classic Berry, Levinsohn and Pakes
(1995) model. In the model, households choose their optimal housing type, which is defined by the
combination of neighborhood (PUMAs or Public Use Microdata Areas), number of bedrooms (0
and up to 4 or more bedrooms), and building structure (single- or multi-family home). As a result,
the housing stock is divided into 676 categories. For all units under a particular housing type, there
is no further differentiation other than idiosyncratic preferences specific to individual households.
Formally, household i choosing unit j of housing type h(j) yields the following utility:
uijh(j) = α
i
ph(j) +
X
l
β
i
lx
h(j)
l + Ψh(j) + ϵij (2.15)
where ph(j)
is the rental price of housing type h(j), x
h(j)
l
’s include both neighborhood and building
characteristics of housing type h(j). Neighborhood characteristics consist of shares of Hispanic,
black, Asian and college-educated population, and average commuting time to work. Building
characteristics include number of bedrooms and an indicator whether the unit is a single- or multifamily home structure. Ψh(j)
is the unobserved characteristics (to the econometrician) of housing
type h(j) that correlates with rental price, and ϵij is the idiosyncratic tastes of household i for unit
37

j. Note that the coefficients α
i and β
i
l
’s are specific to the households, which I assume to vary by
their observable characteristics z
i
k
, namely race/ethinicty, income, education, household size, as well
as whether the household has a child under the age of 16.
Household i maximizes their utility by choosing housing unit j that is at least as good as any other
alternatives j
′
:
y
i = j ⇐⇒ ∀j
′
̸= j : uij ≥ uij′ (2.16)
Hence the long-term rental demand for housing type h is
Dh(ph, ph′) = Z
dh
dFϵ(s)dFz(t) (2.17)
where dh = ∪j:h(j)=h{ϵi·
, zi
| ∀j
′ ̸= j : uij ≥ uij′} is the set of households choosing housing type h,
ph′ denotes the price vector of all housing types h
′ other than h, and Fϵ and Fz are the distribution
functions of idiosyncratic tastes and household characteristics, respectively.
The reduced-form evidence suggests no evidence that the supply of housing stock responds to shortterm rentals, hence I assume housing supply is exogenously determined and fixed. In equilibrium,
the housing market must clear, i.e. demand and supply of each housing type are equal:
∀h : Dh(ph, ph′) = Sh (2.18)
38

The introduction of Airbnb results in Ah ≥ 0 units being reallocated8
to the short-term rental
market, leading to the following market clearing conditions in the long-term rental market:
∀h : Dh(ph, ph′) = Sh − Ah (2.19)
2.7.2 Estimation
Parameterization
In bringing the model to data, I make common assumptions about its components: firstly, idiosyncratic shocks ϵij ’s are independently, identically, and Type I Extreme Value distributed; and
secondly, each of the heterogeneous coefficients α
i
, {β
i
l
}l=1,...,L is equal to the sum of two terms: the
first term is common to all households, and the second term is a linear combination of household
characteristics z
i with the linear weights denoted πα or πβl
. Formally, α
i
, {β
i
l
}l=1,...,L are described
as follows:
α
i = α +
X
k
παkz
i
k β
i
l = βl +
X
lk
πβlkz
i
k
(2.20)
Under such parameterization, I can derive the closed-form expression for the probability pih of
household i choosing housing type h:
pih =
Nhe
Vih
P
h′ Nh′e
V ih′
(2.21)
Vih = δh + λ
i
h
(2.22)
δh = αph +
X
l,k
βlx
h
l
z
i
k + Ψh (2.23)
λ
i
h =
X
k
παkz
i
k
ph +
X
l,k
πβlkx
h
l
z
i
k
(2.24)
8
I follow Calder-Wang (2019) in defining a unit as being re-allocated if it is discovered active on Airbnb.com for
at least 6 months in a given year.
39

where Nh = Sh − Ah is the stock of housing type h, and Vih is the indirect utility of household i
who chooses housing type h. Here the indirect utility term Vih is consisted of two parts, namely
the mean utility δh for housing type h common to all households; and the heterogeneous utility
λ
i
h
specific to household i. Note that δ is only defined up to a constant, i.e. household choices
are not affected if δ is shifted by a constant, I normalize the mean utility of one housing type to
0 by substracting demand characteristics x
h
l with with those of the outside option x
0
l
. Continuous
household characteristics, namely (log) income and household size, are demeaned, so the linear
parameters α and βl are interpreted as marginal utilities of the average household.
Data
In addition to listing data from InsideAirbnb.com, I use individual-level data from the 1% Public
Use Microdata Sample from the 2018 American Community Survey (ACS) to estimate the structural
model. Households’ locations are identified at the Census-defined Public Use Microdata Area, or
PUMA, which has a population of approximately 100,000 to 200,000. There are 69 PUMAs identified
in the County of Los Angeles. I use PUMA as an approximation of neighborhoods and aggregate
microdata to the PUMA level to construct statistics on neighborhood characteristics.
In 2018, 3.28m occupied housing units in Los Angeles are divided into 676 housing types. For each
housing type, I impute the market price using self-reported home values and gross rents following
the procedure described in Bayer, Ferreira and McMillan (2007).
Estimation
The specification above allows me to employ a two-stage estimation procedure to identify the parameters of the model as in Bayer, McMillan and Rueben (2004), Bayer, Ferreira and McMillan
(2007), and Calder-Wang (2019). In the first stage, I use microdata on households’ housing choice
40

to estimate the mean utilities δh’s as well as the heterogeneous, non-linear coefficients παk’s and
πβlk’s via maximum likelihood (MLE). In the second stage, I estimate the linear coefficients α and
βl via instrumental variable (IV).
MLE Stage:
Given individual choice data, I construct its log-likelihood as a function of δ and π. The MLE step
essentially identifies parameters (δ, π) that maximize the probabilities that individual households
choosing the housing types observed in the data:
log L(δ, π|X, Z, p) = logY
i
Y
h
p
I{yi∈h}
ih =
X
i
X
h
I{yi ∈ h} log pih (2.25)
where I{yi ∈ h} is the indicator whether household i chooses housing type h, and pih is the closedform probability in equation (2.21) that household i chooses housing type h. In reality, solving
an optimization problem with over 700 parameters is practically infeasible given the computing
capability of a standard computer. I use the iterative procedure as in Berry, Levinsohn and Pakes
(1995), where the optimization problem is reduced to just searching for the optimal non-linear
coefficients π as mean utilities δ can be approximated via a nested fixed point algorithm. Specifically,
the first-order conditions of the objective with respect to δ implies that the predicted demand of
housing type h must be equal to observed stock Nh for all housing type h:
∀h :
X
i
pih = Nh (2.26)
which gives rise to the usual market share moments that pin down the contraction mapping:
∀h : δ
t+1
h = δ
t
h + log Nh − logX
i
pih (2.27)
41

Proof of convergence is provided in Berry (1994) and Bayer, Ferreira and McMillan (2007). In
practice, I implement a variant iterative procedure called SQUAREM by Varadhan and Roland
(2008) to speed up convergence, which is also suggested in Conlon and Gortmaker (2020). The
identification assumption is that demand characteristics are exogenous to individual households,
which holds as each household is infinitesimally small.
IV Stage:
With mean utlities and heterogeneous coefficients estimated in the MLE step, I proceed to estimate
the linear coefficients in equation (2.23). Since the unobserved housing quality Ψh may correlate with
price and neighborhood characteristics, a simple OLS routine would result in biased estimates of the
paramaters. Assessing welfares requires me to obtain an unbiased estimate for α, the coefficient for
price. To do so, I formulate an instrument as a function of several immutable housing characteristics,
which are assumed to be exogenous, while shutting down unobserved quality by setting Ψh = 0 and
imposing the market clearing conditions. The plausibly exogenous characteristics are number of
bedrooms, unit structure, and average commuting time. Formally, the set of moment conditions are
as follows:
∀h : E[pih|qh, α, β, π, Ψh = 0, x
exog
l
] = sh (2.28)
E[Ψhqh] = 0 (2.29)
∃l : E[Ψhx
h
l
] = 0 (2.30)
where qh and sh are the constructed instrument for price and market share of housing type h.
Intuitively, the market share moments imply that variation in the price instrument depends the
availability other other housing units sharing the same set of exogenous characteristics. To see this,
42

note that a higher market share corresponds to a higher mean utility, which given α and β implies a
lower price. This variation allows for uncontaminated identification of α, the coefficient on price, as
long as there exists some exgeonous housing characteristic x
exog. Here the exogeneity assumption
may be interpreted as that, for example, single-family homes are not always inherently better
than apartments or condos, or 3-bedroom units are not inherently less desirable than 4-bedroom
counterparts.
I solve for the moment conditions iteratively. Starting with an initial guess for the linear coefficient
α, I obtain initial values for β by regressing equation (2.23) with αph on the left-hand side. Given α
and β, the instruments qh are completely determined. Once the optimal instruments are obtained,
I estimate the IV-2SLS estimates of interest. Note I do not estimate coefficients associated with
other endogenous demand characteristics, such as neighborhood racial and ethnic composition and
school quality, despite my reduced-form analysis suggests that these factors do change in response
to Airbnb. In the interest of keeping the model simple and tractable, I forgo the welfare analysis
on these channels and focus entirely on the welfare impact resulted from the rent channel.
2.7.3 Structural model results
Table 2.13 reports estimates for the linear coefficients from the IV step. As expected, the OLS
estimate for price are biased upward as price correlates with unobserved quality. Column 2 shows
the estimates obtained by the instrumented strategy. I observe a negative, statistically significant estimate of -2.14 of price on mean utilities. Taking into consideration heterogeneous coefficients on households’ demographic characteristics reported in Table 2.14, Hispanic, black, as well
as low-income households are more responsive to rent increases, although such the difference is not
statistically significant for Black households, this result is consistent with my prior reduced-form
43

findings, where I observe lower enrollment counts of economically-disadvantaged, Hispanic, and
black students in popular Airbnb neighborhoods.
The remaining linear estimates are of expected signs. All else equal, the average household is willing
to $286, $976, $1,429, and $2,321 more in monthly rent to live in a one-, two-, three-, and fouror-more bedroom dwelling relative to a studio unit. Regardless of number of bedrooms, marginal
willingness to pay to live a house is $235 higher per month than to live in a multi-family residence.
The average household is willing to pay up to $381 per month to live in a location that is 1 standard
deviation faster to commute to work.
In terms of heterogeneous preferences, I find sorting patterns consistent with existing literature
Bayer, McMillan and Rueben (2004), Bajari and Kahn (2005), Bayer, Ferreira and McMillan (2007),
Calder-Wang (2019). For example, larger households value homes with higher number of bedrooms.
Households with higher income prefer a house to an apartment and are willing to pay a higher price.
On average, a 1% increase in household income values would increase households’ willingness to pay
by $40 more for every $1,000 in monthly rents. Households do sort by education. The average
college-educated household is willing to pay $352 dollars more a month to live in neighborhoods
with 1-standard deviation higher share of college-educated population. Across demographics, value
the presence of other racial/ethnic groups in a neighborhood, but have the highest willingness to
pay for locations with higher shares of racial/ethnic group of their own. This may be driven by
either preferences for group amenities (Waldfogel 2008) or discrimination (King and Mieszkowski
1973).
44

2.7.4 Counterfactual analysis
In what follows, I analyze the counterfactual where all short-term rental units are returned to
the long-term rental market. I first discuss the metric for welfare evaluation. I then present the
overall welfare evaluation for renting households, followed by results on distributional impact along
observable households characteristics.
2.7.4.1 Metric for welfare evaluations
The counterfactual exercise boils down to computing the price vector that resolve the system of
market clearing conditions:
∀h : Dh(p
cf
h
, p
cf
h′) = Sh − Ah (2.31)
where the number of re-allocated units Ah = 0. Once equilibrium prices are solved for, I compute
the compensating variation for each renting household i in the sample:
CVi =
1
αi

log X
j∈S\A
e
Vi,j (ph(j)
) − logX
j∈S
e
V
cf
i,j (p
cf
h(j)
)

(2.32)
where S and S\A denote the set of housing supply and housing supply net re-allocated units,
respectively. This difference in consumer surpluses indicates in the dollar amount household i
requires or forgoes in the counterfactual scenario in order to maintain same utility level as in the
observed equilibrium, holding everything else equal but price. Intuitively, we expect that CVi > 0,
as imposing restrictions on the short-term rental market should cause rents to decrease, making it
cheaper to afford the utility level of their status quo.
45

Rewriting equation (2.32), I can further disaggregate the welfare impact as follows:
CVi =
1
αi
log X
j∈S\A
e
Vi,j (ph(j)
) − logX
j∈S
e
Vi,j (ph(j)
)

+

logX
j∈S
e
Vi,j (ph(j)
) − logX
j∈S
e
V
cf
i,j (p
cf
h(j)
)

(2.33)
where the former difference quantifies the effect of changes in the number of housing units available,
and the latter measures the impact due to changes in equilibrium rents.
2.7.4.2 Welfare and distributional impact of Airbnb
When the Airbnb housing stock is returned to the long-term rental market, rents decrease by an
average of 0.66%. Figure 2.7 plots the mean changes in monthly rents across PUMAs. Overall, I
find that the average renting household loses about $15 per month, or $184 per year due to shortterm rentals (i.e., gaining the respective amounts under a complete ban of short-term rentals).
Aggregating across all 1.8m renting households in LA County, I estimate an cumulative annual loss
of $331m. Using equation (2.33) to break down the compensation variatons, I find $66m of the
welfare impact is due to choice set reduction. As a result, rent increases account for $265m of the
annual welfare impact for renters. The deadweight loss due displaced households is simply 1
2∆phAh,
which amounts to $3.6m per year, suggesting landlords receive a welfare transfer of $261.4m per year
from incumbent, non-displaced renters. Restricting to the City of LA, where approximately half of
Airbnb listings are located, I find a welfare impact of $118m on renters. Given that Airbnb hosts
in the city around the same period made a total of $279m in lodging revenue net of the 14-percent
transient occupancy tax (Mitra, De Anda and Ritter-Martinez 2017), a back of evelope calculation
suggests a renters’ welfare impact of $0.33 for every dollar spent by guests in the short-term rental
market.
46

Next, I find substantial heterogeneity in the welfare impact along observable household characteristics, namely income, education, household size, and racial and ethinic group. Specifically, smaller
households incur higher loss than larger households, as their homes are more likely to be reallocated.
Because Airbnb tends to penetrate neighborhoods with better amenities, college-educated, white,
and high-income households incur a higher loss. While low-income, black, and Hispanic households
face smaller a welfare impact, their loss represents a larger portion of their income compared to
more affluent households. The remaining of this subsection discusses these findings in detail.
Household size: Smaller households incur a higher loss as their homes are more likely to be
converted to Airbnb units. Figure 2.8 plots distributions of welfare impact by household size. The
average 1-person household incurs a loss of $202 per year, with an interquartile range between $180
to $220, whereas the average 5-person household loses $157 annually, with an interquartile range
between $155 to $165. Note that the distributions are fairly spread out within each group, and
even more so for smaller households. This may be driven by the variation in unobserved housing
quality, particularly for smaller housing units (e.g., studio units may vary substantially in quality
with other studios relative to 4-bedrooms).
Figure 2.5 and Table 2.15 provide further descriptive evidence that rationalizes the findings. More
than half of Airbnb listings are either studio or one-bedroom units, whereas less than 1 out of 4
long-term rental units are of the same size. On average, a 0- or 1-bedroom unit is 4.6 and 4 times
more likely to converted to Airbnb than a unit with 4 or more bedrooms. Despite the large difference
in reallocation propensity, the impact on welfares does not vary as much. This is attributable to
fixed supply quantity, which results in shortage in the long-term rental market and in equilibrium
puts upward pressure on prices not just for for the reallocated types, but also for housing segments
with a lower chance of reallocation (Calder-Wang 2019).
47

Race and ethnicity: White households incur the largest welfare impact in absolute dollar terms,
followed by Asian, black, and Hispanic households, in that order. On average, the annual welfare
impact for Hispanic households is $167, with an interquartile range between $158 to $175. For
black households, the average loss is $171 per year, with interquartile between $163 to $176. Asian
and white households experience larger welfare impact, at $186 and $214, respectively. This is
due to Airbnb tends to enter neighborhoods that are predominantly white, which likely come with
amenities more desirable to tourists.
Figure 2.6 presents the heatmap of Airbnb activity across neighborhoods in Los Angeles, measured
by the rate of reallocated housing units. One familiar with LA geography may notice that Airbnb
units are concentrated in locations with higher share of whites, such as Mabilu, Santa Monica, and
West Hollywood, where as much as 1.8%, 2.15%, 3.25% of the occupied housing stock are reallocated
to Airbnb compared to the countywide reallocation rate of 0.69%.
Education: Highly educated households, typically those with a college degree or higher, experience
higher loss due to the rent channel compared to those without a college degree. On average,
Airbnb results in an annual loss of $207 for more educated households, compared to $174 for less
educated counterparts. Figure 2.10 shows the distribution of welfare impact within each group. The
interquartile range for non-college households is between $162 to $182, whereas that for collegeeducated households ranges from $187 to $223.
Income: Figure 2.11 plots the welfare distribution within groups defined by household income
quantiles. The bottom 80% experience similar welfare impact with the average per annum loss
between $180 to $187. The top 20% slightly higher loss, at $195 per year. The gap widens,
although does significantly becomes larger, when comparing between the bottom 1% and the top
48

1%. Specifically, the average loss for the bottom 1% is $165 annually, while that for the top 1% is
$210.
Although the loss is largest for high-income households, I find low-income renters are disproportionately burdened when considering the welfare impact relative to income. Specifically, the welfare
impact is as much as 3.8% (1.97%) of mean household income for the bottom decile (quantile),
compared to just 0.1% (0.16%) for those in the top decile (quantile).
Taken together, a perfectly implemented blanket ban on Airbnb would have a nuanced distributional
impact. On the one hand, high-income, educated, and white households experience the largest gain
in absolute dollar amounts. One the other hand, low-income counterparts may experience the most
relative welfare improvements.
2.8 Conclusion
In this paper, I provide evidence for low-income resident displacement, proxied for using student
enrollment, as one of the many consequences of Airbnb rentals. I show that the underlying displacement mechanism is due to rent increase as housing units are being reallocated to serve the Airbnb
market. I also suggest that regulations on short-term rentals, such as the HSOs, can undo the
adverse spillovers caused to the low-income residents, although the long-term effect could be muted,
potentially due to difficulties associated with monitoring and enacting such policies. Furthermore, I
show that Airbnb does not improve neighborhood quality and thus is unlikely to cause gentrification
to these communities.
Even though only an estimated 0.69% of the housing stock is reallocated, my welfare analysis
suggests a substantial annual loss of $330m for renting households in Los Angeles due to housing
49

reallocation. While high-income, college-educated, and white households are disproportionately
impacted as their preferred housing types are more likely be reallocated, low-income, Hispanic, and
black counterparts are increasingly rent burdened as their loss accounts for a larger share of their
income.
As a result, regulating the short-term rentals market may result in affluent households gaining the
most in absolute dollar amount whereas low-income renters may experience a larger gain relative
to household income. On the other hand, mitigating the welfare impact of short-term rentals may
require addressing the inelastic supply of the housing stock. Constrained supply has been shown to
put additional upward pressure on equilibrium rental prices due to substitution from other housing
types that are reallocated (Calder-Wang 2019). This suggests that the welfare impact on certain
households are disproportionately higher than predicted in a partial equilibrium, even though their
preferred housing types are less likely to be reallocated. On this account, recent relaxed legislation
on housing supply pertaining additional dwelling units (ADU) in California could mitigate the
exacerbated upward pressure on rents, and hence a test for this hypothesis could be an avenue for
a future study.
There are limitations that come with this research. The data on Airbnb listings is imperfect. For
example, I am not able to determine the exact dates a listing enters or exits the short-term rental
market. As such, the constructed metrics in this paper should not be interpreted as precise measures
of Airbnb activity. Nevertheless, the fact that my results are consistent with other research using
proprietary data supports the validity of my estimates. Another limitation with the structural
model is that only focuses on the rent channel impact, foregoing the impact from the host channel,
as well as the surplus on homeowners and tourists. Several stylized reduced-form results, such
as increase in crime rates, decrease in local Hispanic and black population, and improvements
50

in public school quality, are omitted from the counterfactual exercise in the interest of keeping
the model simple and tractable. As households have heterogeneous preferences over demographic
neighborhood characteristics and quality of public school, changes in these channels could result in
important distributional consequences. I leave these questions open for future research.
51

Tables and figures
Figure 2.1: Global search volume for Airbnb on Google, 2004-2019
Note: This figure plots the annual global search volume for keyword “Airbnb” on Google, indexed between 0 and 100, between
2004 and 2019. The annual search volume is computed by averaging monthly search volume in each year.
Figure 2.2: Locations of pre-2000 designated historical monuments
Note: Designated historical monuments are designated by the National Registry of Historic Places.
52

Table 2.1: Impact of STR on student enrollment
Dependent variable is log number of students
(1) (2) (3) (4) (5) (6)
VARIABLES Enrollment FRPM-eligible Hispanic White Asian Black
PANEL A: OLS
log airbnbP enetrationsdt -0.0364* -0.0438 -0.0435** -0.102*** -0.0522*** -0.139***
(0.0208) (0.0381) (0.0206) (0.0258) (0.0150) (0.0330)
PANEL B: IV-2SLS
log airbnbP enetrationsdt -0.141*** -0.256*** -0.0986*** 0.0380 -0.162 -0.243***
(0.0477) (0.0922) (0.0352) (0.102) (0.110) (0.0425)
Observations 12,364 8,631 12,364 12,364 12,364 12,364
Kleinbergen-Paap F-Stat 60.84 31.81 60.84 60.84 60.84 60.84
School FE Yes Yes Yes Yes Yes Yes
District-by-Year FE Yes Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes Yes
Years 2008-2018 2012-2018 2008-2018 2008-2018 2008-2018 2008-2018
Note: airbnbP enetrationsdt is the number of active Airbnb units per 1,000 housing units. Neighborhood controls include
interaction terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. For log-transformed
variables, one is added prior to the transformation. Robust standard errors reported in parenthesis are clustered at the school
district level. ***p < 0.01, **p < 0.05, *p< 0.1
53

Table 2.2: Impact of STR on rents and property values
Dependent variable is log
(1) (2) (3) (4)
VARIABLES Rent Rent Home Price Home Price
log airbnbP enetrationzmt 0.0288*** 0.0476*** 0.103*** 0.111***
(0.00997) (0.0122) (0.0244) (0.0179)
log airbnbP enetrationzmt -0.0162*** -0.0147
×highOwnerOccupancyz (0.00263) (0.0151)
Observations 17,000 17,000 23,614 23,614
Kleinbergen-Paap F-Stat 102.2 22.75 201.7 51.67
ZIP Code-by-Month FE Yes Yes Yes Yes
City-by-Month-by-Year FE Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes
Years 2010-2019 2010-2019 2008-2019 2010-2019
Note: airbnbP enetrationzmt is the number of active Airbnb units per 1,000 housing units.
highOwnerOccupancyz equals 1 if pre-Airbnb owner-occupancy rate is at least 66%, or national average. The additional IV is constructed by multiplying the existing instrument the interact term
highOwnerOccupancyz. Rent is ZIP code monthly listed rents reported by the the Zillow Rental
Index (ZRI). Home price is ZIP code monthly home price reported by the Zillow Home Value Index
(ZHVI). Neighborhood controls include interaction terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust standard errors reported in parenthesis are clustered
at the city level. *** p<0.01, ** p<0.05, * p<0.1
54

Table 2.3: Impact of STR on Ellis Act withdrawals and housing supply
Dependent variable is log per 1,000 housing units
(1) (2) (3) (4)
Ellis Act Permits
VARIABLES withdrawn units New buildings Addition Repair
log airbnbP enetrationzmt 0.0621*** -0.0508 -0.0610 -0.176*
(0.0190) (0.0528) (0.0408) (0.103)
Observations 16,008 8,814 8,814 8,814
Kleinbergen-Paap F-Stat 48.04 18.80 18.80 18.80
ZIP Code-by-Month FE Yes Yes Yes Yes
City-by-Month-by-Year FE Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes
Years 2008-2019 2013-2019 2013-2019 2013-2019
Note: airbnbP enetrationzmt is the number of active Airbnb units per 1,000 housing units. Data on Ellis Act
evictions and housing permits are available only for the city of Los Angeles. Neighborhood controls include
interaction terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust
standard errors reported in parenthesis are clustered at the ZIP code level. *** p<0.01, ** p<0.05, * p<0.1
55

Table 2.4: Impact of STR on student performance
Dependent variable is standardized test score
(1) (2) (3) (4) (5) (6)
English Mathematics
All Econ Not Econ All Econ Not Econ
VARIABLES Students Disadvantaged Disadvantaged Students Disadvantaged Disadvantaged
log airbnbP enetrationsdt 4.550** 5.224*** 1.147 4.098* 5.579*** 0.765
(1.894) (1.916) (1.757) (2.062) (1.656) (1.680)
Observations 2,581 2,581 2,581 2,581 2,581 2,581
Kleinbergen-Paap F-Stat 18.61 18.61 18.61 18.61 18.61 18.61
School FE Yes Yes Yes Yes Yes Yes
District-by-Year FE Yes Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes Yes
Years 2014-2018 2014-2018 2014-2018 2014-2018 2014-2018 2014-2018
Note: airbnbP enetrationsdt is the number of active Airbnb units per 1,000 housing units. Neighborhood controls include interaction terms
between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust standard errors reported in parenthesis are
clustered at the school district level. *** p<0.01, ** p<0.05, * p<0.1
56

Table 2.5: Impact of STR on crimes
Dependent variable is log number of incidents per 1,000 residents
(1) (2) (3) (4) (5)
VARIABLES Total Violent Property Other STR-related
PANEL A: LA COUNTY SHERIFF DATA, 2008-2019
log airbnbP enetrationzmt -0.171 0.0862 -0.147** -0.146
(0.119) (0.0622) (0.0618) (0.0886)
log airbnbP enetrationzmt 0.0161 -0.00999 0.0176* 0.00569
× log airbnbP enetrationz,12m (0.0166) (0.00967) (0.0100) (0.0121)
Observations 23,619 23,619 23,619 23,619
Kleinbergen-Paap F-Stat 76.11 76.11 76.11 76.11
ZIP Code-by-Month FE Yes Yes Yes Yes
City-by-Month-by-Year FE Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes
PANEL B: LADP DATA, 2010-2019
log airbnbP enetrationzmt -0.0358 0.0340 -0.0537 -0.0602* -0.00882**
(0.0381) (0.0294) (0.0367) (0.0307) (0.00363)
log airbnbP enetrationzmt 0.0229* 0.00135 0.0273*** 0.0221** 0.00204**
× log airbnbP enetrationz,12m (0.0116) (0.00881) (0.0103) (0.00849) (0.001000)
Observations 13,224 13,224 13,224 13,224 13,224
Kleinbergen-Paap F-Stat 23.65 23.65 23.65 23.65 23.65
ZIP Code-by-Month FE Yes Yes Yes Yes Yes
Month-by-Year FE Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes
Note: airbnbP enetrationzmt is the number of active Airbnb units per 1,000 housing units. airbnbP enetrationz,12mlagged
is the rolling average of Airbnb penetration rate in the past 12 months, i.e. P12
k=1 airbnbP enetrationz,m−k,t(m−k). The
additional IV is constructed by interacting number of HMs with the 12-month rolling average of Google Trends search
volume for the keyword “Airbnb.” One is added to dependent variable before log transformation. Neighborhood controls
include interaction terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust
standard errors reported in parenthesis are clustered at the city level (Panel A) and ZIP code level (Panel B). *** p<0.01,
** p<0.05, * p<0.1
57

Table 2.6: Impact of STR on local amenities and employment count
Dependent variable is log count per 1,000 residents
(1) (2) (3) (4) (5) (6)
Restaurants Retail Healthcare
VARIABLES Establishment Employment Establishment Employment Establishment Employment
log airbnbP enetrationzt 0.00821 0.0206 0.0264 0.0855 -0.0335 -0.0281
(0.0233) (0.0687) (0.0304) (0.0537) (0.0316) (0.116)
Observations 1,986 2,733 1,986 2,733 1,986 2,733
Kleinbergen-Paap F-Stat 41.43 34.80 41.43 34.80 41.43 34.80
ZIP Code FE Yes Yes Yes Yes Yes Yes
City-by-Year FE Yes Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes Yes
Years 2008-2016 2008-2019 2008-2016 2008-2019 2008-2016 2008-2019
Note: airbnbP enetrationzt is the number of active Airbnb units per 1,000 housing units. Establishment count is obtained from the Census ZBP
data, which is available in consistent format until 2016. Employment count is computed using LEHD workplace area job count aggregated to
ZIP code level from block level. The 2-digit NAICS codes of three sectors are 72, 44-45, and 62. Neighborhood controls include interaction terms
between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust standard errors reported in parenthesis are clustered
at the city level. *** p<0.01, ** p<0.05, * p<0.1
58

Table 2.7: Impact of STR on restaurant inspection score
Dependent variable is
(1) (2)
VARIABLES Has Grade A Log Score
log airbnbP enetrationzmt -0.571*** -0.0758**
(0.166) (0.0286)
Observations 4,944 4,944
Kleinbergen-Paap F-Stat 4.615 4.615
ZIP Code-by-Month FE Yes Yes
City-by-Month-by-Year FE Yes Yes
Neighborhood controls Yes Yes
Years 2016-2019 2016-2019
Note: airbnbP enetrationzmt is the number of active Airbnb units
per 1,000 housing units. Neighborhood controls include interaction
terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. Robust standard errors reported in parenthesis are clustered at the city level. *** p<0.01, ** p<0.05, * p<0.1
59

Table 2.8: Impact of STR on street cleanliness
Dependent variable is
(1) (2) (3) (4) (5)
VARIABLES Is Clean Bulky Items Illegal Dumping Loose Litters Weed
log airbnbP enetrationzqt -0.0697 1.535 0.00900 5.931 -0.347
(0.224) (2.683) (0.965) (6.070) (3.856)
Observations 1,117,662 1,117,662 1,117,662 1,117,662 1,117,662
Kleinbergen-Paap F-Stat 8.206 8.206 8.206 8.206 8.206
Street Segment-by-Quarter FE Yes Yes Yes Yes Yes
Quarter-by-Year FE Yes Yes Yes Yes Yes
Years 2016-2019 2016-2019 2016-2019 2016-2019 2016-2019
Note: airbnbP enetrationzqt is the average number of active Airbnb units per 1,000 housing units among 3 months of quarter
q year t. Robust standard errors reported in parenthesis are clustered at the ZIP code level. *** p<0.01, ** p<0.05, * p<0.1
60

Table 2.9: First-stage regression result
Dependent variable is log
active Airbnb listings per 1,000 housing units
VARIABLES (1) (2) (3) (4)
PANEL A: SCHOOL-LEVEL REGRESSION
IV 0.119*** 0.129*** 0.0775*** 0.0315***
(0.00655) (0.00603) (0.00448) (0.00404)
Observations 12,584 12,584 12,518 12,518
Kleinbergen-Paap F-Stat 331.2 457.3 299 60.84
School FE No Yes Yes Yes
District-by-Year FE No No Yes Yes
Neighborhood controls No No No Yes
Years 2008-2018 2008-2018 2008-2018 2008-2018
PANEL B: ZIP CODE-LEVEL REGRESSION
IV 0.105*** 0.115*** 0.0639*** 0.0318***
(0.0108) (0.00867) (0.00274) (0.00222)
Observations 30,906 30,822 24,497 24,497
Kleinbergen-Paap F-Stat 85.44 163.7 433 131.2
ZIP Code-by-Month FE No Yes Yes Yes
City-by-Month-by-Year FE No No Yes Yes
Neighborhood controls No No No Yes
Years 2008-2018 2008-2018 2008-2018 2008-2018
Note: The instrumented endegenous variable is log airbnbP enetrationsdt (Panel A) and
log airbnbP enetrationsdt (Panel B), which are log counts active Airnbnb listings per 1,000
housing units. Neighborhood controls include interaction terms between pre-Airbnb 5-year
2005-9 ACS neighborhood characteristics and time trend. When including fixed effects, singleton observations are automatically dropped. Robust standard errors reported in parenthesis
are clustered at the school district level (Panel A) and city level (Panel B). *** p<0.01, **
p<0.05, * p<0.1 *** p<0.01, ** p<0.05, * p<0.1
61

Table 2.10: Robustness check: Number of HMs and pre-2008 enrollment trend
Dependent variable is difference in log number of students
(1) (2) (3) (4) (5)
VARIABLES Total Hispanic White Asian Black
log yt − log yt−1
log numMonumentsn(s)
-0.0000543 -0.00137 -0.000891 0.00245 -0.00874
(0.00158) (0.00343) (0.00755) (0.00454) (0.00751)
Observations 7,245 7,245 7,245 7,245 7,245
log y2007 − log y2000
log numMonumentsn(s)
-0.000499 -0.00968 -0.00617 0.0170 -0.0614
(0.0123) (0.0269) (0.0592) (0.0356) (0.0589)
Observations 1,005 1,005 1,005 1,005 1,005
ZIP Code FE Yes Yes Yes Yes Yes
District-by-Year FE Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes
Note: Robust standard errors reported in parenthesis are clustered at the school district level. ***
p<0.01, ** p<0.05, * p<0.1
62

Table 2.11: Robustness check: Number of HMs and pre-2008 eviction and ZHVI trends
Dependent variable is difference in log
(1) (2) (3)
VARIABLES log
y
m
− log
y
m
−1 log
y
m
− log
y
m
−12 log
y2007/12
− log
y2000/01
Ellis Act evictions per 1,000 housing units
log numMonumentsz -0.0000543 -0.00137 -0.000891
(0.00158) (0.00343) (0.00755)
Observations 11,020 9,744 116
ZHVI
log numMonumentsz 0.0000802 0.00111 0.00419
0.0000624 (0.00104) (0.00356)
Observations 25,363 22,426 267
City-by-Month-by-Year FE Yes Yes Yes
Neighborhood controls Yes Yes Yes
Note: Robust standard errors reported in parenthesis are clustered at the school district level. *** p<0.01, **
p<0.05, * p<0.1
63

Table 2.12: Robustness check: Impact of STR on student enrollment, alternative metric
Dependent variable is log number of students
(1) (2) (3) (4) (5) (6)
VARIABLES Enrollment FRPM-eligible Hispanic White Asian Black
log airbnbGuestP enetrationsdt -0.169** -0.292* -0.110** 0.0709 -0.169 -0.315***
(0.0837) (0.153) (0.0525) (0.143) (0.165) (0.0683)
Observations 12,364 8,599 12,364 12,364 12,364 12,364
Kleinbergen-Paap F-Stat 16.37 9.300 16.37 16.37 16.37 16.37
School FE Yes Yes Yes Yes Yes Yes
District-by-Year FE Yes Yes Yes Yes Yes Yes
Neighborhood controls Yes Yes Yes Yes Yes Yes
Years 2008-2018 2012-2018 2008-2018 2008-2018 2008-2018 2008-2018
Note: airbnbGuestP enetrationsdt is the number of Airbnb tourist nights per 1,000 resident nights. Neighborhood controls include
interaction terms between pre-Airbnb 5-year 2005-9 ACS neighborhood characteristics and time trend. For log-transformed variables,
one is added prior to the transformation. Robust standard errors reported in parenthesis are clustered at the school district level. ***p
< 0.01, **p
< 0.05, *p
< 0.1
64

Figure 2.3: Robustness check: Dist. of t-statistics from placebo test for spurious trend
3 2 1 0 1 2
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Density
(a) All enrollment
3 2 1 0 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Density
(b) FRPM-eligible
3 2 1 0 1
0.0
0.2
0.4
0.6
0.8
1.0
Density
(c) Hispanic
0.4 0.2 0.0 0.2 0.4
0.0
0.5
1.0
1.5
2.0
2.5
Density
(d) White
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Density
(e) Asian
6 5 4 3 2 1 0 1
0.0
0.2
0.4
0.6
0.8
Density
(f) Black
Notes: This figure plots the t-stats obtained from the placebo tests, where Airbnb penetration rates are randomly assigned
within each year among neighborhoods with positve active Airbnb listings, as in Christian and Barrett (2017), Barron, Kung
and Proserpio (2021) The vertical dashed lines show the corresponding t-statistics obtained from regressions using actual data.
65

Figure 2.4: Impact of home-sharing ordinances of student enrollment
-.05
0 .05 .1 .15 .2
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(a) All enrollment
-.5
0 .5 1 1.5
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(b) FRPM-eligible
-.1
0 .1 .2
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(c) Hispanic
-.2 -.1
0 .1 .2 .3
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(d) White
-.1
0 .1 .2 .3
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(e) Asian
-.2 -.1
0 .1 .2
-3 -2 -1 0 1 2 3 4
Years Relative to HSO
dCdH LPDID WoolDID CSDID
(f) Black
Notes: 95% confidence intervals are plotted, where robust standard errors are clustered at the school district level.
66

Table 2.13: Estimated linear parameters for LTR demand
(1) (2) (3)
VARIABLES OLS IV WTP
Monthly rent 0.219 -2.140**
(0.126) (0.901)
1-bedroom 0.245*** 0.613*** 286.45
(0.0238) (0.149)
2-bedroom 0.707*** 2.089*** 976.17
(0.0786) (0.545)
3-bedroom 0.899*** 3.068*** 1428.97
(0.119) (0.854)
≥ 4-bedroom 0.523* 4.969*** 2321.96
(0.235) (1.715)
SFR 0.0730 0.504** 235.51
(0.0409) (0.216)
Commuting time (std.) -0.0673 -0.816*** -381.31
(0.0438) (0.217)
Constant -0.364*** 0.277 129.439
(0.0446) (0.183)
Observations 676 676 676
Note: This table reports the OLS and IV estimates of equation (2.23).
Monthly rents are in thousand dollars. WTP refers to monthly willingness to pay in 2010 dollar terms. The omitted categories are 0-bedroom
(studio/bachelor unit) and multi-family residence. The constant term
denotes the utility value of not choosing the outside option. Robust
standard errors reported in parenthesis are clustered by the number of
bedrooms. *** p<0.01, ** p<0.05, * p<0.1
67

Table 2.14: Estimated non-linear parameters for LTR demand
(1) (2) (3) (4) (5) (6) (7)
VARIABLES log Income Household size College Asian Black Hispanic Children
Monthly rent 0.085*** 0.013 0.024 -0.046 -0.277 -0.289*** 0.038
(0.041) (0.038) (0.084) (0.130) (0.226) (0.123) (0.118)
1-bedroom 0.01 0.197** 0.089 -0.455 -0.115 -0.004 0.212
(0.043) (0.112) (0.226) (0.302) (0.323) (0.209) (0.362)
2-bedroom 0.012 0.816*** 0.181 -0.398 0.015 -0.28 -0.293
(0.052) (0.111) (0.227) (0.299) (0.347) (0.217) (0.358)
3-bedroom 0.076 1.099*** 0.351** -0.347 0.151 -0.661 -0.663
(0.074) (0.122) (0.256) (0.341) (0.406) (0.249) (0.390)
4-bedroom 0.254*** 1.331*** 0.564** 0.013 0.421 -0.658 -1.017***
(0.120) (0.138) (0.308) (0.411) (0.517) (0.305) (0.445)
SFR 0.084** -0.068 -0.053 -0.343 -0.704 -0.088 -0.017
(0.043) (0.055) (0.136) (0.193) (0.212) (0.123) (0.180)
% Asian (std.) 0.000 0.034 0.003 0.875*** 0.263** 0.337*** -0.08
(0.018) (0.024) (0.058) (0.069) (0.129) (0.064) (0.080)
% Black (std.) -0.006 0.01 0.070 0.254*** 0.999*** 0.304*** -0.019
(0.016) (0.024) (0.062) (0.100) (0.075) (0.059) (0.080)
% Hispanic (std.) 0.013 0.043 0.268*** 0.937*** 0.924*** 1.360*** -0.217
(0.038) (0.052) (0.129) (0.184) (0.246) (0.131) (0.174)
% College (std.) 0.049 -0.135*** 0.754*** 0.194 0.241 0.352*** -0.03
(0.044) (0.060) (0.137) (0.191) (0.280) (0.158) (0.195)
Commuting time (std.) -0.001 -0.019 -0.014 -0.082 0.063 0.036 0.088
(0.020) (0.025) (0.066) (0.096) (0.102) (0.060) (0.084)
Note: This table reports the maximum likelihood estimates for non-linear, heterogeneous coefficients. The y-axis variables are housing
characteristics, x-axis variables are households characteristics. Monthly rents are in thousand dollars. The omitted categories are
0-bedroom (studio/bachelor unit) and multi-family residence. ***p < 0.01, **p < 0.05, *p< 0.1
68

Figure 2.5: Distribution of home size in STR and LTR markets
0 1 2 3 4
Number of bedrooms
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Share of units within the market
Market
STR
LTR
Table 2.15: Count and share of reallocated housing by number of bedrooms
Bedroom Count # Reallocated % Reallocated
0 228,313 3,536 1.5488
1 628,334 8,517 1.3555
2 968,826 6,057 0.6252
3 931,971 2,872 0.3082
≥ 4 523,466 1,752 0.3337
Total 3,280,910 22,734 0.0069
69

Figure 2.6: Share of reallocated housing by neighborhood
Notes: This figure plots the spatial variation share of occupied housing units being reallocated to the short-term rental market
across 69 PUMAs in LA County in 2018. A housing unit is considered reallocated if it is listed on Airbnb for at least 6 months
in 2018.
Figure 2.7: Average counterfactual price changes by neighborhood
Notes: This figure plots average changes ($) in equilibrium monthly rents in the counterfactual where all short-term rental
units are returned to the long-term market.
70

Figure 2.8: Welfare impact by household size
1 2 3 4 5
Household Size
140
160
180
200
220
240
260
Welfare Impacts on Renters ($)
Figure 2.9: Welfare impact by race and ethnicity
Black Hispanic Asian White & other
140
160
180
200
220
240
Welfare Impacts on Renters ($)
Notes: The above figures plot the distribution of welfare impact by household size, and race and ethnicity from the counterfactual
exercise where all short-term rental units are returned to long-term rental market. The inner figure inside each violin plots the
box and whisker plot of the corresponding distribution.
71

Figure 2.10: Welfare impact by education
Non-colllege College
140
160
180
200
220
240
260
Welfare Impacts on Renters ($)
Figure 2.11: Welfare impact by household income
Q1 Q2 Q3 Q4 Q5
Household Income Quantile
140
160
180
200
220
240
260
Welfare Impacts on Renters ($)
Notes: The above figures plot the distribution of welfare impact by household education and income from the counterfactual
exercise where all short-term rental units are returned to long-term rental market. The inner figure inside each violin plots the
box and whisker plot of the corresponding distribution.
72

Chapter 3
Buy Now with 1-Click: Spatial Impacts of
E-commerce
3.1 Introduction
Within the past decade, online shopping has become essential to the lives of many American households. Thanks to technological advances and extensive logistic networks, e-commerce giants like
Amazon and Walmart can now reach consumers in almost every corner of the country in the matter
of a day or two, sometimes even just within a few hours. As of 2019, Amazon alone is able to
provide same- or one-day delivery to 72% of the U.S. population, and 59% of U.S. households subscribe to Prime, Amazon’s signature subscription service that provides expedited shipping, special
deals and discounts, among other member benefits (Kim 2019, La Roche 2019, Amazon Help &
Customer Service 2023). As online shopping increasingly serves as an alternative for consumers
to access retail goods without having to visit the stores, there has been a steady decline in the
number of physical trips taken per day by the average American. Figure 3.1, based on data from
73

the National Household Transportation Surveys (NHTS), illustrates this trend between 1995 and
2017. Overall, the number of daily trips has declined by 26%, from 4.3 in 1995 to 3.2 trips per day
in 2017, primarily due to fewer shopping and errand-related trips.
One of the central focuses of urban economics is studying the spatial arrangements of people, businesses, and economic activity within urban areas and factors affecting these arrangements (Mason
and Quigley 2006). One such factor is how people travel within cities, which in turn depends on
their access to the workplace and consumption. While studying commuting patterns has significantly improved our understanding of the spatial equilibrium, prior literature usually overlook the
fact that the majority of trips traveled within cities is consumption-related (Miyauchi, Nakajima
and Redding 2021). These are trips that city residents take to access consumption amenities, such
as shopping malls, retail stores, and restaurants; in contrast to commuting trips for work. As the
spatial variation in consumption amenities contributes to shaping urban structure (Glaeser, Kolko
and Saiz 2001), the arrival of e-commerce – as it allows for an alternative access to consumption
– may have significant effects on the spatial equilibrium. Despite having important economic and
policy implications, empirical evidence on this question is still limited. In this paper, I seek to fill
this gap in the literature. Specifically, I ask (1) what are the causal effects of e-commerce on the
spatial organization of people and economic activity within cities? ; and (2) what are the welfare and
distributional consequences of e-commerce? As many local governments are considering or have already provided financial incentives for companies like Amazon and Walmart to locate and/or expand
e-commerce facilities within their jurisdictions, addressing these outstanding questions is particularly relevant from a policy standpoint. As of 2022, Amazon alone has received nearly $5 billion in
subsidies and tax cuts from local governments across the U.S., and 52% of which, or $2.7 billion,
is specifically given to distribution centers (Shendruk 2022). Nevertheless, existing literature and
74

policy discussions largely consider only the direct impact of facility openings on local employment
within selective industries, such as warehousing and transportation.
There are several ways through which e-commerce may impact the spatial structure of cities. As
discussed below, the ultimate effect on neighborhood outcomes is ambiguous. One the one hand,
the ability to substitute in-store visits with online shopping may lead to a dispersion force within
cities. As e-commerce likely serves as a stronger substitute for in-store shopping in locations with
lower pre-existing access to retail amenities, it induces a larger improvement in consumption access
in these locations. As a result, historically underserved locations may experience an increase in
residential and economic outcomes, while locations with previously higher access might not remain
as attractive. On the other hand, e-commerce may lower demand for retail space, freeing up
commercial estate for non-retail amenities (e.g. restaurants and other recreational services). If
this replacement effect is strong, locations with historically more retail outlets may experience the
most increase in service amenities and become even more desirable. Both of the dispersion and
agglomeration forces are not mutually exclusive and could be simultenously at play. Additionally,
it is worth noting that the e-commerce effect may extend across cities, in addition to the effect
within cities. While it is less common for people to move from one city to another for the sake of
e-commerce access, the impact across cities could be driven by business relocation and/or entry of
new firms into regions that are previously less connected to the national market. This could be
due to the online marketplace providing firms in such locales with improved access to customers
nationwide. In this paper, I focus on the spatial effects of e-commerce on internal urban structure;
and in a seperate research effort, I analyze the impact of e-commerce on the economic geography
across cities and regions.
75

Throughout this paper, I exploit the staggered rollout of Amazon fulfillment centers (FCs) as a
source of variation in local e-commerce expansion to study the questions of interest. There are
several reasons why Amazon and the FCs are a good proxy for e-commerce and its local expansion.
Firstly, Amazon is the dominant player in the online commerce market, accounting for a market
share of 39.5% of all online sales, compared to a 7% of Walmart, the second largest online retailer
(Repko 2022). Amazon.com is the most popular destination for online shopping, having an average
of 2.7 billion visits per month, compared to 856 and 468 million monthly visits received by their
competitors eBay.com and Walmart.com, respectively (E-commerce Guide 2020). The company’s
status as the largest player in the online market means that its growth and strategies could reflect the
broader development and trends in the e-commerce sector. Moreover, a significant driver behind
the rapid growth of e-commerce in the past decade is the remarkable improvement in businessto-consumer shipping and last-mile logistics. Amazon is the first major online retailer to offer
guaranteed two-day shipping as early as 2005 for a wide variety of products through the introduction
of Amazon Prime. Over the years, the company has worked toward making one-day shipping and
even same-day delivery available to consumers across metropolitan areas, while broadening the set of
products eligible for expedited delivery. At heart of this logistic breakthrough lies Amazon’s complex
and vast network of FCs, which are large warehouse and distribution facilities across the country
dedicated to housing inventory for millions of retail products as well as to efficiently receiving,
packaging, and shipping out orders to customers (Houde, Newberry and Seim 2022). Amazon FCs
are located in the outskirts of major urban hubs and typically serve the population in the nearby
catchment areas. Hence, the arrival of an FC marks a significant milestone in the expansion of
e-commerce in the local context, enhancing consumers’ ability to swiftly access a vast array of retail
products. The sharp improvement to the online shopping experience in terms of both convenience
76

and product variety across cities as the result of the Amazon’s FC network expansion provides an
ideal quasi-experimental design for e-commerce research.
I proceed in the analysis in two parts. In the first part, I test for both of the dispersion and
agglomeration effects by implementing a dynamic difference-in-differences or event-study design,
where I interact the leads and lags relative to the dates of FC openings with a constructed measure
of neighborhood retail accessibility. This identification strategy allows me to estimate the spatial
impact of e-commerce due to neighborhoods’ differential access to pre-existing retail amenities. The
validity of the estimated treatment effects rests on the assumption that neighborhoods differing in
pre-existing retail access would otherwise have similar trends in the absence of FC entry. Additionally, I control for neighborhood proximity to FC entry location to ensure the estimates reflect the
effect of e-commerce expansion and are not confounded by potential spillovers from large warehouse
openings nearby. I conduct the analysis at the ZIP code level, examining the dynamic effects in
the periods leading to and following FC entry on home values, rents, and economic activity (e.g.,
number of restaurants). Under this empirical strategy, if the expansion of e-commerce results in
a dispersion force within cities, we would observe a significant increase in home values, rents, and
economic activity in areas with limited retail access. On the other hand, an e-commerce-induced agglomeration effect would imply a significant increase in the outcome variables in the more accessible
locations.
The reduced-form results are consistent with the hypothesis that e-commerce results in a decentralization effect within urban areas. Specifically, in the years following FC entry, neighborhoods with
historically limited access to retail goods experience an average of 4.2% increase in home values,
4.8% increase in rents, as well as a 33% increase in the number of restaurants. For neighborhoods
with historically more retail access, I detect no significant effects but a short-run decline in the
77

number of retail establishments in the three years following FC entry, suggesting little evidence
for agglomeration into these locations. These estimated effects are robust to alternative econometric specifications and definitions of accessibility measure. Furthermore, I observe no evidence
of differential trends across neighborhood differing in their retail accessibility prior to FC entry,
lending support for the parallel trend assumption and thus credibility for interpreting the estimates
as causal.
As locations with lower retail access appear to gain popularity, one may wonder if the rising housing
costs and improved amenities translate into changes in neighborhood characteristics, such as population size, economic status, and demographic composition. This inquiry holds relevance in assessing
the distributional consequences of e-commerce, as existing housing research and policy debates commonly highlight the association between rising housing costs and phenomena such as gentrification.
I find no such evidence of gentrification or sorting of residents as a result of e-commerce expansion.
Specifically, I do not observe statistically significant differences in trends across neighborhoods, in
terms of their average personal and household income, population size, as well as the employment
ratio of college-educated to non-college-educated workers both before and after FC entry. Again, I
detect no statistically differential trends across neighborhoods in their fundamental characteristics
prior to FC entry, which attests to the validity of the parallel trend assumption. While one might
attribute the increased housing costs and improved amenities in less accessible locations to potential
spillover effects on employment from warehouse openings, these findings do not support this hypothesis. There is no discernible evidence of an increase or decrease in household or personal income,
nor do I observe less skilled (non-college-educated) workers, who are more likely to be employed in
the warehousing and related sectors, sorting into less accessible neighborhoods. Hence, the observed
changes in these locations are unlikely to be driven by the employment effect, but rather by changes
in local consumption access stemming from the entry of FC. Taken together, these findings are
78

consistent with the conclusion e-commerce could stimulate local economic growth, particularly in
historically less attractive or accessible areas, contributing to overall decentralization within cities.
In the second part of the paper, I examine the welfare and distributional implications of e-commerce
in a quantitative spatial framework à la Ahlfeldt et al. (2015), with model components being informed by the reduced-form results. Specifically, I assume a closed city model, and abstract away
the resident’s workplace location and commuting problem. Hence, income is exogenously determined. The model focuses on the location choice problem for residence, retail consumption, and
service consumption. In the model, residents tradeoff housing cost for consumption access, so locations in proximity to consumption amenities would have higher rents and home values. I introduce
e-commerce or FC entry as a distortion to these tradeoff forces through two channels: a decrease in
the non-percuniary cost of online shopping and a reduction in local retail prices. Using data from
periods before and after FC entry, I estimate that FC entry shrinks the average non-percuniary
cost by between 29% and 58%, while local retail price index falls by an average of 0.6 points, the
latter of which is potentially due to hightened competition. As consumers now benefit from the
more convenient and cheaper access to retail goods, the housing and service sectors adjust to the
new equilibrium. Taken together, I calculate that, in the absence of e-commerce, the average city
renter is willing to pay a monthly amount of $113 (in 2017 dollars) and the average homeowner
benefitting from capital gain is willing to pay as much as $200 for its entry, suggesting an overall
welfare increase. Residents in historically less accessible locations enjoy higher surplus than those
in more accessible counterparts, and residents in larger cities see the highest welfare gain. This
result suggests that while e-commerce induces a decentralization effect within cities, it amplifies the
existing agglomeration benefits across cities.
79

The rest of the paper is organized as follows. Section 3.2 discusses related literature. Section 3.3
outlines the data used in the paper, followed by a discussion of identification strategy in Section
3.4. Results are presented in Section 3.5. Section 3.6 reports robustness results. I assess the
welfare impact of e-commerce in a quantitative spatial framework in Section 3.7. Lastly, Section
3.8 concludes.
3.2 Related literature
This paper relates and seeks to contribute to several strands of literature. First, it speaks to the large
urban economics literature studying the agglomeration impact of improved urban access. BaumSnow (2007) and Baum-Snow et al. (2017) document that transportation infrastructures such as
highways and railroads, which enable better access to city centers, contribute to urban spawl and
population decline. Dong, Zheng and Kahn (2020) find that reduced access cost to superstar cities
thanks to high-speed rail enhances productivity and other agglomeration benefits across cities.
While many of these existing studies focus on the role of large transportation projects, which
are mostly funded by federal and local governments, little attention in urban economics has been
given to the digital infrastructures, whose rapid expansion is mainly driven by the private sector.
Recently, Brueckner, Kahn and Lin (2021) and Delventhal and Parkhomenko (2020) suggest that
private sector’s arrangements like work from home could potentially reshape our cities by altering
how people access jobs and the workplace. Gorback (2020) shows that UberX entry reduces travel
cost to destinations that are difficult to reach by public transit and spurs economic activity in these
areas. I add to these recent work by documenting the relationship between online shopping and
consumption access, and how this relationship contributes to changing the spatial outlook within
cities. To the best of my knowledge, this paper is the first attempt to examine e-commerce from a
spatial and urban economic perspective.
80

This paper also joins the literature studying e-commerce as a significant component of the broader
digital economy. Early research in quantifying overall welfare gains from e-commerce starts with
Brynjolfsson, Hu and Smith (2003), who measure consumer surplus from increased product variety
among online booksellers. Recently, Dolfen et al. (2019) take advantage of rich card transactions
data to address the same question, suggesting substantial welfare gains from online shopping as
it offers convenience and more product varieties. Additionally, a growing literature examines the
impact of e-commerce on local economy and the labor market. Among these include Chava et al.
(2022), who show that e-commerce reduces wage and employment in the retail sector, which is
partially offset by growth in restaurant and warehousing jobs. Bauer and Fernández Guerrico
(2023) study the impact of state nexus legislation on retail employment outcomes, suggesting that
e-commerce crowds out big-box retailer and complements brick-and-mortar shops. Outside of the
United States, Couture et al. (2021) study the impact of e-commerce in rural China, finding that it
mostly benefits consumers and has little impact on rural entrepreneurship. To my knowledge, these
prior research have mostly looked at the effects at the aggregate level, such as county and city, and
is little spatial in nature. I take the analysis to the neighborhood level, and document heterogeneous
effects of e-commerce within cities, while accounting the indirect effects on neighborhood housing
and economic outcomes in evaluating overall welfare consequences.
Lastly, this research relates to the literature on quantitative spatial model first introduced by
Ahlfeldt et al. (2015). Extending this framework, many researchers have documented important
spatial implications of changes in transportation network and technology due to, for instance, new
public transit means (Tsivanidis 2019, Tan 2020, Miyauchi, Nakajima and Redding 2021) and the
entry of ride-sharing (Gorback 2020) and thus in job and consumption access. In the welfare section,
81

I adapt a simplified version of the framework and introduce e-commerce entry into the model as another form of accessibility intervention that does not entail the usual improvements in transportation
infrastructures or shocks to travel times.
3.3 Data
To conduct the research, I combine data from multiple sources. I first collect data on Amazon FC
facilities from the website of MWPVL International, a supply chain consulting firm that claims
specialized expertise in logistics, distribution, and warehouse operations. I construct neighborhood
measure of retail accessibility using retail employment data and travel times queried from the Open
Source Routing Machine (OSRM), a free and open-source routing tool that provides efficient shortest
path computation between any input locations. Lastly, data on neighborhood characteristics come
from the ZIP code-level statistics provided by the U.S. Census Bureau, Internal Revenue Service
(IRS), and Zillow. In what follows, I will describe the data and variables used in the analysis in
detail.
3.3.1 E-commerce facilities
I obtain e-commerce facility data from MWPVL International, which provides detailed listing of all
known existing and scheduled facilities within the Amazon Fulfillment Center network. This list
contains 393 current and future locations within the United States, of which 176 were opened prior
to the COVID-19 pandemic and are the locations I focus on in my analysis. For each facility, I
scrape information on the exact street address, estimated square footage, opening date, and a short
description about its main operations. I further restrict the sample to FCs that were opened after
2005, which is when Amazon first introduced two-day shipping and as a result shifted their logistic
strategy from several central FCs to a decentralized network with locations across the country.
82

Lastly, I exclude 14 seasonal and return-processing facilities as they likely have at most a transitory
effect and are thus unlikely to create a persistent direct impact on consumption access for the local
population. My final list includes 162 FC facilities. Figure 3.2 shows the expansion of the Amazon
FC network over time. In Appendix A.2, I provide more background information on the FCs and
Amazon’s warehousing strategy.
3.3.2 Measuring retail accessibility
I hypothesize that e-commerce expansion would have differential impact across locations depending
on their access to pre-existing retail amenities. To measure accessibility, I follow the standard in
the literature using travel times. Specifically, for each ZIP code z, I compute the average travel
time by car to nearby ZIP codes, weighted by the count of retail employment in each destination
ZIP code z
′
:
retailAccessz =
X
z
′
retailEmploymentz
′dzz′ (3.1)
where retailEmploymentz is the pre-FC number of employment in retail sectors that are most likely
affected by e-commerce. These sectors are NAICS 4522 Department Stores and NAICS 4523 General
Merchandise Stores, including Warehouse Clubs and Supercenters.1
I exclude employment count
in retail sectors that are unlikely to be impacted by Amazon expansion, such as Gasoline Stores,
Automobiles and Other Motor Vehicles, and Groceries.2 dzz′ is travel time in minutes by car, which
I query from the OSRM tool, between the census-defined internal points of ZIP codes z and z
′
. An
advantage of the OSRM tool over other commercial routing services like Google Maps is its ability
to work with map snapshot at a specified date, thus providing travel times for historical, rather
1The Census ZIP Code Business Patterns data does not provide exact employment count, but establishment
count by firm size bin. I follow Qian and Tan (2021) using the midpoint of each firm size bin as an approximation
for employment count. For ZIP codes with missing data in 2005, I impute the number by using the first data point
available after 2000.
2
In recent years, Amazon has started to expand their food and groceries delivery service. Volpe and Boland (2022)
suggests Amazon’s food retailing is growing fast, but only in recent years, and thus is unlikely to have a strong impact
during the period of my study.
83

than contemporaneous, routes. This is particularly important for identification since contemporary
street network and routing patterns may endogenously respond to e-commerce expansion or FC
entry. For instance, a local government may invest in expanding road network to accommodate
the FCs or adjust traffic rules in response to changes in travel patterns. To query travel times for
historical routes, I use the OpenStreetMap U.S. map snapshot on January 1, 2010, which is based
on the Census 2005 TIGER data.3 Since the number of all possible origin-destination ZIP code pair
is exponentially large (approx. 1 billion pairwise routes), I restrict to only pairwise permutation
between locations within the same Metropolitan Statistical Area (MSA) that are no more than 100
miles apart by a straight-line distance. In total, I query travel times for 2.8 million pairwise routes
in 360 MSAs. Among these, 71,366 queries (or 2.5%) are returned with nulled results. In Appendix
A.3, I describe a simple imputing method for routes with unidentified travel times.
The weighted average in equation (3.1) is computed for all destination ZIP codes among the set
of top 20% closest by a straight-line distance to the origin ZIP code, which corresponds to a 15 to
20-mile radius. I assume that consumers typically travel no more than such distance to access retail
amenities. This radius choice is also in line with Dolfen et al. (2019), who look at offline purchases
made within a 20-mile radius of a cardholder’s location. To account for within-neighborhood retail
access, I let travel time within each origin ZIP code as one half of the travel time to the closest
neighboring ZIP code.4 The retailAccessz measure has a natural interpretation as the average time
it takes for a resident in z to access retail amenities in their extended neighborhood. The higher
the value, the less accessible a neighborhood is to retail amenities. In Figure 3.4, I illustrate this
constructed variables for ZIP codes in the Los Angeles-Long Beach-Anaheim and San FranciscoOakland-Hayward MSAs.
3The OpenStreetMap project completed importing TIGER data in 2009. 2010 is the earliest snapshot that I could
use to query the most complete travel times.
4This is because dzz = 0 as the coordinates of origin and destination points are the same, so retail access within
each origin ZIP code would not be accounted for if dzz is not assigned a positive value.
84

It is worth noting that the absolute value of this measure depends on each city’s road network
and urban structure. Figure 3.3 plots the distribution of retail access time for various cities in the
sample. I observe a substantial variation in travel times across cities. For example, while the median
driving time it takes for a resident in Los Angeles to access retail goods is 35 minutes, a resident of
New York City would need to drive only for 21 minutes. Hence, a location’s accessibility should be
measured relative to other locations’ in the same city. In the analysis, I divide neighborhoods into
treatment groups based on their access time percentiles within their respective MSA. Specifically,
there are three treatment groups: (1) more accessible, (2) accessible, and (3) less accessible. A
neighborhood is considered less accessible if it takes residents in that neighborhood longer than
the city’s median to access retail amenities. As an example, the median access time among LA
neighborhoods is 35 minutes, so neighborhoods with access metric above that would be considered
less accessible. Among the neighborhoods with retail access better the median, I define the more
accessible group to consist of locations in top 25% most accessible, and accessible the remaining
next 25%. Formally, the group assignments are as follows:
accessBin1
z = I{retailAccessz ≤ retailAccessMSA(z)
p25 } (3.2)
accessBin2
z = I{retailAccessMSA(z)
p25 < retailAccessz ≤ retailAccessMSA(z)
p50 } (3.3)
accessBin3
z = I{retailAccessMSA(z)
p50 < retailAccessz} (3.4)
where accessBini
z
for i ∈ {1, 2, 3} indicates if ZIP code z belongs to one of the three treatment
groups, namely (1) more accessible, (2) accessible, and (3) less accessible, respectively; retailAccessz
is the constructed retail accessibility measure for ZIP code z, as described above; and retailAccessMSA(z)
pl
for l
th percentile of the retail accessibility measure for the city MSA(z) of ZIP code z.
85

3.3.3 Other data
Home value and rent data: I obtain ZIP code-level real estate data from Zillow, which consists
of the Zillow Home Value Index (ZHVI) and the Zillow Rent Index (ZRI). I use the smoothed,
seasonally adjusted ZHVI All Homes (SFR, Condo/Co-op) Time Series. This series provides a
monthly dollar-denominated index between 2001 and 2021 for all homes whose values fall into their
ZIP code’s middle tier price range (33th to 66th percentile). Zillow does not publish the ZHVI data
for the top or bottom tier at the ZIP code level. Similarly, the ZRI All Homes Plus Multifamily
series provide the neighborhood’s typical rents over the time for homes listed for rents that fall in
the middle tercile of ZIP code rent distribution. Compared with the ZHVI, the ZRI is available for
a shorter time span, from 2010 to 2021, and a smaller geographic coverage. As my analysis is at
the annual level, I take the average of all monthly values within each year for each series. I then
normalize each series so that its value in the year prior to FC entry equals 1, providing information
on how local housing dynamic evolves over time relative to the period immediate to treatment.
Local economic patterns: I measure local economic activity on the extensive margin using the
number of full-service restaurants and retail establishments. I collect this data from the Census
ZIP Code Business Patterns database for the period between 2001 and 2016.5
I focus on these
two sectors because restaurants, unlike other non-tradable industries typically have lower entry cost
and thus are more likely to respond quickly to local economic conditions (Gorback 2020), whereas
the retail sector is most likely to experience the direct impact of e-commerce expansion. Similar
to home values and rents, I normalize the establishment count series so that each reflects local
economic patterns relative to the time of FC entry.
5
I use data up to 2016 since the Census changed their reporting standard for the data after 2017, making it
essentially different from previous data.
86

Other neighborhood characteristics: To proxy for possible signs of general gentrification, I
first use the residential Individual Income Tax ZIP Code Data between 2005 and 2019 from the
Internal Revenue Service (IRS). For each ZIP code, I have data on adjusted gross income for all
tax return filings with mailing addresses within its boundary. I then divide the gross income by
the number of returns (which approximates the number of households) and the number of personal
exemptions (which approximates the population) to obtain ZIP code average household income and
average personal income, respectively. Additionally, I turn to the Longitudinal Employer-Household
Dynamics (LEHD)’s Residential Area Characteristics (RAC) data to proxy for neighborhood demographic composition. The LEHD provides detailed job count breakdown by sector and by residents’
characteristics, such as education, race and ethnicity, starting 2009. I collect the LEHD data at the
tract level, and crosswalk them to the ZIP code level using population weight in the relationship
file provided by the Census.
National Household Travel Surveys: The NHTS provides detailed trip data from nationally
representative sample every 8 years at the individual level, breaking down by trip purposes. The
two most recent surveys in 2009 and 2017 were updated to include questions on online shopping
behavior by asking respondents to report the number of Internet deliveries in the month preceding
the survey. I use this information together with data on shopping trips to construct statistics on
online- and offline shopping shares at the city level.
3.4 Empirical strategy
Before estimating the reduced-form effects, I identify the sample for analysis as follows. For each
FC entry, I define a service area k consisting of metropolitan statistical areas (MSAs) whose Central
87

Business District (CBD) is within 100 miles from the entry location.6 This radius choice is informed
by information provided by MWPVL and prior research on Amazon’s warehousing strategy (Houde,
Newberry and Seim 2022). If there are multiple FC openings in a service area over time, I define
the treatment timing as the earliest year that the area receives an FC. As previously mentioned,
I focus on the staggered rollout of FCs between 2005 and 2019. I include service areas that first
receive or are scheduled to receive an FC after 2019 as the never-treated control group as suggested
by Schmidheiny and Siegloch (2019) to reduce potential bias in the estimates of the treatment
effects. Although Amazon’s FC rollout strategy after 2019 may differ from before in response to the
COVID-19 pandemic, my results show no evidence of pre-trends and remain robust when excluding
this never-treated group.
To evaluate the effects of e-commerce expansion, I implement an event-study design exploiting the
staggered rollout of Amazon FCs and pre-period variation in neighborhood retail access as follows:
yzkt = α0 +
X
g̸=2
Xβ
g
d
· I{yearEntryk − t = d} × accessBing
z + δz + δhkt + ϵzkt (3.5)
where yzkt refers to outcome of ZIP code z in service area k in year t. I(yearEntryk − t = d)
is the indicator whether calendar year t is d years relative to the opening date of the first FC in
service area k. accessBing
z’s are the set of dummies indicating the treatment group to which ZIP
code z belongs. δz denotes ZIP code z’s fixed effect, which controls for time-invariant neighborhood
characteristics that may correlate with the timing of FC entry. I normalize the treatment effect
coefficients of the event year immediately preceding FC entry β−1 to 0. Note that this triple
difference design allows me to difference out the service area-wide general equilibrium effects common
to all treatment groups due to FC entry, similar to the design used to study firm entry in Qian and
6
I follow Holian and Kahn (2013) and define CBD location using the longitudinal and latitudinal coordinates
returned by Google Earth when querying each MSA’s principal city.
88

Tan (2021). The coefficient β
1
d
(β
3
d
) hence identifies the dynamic treatment effect d years both before
and after e-commerce expansion on neighborhoods that are most (least) accessible to pre-existing
retail amenities relative to those with median access, the omitted baseline group accessBin2
z
. I
cluster standard errors at the service area level, allowing for outcome variables to correlate over
time and between neighborhoods treated by the same FC.
Since each FC is a large warehouse employing over a thousand workers, interpreting β
g
d
as the
treatment effect of e-commerce expansion could be confounded by the spillovers from increased
employment due to plant openings, especially if accessibility correlates with proximity to FCs. For
instance, the arrival of new warehouse jobs may attract less skilled workers to move in cheap, less
accessible neighborhoods near the FC, which cause rents and home values to go up in these areas.
The bias could also go in the opposite direction: home values and rent may decrease if the arrival
of FCs creates negative externalities like noise or pollution due to the increase in trucking and
warehousing activities. I address this empirical challenge in two steps. Firstly, for each city, I
restrict the sample to the set of top 10% neighborhoods closest to the CBD, which corresponds to
a ring of 5 to 10 miles, where most economic activities take place (Glaeser and Kahn 2001) and is
likely far away enough from the city’s outskirts to experience significant spillovers from warehouse
openings. Secondly, I include a rich set of time fixed effects that vary by neighborhood proximity to
FC and by service area. Specifically, for each service area k, I divide neighborhoods into (10) decile
bins based on their straight-line distance to the FC. I then allow neighborhoods in each decile bin
h in each service area k to trend non-parametrically for each calendar year t, hence the fixed effect
term δhkt in the specification. This fixed effect subsumes the less coarse service area-by-year fixed
effect, absorbing potential unobserved time-varying spillovers due to proximity to FCs in addition
to unobserved time-varying endogenous forces at the service area level. In the robustness section,
89

I provide estimates from an alternative design to rule out warehouse spillovers as a confounding
factor.
The identification assumption is that, conditional on the fixed effects, the timing of FC entry is
uncorrelated with other unobserved neighborhood characteristics that may impact outcomes. In
other words, this is equivalent to neighborhoods across treatment groups having similar trends in
the absence of FC. Note we do not require the timing of FC entry to be exogenous to city- but
neighborhood-level characteristics: while Amazon may prioritize super-star cities like New York and
Los Angeles over smaller cities like Huntsville in their FC expansion strategy, such decision is less
likely be driven by ZIP code-level (time-varying) characteristics within each city, once controlled
for the rich set of fixed effects as discussed. We do not observe the counterfactuals, i.e. how
neighborhoods evolve had there not been FC entry, so the parallel trends assumption is not directly
testable. However, the absence of differential trends across treatment groups in the periods leading
to FC entry, or “pre-trends,” could provide support for the assumption. Throughout the analysis,
I do not observe statistically significant pre-trends, suggesting little differences across treatment
groups and thus lending credibility to interpreting the estimates as causal.
3.5 Results
In this section, I present and discuss the reduced-form estimates for the spatial impact of e-commerce
expansion. The treatment group consists of ZIP codes with average pre-period retail access time
above the city’s median, which I label less accessible for short, and those with average access time
in the top 25%, i.e. more accessible. The baseline group consists of locations with access time in
the next 25%.
90

3.5.1 Impact on housing and local economic activity
In Figure 3.5, I present the event-study coefficient estimates from equation (3.5) for home value and
rent indices in panels (a) and (b), respectively. The estimates suggest consistently similar trends
leading up to FC entry across treatment groups. In the years following FC opening, I find that
locations with less access to pre-existing retail amenities experience an average of 4.2% increase
in home value and 4.8% rise in rent relative to the year prior to FC entry. These estimates are
statistically significant at 95% level. In stark contrast, such impact is absent in neighborhoods with
historically higher access to retail goods.
Figure 3.6 presents the estimates for the effect of e-commerce on the spatial distribution of economic
activity. The estimates for restaurants are plotted in panel (a) and those for retail establishments
are presented in panel (b). Similar to home value and rent indices, there are no significantly different
pre-trends. Following FC entry, I observe an average increase of 33% in the number of restaurants
relative the year prior to FC entry in less accessible neighborhoods, whereas more accessible locations
experience a short-run decrease in the number of retail establishments.
3.5.2 Does e-commerce cause gentrification?
As e-commerce expands and induces higher home prices, rents, and economic activity in areas
with historically limited access, one may ask if the effects are driven by gentrification. Figure
3.8 plots the event-study coefficients for potential gentrification outcomes. In all panels, I observe
similar trends across treatment groups both before and after FC entry, suggesting little evidence
of gentrification or neighborhood sorting by socio-economic characteristics. Specifically, I do not
observe statistically significant changes in per capita income, average household income, population,
as well as employment ratio of college to non-college workers. Taken together, my results are
91

consistent with the dispersion hypothesis, where e-commerce stimulates economic growth in less
accessible locations and reduces economic activity in previously more accessible counterparts. The
effect is not driven by gentrification, but rather due to residents in less accessible locations having
improved access to consumption.
3.5.3 Discussion
What underpins these changes? Online shopping induces a stronger substitution effect on in-store
visits for residents living in neighborhoods with limited access than those in more accessible locations. As the arrival of FC introduces a convenient access to retail products, locations with limited
access featuring more affordable housing costs become more desirable to live in, resulting in higher
housing demand and thus higher property values and rents. Such appreciation in the housing market results in a spillover effect onto the local economy, as seen in the increase in the number of
restaurants in less accessible locations. On the other hand, in locations with higher pre-existing
retail access, competition with online shopping causes a negative effect on the retail industry. The
observation that the impact is rather short-run may be due to businesses gradually adapting to
e-commerce, for instance, by expanding their online presence and improving their warehousing and
delivery handling.
It is worth discussing other potential mechanisms that offer alternative explanations for the observed
effects. This concern would be valid if there are unobserved interventions (e.g., a government policy
or a competitor’s action) that occur around the entry dates of the FCs. Prior literature has disccused
the association between FC openings and the enactments of local nexus laws (Houde, Newberry and
Seim 2022). The nexus legislation typically requires online retailers to collect sales tax on behalf
of local governments, regardless of whether or not the online vendors carry a physical presence in
the state. On this account, my results could be driven by goverments’ place-based policies (e.g.
92

investment in public schools or public infrastructures) favoring locations with historically limited
access, as a result of additional tax revenue. It is unclear why local goverments may target less retail
accessible locations, as opposed to, for example, neighborhoods with limited access to physicians,
where I do not observe similar treatment effects in the placebo test described in Section 3.6. In
some cases, Amazon voluntarily collects sales tax prior to FC entry in exchange for other business
incentives from local governments (Baugh, Ben-David and Park 2018, Houde, Newberry and Seim
2022, Cafcas and LeRoy 2016). For this reason, if increased place-based government spending is the
main driver for the post-FC real estate and economic dynamics observed in less accessible areas, we
should at least have observed some of these effects prior to FC entry, which is not the case.
3.6 Robustness checks
3.6.1 Alternative baseline group
In Section 3.4, I discuss a potential threat to identification concerning the retail access metric
being correlated with proximity to FC facilities, which may rationalize the results as being driven
by spillovers from warehouse openings. This is unlikely because the results in Figure 3.8 suggest
little evidence of negative sorting of residents into less retail-accessible neighborhoods by income
and education. As an additional robustness check, I employ an alternative omitted baseline group,
which consists of neighborhoods with similar retail access time as the previous controls but located
further away yet not too far from the city center so that they are not fundamentally different from
the treatment groups. Specifically, for each city, I restrict to the 15% neighborhoods closest to
the CBD. The set of first 10% ZIP codes closest to the center is assigned to treatment groups as
described in equations (3.2) to (3.4). ZIP codes in the second quartile and located in the next 5%
closest to the CBD constitute the new baseline group. Intuitively, if the results in Section 3.5 are
93

explained by neighborhood proximity to FC locations, then we should not observe increases in home
values and rents in ZIP codes closer to the CBD (i.e., further away from the location of FC entry).
Alternatively, robust results under this specification would suggest that the observed outcomes are
largely explained by pre-period variation in retail access.
Figure 3.9 plots the set of robustness estimates for the main outcome variables. The results are
similar to those presented in Section 3.5 in that there are little differential trends across treatment
groups prior to FC entry, and that neighborhoods with lower pre-period retail access experience
an increase in home values, rents, and restaurant services following e-commerce expansion. These
estimates are all well within the range of the those estimated using the benchmark specification.
Note that with the “outer ring” control group, I am now able to estimate the treatment effect on
the previously omitted “inner ring” control group, which is shorthanded as the “accessible” group
in the figure. More precisely, here the “inner ring” estimates identify the heterogeneous treatment
effect by distance to CBD (and so to FC as well) conditional on neighborhoods having retail access
times within the city’s second fastest quartile. The fact that there is no statistically significant
difference between the inner ring and outer ring neighborhoods both before and after FC entry
suggests that potential spatial spillovers of warehouse openings as well as other common treatment
effects that vary by distance to CBD are differenced out by the set of location-by-time fixed effects
specified in equation (3.5). Taken together, it is likely that pre-period variation in retail access
across neighborhoods is the main mechanism driving the spatial effects of e-commerce expansion.
3.6.2 Alternative measures of accessibility
Recall in equation (3.1) where I define retail access of a ZIP code as the average travel time by car
to nearby ZIP codes, weighted by retail employment at each destination. I employ two additional
measures of accessibility as robustness checks. First, to test for the sensitivity of the results, I weight
94

travel times by the number of retail establishments instead of approximated employment count.
The estimates plotted in Figure A.1 suggest that the results are not sensitive to this alternative
definition. Secondly, as a placebo test, I repeat the regressions using a non-retail accessibility
measure. Specifically, I use the number of physician offices (NAICS 621111), a non-tradable service
that residents have to travel to and cannot be substituted for by e-commerce orders. Figure A.2
plots the estimates for this placebo exercise. Although I observe a decrease in home values in areas
more accessible to physician offices, the fact that the estimates are not statistically significant for
other outcomes supports the explanation that my main results are driven by pre-period variation
in retail access.
3.6.3 Other robustness checks
So far, I have showed robustness results that employ several alternative definitions for baseline
group, sample as well as definitions of accessibility. I outline additional robustness checks in this
section, and Figures A.3, A.4, and A.5 in Appendix A.1 plot the results for all these alternative
specifications. Overall, the results remain similar to what is observed in the benchmark equation.
Specifically, instead of defining the fixed effects at service area δkht in equation (3.5), I show that
the results are extended to when including fixed effects at the MSA level c, or δcht, which allow
cities treated by the same FC to have different non-parametric trends. Additionally, the results are
robust when I estimate specification (3.5) with additional city linear trends or with ZIP code linear
trends.
3.7 Welfare implications of e-commerce
In this section, I present a simplified quantitative spatial framework for welfare evaluation of ecommerce. The analysis features a closed city model in which residents choose where to live and to
95

consume retail and service goods from a set of locations that differ in their consumption quality, as
well as travel costs to other locations. Consumers have standard preference represented by CobbDouglas utility function and derive utility from their consumption of housing, retail goods, and
service goods (restaurants). Since e-commerce is unlikely to directly affect workplace and job access
and because the reduced-form results do not suggest statistically significant changes in neighborhood
average income and demographic composition, I abstract away from the resident’s workplace location
choice as well as worker heterogeneity, taking income as exogenous. Another implicit assumption
of the model is that housing supply is exogenously fixed and inelastically responds to e-commerce
entry, at least in the short run. In equilibrium, housing, retail, and service markets clear at all
locations.
In what follows, I present the resident’s problem in Sections 3.7.1 and 3.7.2 I then describe parameterization and estimation strategy in Section 3.7.3. Lastly, I discuss the welfare metric in 3.7.4 and
the welfare implications in Section 3.7.5.
3.7.1 Resident’s choice problem
I adapt resident’s demand structure similar to that in Ahlfeldt et al. (2015) and in Miyauchi,
Nakajima and Redding (2021). Specifically, individuals derive utility from their consumption of
housing amount hi at location i, retail goods Cj at location j, and service goods (restaurants) nk
at location k following Cobb-Douglas utility function
Uijk =

hi
αh
αh

Cjz
C
ij
αC
αC

nkz
n
ik
αnt
n
ik
αn
(3.6)
where 0 < αh, αC, αn < 1 and αh+αC+αn = 1 are the income shares of housing, retail, and services,
respectively. z
C
ij is the idiosyncratic preference shock to retail consumption at location j conditional
96

on living in i. Similarly, z
n
ik denotes the idiosyncratic quality shock of service consumption at k
when the individual chooses to reside in i. t
n
ik is the iceberg travel cost from i to k for service
consumption.
To introduce online shopping to the model, I define Cj as the “super” retail goods consisting of both
in-store and online purchased goods using a simple CES utility structure:
Cj =

θ
β
ij
t
C
j
cj
σ−1
σ +
(1 − θij )
β
ϕ
ce
σ−1
σ
σ
σ−1
(3.7)
where cj and ce are the amount of offline- and online-purchased retail goods. t
C
ij is the travel disutility
from i to j for offline retail consumption. ϕ is the parameter representing non-price costs associated
with online shopping that do not depend on travel time. The two disutility terms are weighted by
the corresponding shares of in-person and online shopping trips made by the resident, denoted by
θ ∈ [0, 1] and (1−θ) respectively. β > 0 is the consumption elasticity of shopping frequency for each
mode, implying that residents prefer shopping both in-person and online to doing solely either way.
Note that, in the absence of e-commerce, θ takes value 1 and the utility from retail consumption
has the same non-nested structure as the standard model without e-commerce. σ > 1 is the usual
elasticity of substitution between offline and online purchased goods.
I follow the literature convention assuming that z
o
i
for o ∈ {c, n} are independently and identically
Fréchet-distributed across locations and sectors, namely z
o
ij ∼ F(z
o
ij ) = e
−TiEo
j
(z
o
ij )−ϵo
with Ti
, Eo
j >
0 and for all i, j and ϵo > 1. Here the “shape” parameters Ti and Eo
j
determine the average value
of having residence in location i and average quality of consuming goods of sector o at location j,
respectively. The “scale” parameters ϵo determine the degree of substitution for consuming goods of
sector o among locations. If ϵo is high, consumers have diverse idiosyncratic tastes across locations
97

and thus we are more likely to observe consumers traveling a long distance to consume the goods.
Conversely, when ϵo is low, there is less variation in taste among consumers, hence consumption
destination tends to be closer to residence location.
In the model, a resident, given income, prices and travel costs, makes the following sequence of
choices to maximize their utility:
i. Choose residential location, indexed by i
ii. Given on residence at i, observe vectors z
C
i
and z
n
i
of idiosyncratic shocks for all locations and
choose retail and service consumption destinations, indexed by j and k respectively.
iii. Given retail consumption route ij, choose how often to shop in-person for retail goods, denoted
by θij ∈ [0, 1]
iv. Given route ijk and θij , choose optimal consumption bundle {hi
, cj , ce, nk}.
Formally, the utility maximization problem is as follows
max
{hi,cj ,ce,nk}|ijk;θ
Uijk(θ) s.t. qihi + p
c
j
cj + p
e
ce + p
n
knk ≤ Ii (3.8)
where Ii
is the exogenous income of residents living in i, qi
is per unit cost of housing in residence
location i, p
c
j
is the price of retail goods at location j, p
e
is the online price of retail goods common
to all locations, p
n
k
is the price of service goods at location k.
98

3.7.2 Model solution
Given the resident’s sequence of choices described above, I solve the problem backwards starting
with deriving the optimal bundle {hi
, Cj (θ), nk}, conditional on previous location choices ijk and
θ.
hi =
αhIi
qi
Cj (θ) = αCIi
P
C
ij
nk =
αnIi
t
n
ikp
n
k
(3.9)
where P
C
ij is the price per unit of the composite retail goods Cij for residents with location choice
ij:
P
C
ij (θ) =
θ
β
t
c
ij
σ−1
p
c
1−σ
j +
(1 − θ)
β
ϕ
]
σ−1
p
e
1−σ
1
1−σ
(3.10)
which is obtained by maximizing the nested CES utility Cj (θ) subject to budget constraint αCIi
,
the fraction of income spent on retail goods.
Substituting optimal demands into (3.6) yields indirect utility function for residents choosing route
ijk and in-person share θ:
Vijk(θ) = Ii
q
αh
i

z
C
ij
P
C
ij (θ)
αC

z
n
ik
t
n
ik
αn
(3.11)
The indirect utility of route ijk depends positively on income obtained by living in i, negatively
on prices at consumption locations j, k and on travel costs from i to j and k, as well as on the
resident’s idiosyncratic tastes. Note that, given residence choice i, utility from retail consumption
solely depends on the second term, which also determines optimal θ once conditioned on retail
consumption location j. Likewise, the last term fully determines utility from service consumption.
Solving for optimal θ
Maximizing equation (3.11) requires an optimal θ
∗
that minimizes price P
C
ij of the composite retail
goods Cj . In other words, once a resident has decided on where to live and to shop in-person,
99

they determine an optimal trip allocation for in-store and online shopping that achieves the lowest
effective consumption price:
min
θ|ij
P
C
ij (θ) (3.12)
Proposition 1 Given route ij, the following trip allocation θ
∗
ij is optimal if β ≤
1
σ−1
:
θ
∗
ij =
1
1 + p
c
j
p
e
t
cij
ϕ
σ−1
1−β(σ−1)
(3.13)
Proof: Appendix A.4
Equation (3.13) suggests that that share of online shopping is higher if, holding everything else
equal, (1) it is cheaper to shop online, i.e. p
e < pc
j
, and (2) residents are more averse to commuting
to physical stores than to shopping online, t
c
ij > ϕ.
Solving for local demand equations
Conditional on living in i, the resident chooses retail consumption destination by solving
max
j
z
C
ij
P
C
ij (θ
∗
ij )
(3.14)
By standard property of Fréchet distribution, the share of population choosing retail consumption
destination j conditional on living in i has the following closed-form expression:
π
C
ij|i =
EC
j P
C
ij (θ
∗
ij )
−ϵC
P
j
′ EC
j
′P
C
ij (θ
∗
ij )−ϵC
(3.15)
100

Similarly, the share of population living in i choosing service consumption destination k is
π
n
ik|i =
En
k
(p
n
k
t
n
ik)
−ϵn
P
k
′ En
k
′(p
n
k
′t
n
ik′)−ϵn
(3.16)
Equations (3.15) and (3.16) suggest that an individual is more likely to consume at a location if,
relative to other locations, the average quality value E is higher in that location and if travel cost t
from home to that location is lower. The demand for retail (service) goods DC
j
(Dn
k
) at location j
(k) is determined by the aggregate demand of all individuals who travel to j (k) to consume from
all residence locations i:
DC
j =
X
i
θ
∗
ijπ
C
ij|i × αCIi Dn
k =
X
i
π
n
ik|i × αnIi (3.17)
These equations suggest that demand for a sector’s goods at a destination location is the weighted
average of income spent on that sector among all origin locations in the city, where the weights are
determined by how likely the resident from each origin location is to travel there physically.
Residential location choice
Since residential choice is assumed to take place prior to the realization of idiosyncratic shocks
z
c and z
n
, individuals form expectation over these shocks and choose to reside in a location i
101

that maximizes their expected utility. Double-integrating indirect utility function (3.11) over the
idiosyncratic shocks yields the expected utility of living in location i:
Vi ≡ E(Vi) = Ii
q
αh
i
RiSi (3.18)
Ri =

Γ
ϵC
αC
− 1
ϵC
αC
X
j
′
E
C
j
′P
C
ij′(θ
∗
ij′)
−ϵC
αC
ϵC
(3.19)
Si =

Γ
ϵn
αn
− 1
ϵn
αn
X
k
′
E
n
k
′(p
n
k
′t
n
ik′)
−ϵn
αn
ϵn
(3.20)
where Ri and Si are the expected utility from retail and service consumption conditional on living
in i, which are now referred to as retail access and service access for short. Taken together, equation
(3.18) suggests that individuals prefer to live in areas with higher income, lower prices, and higher
level of retail and service access.
3.7.3 Parameterization and estimation
Parameterization
In bringing the model to the data, I first parametrize and make assumptions on the specfied parameters and variables. Specifically, I let t
c
ij = e
τcdij and t
n
ik = e
τndik for service consumption travel
costs, where dzz′ is the driving time in minutes from z to z
′
. τc and τn, which represent the travel
cost elasticities of commuting time for retail and service consumption, along with online shopping
cost ϕ are the parameters to be estimated.
The remaining parameters of the model are externally set according to estimates and calibrated
values from previous literature. I set the consumption share of housing αh = 0.24 as in Davis and
Ortalo-Magné (2011), share of retail consumption αc = 0.56, which is in the range of the estimate by
Moretti (2013), and therefore the implied share of service consumption is αn = 0.20. The Fréchet
102

scale parameters are equated, ϵn = ϵC = 4.4, which is the average value of ϵ = 3 amd ϵ = 6.8
from Tan (2020) and in line with the estimate ϵ = 4.43 by Delventhal and Parkhomenko (2020).
The elasticity of substitution for online- and offline-purchased goods σ is set to 4.3, according to
Dolfen et al. (2019). The elasticities of substitution between retail outlets and restaurant outlets
are taken from Atkin, Faber and Gonzalez-Navarro (2018) and Couture (2016), respectively. Since
there is no prior empirical estimate for β, I let β = 0.15 - the middle of the range [0,
1
σ−1
] required
for Proposition 1 to hold - although the choice does not affect the qualitative results of my welfare
analysis. Table 3.1 summarizes all parameters specified in the model.
Estimating travel time elasticities: τc and τn
Setting θij = 1, substituting commuting shares in equations (3.15) and (3.16) into the demand
schedules in (3.17), and taking log on both sides yield the following equations:
log Dc
j = λcity + log X
i
p
c
−ϵC
j
e
−ϵCτcdijαcIi

+ log E
c
j
(3.21)
log Dn
k = γcity + log X
i
p
n−ϵn
k
e
−ϵnτndik αnIi

+ log E
n
k
(3.22)
where λ = log(P
j
′∈city Ec
j
′p
c
−ϵC
j
′ e
−ϵCτcdij′
) and γ = log(P
k
′∈city En
k
′p
n−ϵn
k
′ e
−ϵnτndik′
) are constant
terms common to all comsumption locations within the city, and Ec and En are the vectors of
idiosyncratic tastes across locations in the city.
Since I do not empirically observe retail and restaurant demands Dc
j
and Dn
k
, I impute their values
using the following formula:
Dc
j = E¯
c × rppMSA(j)
c × R
1
σc−1
j Dn
k = E¯
n × rppMSA(k)
n × S
1
σn−1
k
(3.23)
1

where E¯
c and E¯
n represent national averages of personal spending on retail and restaurant goods in
a given year. I introduce city-level variation by multiplying these national averages with MSA price
parity indices, denoted as rpp
MSA(j)
c and rpp
MSA(k)
n . Furthermore, I consider the local availability
of varieties by incorporating the counts of retail establishments (Rj ) and restaurants (Sj ) raised to
the level of “love for variety,” which is inversely correlated with the elasticities of substitution (σc
and σn). In essence, this approach yields a variety-adjusted, dollar-denominated local price index
for per unit of annual consumption within each sector.
With ϵ’s being externally calibrated to 4.40, I estimate the above equations using pooled national
ZIP code sample via non-linear least squares to recover the travel cost elasticities τc and τn. Note
that without setting θ = 1, I cannot seperately identify τc as both τc and ϕ joinly determine equation
(3.21). As the result, the data used to estimate equation (3.21) must correspond to the state where
θ = 1 (i.e. no online shopping) or is very close to 1. I use data from 2001, when the share of
e-commerce among total retail sales was less than 1%, for recovering τc. In the model, θ = 1 could
be rationalized by the aversion ϕ to e-commerce taking an infinitely large value, which reflects
consumers’ significant reluctance towards online shopping during the early days of the Internet.
The data sources for data estimation are as follows: I use ZIP code’s adjusted gross income of each
origin i divided by the number of personal exemptions, reported by the Internal Revenue Service,
as a measure for average personal income Ii
. dij is travel time in minutes queried from the OSRM
tool. E¯
c is computed by dividing the BEA’s national aggregate value for Personal Consumption
Expenditures (PCE) for Goods – both durable and non-durable – by total U.S. population. Similarly,
E¯
c is the per capita value of national aggregate for Purchased Meals and Beverages, under the Food
Services category. ZIP Business Patterns provide statistcs on local count of retail establishments
and full-service restaurants.
104

Table 3.2 presents the estimates for τc and τn using equations (3.21) and (3.22). Compared to prior
research that primarily focused on commuting data for work, such as by Ahlfeldt et al. (2015) and
Tsivanidis (2019), who reported respective values of approximately τ ≈ 0.011 and 0.012, my higher
estimates suggest that travel for consumption exhibits greater elasticity to travel time compared to
travel for work and are consistent with similar findings by Miyauchi, Nakajima and Redding (2021)
and Tan (2020). This observation underscores the nuanced relationship between travel behavior
and the nature of the trip purpose.
Recovering ϕ: the convenience channel
To recover ϕ, I start with optimal trip allocation θ
∗
ij in equation (3.13). Averaging θ
∗
ij across origindestination pairs ij within the city yields the city-average trip allocation Θ¯
city:
Θ¯
city =
1
#Nij
X
i,j∈city
1
1 + p
c
j
p
e
t
cij
ϕ
σ−1
1−β(σ−1)
(3.24)
where #Nij is the number of possible routes in the city. Using data from the National Household
Travel Surveys, I construct empirical city-level average shares of shopping trips made in-person,
specifically:
Θ¯
city =
# shopping tripscity
# shopping tripscity + # online deliveriescity
(3.25)
where # shopping tripscity is the count of all physical trips taken for shopping and errands-related
purposes by city residents in the survey year, and # online deliveriescity is the number of online
deliveries received by city residents in the corresponding year, where I count each online order as
one shopping “trip”.7
I make the assumption that (1) in-store prices are the same across locations
within the city, and (2) online price p
e
is the same for all locations and equals the national average
7The NHTS individual survey provides information on the number of Internet orders delivered in the past 30 days.
I obtain the number of annual deliveries by multiplying the value by 12.
105

price. Under these two assumptions, the in-store to online price ratio p
c
j
/pe are proxied for using
the BEA’s MSA-level regional price parity for non-service goods.
As e-commerce expansion makes online shopping more convenient (reducing ϕ), I recover ϕ seperately for the periods before and after FC entry to reflect this effect. I restrict the sample to ZIP
codes within 29 metro areas with the first FC entry between 2010 and 2016 as the 2009 and 2017
NHTS trip data allows me to observe both pre- and post-FC in-store shares Θ¯
city for this set of
cities. Additionally, while the constructed in-store share Θ¯
city and observed price ratio p
c
j
/pe
represent values at equilibrium, recovering post-FC value of ϕ requires values exogenous to changes
other than FC entry. To approximate for the exogenous in-store shares and price ratios, I use the
predicted values from the following difference regression:
yˆcity,t = β0 + β1postF Ccity,t + γcity + λt (3.26)
where ycity,t is either the in-store share Θ¯
city,t or price ratio p
c
city,t/pe
t
for city in year t, and
postF Ccity,t is an indicator for the periods following FC entry.
In recovering ϕ, I allow for two specifications. The first pins down ϕ uniquely for each city, allowing
me to observe the impact of e-commerce expansion specific to different regions. Note that there is
only one unknown in each version of equation (3.24) specific to each city, hence ϕ could be solved
uniquely for each one. In the the second specification, I assume a common ϕ across all cities and
estimates its value via non-linear least squares on the pooled national sample.
Table 3.3 presents the results for ϕ for both of the discussed specifications, along with predicted
city-level share of in-person Θ shopping from equation (3.26). In Panel A, I report CBSA-specific
values of ϕ and Θ, whereas Panel B presents the non-linear least squares estimate for a constant
106

ϕ using the pooled sample. Figure 3.10 plots the distribution of city-specific ϕ before and after
FC entry. There is a consistent pattern across cities. Prior to FC entry, consumers exhibit higher
aversion (ϕ) to online shopping, which rationalizes the higher share of shopping trips made in-person
(Θ). Following FC entry, e-commerce becomes more accessible and convenient, resulting in a fall
between 9.72% and 72.4% in the aversion parameters ϕ among the cities and reducing the share of
shopping trips made in-person.
E-commerce effects on local prices
So far, I have shown that e-commerce expansion affects the spatial equilibrium through reducing
ϕ, the convenience or non-price channel. Equation (3.13) suggests a second channel through which
e-commerce may impact the spatial equilibrium: changing local retail prices. This channel is particularly relevant for welfare if the proliteration of online platforms induces competition and price
transparency, causing local prices to fall. Furthermore, growth in the service sector in less accessible locations could cause local price changes in other non-retail sectors and may also have welfare
consequences in the new equilibrium.
In the model, non-housing prices are constant within but different across cities as in Gorback
(2020), hence the analysis requires city-level price changes exogenous to factors other than FC
entry, which are obtained using equation (3.26). To illustrate the e-commerce effects on local prices,
Figure 3.7 plots the event-study estimates using the robust heterogeneous treatment estimator by
De Chaisemartin and d’Haultfoeuille (2020). Following FC entry, I observe a statistically significant
decline in local retail price over time in Panel A, as suggested by an average total effect of 0.6 point
decline in the MSA price parity for non-service goods. This observation is consistent with existing
results in the literature, such as Jo, Matsumura and Weinstein (2019) who find that e-commerce
on average lowers price level by 0.9 percent. In Panel B, I observe an increase in local price for
107

services, as suggested by an average total effect of 0.33 point increase in the MSA price parity for
services excluding housing and utilities. This increase in service price could be rationalized by the
increased demand for consumption in this sector, particularly in less accessible locations as shown
in Section 5.
3.7.4 Welfare metric
In evaluating the welfare impact of e-commerce expansion, I adopt a compensation variation-styled
metric similar to Gorback (2020). Suppose that an individual living in location i enjoys a given
utility level V¯
i prior to FC entry, I am interested in the dollar amount needed by this individual to
reach the same utility in the post period, holding everything else equal.8 Following the expansion,
if the agent requires a smaller amount of income to reach V¯
i
, then we say that she benefits from
e-commerce. Conversely, the individual is worse off if the entry of FC increases the dollar amount
required to reach V¯
i
. The difference in the required incomes before and after is the net welfare gain
(loss) of e-commerce, reflecting an individual’s willingness to pay (receive) for the improvements
available to online shopping. I proceed to computing this quantity as follows.
Recall from equation (3.18) that the expected utility of residing in location i is
Vi =
Ii
q
αh
i
(p
c)
αC (p
n)
αn
× Ri × Si
where p
c and p
n are U.S. city average per unit price for retail and service goods. Here a “unit” refers
to a year of consumption so that the prices could be measured by per capita annual expenditure
in each sector. Note that there are no subscripts in non-housing prices as spatial price variation,
8Throughout the analysis, nominal values in different years are adjusted to 2017 dollars using the Consumer Price
Index for Urban Areas, All Items (CPI-U) published by the U.S. Bureau of Labor Statistics.
108

which is proxied for using the price parities relative to the national average, is encompassed in the
retail Ri and service Si access terms. Taking logarithm of the expression yields
log Vi = log Ii − αh log qi − αC log p
c − αnlogpn + log Ri(κ) + log Si (3.27)
Since the reference utility threshold V¯
i
is just a constant, I normalize it to 1 so that log V¯
i = 0.
Rearranging equation (3.27) gives the logarithm of dollar amount required by a resident in location
i to afford a utility level of 1:
log Ii = αh log qi + αC log p
c + αnlogpn − log Ri − log Si (3.28)
The net welfare gain (loss) of e-commerce for residents living in i is defined as:
NW Gi = I
bef ore
i − I
af ter
i
(3.29)
where I
bef ore
i
and I
af ter
i
are computed according to equation (3.28) using data before and after the
intervention. Quantifying this compensation variation requires me to isolate changes in relevant
variables that are due to FC entry. City-level outcomes, i.e. price levels, are predicted using
equation (3.26), whereas ZIP code-level outcomes are predicted using the simple difference version
of the reduced-form equation (3.5), where discrete access bins are replaced by their continuous
counterpart:
log yit = γ0 + γ1P ostF Ckt × log(retailAccessi) + δi + δhkt + νikt (3.30)
where yit is outcome of ZIP code i in year t, including home value, rent, number of retail establishments, and number of restaurants. Since there are two measures of housing cost qi specific to
homeowner and renter demographics, I analyze the distributional impact on each group. While
109

the annualized housing cost of a renter is simply their monthly rent multiplied by 12, that of a
homewoner requires an imputation of their user cost. In Appendix A.5, I describe this imputation
procedure in detail.
3.7.5 Counterfactual exercise
In what follows, I examine the impact of e-commerce expansion on overall welfare through the main
channels idenfied in the previous empirical exercises: (1) change in convenience ϕ and (2) change
in local housing and consumption prices (qi
, pc
j
, pn
k
), and (3) retail access Ri and service access Si
.
Table 3.4 presents the results of this counterfactual exercise. Each row reports the incremental
willingess to pay (WTP) for each of the aforementioned channels as they progressively enter the
model.
Overall welfare evaluation
Following e-commerce expansion, the cost to online shopping ϕ decreases. This convenience gain
is substantial, at an average of $90-91 for home owners and renters. E-commerce expansion also
introduces competition among sellers and promotes price transparency, causing local retail prices to
fall. The fall in local retail cost p
c
creates an additional $37-39 in welfare for the average city resident.
As online shopping proliterates, physical retail amenities decline, resulting in an average loss of $8
in welfare due to reduced physical retail access in the new spatial equilibrium. As consumers enjoy
the convenience and lower retail price, the spatial distribution of home prices and rents adjust so
that there is no opportunities for arbitrage. Row (4) shows this adjustment yields homeowners an
additional gain of $67 on average due to lower user cost, whereas renters internalize the increased
rent burden of $20 per month. As retail goods are now more affordable and accessible, consumers
adjust their optimal bundle as their purchasing power increases, resulting in higher demand for
110

service goods and prompting the price of service goods to increase. Row (5) suggests that this
increase in price costs the average consumer $7 in welfare. Finally, row (6) allows service amenities
to respond to the increased demand for service goods, adding an additional $20 in monthly gain
for the average consumer. Taken together, this counterfactual analysis suggests that the average
homeowner is willing to pay $200 per month and the average renter $113 per month for the benefits
of e-commerce.
Spatial variation in online shopping refrequency and welfare impact
In this section, I discuss the spatial consequences of e-commerce across space within and across
cities. Figure 3.11 plots the model’s prediction of online shopping share (1 − θ) with ZIP code’s
prexisting access to retail amenities. As expected, within cities, residents in previously less accessible
areas have a significantly higher share of online shopping. The most responsive ZIP codes following
e-commerce entry in adopting online shopping appear to be ones with access time higher than the
city median, consistent with the reduced-form classification of “less accessible” in equation (3.4).
Table 3.5 presents the net welfare gain associated with residents’ access to existing retail amenities.
Both homeowners and renters experience progressively higher net welfare gains as we move from
the most to the least accessible locations within each city. Specifically, the average homeowner in
the 5th quintile of access time is willing to pay up to $375 per month, or about three times higher
than the $115 willingness-to-pay of their counterpart in the most accessible location. Between
homeowners and renters, the gap in welfare gains becomes smaller as we move from the most to
least accessible locations: while the average homeowner 1st quintile earns 2.43 times higher ($115
vs. $48) than the average renter in the same location, the difference reduces to 1.45 times ($375
vs.$258) in the 5th quintile. This suggests the increase in consumption access in less accessible
locations significantly outgrows the increase in housing costs, thus lowering the inequality between
111

homeowners and renters in these locations. On the other hand, the result implies that inequality
may be worsened in more accessible locations as e-commerce continues to penerate urban life.
Figures 3.12 and 3.13 plot the heatmap of welfare gains for New York, Los Angeles, San Francisco,
and Chicago. As seen, locations gain the most in cities are those further away from ciy centers,
which are less accessible to physical retail amenities. Across cities, I observe significant variation in
net welfare gains from e-commerce. Table 3.6 breaks down the net welfare gain by cities. Superstar
cities, such as New York, Washington DC, and San Francisco, enjoy the largest gain from e-commerce
expansion. This suggests that online shopping enhances the observed agglomeration benefits across
cities and does not substitute for urban living.
3.8 Conclusion
In this paper, I study the spatial implications of e-commerce on cities in the United States. To
the best of my knowledge, the paper is the first attempt to study this topic. The main results
highlight the role of e-commerce in improving consumption access for urban residents, particularly
those living in areas with less access to pre-existing retail amenities. Leveraging Amazon’s rollout
of their e-commerce fulfillment facilities across cities over time, I find that, following e-commerce
expansion, less retail-accessible neighborhoods experience an increase in home values, rents, and
restaurant services, whereas more retail-accessible neighborhoods experience a short-run decrease
in the number of retail establishments. I find no evidence of gentrification as result of improved
consumption access and the effects observed in less accessible neighborhoods are not explained the
spillovers due to the openings of new warehouses. Overall, the results lend support to that the
ability to shop online could increase local economic growth, particularly in less accessible areas, and
contribute to the dispersion of urban economic activity. Through a quantitative spatial exercise,
112

I find that e-commerce provides substantial welfare benefits for city residents, with homeowners
benefiting twice as much as renters.
The results could be relevant from policy point of view, especially when thinking about government
policies in the context of private and digital companies having an increasingly important role in
shaping our cities through improving urban access to consumption and jobs. In the short run, many
local governments may or have already encountered the question on what would be the overall
consequences of having companies like Amazon operate in their local areas and what would be
the optimal business incentives for these companies (and their competitors). Over the long run,
whether the digital economy will have positive impact on redistribution may depend on governments’
complementary policies, especially when historically underserved locations become more desirable.
There are several areas for future research. First, this paper is mostly consumer-focused and thus
has not investigated the impact of e-commerce expansion on local businesses. In a parallel work,
I study the spatial implications of e-commerce on firm location choices, pertaining to when firms
have the option to improve their supply chain efficiency and access to the national market through
online platforms. Second, future research agenda could also examine other non-pecuniary spillover
effects of e-commerce, especially when consumers are expected to rely more on online shopping.
One such example would be the impact of increased e-commerce activity on urban environment and
congestion. Lastly, e-commerce efficiency relies on well-developed existing physical infrastructures,
such as freeways and postal services, which are less common features in developing countries. A
natural extension is to see whether the results for the U.S. in this paper also hold in a developing
context.
113

Tables and figures
Figure 3.1: Number of daily physical trips per person, 1995-2017
0.8 0.7 0.6 0.6
2
1.8
1.6 1.3
0.4
0.4
0.4
0.4
1.1
1.1
1
0.9
0
1
2
3
4
5
1995 2001 2009 2017
Number of physical trips per day, 1995-2017
Commuting Shopping and errands School or church Social and recreational
Source: National Household Travel Surveys
Table 3.1: Summary for model parameters
Parameter Description Value Source
αh housing expenditure share of income 0.24 Davis and Ortalo-Magné (2011)
αC retail consumption expenditure share of income 0.56 Moretti (2013)
αn service consumption expenditure share of income 0.20 1 − αC − αn
ϵC Fréchet scale parameter for retail consumption 4.40 Tan (2020), Delventhal and Parkhomenko (2020)
ϵn Fréchet scale parameter for service consumption 4.40 Tan (2020), Delventhal and Parkhomenko (2020)
σ in-store and online elasticity of substitution 4.30 Dolfen et al. (2019)
β consumption elasticity of shopping frequency 0.15 mid-point of [0,
1
σ−1
]
σc retail outlet elasticity of substitution 4.36 Atkin, Faber and Gonzalez-Navarro (2018)
σn restaurant outlet elasticity of substitution 10.0 Couture (2016)
τc travel time elasticity for retail consumption estimated, see Table 3.2
τn travel time elasticity for service consumption estimated, see Table 3.2
ϕ aversion to online shopping recovered, see Table 3.3
114

Figure 3.2: Map of the Amazon FC network over time
(a) 2005 (b) 2010
(c) 2015 (d) 2019
Notes: This figure shows the expansion of the Amazon fulfillment facilities over time. Each dot represents a fulfillment center
facility. Seasonal and return-processing facilities are excluded from the map. Data source: MWPVL International Inc.
Table 3.2: Estimated travel time elasticities
Estimate Standard error
τˆc 0.0168∗∗∗ 0.0016
τˆn 0.0159∗∗∗ 0.0017
Notes: This table presents non-linear least squares estimates travel time elasticities τc for
retail consumption and τn for service consumption using pooled sample of ZIP codes within 29
core-based statistical areas (CBSAs) with first FC entry between 2010 and 2016. Significance
levels: ∗∗∗ p< 0.01,
∗∗ p< 0.05,
∗ p< 0.1
115

Figure 3.3: Distribution of retail access in driving time (minutes)
0 .1 .2 0 .1 .2 0 .1 .2 0 .1 .2 0 .1 .2 0 .1 .2 0 .1 .2 0 .1 .2
0 50 100 0 50 100 0 50 100
0 50 100
Atlanta Austin Baltimore Boston
Charlotte Chicago Columbus, OH Dallas
Hartford Kansas City, MO Los Angeles Minneapolis
Nashville New York Orlando Philadelphia
Providence Richmond, VA Riverside Sacramento
San Antonio San Diego San Francisco San Jose
Seattle St. Louis, MO Tampa Virginia Beach
Washington
Density
Retail Accessibility in Driving Time
116

Figure 3.4: Retail accessibility measure, Los Angeles & San Francisco MSAs
Los Angeles
San Francisco
Notes: This figure plots the retail access time for ZIP codes in the Los Angeles and San
Francisco MSAs using equation (3.1).
117

Figure 3.5: Impact of e-commerce on home values and rents
-.05
0 .05 .1 .15
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
Notes: This figure plots estimates for the dynamic treatment effects β
g
d
in equation (3.5),
where the dependent variable in panel (a) is the normalized Zillow Home Value Index
and that in panel (b) is the normalized Zillow Rental Index described in section 3.3.
Less accessible refers to ZIP codes with average pre-period retail access time higher than
city’s median, whereas More accessible consists of ZIP codes with average pre-period
retail access time in the city’s fastest quartile. Omitted baseline group is the second
fastest quartile. Event years beyond the plotted time window are binned together with
the earliest lead or latest lag. 95% confidence intervals are plotted, where robust standard
errors are clustered at the service area level.
118

Figure 3.6: Impact of e-commerce on local economic activity
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(a) Restaurants
-1
0 1 2
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(b) Retail Establishments
Notes: This figure plots estimates for the dynamic treatment effects β
g
d
in equation (3.5),
where the dependent variable in panel (a) is the normalized restaurant stock (b) is the
normalized stock of retail establishments. Less accessible refers to ZIP codes with average
pre-period retail access time higher than city’s median, whereas More accessible consists
of ZIP codes with average pre-period retail access time in the city’s fastest quartile.
Omitted baseline group is the second fastest quartile. Event years beyond the plotted
time window are binned together with the earliest lead or latest lag. 95% confidence
intervals are plotted, where robust standard errors are clustered at the service area level.
119

Figure 3.7: Impact of e-commerce on local prices
-1.5 -1 -.5
0 .5
-4 -2 0 2 4 6 8
Years Relative to Entry
(a) MSA price parity, non-service goods
-.5
0 .5 1 1.5
-4 -2 0 2 4 6 8
Years Relative to Entry
(b) MSA price parity, services excluding housing and utilities
Notes: This figure plots estimates for the dynamic treatment effects MSA-level price
indices, where the dependent variable in panel (a) is regional price parity for non-service
goods and that in panel (b) the regional price parity for services excluding housing and
utilities. The price parity for an MSA indicates the purchasing power of that region
relative to the national average, which is normalized to have a value of 100. Estimates
are obtained using dCdH estimator (De Chaisemartin and d’Haultfoeuille 2020). 95%
confidence intervals are plotted, where robust standard errors are clustered at the service
area level.
120

Figure 3.8: Impact of e-commerce on gentrification measures
-.2 -.1
0 .1
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(a) Average Personal Income
-.1 -.05
0 .05
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(b) Average Household Income
-.02 -.01
0 .01 .02
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) College to Non-college Employment Ratio
-.1 -.05
0 .05 .1
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Population
Notes: This figure plots estimates for the dynamic treatment effects β
g
d
in equation (3.5), where the dependent variables are
average personal income, average household income, college to non-college employment ratio, and ZIP code population. Less
accessible refers to ZIP codes with average pre-period retail access time higher than city’s median, whereas More accessible
consists of ZIP codes with average pre-period retail access time in the city’s fastest quartile. Omitted baseline group is the
second fastest quartile. Event years beyond the plotted time window are binned together with the earliest lead or latest lag.
95% confidence intervals are plotted, where robust standard errors are clustered at the service area level.
121

Figure 3.9: Robustness check: Alternative baseline group
-.1 -.05
0 .05 .1 .15
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible Accessible More accessible
(a) Home Value Index
-.2 -.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible Accessible More accessible
(b) Rent Index
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible Accessible More accessible
(c) Restaurants
-1.5 -1 -.5
0 .5 1
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible Accessible More accessible
(d) Retail Establishments
Notes: This figure plots robustness estimates for the dynamic treatment effects β
g
d
in equation (3.5), where the dependent
variables are the normalized Zillow Home Value Index, Zillow Rental Index, restaurant count, and retail establishment count.
Less accessible refers to ZIP codes with average pre-period retail access time higher than city’s median, More accessible consists
of ZIP codes with average pre-period retail access time in the city’s fastest quartile, Accessible are locations in the second
fastest quartile. These treatment groups consist of top 10% ZIP codes closest to the CBD. The omitted baseline group is ZIP
codes within the second fastest quartile and the next 5% closest to city center. Event years beyond the plotted time window
are binned together with the earliest lead or latest lag. 95% confidence intervals are plotted, where robust standard errors are
clustered at the service area level.
122

Figure 3.10: Distribution of online shopping cost before and after FC entry
0 5 10 15 20 25
0.0
0.1
0.2
0.3
0.4
Density
Before FC
After FC
Notes: This figure plots results for recovered city-specific ϕ presented in Table 3.3, Panel A
123

Figure 3.11: Model prediction for share of online shopping
0 .2 .4 .6 .8 1
Share of Online Shopping
0 .2 .4 .6 .8 1
Retail Access Time Percentile
(a) Before FC Entry
0 .2 .4 .6 .8 1
Share of Online Shopping
0 .2 .4 .6 .8 1
Retail Access Time Percentile
(b) After FC Entry
0 20 40 60 80
%
0 .2 .4 .6 .8 1
Retail Access Time Percentile
(c) ∆θ (% change)
Notes: This figure plots model prediction for 1 − θ – share of online shopping trips – before and after FC entry. Panel (c) plots
the difference between before and after in percentage terms. The size of each circle is proprotional to the population of the ZIP
code.
124

Figure 3.12: Welfare impact for homeowners in select cities
(a) New York (b) Los Angeles
(c) San Francisco (d) Chicago
Notes: This figure plots monthly welfare gains for homeowners located New York, Los Angeles, San Francisco, and Chicago
ZIP codes. All values are in 2017 dollars.
125

Figure 3.13: Welfare impact for renters in select cities
(a) New York (b) Los Angeles
(c) San Francisco (d) Chicago
Notes: This figure plots monthly welfare gains for renters located New York, Los Angeles, San Francisco, and Chicago ZIP
codes. All values are in 2017 dollars.
126

Table 3.3: Estimated and recovered costs of online shopping before and after FC entry
Panel A: City-specific ϕ Entry Pre-FC Post-FC %∆
CBSA Year ϕ Θ ϕ Θ in ϕ
Atlanta-Sandy Springs-Marietta, GA 2015 3.896 0.962 2.987 0.902 -23.34
Austin-Round Rock-San Marcos, TX 2013 13.620 0.955 5.210 0.898 -61.75
Baltimore-Towson, MD 2010 2.793 0.953 2.300 0.894 -17.65
Boston-Cambridge-Quincy, MA-NH 2016 2.546 0.946 2.295 0.888 -9.87
Charlotte-Gastonia-Rock Hill, NC-SC 2011 2.837 0.960 2.354 0.903 -17.03
Chicago-Joliet-Naperville, IL-IN-WI 2015 3.518 0.964 2.719 0.904 -22.70
Columbus, OH 2016 2.742 0.965 2.270 0.910 -17.21
Dallas-Fort Worth-Arlington, TX 2010 3.938 0.963 2.871 0.903 -27.08
Hartford-West Hartford-East Hartford, CT 2016 2.396 0.956 2.027 0.900 -15.38
Kansas City, MO-KS 2016 5.596 0.984 3.872 0.942 -30.81
Los Angeles-Long Beach-Santa Ana, CA 2012 24.846 0.969 7.394 0.908 -70.24
Minneapolis-St. Paul-Bloomington, MN-WI 2016 3.888 0.948 2.946 0.890 -24.21
Nashville-Davidson–Murfreesboro–Franklin, TN 2012 7.035 0.961 4.324 0.906 -38.54
New York- New Jersey-Long Island, NY-NJ-PA 2010 3.077 0.946 2.499 0.882 -18.79
Orlando-Kissimmee-Sanford, FL 2014 2.989 0.956 2.443 0.899 -18.27
Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 2010 2.646 0.947 2.122 0.887 -19.82
Providence-New Bedford-Fall River, RI-MA 2016 2.653 0.971 2.172 0.918 -18.14
Richmond, VA 2012 3.860 0.960 2.998 0.904 -22.33
Riverside-San Bernardino-Ontario, CA 2012 6.920 0.970 4.593 0.914 -33.62
Sacramento–Arden-Arcade–Roseville, CA 2013 3.848 0.966 3.057 0.911 -20.56
San Antonio-New Braunfels, TX 2013 26.110 0.965 7.195 0.908 -72.44
San Diego-Carlsbad-San Marcos, CA 2012 10.204 0.965 4.868 0.909 -52.29
San Francisco-Oakland-Fremont, CA 2013 3.708 0.963 2.732 0.905 -26.33
San Jose-Sunnyvale-Santa Clara, CA 2013 3.510 0.957 2.632 0.897 -25.03
Seattle-Tacoma-Bellevue, WA 2011 2.676 0.952 2.054 0.895 -23.25
St. Louis, MO-IL 2016 3.311 0.934 2.697 0.880 -18.54
Tampa-St. Petersburg-Clearwater, FL 2014 3.300 0.961 2.497 0.904 -24.32
Virginia Beach-Norfolk-Newport News, VA-NC 2012 6.345 0.959 3.530 0.903 -44.36
Washington-Arlington-Alexandria, DC-VA-MD-WV 2010 3.757 0.946 2.798 0.887 -25.53
Mean 5.812 0.959 3.257 0.902 -28.947
Panel B: Constant ϕ via NLS estimator Pre-FC Post-FC %∆
Parameter estimate standard error estimate standard error in ϕ
ϕ 8.143 2.046 3.382 0.239 -58.47
Notes: Panel A reports recovered values of ϕ, the aversion parameter to online shopping, and Θ, share of shopping trips made
in-person, for 29 core-based statistical areas (CBSAs) with the first FC entry between 2010 and 2016 before and after FC entry.
ϕ is recovered by solving equation (3.24) for each city. Panel B presents NLS estimate for ϕ, assuming that it is common for all
cities, using sample of 29 CBSAs with the first FC entry between 2010 and 2016.
127

Figure 3.14: Welfare impact by CBSA population
Atlanta
Austin
Baltimore
Boston
Charlotte
Chicago
Columbus, OH
Dallas
Hartford
Kansas City, MO
Los Angeles
Minneapolis
Nashville
New York
Orlando
Philadelphia
Providence
Richmond, VA
Riverside
Sacramento
San Antonio
San Diego
San Francisco San Jose
Seattle
St. Louis, MO
Tampa
Virginia Beach
Washington
100 200 300 400
Net Welfare Gain
5000 10000 15000 20000
CBSA Population (thousands)
(a) Homeowner
Atlanta
Austin
Baltimore
Boston
Charlotte
Chicago
Columbus, OH
Dallas
Hartford
Kansas City, MO
Los Angeles
Minneapolis
Nashville
New York
Orlando
Philadelphia
Providence
Richmond, VA
Riverside
Sacramento
San Antonio
San Diego
San Francisco
San Jose
Seattle
St. Louis, MO
Tampa
Virginia Beach
Washington
100 200 300
Net Welfare Gain
5000 10000 15000 20000
CBSA population (thousands)
(b) Renter
Notes: This figure plots the relationship between average net welfare gain from FC entry by CBSA population. All variables
are log-scaled.
Table 3.4: Welfare impact of e-commerce (2017 dollars)
Homeowner Renter
p25 p50 p75 mean p25 p50 p75 mean
Changes in (1) (2) (3) (4) (5) (6) (7) (8)
Convenience: ϕ 10 40 102 91 10 40 102 90
Retail price: p
c 32 36 46 37 32 36 43 37
Retail amenities: Ri -6 -9 -11 -8 -6 -8 -10 -8
Housing cost: qi 50 63 80 67 -26 -22 -13 -20
Service price: p
n
-5 -7 -10 -7 -6 -7 -8 -7
Service amenities: Si 15 20 25 20 14 17 23 20
Total WTP 95 143 233 200 19 57 137 113
Notes: This table summarizes the distribution of WTP of homeowners and renters, computed using
equation (3.29) across N = 4, 179 ZIP codes in 29 CBSAs with the first FC entry between 2010 and
2016. Each row represents the incremental WTP specific to the mentioned factor. Holding everything
else fixed at the pre-FC level (t = −1), each model component is progressively adjusted to its post-FC
(t = +3) value until the full model is implemented. Changes in model components are computed using
the following sources: ϕ from Panel A of Table 3.3, p
c and p
n from equation (3.26), qi from equation
(3.30), Ri and Si from equations (3.18).
128

Table 3.5: Welfare impact by pre-FC retail access time (2017 dollars)
Homeowner Renter
p25 p50 p75 mean p25 p50 p75 mean
Quintile (1) (2) (3) (4) (1) (2) (3) (4)
1 71 106 145 115 15 35 67 48
2 77 114 165 130 13 38 78 56
3 92 133 210 164 19 54 119 80
4 121 168 288 220 34 83 208 124
5 167 278 516 375 72 173 390 258
Notes: This table presents net dollar gains for homeowners and renters, seperated by
access time quintiles within cities. The most accessible group is Quintile 1, the least
accessible group is Quintile 5.
Table 3.6: Welfare impact by CBSA (2017 dollars)
Homeowner Renter
CBSA p25 p50 p75 mean p25 p50 p75 mean
New York-New Jersey-Long Island, NY-NJ-PA 232 323 488 407 133 196 342 276
Washington-Arlington-Alexandria, DC-VA-MD-WV 134 189 372 288 66 112 275 207
Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 168 239 333 272 125 186 269 216
San Francisco-Oakland-Fremont, CA 137 153 279 262 49 64 151 144
San Jose-Sunnyvale-Santa Clara, CA 135 159 229 250 77 92 153 169
Chicago-Joliet-Naperville, IL-IN-WI 128 165 231 200 37 63 121 98
Boston-Cambridge-Quincy, MA-NH 150 174 221 193 45 57 80 71
Seattle-Tacoma-Bellevue, WA 105 131 205 182 32 54 109 97
Minneapolis-St. Paul-Bloomington, MN-WI 87 113 237 176 20 33 135 94
Baltimore-Towson, MD 97 128 192 171 58 76 131 117
San Diego-Carlsbad-San Marcos, CA 83 124 204 165 29 46 104 90
Dallas-Fort Worth-Arlington, TX 88 111 180 160 49 62 126 116
Los Angeles-Long Beach-Santa Ana, CA 125 143 165 148 5 8 11 8
Hartford-West Hartford-East Hartford, CT 90 125 171 132 48 77 108 79
Tampa-St. Petersburg-Clearwater, FL 73 99 152 126 29 52 92 74
Providence-New Bedford-Fall River, RI-MA 89 119 148 120 35 59 78 61
Atlanta-Sandy Springs-Marietta, GA 78 98 135 117 16 23 62 46
Orlando-Kissimmee-Sanford, FL 70 86 118 101 30 38 62 53
Richmond, VA 66 91 133 101 14 24 48 37
Charlotte-Gastonia-Rock Hill, NC-SC 77 87 114 100 26 41 72 50
St. Louis, MO-IL 76 91 114 98 6 8 11 8
Kansas City, MO-KS 72 88 108 97 11 13 17 21
Virginia Beach-Norfolk-Newport News, VA-NC 57 68 93 93 10 12 20 31
Columbus, OH 67 92 117 93 25 38 57 45
Sacramento–Arden-Arcade–Roseville, CA 58 85 114 91 13 17 29 26
Riverside-San Bernardino-Ontario, CA 62 73 96 84 7 11 29 23
Nashville-Davidson–Murfreesboro–Franklin, TN 67 80 96 82 8 9 20 16
Austin-Round Rock-San Marcos, TX 64 71 93 81 6 7 9 10
San Antonio-New Braunfels, TX 47 58 83 75 7 9 14 16
Notes: This table presents net dollar gains for homeowners and renters, seperated by CBSA.
129

Bibliography
Ahlfeldt, Gabriel M, Stephen J Redding, Daniel M Sturm, and Nikolaus Wolf. 2015.
“The economics of density: Evidence from the Berlin Wall.” Econometrica, 83(6): 2127–2189.
Akerman, Anders, Ingvil Gaarder, and Magne Mogstad. 2015. “The skill complementarity
of broadband internet.” The Quarterly Journal of Economics, 130(4): 1781–1824.
Almagro, Milena, and Tomás Domínguez-Iino. 2022. “Location sorting and endogenous
amenities: Evidence from amsterdam.” Available at SSRN 4279562.
Amazon Help & Customer Service. 2023. “Amazon Prime.”
Amazon Press Room. 2009. “Amazon launches same day delivery in seven major cities and
expands Saturday delivery options.”
Atkin, David, Benjamin Faber, and Marco Gonzalez-Navarro. 2018. “Retail globalization
and household welfare: Evidence from mexico.” Journal of Political Economy, 126(1): 1–73.
Autor, David H, Christopher J Palmer, and Parag A Pathak. 2014. “Housing market
spillovers: Evidence from the end of rent control in Cambridge, Massachusetts.” Journal of
Political Economy, 122(3): 661–717.
Bajari, Patrick, and Matthew E Kahn. 2005. “Estimating housing demand with an application
to explaining racial segregation in cities.” Journal of business & economic statistics, 23(1): 20–
33.
Barron, Kyle, Edward Kung, and Davide Proserpio. 2021. “The effect of home-sharing on
house prices and rents: Evidence from Airbnb.” Marketing Science, 40(1): 23–47.
Basuroy, Suman, Yongseok Kim, and Davide Proserpio. 2020. “Estimating the impact of
Airbnb on the local economy: Evidence from the restaurant industry.” Available at SSRN
3516983.
Bauer, Anahid, and Sofía Fernández Guerrico. 2023. “Effects of e-commerce on local labor
markets.” IZA Discussion Paper.
130

Baugh, Brian, Itzhak Ben-David, and Hoonsuk Park. 2018. “Can taxes shape an industry?
Evidence from the implementation of the “Amazon tax”.” The Journal of Finance, 73(4): 1819–
1855.
Baum-Snow, Nathaniel. 2007. “Did highways cause suburbanization?” The quarterly journal of
economics, 122(2): 775–805.
Baum-Snow, Nathaniel, Loren Brandt, J Vernon Henderson, Matthew A Turner, and
Qinghua Zhang. 2017. “Roads, railroads, and decentralization of Chinese cities.” Review of
Economics and Statistics, 99(3): 435–448.
Bayer, Patrick, Fernando Ferreira, and Robert McMillan. 2007. “A unified framework
for measuring preferences for schools and neighborhoods.” Journal of political economy,
115(4): 588–638.
Bayer, Patrick, Robert McMillan, and Kim Rueben. 2004. “An equilibrium model of sorting
in an urban housing market.”
Becker, Gary S. 1968. “Crime and punishment: An economic approach.” In The economic dimensions of crime. 13–68. Springer.
Berry, Steven T. 1994. “Estimating discrete-choice models of product differentiation.” The RAND
Journal of Economics, 242–262.
Berry, Steven T, James A Levinsohn, and Ariel Pakes. 1995. “Automobile prices in market
equilibrium.”
Bloom, Nicholas, Raffaella Sadun, and John Van Reenen. 2012. “Americans do IT better:
US multinationals and the productivity miracle.” American Economic Review, 102(1): 167–201.
Brueckner, Jan, Matthew E Kahn, and Gary C Lin. 2021. “A new spatial hedonic equilibrium
in the emerging work-from-home economy?” National Bureau of Economic Research.
Brummet, Quentin, and Davin Reed. 2019. “The effects of gentrification on the well-being and
opportunity of original resident adults and children.”
Brynjolfsson, Erik, and Adam Saunders. 2010. “Wired for innovation.” How Information technology in reshaping the economy. Massachusetts Institute of Technology. USA.
Brynjolfsson, Erik, Yu Hu, and Michael D Smith. 2003. “Consumer surplus in the digital
economy: Estimating the value of increased product variety at online booksellers.” Management
science, 49(11): 1580–1596.
Cafcas, Thomas, and Greg LeRoy. 2016. “Will Amazon Fool Us Twice? Why State and Local
Governments Should Stop Subsidizing the Online Giant’s Growing Distribution Network.”
Good Jobs First.
131

Calder-Wang, Sophie. 2019. “The distributional impact of the sharing economy on the housing
market.” Job Market Paper, Joint Center for Housing Studies of Harvard University, Cambridge, MD.
Callaway, Brantly, and Pedro HC Sant’Anna. 2021. “Difference-in-differences with multiple
time periods.” Journal of econometrics, 225(2): 200–230.
Chava, Sudheer, Alexander Oettl, Manpreet Singh, and Linghang Zeng. 2022. “Creative
Destruction? Impact of E-Commerce on the Retail Sector.” National Bureau of Economic
Research.
Christian, Paul, and Christopher B Barrett. 2017. “Revisiting the effect of food aid on conflict:
A methodological caution.” World Bank Policy Research Working Paper.
Cohen, Peter, Robert Hahn, Jonathan Hall, Steven Levitt, and Robert Metcalfe. 2016.
“Using big data to estimate consumer surplus: The case of uber.” National Bureau of Economic
Research.
Conlon, Christopher, and Jeff Gortmaker. 2020. “Best practices for differentiated products
demand estimation with pyblp.” The RAND Journal of Economics, 51(4): 1108–1161.
Couture, Victor. 2016. “Valuing the consumption benefits of urban density.” University of California, Berkeley, Working Paper.
Couture, Victor, Benjamin Faber, Yizhen Gu, and Lizhi Liu. 2021. “Connecting the countryside via e-commerce: evidence from China.” American Economic Review: Insights, 3(1): 35–
50.
Couture, Victor, Cecile Gaubert, Jessie Handbury, and Erik Hurst. 2019. “Income growth
and the distributional effects of urban spatial sorting.” National Bureau of Economic Research.
Davis, Morris A, and François Ortalo-Magné. 2011. “Household expenditures, wages, rents.”
Review of Economic Dynamics, 14(2): 248–261.
De Chaisemartin, Clément, and Xavier d’Haultfoeuille. 2020. “Two-way fixed effects estimators with heterogeneous treatment effects.” American Economic Review, 110(9): 2964–2996.
De Chaisemartin, Clément, and Xavier d’Haultfoeuille. 2023. “Two-way fixed effects and
differences-in-differences with heterogeneous treatment effects: A survey.” The Econometrics
Journal, 26(3): C1–C30.
Delventhal, Matt, and Andrii Parkhomenko. 2020. “Spatial implications of telecommuting.”
Available at SSRN 3746555.
Delventhal, Matthew J, Eunjee Kwon, and Andrii Parkhomenko. 2022. “JUE Insight: How
do cities change when we work from home?” Journal of Urban Economics, 127: 103331.
132

Ding, Lei, Jackelyn Hwang, and Eileen Divringi. 2016. “Gentrification and residential mobility
in Philadelphia.” Regional science and urban economics, 61: 38–51.
Dolfen, Paul, Liran Einav, Peter J Klenow, Benjamin Klopack, Jonathan D Levin,
Larry Levin, and Wayne Best. 2023. “Assessing the gains from e-commerce.” American
Economic Journal: Macroeconomics, 15(1): 342–370.
Dolfen, Paul, Liran Einav, Peter J Klenow, Benjamin Klopack, Jonathan D Levin,
Laurence Levin, and Wayne Best. 2019. “Assessing the gains from e-commerce.” National
Bureau of Economic Research.
Dong, Xiaofang, Siqi Zheng, and Matthew E Kahn. 2020. “The role of transportation speed
in facilitating high skilled teamwork across cities.” Journal of Urban Economics, 115: 103212.
Draca, Mirko, Raffaella Sadun, and John Van Reenen. 2009. “Productivity and ICTs: A
review of the evidence.”
Dube, Arindrajit, Daniele Girardi, Oscar Jorda, and Alan M Taylor. 2023. “A local
projections approach to difference-in-differences event studies.” National Bureau of Economic
Research.
Duso, Tomaso, Claus Michelsen, Maximilian Schäfer, and Kevin Tran. 2020. “Airbnb and
Rents: Evidence from Berlin.” Working paper.
E-commerce Guide. 2020. “Top 10 Ecommerce Sites in the USA.”
Eichengreen, Barry, Romain Lafarguette, and Arnaud Mehl. 2016. “Cables, sharks and
servers: Technology and the geography of the foreign exchange market.” National Bureau of
Economic Research.
Ellen, Ingrid Gould, Keren Mertens Horn, and Davin Reed. 2019. “Has falling crime invited
gentrification?” Journal of Housing Economics, 46: 101636.
Figlio, David N, Paola Giuliano, Riccardo Marchingiglio, Umut Özek, and Paola
Sapienza. 2021. “Diversity in Schools: Immigrants and the Educational Performance of US
Born Students.” National Bureau of Economic Research.
Filippas, Apostolos, and John J Horton. 2020. “The tragedy of your upstairs neighbors: The
externalities of home-sharing platforms.” Working Paper.
Fontana, Nicola. 2020. “Backlash against Airbnb: Evidence from London.” mimeo LSE.
Forman, Chris, Avi Goldfarb, and Shane Greenstein. 2012. “The Internet and local wages:
A puzzle.” American Economic Review, 102(1): 556–575.
Fradkin, Andrey. 2015. “Search frictions and the design of online marketplaces.”
133

Fradkin, Andrey, Elena Grewal, and David Holtz. 2021. “Reciprocity and Unveiling in Twosided Reputation Systems: Evidence from an Experiment on Airbnb.” Marketing Science.
Garate, S, A Pennington-Cross, and W Zhao. 2020. “The Effect of the Shared Economy on
Crime: Evidence from Airbnb.”
Garcia-López, Miquel-Àngel, Jordi Jofre-Monseny, Rodrigo Martínez-Mazza, and Mariona Segú. 2020. “Do short-term rental platforms affect housing markets? Evidence from
Airbnb in Barcelona.” Journal of Urban Economics, 119: 103278.
Gaspar, Jess, and Edward L Glaeser. 1998. “Information technology and the future of cities.”
Journal of urban economics, 43(1): 136–156.
Glaeser, Edward L, and Matthew E Kahn. 2001. “Decentralized employment and the transformation of the American city.”
Glaeser, Edward L, Jed Kolko, and Albert Saiz. 2001. “Consumer city.” Journal of economic
geography, 1(1): 27–50.
Glaeser, Edward L, Michael Luca, and Erica Moszkowski. 2020. “Gentrification and Neighborhood Change: Evidence from Yelp.”
Goldfarb, Avi, and Catherine Tucker. 2019. “Digital economics.” Journal of economic literature,
57(1): 3–43.
Goldsmith-Pinkham, Paul, Isaac Sorkin, and Henry Swift. 2020. “Bartik instruments:
What, when, why, and how.” American Economic Review, 110(8): 2586–2624.
Goolsbee, Austan, and Peter J Klenow. 2006. “Valuing consumer products by the time spent
using them: An application to the Internet.” American Economic Review, 96(2): 108–113.
Gorback, Caitlin. 2020. “Your uber has arrived: Ridesharing and the redistribution of economic
activity.” Job Market Paper.
Guerrieri, Veronica, Daniel Hartley, and Erik Hurst. 2013. “Endogenous gentrification and
housing price dynamics.” Journal of Public Economics, 100: 45–60.
Holian, Matthew J, and Matthew E Kahn. 2013. “The rise of the low carbon consumer city.”
National Bureau of Economic Research.
Houde, Jean-François, Peter Newberry, and Katja Seim. 2022. “Nexus Tax Laws and
Economies of Density in E-Commerce: A Study of Amazon’s Fulfillment Center Network.”
NBER Working Paper.
InsideAirbnb.com. 2021. “Inside airbnb data dictionary.”
134

Jain, Shomik, Davide Proserpio, Giovanni Quattrone, and Daniele Quercia. 2021. “Nowcasting gentrification using Airbnb data.” Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1): 1–21.
Jo, Yoon J, Misaki Matsumura, and David E Weinstein. 2019. “The impact of e-commerce
on relative prices and consumer welfare.” National Bureau of Economic Research.
Ke, Laiyang, Daniel T. O’Brien, and Babak Heydari. 2021. “Airbnb and neighborhood crime:
The incursion of tourists or the erosion of local social dynamics?” PLoS one, 16(7): e0253315.
Kim, Eugene. 2019. “Amazon can already ship to 72% of US population within a day, this map
shows.”
King, A Thomas, and Peter Mieszkowski. 1973. “Racial discrimination, segregation, and the
price of housing.” Journal of political economy, 81(3): 590–606.
Koster, Hans RA, Jos van Ommeren, and Nicolas Volkhausen. 2021. “Short-term rentals
and the housing market: Quasi-experimental evidence from Airbnb in Los Angeles.” Journal
of Urban Economics, 124: 103356.
La Roche, Julia. 2019. “59% of US households are Amazon prime members, according to analyst.”
Lee, Dayne. 2016. “How Airbnb short-term rentals exacerbate Los Angeles’s affordable housing
crisis: Analysis and policy recommendations.” Harv. L. & Pol’y Rev., 10: 229.
Lipton, Alex. 2014. “How to Sublet Without Breaking the Law.”
Mason, Carl, and John M Quigley. 2006. “Program on Housing and Urban Policy.” Berkeley
Program on Housing and Urban Policy.
McDonald, Scott C. 1986. “Does gentrification affect crime rates?” Crime and Justice, 8: 163–201.
Mitra, Somjita, Rafael De Anda, and Kim Ritter-Martinez. 2017. “Airbnb in Los Angeles:
An Economic Impact Analysis.”
Miyauchi, Yuhei, Kentaro Nakajima, and Stephen J Redding. 2021. “Consumption access
and the spatial concentration of economic activity: evidence from smartphone data.” NBER
Working Paper.
Moretti, Enrico. 2013. “Real wage inequality.” American Economic Journal: Applied Economics,
5(1): 65–103.
Newman, Sandra J, and C Scott Holupka. 2014. “Housing affordability and investments in
children.” Journal of Housing Economics, 24: 89–100.
Newman, Sandra J, and C Scott Holupka. 2015. “Housing affordability and child well-being.”
Housing Policy Debate, 25(1): 116–151.
135

Orfield, Gary. 2001. “Schools more separate: Consequences of a decade of resegregation.”
Pennington, Kate. 2021. “Does Building New Housing Cause Displacement?: The Supply and
Demand Effects of Construction in San Francisco.” The Supply and Demand Effects of Construction in San Francisco (June 15, 2021).
Qian, Franklin, and Rose Tan. 2021. The effects of high-skilled firm entry on incumbent residents. Stanford Institute for Economic Policy Research (SIEPR).
Repko, Melissa. 2022. “Walmart is using its thousands of stores to battle Amazon for e-commerce
market share.”
Roth, Jonathan, Pedro HC Sant’Anna, Alyssa Bilinski, and John Poe. 2023. “What’s
trending in difference-in-differences? A synthesis of the recent econometrics literature.” Journal
of Econometrics.
Ryan, Chris. 1993. “Crime, violence, terrorism and tourism: an accidental or intrinsic relationship?” Tourism Management, 14(3): 173–183.
Sampson, Robert J, Stephen W Raudenbush, and Felton Earls. 1997. “Neighborhoods and
violent crime: A multilevel study of collective efficacy.” science, 277(5328): 918–924.
Savage, Scott James, and Donald M Waldman. 2009. “Ability, location and household demand
for Internet bandwidth.” International Journal of Industrial Organization, 27(2): 166–174.
Schmidheiny, Kurt, and Sebastian Siegloch. 2019. “On event study designs and distributed-lag
models: Equivalence, generalization and practical implications.” CESifo Working Paper.
Shendruk, Amanda. 2022. “Amazons’s $5 billion discount: See all its tax cuts and other US
subsidies.”
Tan, Brandon. 2020. “Urban Transit Infrastructure and Inequality: The Role of Access to NonTradable Goods and Services.” Available at SSRN 3750438.
Tsivanidis, Nick. 2019. “Evaluating the impact of urban transit infrastructure: Evidence from
bogota’s transmilenio.” Unpublished manuscript.
Varadhan, Ravi, and Christophe Roland. 2008. “Simple and globally convergent methods
for accelerating the convergence of any EM algorithm.” Scandinavian Journal of Statistics,
35(2): 335–353.
Volpe, Richard, and Michael A Boland. 2022. “The Economic Impacts of Walmart Supercenters.” Annual Review of Resource Economics, 14.
Wachsmuth, David. 2021. “Short-term rentals in Los Angeles: Are the city’s regulations working?:
Upgo-mcgill.”
136

Wachsmuth, David, and Alexander Weisler. 2018. “Airbnb and the rent gap: Gentrification
through the sharing economy.” Environment and Planning A: Economy and Space, 50(6): 1147–
1170.
Waldfogel, Joel. 2008. “The median voter and the median consumer: Local private goods and
population composition.” Journal of urban Economics, 63(2): 567–582.
Wooldridge, Jeffrey M. 2021. “Two-way fixed effects, the two-way mundlak regression, and
difference-in-differences estimators.” Available at SSRN 3906345.
World Bank. 2023. “Urban Development Overview.”
World Economic Forum. 2023. “Digital Transition Framework 2023.”
Zahniser, David. 2022. “Nearly 1 in 5 Airbnb listings in L.A. violated city law, advocacy group
says.” Los Angeles Times.
137

Appendix A
Appendix to “Buy Now with 1-Click: Spatial
Impacts of E-commerce”
A.1 Figures for additional robustness checks
138

Figure A.1: Robustness check: Alternative retail accessibility measure
-.1 -.05
0 .05 .1 .15
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.2 -.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) Restaurants
-1 -.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Retail Establishments
Notes: This figure plots estimates for the dynamic treatment effects β
g
d
in equation (3.5), where the dependent variables are
normalized Zillow Home Value Index, Zillow Rent Index, restaurant count, and retail establishment counts. The alternative
accessibility measure uses the number of retail establishment to weight travel times instead of employment count. Less accessible
refers to ZIP codes with average pre-period retail access time higher than city’s median, whereas More accessible consists of
ZIP codes with average pre-period retail access time in the city’s fastest quartile. Omitted baseline group is the second fastest
quartile. Event years beyond the plotted time window are binned together with the earliest lead or latest lag. 95% confidence
intervals are plotted, where robust standard errors are clustered at the service area level.
139

Figure A.2: Robustness check: Placebo test with physician accessibility measure
-.1 -.05
0 .05 .1
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.2 -.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) Restaurants
-1 -.5
0 .5 1
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Retail Establishments
Notes: This figure plots estimates for the dynamic treatment effects β
g
d
in equation (3.5), where the dependent variables
are normalized Zillow Home Value Index, Zillow Rent Index, restaurant count, and retail establishment counts. The placebo
accessibility measure uses the number of physician offices to weight travel times. Less accessible refers to ZIP codes with average
pre-period physician office access time higher than city’s median, whereas More accessible consists of ZIP codes with average
pre-period physician office access time in the city’s fastest quartile. Omitted baseline group is the second fastest quartile. Event
years beyond the plotted time window are binned together with the earliest lead or latest lag. 95% confidence intervals are
plotted, where robust standard errors are clustered at the service area level.
140

Figure A.3: Robustness check: MSA fixed effects
-.05
0 .05 .1 .15
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
-.5
0 .5 1
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) Restaurants
-1
0 1 2
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Retail Establishments
Notes: This figure plots robustness estimates for the dynamic treatment effects β
g
d
in equation (3.5), where the service area-level
fixed effects are replaced by MSA level fixed effects. Dependent variables are the normalized Zillow Home Value Index, Zillow
Rental Index, restaurant count, and retail establishment count. Less accessible refers to ZIP codes with average pre-period
retail access time higher than city’s median, whereas More accessible consists of ZIP codes with average pre-period physician
office access time in the city’s fastest quartile. Omitted baseline group is the second fastest quartile. Event years beyond the
plotted time window are binned together with the earliest lead or latest lag. 95% confidence intervals are plotted, where robust
standard errors are clustered at the service area level.
141

Figure A.4: Robustness check: MSA linear time trends
-.05
0 .05 .1 .15
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.1
0 .1 .2
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) Restaurants
-1
0 1 2
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Retail Establishments
Notes: This figure plots robustness estimates for the dynamic treatment effects β
g
d
in equation (3.5) with additional MSA
linear trends, where the dependent variables are the normalized Zillow Home Value Index, Zillow Rental Index, restaurant
count, and retail establishment count. Less accessible refers to ZIP codes with average pre-period retail access time higher than
city’s median, whereas More accessible consists of ZIP codes with average pre-period physician office access time in the city’s
fastest quartile. Omitted baseline group is the second fastest quartile. Event years beyond the plotted time window are binned
together with the earliest lead or latest lag. 95% confidence intervals are plotted, where robust standard errors are clustered at
the service area level.
142

Figure A.5: Robustness check: ZIP code linear time trends
-.1
0 .1 .2 .3
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(a) Home Value Index
-.2
0 .2 .4 .6
-4 -2 0 2 4 6 8
Years Relative to Entry
Less accessible More accessible
(b) Rent Index
-.5
0 .5 1 1.5
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(c) Restaurants
-2 -1
0 1 2
-4 -2 0 2 4 6
Years Relative to Entry
Less accessible More accessible
(d) Retail Establishments
Notes: This figure plots robustness estimates for the dynamic treatment effects β
g
d
in equation (3.5) with additional ZIP code
linear trends, where the dependent variables are the normalized Zillow Home Value Index, Zillow Rental Index, restaurant
count, and retail establishment count. Less accessible refers to ZIP codes with average pre-period retail access time higher than
city’s median, whereas More accessible consists of ZIP codes with average pre-period physician office access time in the city’s
fastest quartile. Omitted baseline group is the second fastest quartile. Event years beyond the plotted time window are binned
together with the earliest lead or latest lag. 95% confidence intervals are plotted, where robust standard errors are clustered at
the service area level.
143

A.2 Background on Amazon fulfillment centers
In its early days, Amazon’s fulfillment network featured several central locations, with its first
warehouses opened in 1997 in Washington state and Delaware. Over the years, especially since
2005, Amazon has engaged in aggressive expansion strategy of the logistic network in order to fulfill
its ambitious promise of two-day delivery to consumers. Although several factors such as access
to urban consumer base and access to transportation infrastructures do impact where and when
Amazon may locate their FC facilities, recent work by Houde, Newberry and Seim (2022) notes
local nexus tax laws as an important determinant of Amazon’s location decision.
To provide some context, the U.S. Supreme Court’s rule on the 1992 case of Quill v. North Dakota
stated a state cannot collect sales tax from online retailers unless the retailers have a physical
presence, or “nexus,” in the state. This ruling effectively created a price advantage for e-commerce
sellers who are located out of state. As a result, Amazon in the early years set up warehouse
facilities in jurisdictions with little to no sales tax, such as Delaware and New Hampshire, yet still
in close proximity to major urban markets in other states. As state governments’ become concerned
about lost tax dollars, many as early as 2007 have revised the definition of “nexus,” from having
in-state physical presence to the some equivalence of having “constitutionally sufficient” business
transactions with the states. These nexus laws created incentives for Amazon to decentralize its
network to lower shipping costs as a cost-saving strategy in response to no longer having a price
advantage. As the company expands its fulfillment network to the fringe of many cities, consumers
in these areas enjoy the benefits of fast delivery, including one-day and same-day delivery in addition
to the premium two-day shipping standard, as well as the option to receive packages in the weekend
(Amazon Press Room 2009).
144

The hypothesized correlation between nexus tax laws and FC entry can raise alternative interpretations of my results. For instance, FC entry may raise local tax revenue and result in local
governments investing more in revitalizing less accessible areas, making these neighborhoods more
attractive and thus explaining the observed increase in housing values, rents, and service amenities.
On this account, it is worth noting that, in many cases, the enactment of nexus laws typically
precedes an FC entry. In some cases, Amazon voluntarily collects sales tax prior to FC entry in
exchange for other business incentives from local governments (Cafcas and LeRoy 2016, Houde,
Newberry and Seim 2022). If increased place-based government spending is the main driver for the
post-FC real estate and economic dynamics observed in less accessible areas, we should at least
have observed some of these impacts prior to the arrival of FC. The fact that I do not observe
any significant pre-trends suggest that it is unlikely that my results are rationalized by increased
government spending.
A.3 Handling routes with null travel times
As previously mentioned in section 3.2, I use the U.S. map snapshot on January 10, 2010 with
the OSRM tool to query historical travel times between ZIP code pairs. Restricting to all possible
pairwise permutation between ZIP codes that are at most 100 miles apart within each of the 360
MSAs, I query a total of 2.8 million origin-destination pairs. Although most the routes are identified,
71,366 queries (or 2.5%) are returned with nulled results. For these cases, the OSRM tool is not
able to provide a routing itinerary and thus an estimated driving time given the input coordinates.
I proceed to impute the missing driving time using the following regression with complete queried
data:
travelT imezz′ = const + βstraightlineDistancezz′ + δz + δz
′ + ϵ
145

where straightlineDistancezz′ is the straight-line distance between origin ZIP code z and destination ZIP code z
′
, δz is the origin ZIP code fixed effect, and δz
′ is the destination ZIP code fixed
effect.
A.4 Proof for Proposition 1
Proof for Proposition 1:
Given Cobb-Douglas utility structure, the amount income spent on retail consumption Cij is αCIi
.
Hence, the optimal demands for offline- and online-purchased goods (cj , ce
) is the solutionn to the
following problem:
max
cj ,ce|θ

θ
β
ij
t
C
j
cj
σ−1
σ +
(1 − θij )
β
ϕ
ce
σ−1
σ
σ
σ−1
s.t. p
c
j
cj + p
e
c
e ≤ αCIi (A.1)
Solving for the first-order conditions gives
c
∗
j =
αCIi(θ
β
ij t
c−1
ij )
σ−1
p
σ
j P
C1−σ
j
c
e∗ =
αCIi
[(1 − θij )
βϕ
−1
]
σ−1
p
σ
e PC1−σ
e
(A.2)
Substituting optimal demands to equation (3.7) and simplifying give Cij (θ) = αCIi
P
C
j
(θ)
. Finding the
optimal θ means that we maximize αCIi
P
C
j
(θ)
. Note that P
C
j
(θ) is strictly positive, so P
C
j
(θ) increases
(decreases) and 1
P
C
j
(θ)
decreases (increases) in θ. Hence, the maximization problem is equivalent to
min
θ|ij
P
C
ij (θ) (A.3)
146

First order condition implies:
P
C′
ij (θ) = −P
Cσ

θ
β(σ−1)−1
(pj t
c
ij )
σ−1
−
(1 − θ)
β(σ−1)=1
(p
eϕ)
σ−1

= 0
⇐⇒
θ
β(σ−1)−1
(pj t
c
ij )
σ−1
−
(1 − θ)
β(σ−1)=1
(p
eϕ)
σ−1
= 0
⇐⇒
θ
1 − θ
β(σ−1)−1
=

pj t
c
ij
p
eϕ
σ−1
⇐⇒ θ
∗
ij =
1
1 + p
c
j
p
e
t
cij
ϕ
σ−1
1−β(σ−1)
θ
∗
ij is optimal if the second condition holds, i.e. P
C′′
ij (θ) > 0:
P
C′′
ij (θ) = −
σ
1 − σ
P
C2(σ−1)
ij
θ
β(σ−1)−1
(pj t
c
ij )
σ−1
−
(1 − θ)
β(σ−1)=1
(p
eϕ)
σ−1
2
− P
Cσ
ij (β(σ − 1) − 1)
θ
β(σ−1)−2
(pj t
c
ij )
σ−1
+
(1 − θ)
β(σ−1)−2
(p
eϕ)
σ−1

> 0
The first term is positive since σ > 1, the second is positive if β(σ −1)−1 < 0 ⇐⇒ β < 1
σ−1 QED.
A.5 Imputing user cost for homeowners
In section 7, I do not explicitly incorporate a housing sector in the model when analyzing the welfare
consequences of e-commerce. Instead, I leverage the reduced-form findings in that e-commerce entry
differentially impacts home values and rents depending on neighborhood access to pre-existing
retail amenities. For renters, I predict the e-commerce-induced outcomes for qi using the following
regression:
log qi = γ0 + γ1P ostF Ckt × log(retailAccessi) + δi + δhkt + νikt
147

where qi
is the dollar-nominated ZRI. To compute housing costs for homeowners, I first predict
changes in home values using the ZHVI, and impute the implicit rents using the following formula:
q
ho
it = (r + δ + µ)ZHV I \i,2010 − ηZHV I \it
where r is the interest rate, δ is the depreciation cost, µ is the maintenance cost, and η is the
expected net nominal capital gain. ZHV I \i,2010 is the predicted home value in neighborhood i in
2010, which is before all FC entry in the restricted sample, and ZHV I \it is the predicted home
value in year t. Intuitively, the effective cost of an individual to own a home in year t is their
forgone opportunity cost rental rate r plus depreciation and maintenance cost of the home less their
expected capital gains in that year. Note that if η > 0, an increase in home value will translate
into lower user cost for homeowners. Following Gorback (2020), I set r+δ+µ = 0.088 and η = 0.024.
148

Abstract (if available)

Abstract This dissertation studies the effects of digital transformation on cities and neighborhoods, as well as their welfare and distributional implications. The author focuses on two of the fastest growing technologies in the past decade: short-term rentals and e-commerce, which have prompted many critical questions of both academic and policy interest. The first chapter motivates the research questions underlying the author’s work and summarizes the main findings. The second chapter examines the impacts of short-term rentals in Los Angeles County, while the last chapter investigates the spatial implications of e-commerce expansion across cities in the United States.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on urban and real estate economics

PDF

Location choice and the costs of climate change

PDF

Essays on the economics of climate change adaptation in developing countries

PDF

Essays on the dual urban-rural system and economic development in China

PDF

The impact of minimum wage on labor market dynamics in Germany

PDF

Essays on applied microeconomics

PDF

Sustaining open source software production: an empirical analysis through the lens of microeconomics

PDF

Exploration of human microbiome through metagenomic analysis and computational algorithms

PDF

Essays on wellbeing disparities in the United States and their social determinants

PDF

Taking the temperature of the Columbia Card Task

PDF

Essays in climate change adaptation: role of market power in incentivizing adaptation behavior

PDF

The effects of fast walking, biofeedback, and cognitive impairment on post-stroke gait

PDF

Migration, location, and economic opportunity

PDF

More than a game: understanding the value of funding that corporate partnership decision-makers can offer clubs within the National Women’s Soccer League

PDF

Osseous and integuemental cephalometric study of the Puerto Rican ethnic group utilizing public opinion in the concept of esthetics

PDF

On the quest for global competencies

PDF

Urban consumer amenities and their accessibility

PDF

Aggregation and the structure of value

PDF

The impact of mobility and government rental subsidies on the welfare of households and affordability of markets

PDF

Essays on the economics of cities

Asset Metadata

Creator Hoang, Tri (author)

Core Title The impact of digital transformation on urban communities, welfare, and distributional outcomes

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2024-05

Publication Date 03/29/2024

Defense Date 03/22/2024

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag Airbnb,Amazon,digital economics,e-commerce,economics,OAI-PMH Harvest,short-term rentals,urban economics

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kahn, Matthew E. (committee chair), Green, Richard K. (committee member), Oliva, Paulina (committee member)

Creator Email ngocminh@usc.edu,nmtrihoang@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113862155

Unique identifier UC113862155

Identifier etd-HoangTri-12734.pdf (filename)

Legacy Identifier etd-HoangTri-12734

Document Type Dissertation

Format theses (aat)

Rights Hoang, Tri

Internet Media Type application/pdf

Type texts

Source 20240401-usctheses-batch-1133 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu