University of Southern California Dissertations and Theses

Insights into residential mobility and pricing of rental housing: the role of gentrification, home-ownership barriers, and market concentrations in low-income household welfare

(USC Thesis Other)

Insights into residential mobility and pricing of rental housing: the role of gentrification, home-ownership barriers, and market concentrations in low-income household welfare

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Insights into Residential Mobility and Pricing
of Rental Housing
The Role of Gentrication, Home-Ownership Barriers, and Market
Concentrations in Low-Income Household Welfare
by
Evgeny Burinskiy
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
Urban Planning and Development
August 2021
Copyright 2021 Evgeny Burinskiy
Dedication
I dedicate this thesis to all of the practitioners and researchers who aim to make our world a
more equitable and fair society. We may not be able to single-handedly solve all of the world's prob-
lems, but in leveraging out comparative advantages, perhaps we can one day erase the distinctions
professed by luck and privilege.
ii
Acknowledgements
I want to thank my advisor Richard Green and committee members Jorge De la Roca and
Marlon Boarnet for their continued support, mentorship, and advice throughout the PhD program.
Your ideas, projects, shared resources, and ceaseless prodding to aim higher, especially from Jorge,
enabled me to fulll my aspiration of performing impactful research.
To the entire cohort above me, including Julia Harten, Carmen Mooradian, Hue-Tam Jamme,
and countless others, thank you for the friendship and laughter that allowed me to gain my footing
in Los Angeles and survive the initial years.
I want to thank my colleagues and peers, especially Linna Zhu and Sahil Gandhi, for pushing
me, for your friendship, and for creating a forum in which we can bounce ideas. You are the giants
whose footsteps I followed and shoulder I sat on.
To Adam Roeder, Calum Rickard, Mark Herms, Kylie Trettner, the entire cycling team board,
and countless others from the triathlon and cycling teams, nding you was the biggest blessing in
Los Angeles. The ability to continue pretending to be an athlete despite the concrete jungle of Los
Angeles was crucial in the maintenance of my sanity and identity.
I want to thank the Lusk Center for Real Estate Research for their nancial support of my
research.
Finally, I want to thank my parents and sister for the love and support they provided throughout
the program.
iii
Contents
Dedication ii
Acknowledgements iii
List of Tables vi
List of Figures vii
Abstract viii
Chapter 1 1
1 Introduction 2
2 Prior Literature 3
3 Methodology 4
3.1 Neighborhood
ow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Statistical Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3 Dening Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.4 Treatment Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4 Data 9
4.1 Identifying Movers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2 Treatment and Control Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3 Income Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.4 Station Inclusion Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5 Results 17
5.1 Income Distribution Descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.2 Mobility Rate Descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.3 Parallel Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4 Impact on mobility rates across the income distribution . . . . . . . . . . . . . . . . 24
5.5 Robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.6 Incumbent Households . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.7 Long-term impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.8 Impact of mobility changes on income distribution . . . . . . . . . . . . . . . . . . . 34
6 Conclusion & Discussion 35
Chapter 2 38
1 Introduction 39
2 Motivation 40
iv
3 Housing Search Frictions 42
4 Method 43
4.1 Herndahl-Hirschmann Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Rents, Market Concentration, and Instruments . . . . . . . . . . . . . . . . . . . . . 44
4.3 Proxy Frictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5 Data 45
6 Results 51
6.1 Rent extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Impact estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7 Discussion 54
7.1 Milwaukee is not unique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
7.2 Policy Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Chapter 3 56
1 Introduction 57
2 Related Literature 59
3 Theory 60
4 Methodology 61
4.1 Income-to-rent slopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 Access to Ownership . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5 Data 62
6 Results 65
6.1 Role of Access to Ownership Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.2 Top and bottom results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3 Robustness Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
7 Discussion and Conclusion 69
References 70
Appendix 75
v
List of Tables
1 Station Summary Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Number of years households are observed for . . . . . . . . . . . . . . . . . . . . . . 11
3 Observable means between control and treatment groups rounded to 0 decimals . . . 13
4 Deviations from LTM income . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5 Stations dropped due to small populations (< 100) . . . . . . . . . . . . . . . . . . . 16
6 Test on income distribution trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
7 Placebo tests on rst and second lags on 5+ year households . . . . . . . . . . . . . 24
8 Base DID Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
9 DID Robustness Results with Sub-Sample Variation . . . . . . . . . . . . . . . . . . 30
10 DID Robustness Results with Housing Variables . . . . . . . . . . . . . . . . . . . . 31
11 DiD Results on Incumbent Households . . . . . . . . . . . . . . . . . . . . . . . . . . 32
12 Impact of mobility changes on income distribution . . . . . . . . . . . . . . . . . . . 35
1 Estimated Mark-Ups by Neighborhood Income . . . . . . . . . . . . . . . . . . . . . 40
2 Income and online search in AHS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3 Income and method of apartment search (National) . . . . . . . . . . . . . . . . . . . 43
4 Income and method of apartment search (Milwaukee - AHS) . . . . . . . . . . . . . 43
5 Sample Means with Co-Star Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6 Relevance Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
7 Exogeneity Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8 Outside ownership shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
9 OlS Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
10 IV Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
11 Rent to income slopes for 25 MSAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
12 Variable Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
13 Variable Summary Statistics for Estimation Sub-sample . . . . . . . . . . . . . . . . 64
14 First Stage on denials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
15 Estimate results with denial rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
16 Results on top and bottom 25% of renter income subsample . . . . . . . . . . . . . . 67
17 Estimate results with downpayment to income ratio . . . . . . . . . . . . . . . . . . 68
18 Observable characteristic medians by treatment group . . . . . . . . . . . . . . . . . 76
19 annual AMI Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
20 Annual number of households in treated basin . . . . . . . . . . . . . . . . . . . . . . 78
21 Parallel trend test estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
22 Parallel trend test estimates by income . . . . . . . . . . . . . . . . . . . . . . . . . . 81
23 Parallel trend test estimates on incumbent households . . . . . . . . . . . . . . . . . 82
24 Share of lers included . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
25 FTB Sub-sample and Census data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
26 Level-level rent regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
27 Log-log rent regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
vi
List of Figures
1 Annual share of last-observed households . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Number of households by treatment group . . . . . . . . . . . . . . . . . . . . . . . . 14
3 LTM Incomes for Dense parts of LA County . . . . . . . . . . . . . . . . . . . . . . . 18
4 LTM Incomes Near LA Metro Rail Transit . . . . . . . . . . . . . . . . . . . . . . . 18
5 Mean mobility rates for moves beyond stations by income . . . . . . . . . . . . . . . 21
6 Parallel trends by line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
7 Parallel trends by line for moves beyond stations . . . . . . . . . . . . . . . . . . . . 23
8 Income distribution of Expo line stations . . . . . . . . . . . . . . . . . . . . . . . . 26
9 Income distribution of Gold line 2003 stations . . . . . . . . . . . . . . . . . . . . . . 27
10 Income distribution of Gold line 2009 stations . . . . . . . . . . . . . . . . . . . . . . 27
11 Income distribution of Red line 1999 stations . . . . . . . . . . . . . . . . . . . . . . 27
12 Income distribution of Red line 2000 stations . . . . . . . . . . . . . . . . . . . . . . 27
13 DiD Results with Deep Lags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
14 DiD Results with Deep Lags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1 Milwaukee City Market Concentration Indexes . . . . . . . . . . . . . . . . . . . . . 46
2 Higher HHI, more recent transactions . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3 Estimated Impact on Lowest Income Neighborhoods . . . . . . . . . . . . . . . . . . 54
5 Histogram of demeaned 2-bedroom rents . . . . . . . . . . . . . . . . . . . . . . . . . 58
7 Map of Los Angeles Metro Rail Lines (2013) . . . . . . . . . . . . . . . . . . . . . . . 75
8 Mean mobility rates by income . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
9 Annual LTM income distribution in dense parts of LA County . . . . . . . . . . . . 83
10 Annual LTM Income Distribution by Treatment Group . . . . . . . . . . . . . . . . 84
11 Station-to-station mobility rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
vii
Abstract
This dissertation is a collection of three empirical essays in Urban and Housing Economics
that explore the factors behind residential security and housing rents among low-income renters.
The rst chapter examines the impact that the introduction of rail-transit has on households'
residential mobility and, in turn, the neighborhood income distribution. In essence, it answers
whether transit-induced neighborhood shocks cause neighborhoods to gentrify through an in
ux
of high income households and an out
ow of low-income residents. The chapter uses 20 years
of micro-data on household income and place of residence provided by the California Franchise
Tax Board to assess residential mobility
ows before and after a transit station opens in Los
Angeles County using a Dierence-in-Dierence approach. The study nds substantial variation
in impact across low-income households and transit lines. Residential patterns among extremely
low-income households suggest a preference for locating near public transit opportunities while
patterns among very low-income households suggest that positive neighborhood shocks may
reduce their ability to stay in transit neighborhoods. Residential
ows among low and higher-
income households do not seem to be impacted by the opening of transit stations. Although
the study nds that station-opening causes some changes in residential mobility patterns, these
shifts are largely drowned out by the same macro-factors that drive distributional shifts across
urban Los Angeles County. The second chapter proposes and empirically supports the theory
that the ability of households to become homeowners has an impact on rent distributions. The
study uses American Housing Survey and HDMA mortgage data along with an instrumented
variable approach to show that markets in which households have a more dicult time purchas-
ing a house have substantially wider rent distributions than markets in which access to ownership
is easier. Moreover, the impact is not uniform across renter households as high-income renters
are impacted substantially more than low-income renters. This suggests that policies that de-
crease barriers to ownership primarily benet higher-income renters. The third chapter shows
the existence of high markups in Milwaukee's low-income rental market and tests whether high
ownership concentrations enable landlords to maintain the markups. The study pairs ownership
records from Milwaukee's tax assessor database with rent and CoStar multi-family housing data
to implement an instrumented variable approach that tests the impact of ownership concen-
trations on rents. Estimates show that higher ownership concentrations raise rents and that
dierences in ownership concentrations between high and low income neighborhoods explain
about a third of landlords' ability to maintain high markups.
viii
Chapter 1
Insights into residential displacement: A micro-data analysis of household
moves near Los Angeles Metro rail stations
by
Evgeny Burinskiy, Seva Rodnyansky, Marlon G. Boarnet, Raphael Bostic, Allen Prohofsky
Hitherto, studies on gentrication and displacement have not had the requisite data to
directly asses the impact of neighborhood shocks on residential mobility rates and the neigh-
borhood income distribution. We address that gap through our partnership with the California
Franchise Tax Board. We leverage a novel panel dataset on tax-lers in Los Angeles (LA)
County between 1993 and 2015 to compute residential mobility rates across income distribu-
tions for households residing within .5 miles of LA Metro rail stations. By comparing mobility
rates between households residing inside the .5 mile radius against those just outside of it, we
test whether the opening of LA Metro rail stations coincides with mobility rate shifts consistent
with displacement patterns. Our assessment of income distributions near LA Metro stations
shows that income inequality and share of extremely-poor households has increased and that
these patterns are similar to those of LA County and do not statistically dier from those of
neighborhoods adjacent to transit basins. Our results on residential mobility shifts nd modest,
if heterogeneous, eects on extremely-poor and very-poor households. Along most lines, the
move-out rate among extremely-poor households either did not change or decreased which is
consistent with the hypothesis that low-income households prefer to locate near transit (Glaeser,
Kahn, and Rappaport 2008). On the other hand, where aected, very-poor household mobility
shifts show decreased in
ow which is consistent with the gentrication narrative. Mobility shifts
among poor, middle, and high income households were not aected by station-opening. Results
on households who resided near rail station areas prior to station-opening suggest incumbents
reduce the rate at which they move out of rail stations. Overall, we nd rail stations impact res-
idential mobility patterns but these changes do not seem to impact income distribution trends.
This suggests that rail-station opening can alter
ows but these
ow-changes are obscured by
the larger factors that drive increasing inequality throughout Los Angeles County. We note
that the event that we study, rail station openings, can bring a mix of housing price pressure
and increased transportation access that can have countervailing eects on residential mobility.
Variations by rail lines and across income levels indicate that the eect of station opening on
residential mobility is heterogenous across places and income levels.
The views expressed herein are those of the authors and do not necessarily re
ect those of
the Board of Governors of the Federal Reserve System or other System ocials.
Any opinions expressed in this report are those of the authors, not ocial positions of the
California Franchise Tax Board.
1
1 Introduction
Rail transit expansion, nearby increasing housing prices, and shifts in demographic and socioe-
conomic neighborhood compositions are linked in the public mind. Examples where rail transit
has been associated, at least anecdotally, with neighborhood gentrication abound. In Washing-
ton, D.C., the Green and Yellow lines are associated with neighborhood transition north and east
of downtown. In the Los Angeles (LA) area, the Gold, Expo, and Red/Purple lines have been
associated with gentrication concerns (Zuk and Chapple 2015), and similar concerns have been
raised regarding the soon-to-open Crenshaw Line. On balance, these same concerns are present
in most large metropolitan areas that are building or expanding rail transit (Ong, Zuk, Pech, and
Chapple 2017, Rayle 2015).
However, the use of aggregated data to study displacement and gentrication confounds the
mechanism behind neighborhood change and cannot clearly identify displacement. We address this
issue by using household-level panel data on income and residential locations between 1993 and
2015 to measure changes in income distributions and mobility rates near Los Angeles Metropolitan
Transit Authority (LA Metro) rail stations. For our key contribution, we exploit location and
income data to infer residential changes for households across dierent income brackets and run
a dierence-in-dierence (DiD) model to test whether the opening of rail stations coincides with
changes in mobility rates.
The Los Angeles metropolitan area presents an ideal study area for analyzing transit-oriented
development (TOD) and potential displacement. Prior to 1990, Los Angeles had not had any
intra-urban rail transit service for decades. Since then, 93 new rail-transit stations were opened by
the Los Angeles County Metropolitan Transportation Authority (L.A. Metro) and an additional 17
are currently under construction (Boarnet, Bostic, Evgeny Burinskiy, and Prohofsky 2018). This
buildout amounts to about half of the U.S. spending on new rail transit. Within L.A. Metro,
21% of its budget from 2005-2040 will go toward rail transit capital and operations expenditures
(Metro LA 2009)/ Concurrently, regional and local plans envision that over half of new housing and
employment to occur within a half-mile of a well-serviced transit corridor, including rail (Metro
LA 2009, SCAG 2012).
Our results suggest that the large investments in LA rail transit did impact mobility rates
among nearby low-income groups. We nd that the average annual mobility rate after stations
open either does not change or decreases among extremely-poor households (0-30% of Area Median
Income (AMI)) which is consistent with the ndings of Glaeser, Kahn, and Rappaport (2008) who
suggest that low-income households prefer to locate near public transit. On the other hand, the rate
at which very-poor households (30-50% of AMI) move into rail neighborhood decreased while the
move-out rate remained unchanged. This increased out
ow of very-poor households is consistent
with gentrication narratives. Households earning more than 50% did not seem to experience shifts
in their move-in and move-out rates.
Although the opening of rail stations impacts mobility rates, estimates on income distribution
trends suggest that the local transit-induced impact on mobility rates is drowned out by other
factors that drive increasing inequality in Los Angeles County. Between 1994 and 2014, we nd
that the income distribution of rail-transit neighborhoods follows similar trends to LA County and
that this trend is not statistically distinguishable from adjacent neighborhoods. All three areas saw
increased income inequality as the share of extremely-poor households grew while the share of high-
earning households (over 200% of AMI) largely remained constant across time. Since increasing
inequality is not unique to rail stations, shifts in the income-composition of rail-transit neighbor-
2
hoods are more likely due to larger macro-economic trends aecting the Los Angeles metropolitan
area than local transit-induced shocks.
Our study proceeds as follows. In section 2, we summarize prior ndings on the link between
neighborhood composition and rail transit access, and describe how ours is a novel contribution.
In section 4, we describe our framework for assessing neighborhood composition and show how we
link it to rail transit access. In section 5, we describe the California Franchise Tax Board data.
Section 6 contains our key results on mobility rates, income distributions, and their relationship to
rail transit access while section 6 describes the limitations of our study and future steps for studies
on residential displacement.
2 Prior Literature
Studies in urban planning, public policy, sociology, and other disciplines examined Census
and administrative data to see whether neighborhood shocks lead to changes in neighborhood
demographic characteristics (Zuk 2018, Zuk, Bierbuaum, Chapple, Gorska, Sideris, Ong, and
Thomas 2015). Some nd that neighborhoods aected by a positive shock showed signs of gentri-
cation such as increases in the share of white households, rents, and in
ows of higher educated
households (Baker and Lee 2017, Kahn 2007, GrubeCavers and Patterson 2015, Heilmann 2018,
Chapple 2009, Glaeser, Kahn, and Rappaport 2008), though others nd little to no evidence of gen-
trication (Barton and Gibbons 2018, Deka 2017, Dong 2017, Nilsson and Delmelle 2018, Pathak,
Wyczalkowski, and Huang 2017, Wang and Woo 2017). However, these studies cannot identify a
causal tie between neighborhood-level gentrication and shocks.
Our study focuses on the introduction of new rail transit service as the neighborhood shock and
resultant eects on residential mobility and neighborhood composition.
New transit station openings have been associated with increased housing prices, commercial
rents, and amenities which may lead to neighborhood gentrication (Bardaka, Delgado, and Florax
2018, Chapple 2009, Hess 2020, Heilmann 2018). These studies suggest that the opening of a new
transit station will lead to an in
ow of higher-income households attracted by the positive amenities
and an out
ow of out-priced low-income households.
In the U.S., many lower-income households rely on public transit for commuting and travel,
regardless of where they live. Glaeser, Kahn, and Rappaport (2008) and Kahn (2007) provide
evidence of low-income households' preference for locating near public transportation. This includes
two ndings about urban areas: a) low-income households tend to reside in neighborhoods that are
close to public transportation and b) the introduction of public transportation leads an increase in
the presence of low-income households. Moreover, Bereitschaft (2020) and Bardaka, Delgado, and
Florax (2018) show that even when incomes go up in some transit neighborhoods, especially near
downtown cores, lower-income households migrate toward more peripheral stations and still exhibit
high ridership. These studies suggest that the introduction of rail-transit need not necessarily lead
to an out
ow of low-income households.
This paper most closely aligns with studies that use micro-level data that track residential mo-
bility to understand gentrication and neighborhood change (Atkinson, 2000; Billings et al., 2018;
Ding et al., 2016; Freeman, et al., 2016; Freeman, 2005; Gamper-Rabindran and Timmins 2011;
Lee, 2014; Martin and Beck 2018; McKinnish et al., 2010). The studies vary in their ndings. Some
nd evidence of increased in-movement of higher-income, whiter, or more educated households {
i.e., gentrication (Atkinson 2000, Freeman, Cassola, and Cai 2015, Freeman 2005, ?, Gamper-
3
Rabindran and Timmins 2011), while others do not detect such increased in
ows (Lee 2014, McK-
innish, Walsh, and White 2010). Similarly, while Atkinson (2000) nds increased out-movement
of lower-income households, a whole set of other studies do not detect increased out
ows (Ding,
Hwang, and Divringi 2016, Freeman, Cassola, and Cai 2015, Martin and Beck 2018).
When it comes to rail transit induced shocks Nilsoon and Delmelle (2020) examine movers
to and from neighborhoods with rail station openings from 1970 to 2013 using the Panel Study
of Income Dynamics (PSID). They nd that nd that low-income households are more likely to
move in general, but no more likely to move from transit neighborhoods. Their results are robust
to decade, housing tenure, time span, and dierent types of transit. After transit stations open,
Nilsoon and Delmelle (2020) nd that mover destination varies by income and station accessibility
from the neighborhood. Low-income movers from neighborhoods with highly accessible stations
tend to move within the same or to similar neighborhoods which suggests they value rail transit.
On the other hand, low-income movers from neighborhoods with poorly accessible stations tend to
move to more disadvantaged neighborhoods (Nilsoon and Delmelle 2020).
Our study contributes in two distinct ways to micro-data founded literature. This is the rst
study to examine both in- and out-mobility rates because the magnitude of in and out
ows both
aect neighborhood composition. This is also the rst study to directly tie residential
ows to
neighborhood income composition.
3 Methodology
The key question we wish to answer is whether there were any changes in residential mobility
rates near LA Metro stations that coincide with the opening of LA Metro rail stations and how
these changes relate to the income distributions in these neighborhoods. Unlike much of prior
literature, we focus in on residential mobility's impact on the income distribution by eliminating
the impact of household income changes. We leverage our longitudinal FTB data on household
locations and spatial variation to run dierence-in-dierence (DiD) tests on the mobility rates
across the income distribution. We do this by allocating households into income bins based on
their observed mean real income (MRI) and test each income group separately. The use of MRI
allows us to relate residential mobility changes back to income distribution trends, and; thus, see
whether neighborhood income distributions were aected by changes in residential
ows.
3.1 Neighborhood
ow
Changes in the income distribution of a neighborhood can be explained by two
ow equations:
1) the
ow of households into and out of the neighborhoods and the
ow of household-incomes over
time, and 2) the stock of available housing in a neighborhood in yeart. Letg index neighborhoods,
t index years, andb index income categories or bins. In equation 1, letn
b
gt
be the number of house-
holds in neighborhood g in year t in income bin b, mi
b
gt
the number of households newly observed
in neighborhood g in year t, mo
b
gt
the number of households no longer observed in neighborhood
g in year t, and di
b
and do
b
describe the number of households whose incomes shift into income
bin b and number of households whose incomes shift out of income category b in year t. Assuming
h
b
gt
housing units are available to income group b, the number of households in a neighborhood in
a particular income category is constrained by the previous year's housing stock h
b
g;t1
, plus the
number of units nh
b
gt
that were added in t minus the number of units that were decommissioned
4
dh
b
gt
minus the number of vacant units v
b
gt
in equation 2
1
. We can also place constraints on the
number of households that can move in and the number of households that can move out. This is
particularly important in regard to households that can move in since ability to inhabit a location
g is physically constrained by the availability of units. As such, in equation 3, we see that number
of households that can move-in is constrained by the availability of units and vacancy rates. In
a continuous setting, it would also make sense to include the number of households that moved
out of location g but since we are operating in a discrete time space and moved-out households
are ones that were last seen in location g in year t 1, we omit the mo
b
gt
term. The number of
households who can move out is less constrained since it is only constrained by the total number
of householdsn
b
gt
in locationg minus the number of units that are vacant. Because we do not have
accurate annual data on number of units in our areas of interest, it makes it dicult to incorporate
the constraints outlined in 3 and 4 so we omit these constraints from our base-line analysis. We do
however incorporate housing variables as robustness checks. However, it is imperative that future
research addresses the question of housing-stock
ow and quality thereof to accurately describe
income distribution changes and the concrete shocks that underlie gentrication in neighborhoods.
n
b
gt
=n
b
g;t1
+ [mi
b
gt
mo
b
gt
] + [di
b
gt
do
b
gt
] (1)
n
b
gt
=h
b
g;t1
+ [nh
b
gt
dh
b
gt
]v
b
gt
(2)
mi
b
gt
[nh
b
gt
dh
b
gt
] +v
b
gt
(3)
mo
b
gt
n
b
gt
v
b
gt
(4)
We simplify equation 1 by using long-term average real household incomes. This means that
the term [di
b
gt
do
b
gt
] = 0 falls out of equation 1 and simplies down to the number of households in
the year before and the net
ow of households. Subtractingn
b
g;t1
from equation 1 and dividing by
n
b
gt
gets us to seeing changes in income distribution through mobility rates as dened in equation
5.
n
b
gt
n
b
gt
=
n
b
g;t1
n
b
gt
+
mi
b
gt
n
b
gt

mo
b
gt
n
b
gt
(5)
In this setup, to see whether neighborhood composition changed as a result of a particular
shock Z
gt
in year t, we check whether mi
b
gt
or mo
b
gt
changed as a result of a shock. Given that
a neighborhood's mobility rate is a set of household i decisions determined by a household's X
igt
characteristics and shock Z
gt
, the mobility rates are
mi
b
gt
n
b
gt
=
1
n
b
gt
X
i2b
1(move-in)
it
(X
it
;Z
gt
) (6)
mo
b
gt
n
b
gt
=
1
n
b
gt
X
i2b
1(move-out)
it
(X
it
;Z
gt
) (7)
1
Note, it is also possible to account for crowding by adding a corresponding multiplier in front of housing unit
variables but since data on crowding at neighborhood levels is not available, we omit this from our analysis.
5
Assuming a linear form for the decision function to move-in or move-out, 1(move in)
itb
=
f(X
it
;Z
it
)
itb
and an error term "
itb
, estimating the above is equivalent to the following two re-
gressions where re
ects the average annual change in a mobility rate after a shock occurs in a
neighborhood.
1(move-in)
igtb
=c
in
+
in
X
i
+
in
Z
gt
+ +"
igtb
(8)
1(move-out)
igtb
=c
out
+
out
X
i
+
out
Z
gt
+ +"
igtb
(9)
A keen eye will note that there is a particularly important variable missing in all of the above
setup, namely price. Because we do not have accurate geospatial and longitudinal data on unit-
level rents, we omit price from our setup. Crucially, this restricts our ability to distinguish between
various components of shock Z
g
. In our specic case, this means that we cannot dierentiate
between the eect of household preferences for access to rail transit, neighborhood amenities, or
housing prices. Moreover, because we do not possess accurate longitudinal data on housing units,
shocks are also confounded with the availability of housing stock in location g. As a result, it is
useful to think ofZ as a composite shock that encompasses all changes associated with the opening
of a rail station.
3.2 Statistical Test
For our statistical tests, we lean on the above framework but in a dierence-in-dierence setup
to more cleanly isolate the impact of rail station opening. To see the impact of mobility rates on
the income distribution, the easiest method would be to estimate model 5 on our data but that runs
into issues of interpretation. Suppose we estimate model 10 in which
g
is the treatment xed-eect,
open
t
is a dummy for when stations are open. A statistically signicant
4
would tell us whether
the share of a particular income bin b was impacted by mobility-rate changes due to rail-station
opening. However, although the sign on
4
is straightforward to interpret, the magnitude of
4
is a composite eect of the impact of station-opening on the mobility-rate and the impact of the
marginal change in the rate on the income share. Although we estimate model 5, our main results
use an alternative setup based on equation 8.
n
b
gt
n
b
gt
=c
n
b
g;t1
n
b
gt
+
n
b
g;t1
n
b
gt

g
+
X
m2fmi
b
gt
;mo
b
gt
g

m
1
m
n
b
gt
+
m
2
open
t
+
m
3

g
+
m
4

g
open
t
+
tg
(10)
For our statistical test, we place equation 9 into a dierence-in-dierence setup by comparing
our transit neighborhoods to their respective control groups. Let i index households, s index
stations, t index years, p index income categories, and l index LA Metro lines or stations by year
of opening. In model 11, 1(move)
its
is a dummy for whether household i living at station s moved
in year t.
2
We decompose moves into households that move into and out of a station but for
succinctness, we omit the in and out labels from the model specications. is the constant,
s
is
the station xed-eect,
is
is the xed-eect for households located within .5 miles of a Metro rail
station, open
st
is the dummy for whether a station is open in year t, and X
its
is a set of controls
2
Since the outcome variable is binary and is estimated using OLS, this model can also be viewed as a linear
probability model (LPM).
6
that account for observable factors that can impact a household's decision to move. X includes
dummies for the status under which household i les taxes in year t (ie single, married, joint but
separately), a variable for the number of dependents in the household in year t, two dummies that
indicate whether the ling status or number of dependents changed between years t 1 and t, the
household's age, marital status, number of dependents, income bracket, and whether any of these
traits changed relative to the prior year. Although households are stratied by their estimated LTM
incomes, we still incorporate year-to-year changes in X
its
as a discrete variable if the household's
real income changes by more than 20% between years t 1 and t. The coecient of interest, ,
on the interaction between the dummy variable designating open stations and the xed eect on
treated stations provides the marginal change in neighborhood mobility rates that coincides with
station openings along the lines. Since the outcome variable is binary, captures the average annual
shift in residential mobility once the stations open.
Since mobility along each line trends dierently depending on line and station-specic traits,
we estimate separately for each combination of line and year of opening in model 11. Stations
within our panel timespan opened along the Red line in 1999 and 2000, along the Gold line in 2003
and 2009, and along the Expo line in 2012. As a result, we have a total of 5 s estimated for in
and out mobility rates. All errors are clustered at the station level.
1(move)
its
= +
s
+
i
+ open
st

i
+X
its
+"
its
(11)
We estimate model 11 separately for each income group because we aim to see the heterogene-
ity of the impact across the income distribution. The canonical the gentrication story suggests
that out-mobility rates for poor households and in-mobility rates for non-poor households increase
in gentrifying neighborhoods; thereby, skewing the income distribution away from lower income
households. This story is one combination of mobility-rate changes that can skew the neighbor-
hood distribution. Separate testing of
ow rates across income groups allows us to see magnitude
changes that can capture a far richer set of
ow changes that could lead to shifting income distri-
butions.
Aron-Dine and Bunten (2019) suggest that eorts to nd gentrication impacts fail because
there is selection among households who move in and out of gentrifying areas. To lter out the
impact of the neighborhood's new residents, they estimate the impact on residents who resided in
the area prior to a particular shock and nd signicant impact on said households. We follow their
footsteps and estimate model 11 only on households who resided in the station and control areas
prior to the opening of rail stations.
The key identifying assumption of our method are parallel mobility rate trends between control
and treated households prior to station opening. We test this assumption both numerically and
visually on mobility rates and population gures. Per Table 1, most Metro lines opened their
stations in multiple years making a simple visual inspection of parallel trends dicult. To get
around this, we make lagged dummies for 1-2 years prior to the opening of a station for treated
stations and plug the lags into model 11. The lags are falsication tests because if parallel trends
hold then the coecients on the treated lags should be statistically indistinguishable from 0.
7
Table 1: Station Summary Table
num of stations num of stations opened in time span years stations opened years opened in sample num stations in sample
Blue 22 [1990]
Expo 17 10 [2012, 2016] [2012] 10
Gold 26 20 [2003, 2016, 2009] [2003, 2009] 19
Green 13 13 [1995]
Red 15 11 [1993, 1996, 1999, 2000] [1999, 1996, 2000] 11
Since mobility rates can be decomposed between households who move station-to-station and
who move between non-station and station neighborhoods, our coecient estimates on move-inin
and move-outout variables show us the share of mobility that is explained by movers of the latter
type. Namely, if the overall mobility rate is a while the non-station and station rate is b then the
station-to-station rate is c = ab. As a result, if station opening is associated with, say, a 10
percentage point increase in the move-out rate and an 8 percentage point increase in move-outout
rate then the station-to-station mobility rate increased by 2 percentage points and 80% of the
move-out increase is explained by households that moved out of rail neighborhoods.
As outlined in equation 1, in theory it is possible to use our impact estimates on mobility rates
to directly compute whether changed residential
ows result in an increase or decrease in the share
of an income bin within the neighborhood income distribution When we run model 11 for each
income bin, we estimate the additional average annual share of households that move in or out as a
result of the station opening (ie but would not have moved otherwise). Since out-
ow and in-
ow
largely depend on the magnitude of changes (ie net rate = in
ow rateout
ow rate), simultaneous
increases or decreases in both rates could result in net in
ows or out
ows while matched out-

ow/in
ow magnitudes can simply mean greater turnover. However, using our
^
estimates directly
would give the wrong impression because we know our in-mobility rates are systematically under-
estimated. This is readily apparent from the fact that the number of households in LA County and
transit neighborhoods is generally increasing over the 2+ decades of observation yet the move-out
rates are consistently 1-2% points higher than the move-in rates which would suggest a decreasing
population. Although we believe that we estimate the change in mobility rates accurately and;
therefore, the change in
ows is also accurately estimated, there is the possibility that at least for
move-in rates,
^

in

in
3.3 Dening Mobility
To determine whether a household moved, for every year-pairt andt1 between 1993 and 2015,
we check if its ling location changed between consecutive years. Households labeled as movers
must meet any of the following criteria: 1) not in a station area in yeart1 but is in a station area
in yeart, 2) in a station area in year t 1 and is not in year t, 3) at station A in yeart 1 and at
stationB in yeart. Households whose ling location changed between the two years but remained
in the same station area are not considered movers. Since we do not know the exact year in which
a ler changed her location, for practical reasons, she is labeled as an out-mover in year t 1 or
the last year seen at a station provided that she is labeled as a mover. Conversely, a ler is labeled
as an in-mover in yeart, or the rst year she is seen at a station, given that the ler is labeled as a
mover between yearst 1 andt. For a detailed discussion on why these choices were made, please
see the Appendix in section 7. Because we require households to le in two consecutive years, the
number of households who move-in is inherently underestimated if the population grew between
8
years t and t+1 due to out-of-state movers (including immigrants) or entrants into the labor force
(ie college students).
When we discuss the rail-station income distribution, we mean the income distribution of house-
holds that reside within the half-mile radius of all rail stations we study. Households that move
station-to-station do not alter this income distribution, only those coming in from the outside or
going outside impact the distribution. To discern between households who moved station to sta-
tion and those who moved beyond station areas, we compute another set of mobility rates that,
for brevity, we call move-outout and move-inin. Move-outout households are those that were in a
station area in year t 1 and are not in any station area in year t while move-inin households are
those that were not near a rail-transit station in year t 1 and are near a rail transit station in
year t.
3.4 Treatment Groups
We allocate households into control and treatment groups based on their distance from rail
stations. We determine a tax ler's proximity to Metro rail stations by calculating the straight-line
distance between the ler's associated 9-digit zip code centroid and the nearest Metro station. We
dene a treated rail station neighborhood as the set of lers within a 0.5 mile (or 800 meter) radius
around a station. This represents an approximate 10-15 minute walking distance from a station
to the furthest extent of the neighborhood and is a common neighborhood distance used in the
planning of land uses surrounding rail stations and transit oriented development. Depending on
the street network, however, actual walking times from a station to the edge of the rail-station area
may be longer than 15 minutes.
We associate households that reside just outside of rail-station neighborhoods with the control
group because they are likely to be similar to the treated households prior to rail station opening.
Initially, the criteria for the distance bands that dene the control group was a band between
801m and 1600m away from rail stations whose residents would t the parallel trends assumption
prior to station opening. However, regardless of the band's size (801-900m) or (900-1,500m), no
control group unambiguously passed parallel trends tests. Since the dierences in the income
distributions between control and treatment groups are robust to the choice of distance band,
we instead picked distances from rail stations in a way that would yield a control group roughly
similar in size to the treated group. The distance band 850-1250m yields a set of household-year
observations sized 1445741 and 1405851 for treatment and control groups respectively.
3
To avoid
contamination between the two groups, we exclude roughly 134,000 household-year observations
that move between the treatment and control areas.
4 Data
4.1 Identifying Movers
In order to assess mobility rates, we leverage annual data on household locations and incomes
from 1993 to 2015. This dataset is constructed using data from income tax lings obtained from
the California Franchise Tax Board (FTB). The data universe contains anonymized information on
all households who led taxes in Los Angeles County in any year between 1993 and 2015, even if
3
These gures include, among others, households residing near stations that were opened outside of the tax data's
time range so the number of household-year observations used in the regressions is smaller.
9
they lived outside the County or California for some of the years during this period provided that
they still led California taxes.
For each year a ler appears in our data, we know the ler status (single, head of household,
married ling jointly, married ling singly, qualied widower), whether another tax-ler can claim
this person as a dependent, the number of dependents, the ler's age starting in 1997, and their
federal and state AGI and state taxes paid. For household income, we use federal wages, reported
by the California FTB as Federal adjusted gross income (AGI). Federal AGI is typically income
from all sources (wages, interest, dividends), less deductions such as contributions to retirement
accounts or business expenses/losses, and is a good representation of available disposable income.
Itemized deductions, such as charitable contributions or home mortgage deductions, or dependent
tax credits are applied after calculating adjusted gross income, and hence are not re
ected in our
income measure. California uses federal AGI as a starting point in state tax calculations, hence
the income we use is, in eect, federal AGI.
Location in the FTB data are designated at either the 5 or 9-digit zip codes of the address
at which a household led taxes. Within the hierarchy of the U.S. Postal Service delivery routes,
9-digit zip codes represent one block, one block-face, or large buildings. To preserve household
condentiality, the 9-digit zip code is included by the FTB only if that zip code's population
exceeds a certain threshold of households. Household residing in a zip codes that do not meet
the threshold have their zip code reported at the 5-digit level. Because 5-digit zip codes are too
large for our analysis, we exclude an average of 51% of households who are geocoded to 5-digit
zip codes though the share varies year-to-year. The ability to measure the distance between the
centroid of a 9-digit zip code to an L.A. Metro station is of prime importance because it is the
most granular geographical representation of households relative to prior work. Adding 5-digit zip
codes to this analysis would dilute the geographic specicity necessary for our neighborhood-level
analysis. Because higher residential densities are needed to report the 9-digit zip code for a tax
ler, it is likely that ling households with 9-digit zip code over-represent more dense regions of
Los Angeles County and under-represent mountainous, forested, and desert regions of the County.
Our choice of 9-digit zip codes to represent households near L.A. Metro rail stations is appropriate
because stations tend to be located in the County's most populated and dense areas. Thus, even
though more households are geocoded by FTB to the 5-digit zip code level, we choose to use 9-digit
zip codes because they allow us to track household residence in dense urban environments with
greater geographic specicity than 5-digit zip codes.
Although in theory we have data on every household that has led taxes in LA County between
1993 and 2015, the share of lers that we can use in our analysis is smaller. In the county row of
table 2, we see statistics on the distribution of the number of years for which we observe households.
For brevity, we call the number of years for which we observe a household the longevity. To start,
let us keep in mind that we are starting with only households that we observed at least in two
consecutive years so that we can determine mobility. Given this, about 20% of our data consists of
households that we only observe for 2 years though that also means that for 80% of our data, we
observe households for 3 or more years. The other thing to keep in mind is that most lers simply
disappear for reasons that we do not know and only about 6% of households in the LA County data
exit our sample because they move. Of course, given how we determine mobility, namely a change
in 9 digit Zip-code and a change in a station, this likely undercounts the number of households
that actually moved out of LA County. However, the share of households we last observe in any
given year is a fairly uniform 6-8% every year until the 2011 (gure 1). This phenomenon is not
10
surprising given that some households take a few years to le their taxes and since our data only
go up to 2015, this 2-3 year gap makes sense. We do not show years 1993 and 2015 because by
sample construction, the share of households lost in those years is 0% and 100%, respectively.
In the control and treatment rows under "overall" of table 2, we also show household longevity
statistics after we remove households that do not reside in either the control or treatment basins
(ie near rail transit). We see that our sample tends to consist of households who were observed for
slightly fewer years than the larger county sample. When we examine how long households tend to
be near rail transit under rows "near transit", we see that about half of households seem to only
be observed near transit for less than 3 years.
To lower the noise incurred from the presence of short-lived lers, we perform our analysis only
on households who are observed for 3 or more years in our overall sample though our preferred
sample is on households who are observed for 5 or more years. This means that we only include
households who have led taxes in a 9-digit zip code within LA County for 3 or more years so that
the household's long-term mean income is based on the average of at least 3 years of income data.
To avert confusion, it is possible that we observe a household for 3 or more years but that we only
observe that household in a treatment or control area for 1-2 years. All income distribution and
mobility statistics are computed only on households we observe for 3 or more years.
Table 2: Number of years households are observed for
count mean std min 10% 20% 30% 40% 50% 60% 70% 80% 90% max
treated
overall 0.0 321732.0 6.77 5.32 2.0 2.0 2.0 3.0 4.0 5.0 6.0 8.0 11.0 15.0 23.0
1.0 374757.0 6.22 4.95 2.0 2.0 2.0 3.0 3.0 4.0 6.0 7.0 10.0 14.0 23.0
county 6474472.0 7.60 5.70 2.0 2.0 3.0 3.0 4.0 6.0 7.0 9.0 12.0 17.0 23.0
near stations 0.0 490406.0 4.63 4.22 1.0 1.0 2.0 2.0 2.0 3.0 4.0 5.0 7.0 10.0 23.0
1.0 538024.0 4.51 4.13 1.0 1.0 2.0 2.0 2.0 3.0 4.0 5.0 7.0 10.0 23.0
11
Figure 1: Annual share of last-observed households
4.2 Treatment and Control Comparisons
To make sure households within .5 miles of rail stations are similar to those residing just outside
in our control group, we compute the mean
4
household's income, age, number of dependents, and
ling status (Table 3).
The mean statistic for number of dependents, and ling status align for every year (Table 3).
The mean birth year between the control and treated groups are within 1 year of each other 2001
onwards but exhibit some odd behaviour prior to 2000. Namely, some of the mean birth years
between 1996 and 2000 include households who are, allegedly, over 130 years in age. However, even
despite these irregularities in birth year records, the birth years between control and treatment
groups are still very similar.
4
We also provide medians in Table 18
12
Table 3: Observable means between control and treatment groups rounded to 0 decimals
adjIncome tpdobyy depndnts lestat
treated control treated control treated control treated control treated
year
1994 28302.0 38666.0 NaN NaN 1.0 1.0 2.0 2.0
1995 7199.0 9526.0 NaN NaN 1.0 1.0 2.0 2.0
1996 29487.0 24903.0 1955.0 1958.0 1.0 1.0 2.0 2.0
1997 26687.0 41187.0 1890.0 1839.0 1.0 1.0 2.0 2.0
1998 32222.0 27688.0 1903.0 1901.0 1.0 1.0 2.0 2.0
1999 33507.0 27655.0 1934.0 1929.0 1.0 1.0 2.0 2.0
2000 33577.0 27738.0 1947.0 1945.0 1.0 1.0 2.0 2.0
2001 33037.0 26995.0 1960.0 1960.0 1.0 1.0 2.0 2.0
2002 32715.0 26307.0 1960.0 1961.0 1.0 1.0 2.0 2.0
2003 32828.0 26514.0 1961.0 1962.0 1.0 1.0 2.0 2.0
2004 33392.0 26732.0 1962.0 1963.0 1.0 1.0 2.0 2.0
2005 33840.0 26914.0 1963.0 1964.0 1.0 1.0 2.0 2.0
2006 34403.0 27159.0 1964.0 1965.0 1.0 1.0 2.0 2.0
2007 35163.0 28187.0 1965.0 1965.0 1.0 1.0 2.0 2.0
2008 32345.0 26968.0 1966.0 1967.0 1.0 1.0 2.0 2.0
2009 30572.0 26462.0 1966.0 1967.0 1.0 1.0 2.0 2.0
2010 31332.0 26827.0 1968.0 1969.0 1.0 1.0 2.0 2.0
2011 29375.0 26600.0 1968.0 1969.0 1.0 1.0 2.0 2.0
2012 32197.0 27327.0 1969.0 1969.0 1.0 1.0 2.0 2.0
2013 31835.0 27210.0 1970.0 1970.0 1.0 1.0 2.0 2.0
2014 34166.0 29605.0 1970.0 1971.0 1.0 1.0 2.0 2.0
Growth in the number of households in treatment and control areas follow similar trends though
the half-mile rail-transit basin tends to grow slightly faster (Figure 2). This is expected if the areas
within .5 mile of rail stations are receiving more housing development that enables more households
to live near transit. These trends hold regardless of the distance bandwidth we choose for the control
area. Moreover, we also see that dense parts of LA County grew at rates comparable to rail-station
and adjacent neighborhoods.
13
Figure 2: Number of households by treatment group
Beside the above, we also compare income movements and sample longevity between treatment
and control households. As noted above in table 2, the treatment and control group households
have very similar sample longevity distributions. The same holds when we compare their income
deviations in the next section in table 4.
4.3 Income Categories
To assess mobility impact on households across the income distribution, we allocate households
into income bins based on the household's average real income across the years it is present in the
data. We use a seasonally-unadjusted all-item urban West CPI normalized to 1994 to de
ate house-
hold incomes and estimate long-term mean incomes using the mean of adjusted incomes. Because
gentrication and displacement are usually framed as a housing-related issue aecting low-income
households, we assign households to income categories that correspond with U.S. Department of
Housing and Urban Development (HUD) poverty denitions based on the Los Angeles County Area
Median Income (AMI). Three category cutos correspond to HUD's designation of Extremely Poor
(0-30% AMI), Very Poor (30-50% AMI), and Poor (50-80% AMI) households. We use 1994 AMI
to allocate households into income categories though any other starting point would do too. To
determine non-poor income bins, we observe that just over 22% of households fall into the 0-30%,
30-50%, and 50-80% of AMI bins around the mid-point of our study. To keep sample sizes similar
for every income category regression, we divided higher-earning households into two bins: 80-200%
and 200-500% of AMI. The former also contains about 22% of the sample while the 200-500%
contains about 5%. Every year, about 1% of households report incomes below $0 which suggests
that they are self-employed and had more expenses than income in a particular year. Another 1%
of households in our sample have incomes over 500% of AMI. Although households with incomes
below $0 and above 500% of AMI may be of interest in particular scenarios, we believe that our
0-500% of AMI bins capture the majority of the households and provide a simpler and more ac-
cessible snapshot of what is happening in the neighborhood. Although HUD adjusts the denition
14
of poverty for a household's size and number of earners, we eschew such a granular allocation of
households to poverty status in favor of our simpler approach because our key goal is to assess
mobility rates across the income distribution rather than just poor households. Table 19 shows the
yearly income cutos for the 5 income categories though since incomes are de
ated to 1994 dollars,
we use 1994 AMI gures to allocate households to income groups.
There are many advantages to using LTM incomes as opposed to nominal annual incomes. First,
it accounts for the fact that the Area Median Income (AMI) is a relative measure of income that
changes as the median income of the metropolitan area changes. A household's income may increase
or decrease across time but its relative status within the income distribution does not necessarily
have to change. Conversely, even if a household's income does not change over time then it may
still step up or down along the income distribution depending on changes in the AMI. Moreover,
household incomes
uctuate across time and households with incomes near our AMI cutos may
switch categories despite small income
uctuations. For example, if a household earns $28,000 is
year 2005 it will be categorized as Very Poor but if it earns just $100 more in year 2006 then it will
be categorized as Poor in year 2006. However, if the household's LTM income is actually in the
Very Poor category then its hops across categories will create noise and dilute our model estimates.
To avoid these pitfalls, we use the household's real
5
long-term mean (LTM) income and allocate
to income categories based on the 1994 AMI
6
. The last perk of using LTM income to categorize
households is it enables us to compare a more homogenous set of households year to year. Were we
to use annual income to re-categorize a household's income bin in every year, the set of households
we compare year-to-year would change. This would likely add noise to our measures of residential
mobility changes because we are no longer able to distinguish between residential mobility trends
and sample composition.
To see how generally households adhere to their LTM incomes, we examine the share of house-
holds whose incomes are either in their average income bin, below it, and above it in table 4. In any
given year, over 50% of households earned within their average income bin. We also observe that
the patterns between the control and treatment groups are nigh identical which gives us further
condence in the validity of our control household choice.
5
Incomes adjusted using West Urban CPI
6
The West Urban CPI series we use has a 1982-1984 base year but we normalize to 1994.
15
Table 4: Deviations from LTM income
InAvgInc belowAvgInc AboveAvgInc
treated 0.0 1.0 0.0 1.0 0.0 1.0
year
1994 0.57 0.58 0.11 0.10 0.32 0.31
1995 0.57 0.58 0.12 0.12 0.30 0.30
1996 0.57 0.57 0.11 0.11 0.31 0.31
1997 0.58 0.58 0.12 0.12 0.29 0.29
1998 0.58 0.58 0.12 0.12 0.29 0.29
1999 0.59 0.59 0.13 0.13 0.28 0.28
2000 0.60 0.60 0.15 0.15 0.25 0.25
2001 0.60 0.60 0.13 0.13 0.26 0.26
2002 0.60 0.60 0.14 0.14 0.25 0.25
2003 0.58 0.58 0.23 0.23 0.19 0.19
2004 0.59 0.60 0.20 0.20 0.20 0.20
2005 0.59 0.59 0.23 0.23 0.18 0.17
2006 0.59 0.59 0.24 0.25 0.16 0.16
2007 0.58 0.58 0.28 0.28 0.13 0.13
2008 0.60 0.60 0.25 0.26 0.14 0.14
2009 0.62 0.63 0.19 0.20 0.18 0.17
2010 0.63 0.63 0.19 0.20 0.17 0.16
2011 0.63 0.64 0.20 0.20 0.16 0.15
2012 0.62 0.63 0.22 0.22 0.15 0.14
2013 0.59 0.60 0.28 0.28 0.12 0.12
2014 0.55 0.56 0.35 0.34 0.10 0.09
4.4 Station Inclusion Criteria
To get the panel on which we run our tests, we lter out some of the 93 LA Metro rail stations.
We can only t a DiD model on stations that opened between 1995 and 2014 to allow for before
and after observations but we also exclude observations prior to 1996 due to data quality concerns.
All stations along the Blue line opened before 1993 so we drop the Blue line entirely. Due to the
large overlap between Red and Purple stations, we treat the Purple line as the Red line. Due to
privacy and measurement concerns, we also drop stations with fewer than 100 observations across
the control and treated groups in any given year. This size-restriction eliminates the El Segundo
station along the Green line that opened in 1995 and the Little Tokyo and Pico/Aliso stations along
the Gold line that opened in 2009 (Table 5). Since our statistical tests are run 1996-onward, the
elimination of the El Segundo station does not impact our statistical results. In the end, we are left
with 36 stations between 1996 and 2014 or 684 station-year pairs. The median station has about
2,500 households which amounts to about 1.7 million household-year pair observations available for
our models.
Table 5: Stations dropped due to small populations (< 100)
stop name Line Open year Announce year
#
48 El Segundo Green 1995 1991
70 Little Tokyo / Arts District Gold 2009 2004
In addition to sample-sizes, we also consider parallel trends and sample availability for every
station
7
. Along the Red line stations, almost every tested station had a large deviation or spike, in
either the move-in or the move-out rate at least in one year before the station opened. For example,
along station 31 that opened in 1999, the dierence between the treated and control move-in rates
in 1998 is over 5 percentage points and along station 32, also opened in 1999, the dierence in the
move-out rates in 1997 is over 3 percentage points. Since the pre-treatment along the Red line is
fairly short (3-4 years), many of these deviations show up when we examine lags in our parallel
7
Figure is available upon request as it is too large to be included in the paper.
16
trends test. Despite these large deviations before station opening, along many of such stations,
the treatment and control trends either converge or intermingle in the 2000s. We include all Red
line stations in our analysis but caution the reader to these quirks. We do, however, exclude one
station opened in 2009 along the Gold line, station 71, because its control area does not have tax
data for all years in our period of analysis.
5 Results
5.1 Income Distribution Descriptives
When using a household's estimated LTM real income, both LA County and rail-station areas
experienced increases in their share of extremely-poor households (0-30% of AMI)
8
. The share of
extremely-poor households went up in LA County's denser parts from 20% in 1994 to 27% in 2014
albeit with a bit of ebb and
ow in the years between (Figure 3). The income distributions for
households residing within .5 miles of rail stations and those just outside follow the pattern of
LA County. In the rail station neighborhoods, the share of extremely-poor households increased
from 28% in 1995 to 36% in 2014 while in neighborhoods just outside of rail stations, the shares
increased from 25% to 34% (Figure 4). Like LA County more generally, both the treated and
control income distributions maintained the share of very-poor (30-50% of AMI) households at
20-26% throughout the two decades of observation. The share of poor (50-80%) households, on
the other hand, decreased by 3-4% percentage points from a start of 20% in LA County and 26%
and 24% in the treated and control areas respectively. At the higher end of the income spectrum,
households with incomes over 200% of AMI retained their 10%-11% share in LA County and their
2-3% in the treated areas and 5% in the control across the period.
The above patterns suggest three things. First, households across the income distribution who
reside within .5 miles of rail transit stations tend to have lower mean incomes than LA County more
generally. Second, the income distributions across all three areas experienced nigh identical trends
which implies that income distribution shifts near LA Metro areas are likely due to reasons that
prevail more widely throughout the County. Finally, we see that income-inequality, as measured
by LTM incomes, in rail-transit neighborhoods and LA County increased across our two decades
of observation.
8
For these income distributions, we only use lers observed for over 3 years. You can see the income distribution
for the full sample for LA County in gure 9 and for LA Metro areas in gure 10
17
Figure 3: LTM Incomes for Dense parts of LA County
Figure 4: LTM Incomes Near LA Metro Rail Transit
In gure 4, we see the patterns are similar between our treatment and basin areas so we test this
directly by estimating the equation 12 on each of our line-years. The estimates tells us whether
the income-share trend of households in a transit neighborhood deviated from the control group
once the station opened.
18
n
b
tg
N
tg
=c +
g
+ open
t
+
g
open
t
+
tg
(12)
According to our estimates in table 6, the income distribution trends between the treatment
and control basins are, for the most part, statistically indistinguishable. This supports our hypoth-
esis that the income trends are not unique to LA Metro rail transit neighborhoods and are likely
re
ections of macro-trends in LA County.
However, there are 3 distinctions in table 6. Along the Gold-line stations opened in 2003, the
share of very-poor (30-50% of AMI) households went down by 1.3 percentage points (pp) more
than in the control group. On the other hand, the share of high-income households (200-500% of
AMI) went up by .4pp more than it did in the control group. These two shifts are consistent with
gentrication narratives in which low-income households move out and high-income households
move in. Conversely, along the Red 2000 stations, we see that the share of high-income households
went down by .8pp more than it did in the control group once station opened.
19
Table 6: Test on income distribution trends
dep var dstrShare
ind var condNum nobs open:treated rsqrd
model line-year AMI
share Expo2012 30 7.3853 38 0.008 (0.0217) 0.3773
50 7.3853 38 0.001 (0.0058) 0.1789
80 7.3853 38 -0.005 (0.0130) 0.4413
200 7.3853 38 -0.007 (0.0130) 0.1544
500 7.3853 38 0.001 (0.0011) 0.8993
Gold2003 30 7.99823 38 -0.003 (0.0116) 0.6124
50 7.99823 38 -0.013** (0.0044) 0.67
80 7.99823 38 -0.001 (0.0063) 0.6205
200 7.99823 38 0.01 (0.0051) 0.636
500 7.99823 38 0.004* (0.0017) 0.9725
Gold2009 30 6.32765 38 -0.006 (0.0133) 0.7758
50 6.32765 38 0.002 (0.0068) 0.2303
80 6.32765 38 -0.001 (0.0072) 0.7538
200 6.32765 38 0.005 (0.0086) 0.5535
500 6.32765 38 -0.0 (0.0004) 0.5022
Red1999 30 12.6855 38 -0.013 (0.0175) 0.728
50 12.6855 38 0.002 (0.0071) 0.7992
80 12.6855 38 -0.004 (0.0073) 0.1678
200 12.6855 38 0.014 (0.0076) 0.9014
500 12.6855 38 -0.002 (0.0025) 0.9701
Red2000 30 10.8604 38 -0.001 (0.0159) 0.3887
50 10.8604 38 0.004 (0.0047) 0.8073
80 10.8604 38 0.009 (0.0092) 0.2007
200 10.8604 38 -0.0 (0.0073) 0.7749
500 10.8604 38 -0.008* (0.0039) 0.5579
5.2 Mobility Rate Descriptives
When we examine the average annual mobility rates across income groups of households residing
in the treated areas, we nd the low-income households move out of rail station areas at rates lower
than the higher-income households but move into rail station areas at rates similar to the highest-
income groups. In gure 5, we show the average rate at which households across the income
distribution in the treated group move out (ie move-outout) of our analysis basin and the average
20
rate at which households move into (move-inin) rail transit areas from outside of our analysis
basin. We present move-inin and move-outout rates because station-to-station moves are less likely
to impact the income distribution of near-transit neighborhoods
9
. We nd that the out-mobility
rate is similar across our low-income households (0-80% of AMI) as it ranges from 4.3% to 4.6%.
These gures are lower than the out-mobility rates of the middle-income (80-200% of AMI) and
high-income (200-500% of AMI) households of 5.1% and 6.3% respectively. The in-mobility rates are
somewhat more similar across income groups. In the same gure 5, we nd that low-income (0-80%
of AMI) households move into our basin at an annual average of 3.4%-3.5% which is comparable
to that of the middle-income (80-200%) group of 3.4%. The high-income households (200-500%)
have the highest move-in rate of 3.9%.
Given that the share of extremely-poor households has been growing while the share of high-
income households has been steady, the net eect of move-in and move-out rates in gure 5 would
suggest an out
ow of households across all income groups. However, we caution the reader to recall
that our move-in rates are likely understated because we miss out-of-county immigrants and new
entrants into the labor force. Nonetheless, the relative patterns are consistent with the observed
trends in our income distributions (gure 4). We can also get a rough approximation of the amount
by which our annual move-inin rates are underestimated. Very-poor households have more or less
maintained their share of 26-27%. Given that the move-outout and move-inin rates would need to
balance for this to happen and given a move-outout rate of 4.3% and a move-inin rate of 3.4%, the
gap suggests that our move-inin rates are underestimated by about .9 percentage points.
Figure 5: Mean mobility rates for moves beyond stations by income
Whiskers are 1 standard deviation above and below mean
5.3 Parallel Trends
Mobility rates among households near and just outside of rail stations follow similar trends
(Figure 6). In case of the Expo line, the move-out and move-in rates follow each other fairly closely
but they also frequently criss-cross suggesting little to no dierence between the trends of control
9
We show overall mobility rates in gure 8.
21
and treatment groups.There are also fairly large deviations from the control group in the late 90s
and the Great Recession though the opposing signs on the deviations likely cancel out in the pre-
treatment period. Pre-trends along the Gold line stations opened in 2003 are similar for most of the
period while along the 2009 stations, the trends often criss-cross both before and after suggesting
we likely won't nd an impact. Along the Red lines, the pre-trends always start out slightly higher
and converge with the control group right after station opening. These patterns hold regardless of
the choice of distance bandwidth for control groups or ler longevity. When we exclude households
that moved between stations (gure 7), we see very similar patterns.
Figure 6: Parallel trends by line
22
Figure 7: Parallel trends by line for moves beyond stations
In anticipation of results, we can compare regression coecient estimates to expectations laid
out by gures 6 and 7. Since mobility trends along the Expo line and Gold-line stations opened
in 2009 largely overlap between the control and treatment basins, both before and after station
opening, we would not expect to see large or signicant coecient estimates along these lines. On
the other hand, we see some evidence of trend divergence along the Red line stations opened in
2000 and Gold line stations in 2003 which leaves a greater possibility seeing an impact.
We run model 11 with dummies on one and two years prior to the opening of rail stations to
check whether pre-trends hold prior to opening. If parallel trends hold, then the coecients on the
dummies should be statistically indistinguishable from 0. Results for households observed for 5+
years are table 7 while the results for 3+ year households and incumbents can be found here 21.
Most 1st and 2nd lags in table 7 are insignicant but our setup does not
awlessly pass the lag
23
tests. Along the Expo line, all lag coecients on move-out and move-outout lags are insignicant
though second lags are
agged as signicant for extremely-poor and very-poor household move-in
and move-inin rates. A similar pattern holds along the Gold 2003 stations where no move-out and
move-outout lags are
agged but a few rst lags on move-in rates are
agged. Along the Gold 2009,
Red 1999, and Red 2000 stations, we do see considerably more 1st lags and occasional 2nd lags

agged for both in
ow and out
ow rates.
Nonetheless, we believe there is cause to be optimistic. First, of 200 estimated 1st and 2nd
lag coecients from 20 (5 lines across 5 income groups and 4 mobility rates) regressions, only 20
coecients are signicant at a 5% or lower p-value. In a purely random setting, even if the null
were true, we would still expect about 10 coecients to come up signicant. Second, given the
nuisance of rail-transit construction, some of the decrease in move-in rates and increase in move-out
rates make intuitive sense in this context. Finally, in gure 6 we see that mobility rates tend to be
noisy and erratic. Given that when we perform similar tests across our other sub-samples we nd
that no consistent patterns emerge in sign and coecient, we believe that many of these
agged
results are, indeed, results of randomness. No choice of control bandwidth resolves these issues.
Table 7: Placebo tests on rst and second lags on 5+ year households
dep var movedOut movedOutOut movedIn movedInIn
ind var lag1:trtd lag2:trtd lag1:trtd lag2:trtd lag1:trtd lag2:trtd lag1:trtd lag2:trtd
line-year AMI
Expo 30 0.002 (0.0071) -0.001 (0.0040) -0.001 (0.0055) -0.0 (0.0048) 0.003 (0.0063) 0.016*** (0.0050) 0.001 (0.0061) 0.014** (0.0053)
50 0.002 (0.0064) 0.006 (0.0074) 0.006 (0.0058) 0.005 (0.0067) 0.009 (0.0073) 0.008* (0.0038) 0.005 (0.0056) 0.006* (0.0028)
80 0.012 (0.0067) 0.004 (0.0080) 0.01 (0.0075) 0.003 (0.0072) -0.002 (0.0052) 0.006 (0.0069) -0.005 (0.0062) 0.004 (0.0061)
200 0.005 (0.0063) 0.01 (0.0056) 0.002 (0.0056) 0.006 (0.0058) 0.007 (0.0064) 0.004 (0.0056) 0.004 (0.0057) 0.003 (0.0061)
500 -0.01 (0.0098) -0.019 (0.0176) -0.004 (0.0068) -0.017 (0.0175) 0.015 (0.0110) 0.007 (0.0227) 0.01 (0.0096) 0.014 (0.0224)
Gold2003 30 -0.005 (0.0060) -0.006 (0.0073) -0.003 (0.0050) -0.004 (0.0060) 0.011* (0.0050) -0.008 (0.0052) 0.013** (0.0045) -0.009* (0.0039)
50 0.001 (0.0123) -0.007 (0.0064) 0.003 (0.0096) -0.004 (0.0052) -0.007 (0.0042) -0.002 (0.0076) -0.004 (0.0051) -0.003 (0.0060)
80 -0.015 (0.0126) 0.002 (0.0091) -0.015 (0.0116) 0.001 (0.0102) 0.015 (0.0108) 0.005 (0.0079) 0.006 (0.0088) 0.004 (0.0062)
200 0.0 (0.0080) -0.009 (0.0069) 0.002 (0.0076) -0.007 (0.0064) 0.001 (0.0064) 0.01 (0.0058) 0.001 (0.0060) 0.007 (0.0051)
500 0.011 (0.0194) -0.012 (0.0147) 0.002 (0.0168) -0.013 (0.0119) 0.012 (0.0119) -0.008 (0.0162) 0.011 (0.0109) -0.01 (0.0152)
Gold2009 30 -0.001 (0.0037) 0.002 (0.0068) -0.006 (0.0032) 0.004 (0.0053) 0.002 (0.0074) -0.004 (0.0065) 0.003 (0.0032) -0.003 (0.0044)
50 -0.013*** (0.0040) -0.004 (0.0054) -0.012** (0.0044) -0.001 (0.0048) -0.007*** (0.0009) -0.0 (0.0048) -0.006*** (0.0013) -0.002 (0.0031)
80 0.005 (0.0069) 0.002 (0.0031) 0.004 (0.0051) 0.001 (0.0039) 0.003 (0.0050) -0.007 (0.0056) 0.007* (0.0034) -0.005 (0.0064)
200 -0.009 (0.0069) -0.003 (0.0104) -0.009 (0.0068) -0.007 (0.0102) -0.002 (0.0088) -0.001 (0.0073) -0.005 (0.0091) -0.002 (0.0062)
500 -0.181*** (0.0547) 0.095 (0.0681) -0.178*** (0.0543) 0.098 (0.0665) -0.035 (0.0244) -0.109* (0.0495) -0.032 (0.0237) -0.061 (0.0472)
Red1999 30 -0.007 (0.0170) 0.008 (0.0130) -0.006 (0.0184) 0.008 (0.0083) 0.006 (0.0141) -0.003 (0.0041) 0.003 (0.0133) -0.003 (0.0078)
50 0.001 (0.0229) 0.008 (0.0185) -0.002 (0.0188) 0.011 (0.0162) -0.001 (0.0192) 0.005 (0.0168) -0.0 (0.0138) 0.008 (0.0154)
80 -0.006 (0.0199) -0.02 (0.0220) -0.001 (0.0107) -0.01 (0.0130) -0.028* (0.0138) -0.018 (0.0176) -0.014** (0.0049) -0.015 (0.0127)
200 -0.014 (0.0341) -0.017 (0.0294) 0.002 (0.0214) 0.001 (0.0231) -0.007 (0.0105) 0.011 (0.0133) 0.001 (0.0103) 0.017 (0.0108)
500 -0.039 (0.0199) -0.001 (0.0261) -0.028 (0.0183) -0.02 (0.0204) -0.012 (0.0306) -0.006 (0.0187) -0.02 (0.0158) -0.023 (0.0272)
Red2000 30 -0.034 (0.0208) -0.017 (0.0220) -0.016 (0.0152) -0.01 (0.0232) -0.016 (0.0094) -0.002 (0.0249) -0.011 (0.0073) -0.005 (0.0138)
50 -0.014 (0.0148) 0.007 (0.0169) -0.016 (0.0143) 0.003 (0.0134) -0.023 (0.0149) -0.026** (0.0095) -0.02 (0.0167) -0.02* (0.0102)
80 -0.018* (0.0073) 0.014 (0.0126) -0.009 (0.0152) 0.029 (0.0216) -0.023 (0.0127) -0.006 (0.0228) -0.012 (0.0080) 0.0 (0.0186)
200 0.013** (0.0043) 0.015** (0.0057) 0.011 (0.0104) 0.016*** (0.0039) 0.016 (0.0128) 0.002 (0.0026) 0.017 (0.0131) 0.0 (0.0060)
500 0.008 (0.0386) -0.024 (0.0340) -0.001 (0.0406) -0.017 (0.0274) -0.019 (0.0117) 0.023 (0.0138) -0.021 (0.0132) 0.019 (0.0105)
standard error in parentheses (robust errors clustered at station level) *p<.05; **p<.01;
***p<.001
Units are proportions (not percents)
5.4 Impact on mobility rates across the income distribution
Along the Expo line stations opened in 2012, we nd no changes in the average annual mobility
rates once stations open among households observed for 5+ years except for very-poor households
(table 8). In their case, the move-in rate decreased by .7 percentage points (pp) suggesting an
increased out
ow to non-Expo rail stations. In context of the expo-line income distribution in
gure 8, the lack of impact on mobility rates makes sense since the control and treated income
distributions move nigh-identically. In case of very-poor households, we see in the same gure that
in the control area, the share of very-poor households went down from 28% to 27% a year earlier
24
than in the transit basin which may coincide with the increased move-out rate among very-poor
households.
Along the Gold line stations opened in 2003, we see that the move-out (move-outout) rate
among extremely-poor households decreased by .7pp (.5pp) while the move-in rate increased by
.6pp suggesting a slight increase in the net-in
ow of extremely-poor households once the stations
opened (table 8). Among very-poor households, the move-out (move-outout) rate increased by 1pp
(.8pp) suggesting the opposite eect. In (gure 9), we observe a slightly accelerated pace in the
growth of extremely low-income households and a decreased share in very-poor households relative
to the control area 2003 onwards which is consistent with the changed residential
ows.
Along the Gold line stations opened in 2009, we see a .3pp decrease in the move-inin rate
of extremely-poor households suggesting a slower in
ow. In gure 10), we see that the share
of extremely-poor households grew more in the control area which is consistent with the impact
estimates.
Along the Red line stations opened in 1999, the annual average move-in (move-inin) rate among
poor households (50-80%) increased by 1.4pp (1.6pp) and among the high-income households (200-
500%), the move-in (move-inin) rate increased by 2.9pp (3pp). In context of the income distribution
for these stations (gure 11), these result suggest that the opening of the rail stations may have
slowed the out
ow of poor households relative to the control area. Namely, between 1996 and 1998,
the share of poor households in the treatment basin increased by 1pp in the control area. However,
between 1999 and 2014, the control area experienced a 4pp drop in the share of poor households
whereas the treated area only experienced a 2pp drop.
Along the Red line stations opened in 2000, the move-in rate among very-poor (30-50%) house-
holds increased at an annual average by 2pp while the move-in (move-inin) decreased by 1pp (.9pp)
among the poor households. In the context of the income distribution (gure 12), we see that in the
treated group, the share of poor households dropped slightly earlier and by a slightly larger margin
than in the control neighborhood suggesting that the opening of rail stations may have played a
part. Conversely, among the very-poor households, the treated basin sustained a slight increase in
the share of very-poor households for longer than the control neighborhoods which may have been
due to the increased in
ow rate among these households.
When we consider the magnitude of the estimated impacts on annual residential mobility pat-
terns, we see that the changes are fairly large relative to average mobility rates but small relative
to the median number of households at LA Metro stations. In our base results (table 8), the largest
estimated impact on low-income households (<80% of AMI) is 1.4pp for poor households, 2pps for
very-poor households, and .6pp for extremely-poor households. Consider that between 2000 and
2014, the median number of lers ranged from 1,000 to 1,400 (table 20) and that extremely-poor,
very-poor, and poor households compose a maximum of 37%, 28%, 21% of treated station house-
holds in that same period. The estimated impacts would suggest that once the stations opened,
a high estimate of the number of additional extremely-poor households who move into the tran-
sit neighborhood is 3.1 (=1; 400 37%:006) households. For the very-poor households, a high
estimate of the number of additional households who otherwise would not have moved-in is 8 house-
holds. For poor households, a higher estimate of the number of additional households who move
into a neighborhood because of the station is 4 households. For perspective, the estimated average
number of extremely-poor households that move into a median station is 24 (=1; 40037%4:6%)
given an average move-inin rate of 4.6% (gure 8). The comparable number of very-poor and poor
households that move in is 17 and 13 given average annual move-inin rates of 4.7% and 4.4% respec-
25
tively. In this light, we see that the marginal impact of rail stations on mobility rates relative to
average mobility rates is large. If we consider that in a median station, there are 518(=37%1; 400)
extremely-poor households and over 10 years a cumulative 30 additional extremely-poor households
move-in as a result then these 20 additional households will constitute a 4% increase in the share
of extremely-poor households. Of course, we must keep in mind that we underestimate in-mobility
so the true impact on the number of households is likely larger.
Table 8: Base DID Estimates
dep var movedOut movedOutOut movedIn movedInIn
ind var nobs open:treated nobs open:treated nobs open:treated nobs open:treated
line-year AMI
longobsExpo 30 93688 -0.0 (0.0034) 93688 -0.001 (0.0028) 93688 -0.003 (0.0031) 93688 -0.001 (0.0029)
50 95875 -0.004 (0.0039) 95875 -0.003 (0.0040) 95875 -0.007* (0.0029) 95875 -0.003 (0.0027)
80 71398 0.006 (0.0038) 71398 0.005 (0.0033) 71398 0.0 (0.0035) 71398 -0.001 (0.0031)
200 66111 0.004 (0.0032) 66111 0.003 (0.0028) 66111 -0.0 (0.0018) 66111 -0.0 (0.0009)
500 8330 -0.004 (0.0058) 8330 -0.01 (0.0065) 8330 0.001 (0.0105) 8330 0.003 (0.0109)
longobsGold2003 30 91445 -0.007** (0.0027) 91445 -0.005** (0.0021) 91445 0.006* (0.0025) 91445 0.005 (0.0026)
50 83262 0.01** (0.0034) 83262 0.008* (0.0035) 83262 0.003 (0.0022) 83262 0.003 (0.0019)
80 73071 0.001 (0.0044) 73071 0.0 (0.0041) 73071 -0.006 (0.0039) 73071 -0.006 (0.0032)
200 95051 0.005 (0.0047) 95051 0.004 (0.0043) 95051 0.001 (0.0024) 95051 0.0 (0.0024)
500 28933 -0.0 (0.0064) 28933 -0.003 (0.0064) 28933 -0.002 (0.0051) 28933 -0.007 (0.0040)
longobsGold2009 30 101332 -0.001 (0.0023) 101332 0.0 (0.0019) 101332 -0.002 (0.0017) 101332 -0.003* (0.0013)
50 101573 0.001 (0.0028) 101573 0.0 (0.0018) 101573 -0.001 (0.0039) 101573 0.001 (0.0026)
80 70927 0.003 (0.0037) 70927 0.001 (0.0038) 70927 0.0 (0.0023) 70927 -0.001 (0.0021)
200 34887 0.001 (0.0043) 34887 0.002 (0.0047) 34887 0.004 (0.0025) 34887 0.002 (0.0031)
500 956 -0.006 (0.0191) 956 -0.013 (0.0192) 956 -0.015 (0.0227) 956 -0.014 (0.0179)
longobsRed1999 30 78394 -0.001 (0.0059) 78394 0.003 (0.0062) 78394 -0.0 (0.0088) 78394 0.001 (0.0058)
50 68841 -0.012 (0.0087) 68841 -0.009 (0.0084) 68841 -0.003 (0.0079) 68841 0.002 (0.0082)
80 50561 0.009 (0.0054) 50561 0.007 (0.0072) 50561 0.014* (0.0058) 50561 0.016*** (0.0039)
200 44592 0.003 (0.0067) 44592 0.001 (0.0074) 44592 0.0 (0.0031) 44592 0.0 (0.0050)
500 8730 0.022 (0.0211) 8730 0.021 (0.0193) 8730 0.029*** (0.0079) 8730 0.03* (0.0125)
longobsRed2000 30 26903 -0.003 (0.0044) 26903 -0.005 (0.0037) 26903 -0.009 (0.0070) 26903 -0.01 (0.0069)
50 28066 0.008 (0.0097) 28066 0.006 (0.0063) 28066 0.02* (0.0098) 28066 0.015 (0.0089)
80 28289 -0.012 (0.0099) 28289 -0.011 (0.0088) 28289 -0.01** (0.0035) 28289 -0.009*** (0.0025)
200 36286 -0.003 (0.0022) 36286 -0.003 (0.0016) 36286 -0.009 (0.0105) 36286 -0.016 (0.0116)
500 9141 0.022 (0.0160) 9141 0.027 (0.0244) 9141 0.016 (0.0151) 9141 0.013 (0.0105)
Figure 8: Income distribution of Expo line stations
26
Figure 9: Income distribution of Gold line 2003 stations
Figure 10: Income distribution of Gold line 2009 stations
Figure 11: Income distribution of Red line 1999 stations
Figure 12: Income distribution of Red line 2000 stations
27
5.5 Robustness checks
We run robustness checks on our estimates by checking the relevance of sub-sample selection
and housing variables on our impact estimates and present the results in tables and 9 & 10.
In table 9, we see that the sample composition, in terms of the number of years for which a
household is observed, does not categorically alter the results observed in table 8 in terms of sign
and magnitude. In table 9's base model, we t model 11 to households who are observed for 3+
years instead of our preferred 5+ years. The estimated eects are not entirely stable so we will
focus on non-overlapping results. Along the Expo line, using 3+ year households shows us that the
move-out (move-outout) rates among high-income households decreased by 1.3pp (1.7pp) whereas
with 5+ year households the estimated magnitudes were much smaller; hence, the impact was
not statistically signicant. Along the Gold-line stations opened in 2003, the decreased move-out
(move-outout) results for extremely-poor households are no longer signicant because the estimated
magnitudes are a bit smaller with the 3+ year lers. Along the Gold 2009 stations, the decreased
move-inin rate of .3pp is no longer signicant because of a slightly larger error. Along the Red 1999
stations, the positive eect on move-in (move-inin) rates among poor households are no longer
signicant as the estimated impact coecients are much smaller. For high-income households, the
impact coecient on move-inin rates is no longer signicant because of a larger error. The Red
2000 stations is where results look the most dierent. Using the 3+ sample, the estimates suggest
that the move-out (move-outout) rate among extremely-poor households decreased by 1.2pp (.9pp)
and the move-in (move-inin) rate decreased by .9pp (.7pp) suggesting lower turn-over among these
households and a net-increased in
ow into the area. The 5+ year sample estimates much smaller
move-out magnitudes and larger move-in errors. Conversely, the 3+ year sample no longer suggests
that the move-in rate among very-poor households signicantly increased due to a large shift in
magnitudes though still the same sign. Finally, we also observe that the move-out (move-outout)
rates among high-income households increased by 2.1pp (2.3pp). The statistically signicant results
unique to the 3+ year sample make sense in the context of the income-distributions changes.
In the \superlong" model, we also show you the results when we use households who are
observed for 10+ years in table 9. With this sample, much fewer statistically signicant deviations
are found though this sample is likely prone to selection bias among households. Along the Expo
line, no impact on mobility rates was found. Along the Gold 2003 stations, the signicant decrease
in move-out (move-outout) rates is estimated at a larger magnitude but the impact on very-poor
households and extremely-poor move-in rates are no longer deemed signicant. Along the Red 1999
stations, the increased move-inin rates among poor households is still signicant but the impact
on high-income households is not. Along the Red 2000 stations, the 10+ year sample picks up a
new result of decreased move-out mobility among poor households while the signicant impact on
increased and decreased move-in rates among very-poor and poor households respectively continues
to hold.
Because we only have housing variables starting 2004, we can only run our robustness checks on
the Expo line stations that opened in 2012 and the Gold line stations that opened in 2009. As such,
we rst test whether a start of the pre-treatment year of 2004 instead of 1996 makes a dierence on
our estimates. This model we call the "late-base" and relative to our base results (table 8), we see
that the coecient magnitudes are slightly larger in all cases except for the high-income households
whose coecients are similar or smaller. Regarding statistically-signicant coecients, we see two
things relative to the base-line. Unlike in the sample that starts in 1996 that only
ags the move-
in coecient among extremely poor households as signicant, the one that starts in 2004 shows
28
signicant coecients for both move-in and move-inin rates that are also 2pp larger in magnitude
than the base-line results suggesting average annual decreases of .9pp and .5pp, respectively. Along
the Gold line, the coecient magnitudes are larger but so are the errors so no coecients are
agged
as signicant at any level anymore.
When we incorporate variables on number of units and vacancy-rate data from the 5-year ACS
into the model (model lineYear-housing), predictably, we see the magnitudes on the treatment
coecients drop among all mobility rates and incomes except the high-income (200-500% of AMI)
households. This result is expected since the development of LA Metro stations often coincided with
simultaneous construction of nearby housing units. Since our coecient captures the amalgamation
of various factors that coincide with the opening of rail stations, controlling for one of the factors
should reduce the magnitude of our estimates. However, we also note that the inclusion of housing
variables does not alter the signicance of estimated eects except in case of 200-500% of AMI
households. Along the Gold line stations opened in 2009, once we incorporate housing variables,
we see that their out-mobility rates decreased by 4% suggesting that, at least along the Gold line,
the additional amenities outside of housing, such as access to public transit, may have induced
more high-income households to stay in that area. Note, this does not hold along the Expo line.
29
Table 9: DID Robustness Results with Sub-Sample Variation
dep var movedOut movedOutOut movedIn movedInIn
ind var nobs open:treated nobs open:treated nobs open:treated nobs open:treated
model line-year AMI
base Expo 30 120642 -0.002 (0.0033) 120642 -0.001 (0.0027) 120642 0.0 (0.0030) 120642 0.002 (0.0027)
50 108699 -0.003 (0.0036) 108699 -0.001 (0.0032) 108699 -0.007* (0.0035) 108699 -0.004 (0.0022)
80 79095 0.005 (0.0045) 79095 0.003 (0.0036) 79095 0.0 (0.0029) 79095 -0.0 (0.0029)
200 72142 0.005 (0.0037) 72142 0.003 (0.0034) 72142 -0.0 (0.0018) 72142 -0.0 (0.0015)
500 9099 -0.013* (0.0051) 9099 -0.017*** (0.0045) 9099 -0.0 (0.0111) 9099 0.002 (0.0114)
Gold2003 30 117458 -0.005 (0.0026) 117458 -0.004 (0.0022) 117458 0.006* (0.0029) 117458 0.005 (0.0029)
50 96889 0.011*** (0.0028) 96889 0.01*** (0.0025) 96889 0.004 (0.0024) 96889 0.004* (0.0017)
80 83683 0.001 (0.0041) 83683 0.001 (0.0032) 83683 -0.005 (0.0039) 83683 -0.005 (0.0033)
200 107830 0.007 (0.0045) 107830 0.004 (0.0038) 107830 0.003 (0.0026) 107830 0.001 (0.0023)
500 32243 0.006 (0.0054) 32243 0.001 (0.0047) 32243 0.002 (0.0053) 32243 -0.003 (0.0045)
Gold2009 30 127608 -0.001 (0.0017) 127608 0.0 (0.0012) 127608 -0.002 (0.0020) 127608 -0.003 (0.0015)
50 113343 0.001 (0.0027) 113343 0.001 (0.0011) 113343 0.001 (0.0037) 113343 0.002 (0.0025)
80 76919 0.003 (0.0031) 76919 0.001 (0.0034) 76919 0.0 (0.0022) 76919 -0.001 (0.0023)
200 37858 0.003 (0.0042) 37858 0.002 (0.0045) 37858 0.005 (0.0032) 37858 0.003 (0.0037)
500 1062 -0.016 (0.0146) 1062 -0.02 (0.0143) 1062 -0.006 (0.0191) 1062 -0.009 (0.0176)
Red1999 30 105259 -0.003 (0.0046) 105259 0.002 (0.0038) 105259 -0.002 (0.0079) 105259 -0.001 (0.0065)
50 82665 -0.009 (0.0056) 82665 -0.007 (0.0054) 82665 -0.003 (0.0067) 82665 0.001 (0.0061)
80 59384 0.004 (0.0039) 59384 0.004 (0.0035) 59384 0.005 (0.0061) 59384 0.009 (0.0049)
200 51450 0.0 (0.0046) 51450 0.002 (0.0042) 51450 0.004 (0.0048) 51450 0.005 (0.0063)
500 9808 0.016 (0.0187) 9808 0.013 (0.0184) 9808 0.02** (0.0066) 9808 0.02 (0.0103)
Red2000 30 38179 -0.012* (0.0053) 38179 -0.009*** (0.0017) 38179 -0.009*** (0.0008) 38179 -0.007* (0.0030)
50 35896 0.003 (0.0032) 35896 -0.001 (0.0007) 35896 0.003 (0.0123) 35896 -0.001 (0.0094)
80 35307 -0.003 (0.0070) 35307 0.001 (0.0067) 35307 -0.01 (0.0085) 35307 -0.007 (0.0058)
200 43753 0.004 (0.0050) 43753 0.003 (0.0032) 43753 -0.004 (0.0094) 43753 -0.011 (0.0091)
500 10399 0.021*** (0.0043) 10399 0.023* (0.0114) 10399 0.012 (0.0127) 10399 0.009 (0.0082)
superlong Expo 30 47763 0.002 (0.0072) 47763 0.001 (0.0064) 47763 -0.008 (0.0040) 47763 -0.006 (0.0036)
50 66200 -0.002 (0.0043) 66200 -0.003 (0.0042) 66200 -0.004 (0.0033) 66200 0.0 (0.0028)
80 56136 0.008 (0.0043) 56136 0.006 (0.0037) 56136 0.004 (0.0027) 56136 0.003 (0.0024)
200 54969 -0.0 (0.0047) 54969 0.001 (0.0036) 54969 -0.004 (0.0026) 54969 -0.003 (0.0025)
500 6937 -0.004 (0.0056) 6937 -0.005 (0.0058) 6937 0.0 (0.0123) 6937 -0.001 (0.0113)
Gold2003 30 45515 -0.012*** (0.0025) 45515 -0.008** (0.0027) 45515 0.001 (0.0029) 45515 0.002 (0.0026)
50 53643 0.008 (0.0050) 53643 0.007 (0.0047) 53643 -0.003 (0.0026) 53643 -0.003 (0.0022)
80 52669 0.004 (0.0065) 52669 0.003 (0.0058) 52669 -0.004 (0.0042) 52669 -0.004 (0.0028)
200 70658 -0.001 (0.0053) 70658 -0.0 (0.0051) 70658 0.0 (0.0026) 70658 0.001 (0.0024)
500 22537 -0.001 (0.0072) 22537 -0.001 (0.0072) 22537 -0.004 (0.0081) 22537 -0.005 (0.0075)
Gold2009 30 53581 -0.001 (0.0039) 53581 -0.001 (0.0024) 53581 -0.003 (0.0030) 53581 -0.005 (0.0026)
50 72537 0.0 (0.0028) 72537 -0.0 (0.0029) 72537 0.0 (0.0030) 72537 0.003 (0.0025)
80 58681 0.003 (0.0029) 58681 0.002 (0.0028) 58681 -0.001 (0.0026) 58681 -0.002 (0.0020)
200 30742 0.002 (0.0048) 30742 0.003 (0.0054) 30742 0.005 (0.0038) 30742 0.003 (0.0042)
500 794 -0.017 (0.0199) 794 -0.019 (0.0249) 794 -0.021 (0.0285) 794 -0.018 (0.0214)
Red1999 30 36783 -0.001 (0.0075) 36783 0.0 (0.0049) 36783 -0.015 (0.0109) 36783 -0.012 (0.0112)
50 40971 -0.013 (0.0072) 40971 -0.006 (0.0073) 40971 -0.008 (0.0067) 40971 -0.005 (0.0098)
80 33038 0.009 (0.0092) 33038 0.009 (0.0098) 33038 0.011 (0.0075) 33038 0.018*** (0.0036)
200 31388 -0.003 (0.0114) 31388 -0.003 (0.0111) 31388 -0.005 (0.0047) 31388 -0.004 (0.0042)
500 6525 0.016 (0.0168) 6525 0.022 (0.0145) 6525 0.035 (0.0226) 6525 0.041 (0.0277)
Red2000 30 11929 0.004 (0.0124) 11929 0.001 (0.0122) 11929 -0.009 (0.0189) 11929 -0.014 (0.0163)
50 14791 0.003 (0.0053) 14791 0.001 (0.0047) 14791 0.014*** (0.0044) 14791 0.007 (0.0057)
80 16507 -0.012*** (0.0031) 16507 -0.009 (0.0055) 16507 -0.01*** (0.0020) 16507 -0.005 (0.0032)
200 23277 -0.008 (0.0050) 23277 -0.008 (0.0049) 23277 -0.007 (0.0079) 23277 -0.013 (0.0093)
500 6587 0.036* (0.0164) 6587 0.037 (0.0204) 6587 0.011 (0.0205) 6587 0.01 (0.0186)
30
Table 10: DID Robustness Results with Housing Variables
dep var movedOut movedOutOut movedIn movedInIn
ind var model line-year AMI
open:treated late-base Expo 30 -0.004 (0.0033) -0.002 (0.0025) -0.001 (0.0028) 0.001 (0.0027)
50 -0.006 (0.0035) -0.004 (0.0033) -0.009* (0.0042) -0.005* (0.0022)
80 0.0 (0.0051) -0.001 (0.0043) -0.002 (0.0031) -0.001 (0.0026)
200 0.001 (0.0038) 0.001 (0.0029) -0.003 (0.0029) -0.002 (0.0017)
500 -0.01 (0.0090) -0.015 (0.0085) 0.001 (0.0176) 0.003 (0.0174)
Gold2009 30 -0.002 (0.0026) 0.001 (0.0022) -0.002 (0.0042) -0.001 (0.0023)
50 0.004 (0.0031) 0.003 (0.0019) 0.002 (0.0049) 0.005 (0.0032)
80 0.001 (0.0045) 0.0 (0.0044) 0.0 (0.0017) 0.001 (0.0024)
200 0.003 (0.0044) 0.004 (0.0043) 0.006 (0.0050) 0.006 (0.0051)
500 -0.037 (0.0217) -0.046 (0.0303) -0.023 (0.0310) -0.039 (0.0285)
lineYear-housing Expo 30 -0.003 (0.0041) -0.001 (0.0030) -0.0 (0.0028) 0.001 (0.0026)
50 -0.006 (0.0040) -0.003 (0.0037) -0.009* (0.0040) -0.005* (0.0020)
80 0.0 (0.0054) -0.001 (0.0047) -0.001 (0.0032) -0.001 (0.0027)
200 0.001 (0.0039) 0.002 (0.0029) -0.004 (0.0032) -0.003 (0.0022)
500 -0.012 (0.0092) -0.018 (0.0093) 0.003 (0.0183) 0.005 (0.0177)
Gold2009 30 -0.001 (0.0026) 0.001 (0.0019) -0.001 (0.0045) -0.002 (0.0029)
50 0.003 (0.0040) 0.003 (0.0026) 0.001 (0.0050) 0.005 (0.0031)
80 0.002 (0.0041) 0.001 (0.0043) 0.0 (0.0014) 0.001 (0.0024)
200 0.002 (0.0050) 0.003 (0.0050) 0.006 (0.0047) 0.006 (0.0048)
500 -0.041* (0.0201) -0.048 (0.0325) -0.015 (0.0339) -0.032 (0.0322)
5.6 Incumbent Households
When we estimate model 11 only on households who have been observed for 3+ years and
resided in the transit area before the lines opened, we nd that transit-opening either did not aect
move-out rates or induced households to remain in the station area (table 11). Along the Expo line,
our estimates suggest that incumbent household move-out rates did not change once the stations
opened. Along the Red 1999 stations, our estimates suggest that very-poor incumbents reduced the
rate at which they move out of rail by 1.5pp. Along the Red 2000 stations, extremely-poor, poor,
and middle income households reduced their move-out rates by 1.8pp, 1.5pp, and 1.5pp respectively.
Along the Gold 2003 and 2009 stations, extremely-poor households reduced their move-out rates
by 1pp and .3pp respectively. The only anomaly are the very-poor households along Gold 2003
stations whose move-out (move-outout) rates increased by .7pp (.6pp).
31
Table 11: DiD Results on Incumbent Households
dep var movedOut movedOutOut
ind var model line-year AMI
open:treated incumbents Expo 30 0.002 (0.0026) 0.002 (0.0024)
50 -0.003 (0.0037) -0.001 (0.0032)
80 0.005 (0.0041) 0.003 (0.0036)
200 0.003 (0.0035) 0.002 (0.0030)
500 -0.005 (0.0046) -0.01 (0.0053)
Red1999 30 -0.003 (0.0056) 0.004 (0.0047)
50 -0.015* (0.0071) -0.01 (0.0073)
80 0.004 (0.0050) 0.007 (0.0061)
200 -0.008 (0.0072) -0.004 (0.0042)
500 -0.001 (0.0148) -0.002 (0.0141)
Red2000 30 -0.018*** (0.0031) -0.012*** (0.0030)
50 -0.006 (0.0090) -0.008 (0.0067)
80 -0.015*** (0.0037) -0.008 (0.0067)
200 -0.015*** (0.0040) -0.012** (0.0043)
500 0.004 (0.0087) 0.006 (0.0145)
Gold2003 30 -0.01*** (0.0026) -0.008*** (0.0019)
50 0.007** (0.0026) 0.006* (0.0025)
80 -0.001 (0.0033) 0.002 (0.0027)
200 -0.003 (0.0036) -0.002 (0.0036)
500 -0.009 (0.0050) -0.01* (0.0050)
Gold2009 30 -0.003* (0.0015) -0.002 (0.0011)
50 0.001 (0.0019) 0.0 (0.0016)
80 0.003 (0.0036) 0.001 (0.0039)
200 0.001 (0.0044) 0.001 (0.0049)
500 -0.012 (0.0225) -0.018 (0.0178)
5.7 Long-term impact
One argument on lack of evidence of gentrication from positive shocks to neighborhoods is that
it may take over 10 years to see the impact on a neighborhood. To test this hypothesis, we re-run
model 11 with 2-3 years of lags and 14-15 years of leads on treatment for the Red line stations
opened in 1999 and 2000. Note, all impact estimates are relative to one-year prior to station
opening. Results can be observed in gures 13 and 14 for the 1999 and 2000 stations respectively.
For this analysis, however, we reference gures 6 and 7 as a reminder that pre-trends along the
Red-line stations opened in 2000 are not always parallel so any lag will pick those deviations up.
As a result, we caution the reader to not read too much into the below results.
Most estimated trends show little to no time-variation in the station opening's eect on mobility
rates. Along Red 1999 stations (gure 13), the move-in rates among very-poor and middle-income
households begin to increase 9 and 6 years after station opening. Along Red 2000 stations (14),
the very-poor household move-out rates seem to increase 4-5 years after a rail-station opens.
32
Figure 13: DiD Results with Deep Lags
33
Figure 14: DiD Results with Deep Lags
5.8 Impact of mobility changes on income distribution
We estimated whether mobility-rates had any impact on the income distribution trends using
model 10 and show the results in table 12. As discussed before, the diculty of this model stems
from the confounded nature of
4
as it captures both changes on mobility rates and their impact on
the income distribution. Although our results on residential mobility rates imply modest impact
from transit stations, table 12 suggests that even our modest impact estimates did not alter the
income distribution trends of rail transit basins. The only case in which we see a change are
middle-income households along Gold 2003 stations. However, neither direct estimates on the
income distribution trends (table 6) nor mobility patterns (tables 8 & 10) show any shifts in
income distribution trends or mobility rates for these middle-income households. This implies
that although mobility-rates did not change once the station opened, the impact of middle-income
move-out patterns doubled once stations opened.
34
Table 12: Impact of mobility changes on income distribution
dep var dstrShare
ind var condNum movedIn:open:treated movedOut:open:treated nobs rsqrd
model line-year AMI
bare bareExpo2012 30 2825.18 -0.829 (1.1284) -1.635 (1.2843) 38 0.9709
50 2958.26 -0.717 (1.1877) 0.354 (0.9834) 38 0.7719
80 3753.78 1.019 (1.2904) 0.141 (0.6709) 38 0.9745
200 28387.2 -4.796 (8.0579) -3.235 (6.1575) 38 0.9467
500 5320.26 0.35 (0.5001) -0.094 (0.1589) 38 0.9453
bareGold2003 30 2654.52 0.263 (1.2513) 1.144 (1.1234) 38 0.9509
50 3335.61 1.462 (0.9770) -0.178 (0.5406) 38 0.8816
80 2650.25 -0.669 (0.4065) -0.014 (0.2380) 38 0.979
200 3001.61 1.0 (0.6199) -2.76** (0.9618) 38 0.8927
500 2052.73 -0.117 (0.2579) 0.171 (0.1955) 38 0.9828
bareGold2009 30 3297.59 2.612 (2.6214) 2.813 (2.7437) 38 0.9598
50 2719.49 -0.352 (1.0894) -0.224 (0.7446) 38 0.874
80 1949.16 -0.297 (0.7526) -0.244 (0.5154) 38 0.9647
200 2432.73 -0.607 (0.6119) 0.719 (0.3778) 38 0.9829
500 5123.38 0.003 (0.0103) 0.0 (0.0070) 38 0.8447
bareRed1999 30 8423.79 1.91 (1.8830) -0.498 (1.7357) 38 0.9827
50 11585.6 0.901 (2.8405) -1.86 (2.5549) 38 0.9376
80 3195.5 0.772 (0.4321) -0.818 (0.4183) 38 0.9233
200 8843.76 0.387 (2.3185) 0.348 (0.6660) 38 0.9825
500 4192.83 0.015 (0.1868) -0.338 (0.3367) 38 0.9913
bareRed2000 30 3162.29 -0.342 (0.8672) -1.532 (0.9975) 38 0.9416
50 1750.63 0.623 (0.5098) 0.001 (0.4414) 38 0.8747
80 3920.68 1.683 (1.2007) -0.78 (0.6740) 38 0.8218
200 5640.37 0.794 (1.3753) -1.835 (0.9840) 38 0.9398
500 1627.92 -0.198 (0.1937) -0.287 (0.2233) 38 0.9093
6 Conclusion & Discussion
Although the robustness checks show a bit of instability on the signicance of our estimated
residential mobility impacts, amid the consistent signs and magnitudes there are a few consistent
patterns that emerge within lines.
For extremely-poor households, mobility pattern shifts are consistent with both preference for
transit and gentrication though along most stations, they mobility patterns did not seem to
change. Along Gold 2003 stations, their move-in rate increased which is consistent with preferences
for transit while along Red 2000 stations, both the move-out rate and the move-in rate decreased
by similar magnitudes. The slow in household turnover and magnitudes are both suggestive of a
preference for public transit access. Only along the Gold 2009 stations do we see a decrease in
move-in rates which is consistent with gentrication.
For very-poor households, we see mobility pattern shifts consistent with both gentrication and
preference for transit. Along the Expo line, very-poor households slowed their move-in rate once
the stations opened which is consistent with gentrication. Along the Gold 2003 stations, their
move-out rate increased which is also consistent with gentrication. However, along the Red 2000
stations, their move-in rate increased which is more consistent with preferences for public transit.
35
Poor household mobility patterns were mostly unreactive to station opening. Along the Red
line 1999 stations, the move-in rate among poor households increased which is a sign of preference
for transit but it decreased along Red line 2000 stations which is consistent with gentrication.
Among high-income households, we see mixed results on move-out rates as it increases along
some lines-years but decreases along others. Move-in rates among high-income households, however,
either do not change or increase.
Categorically, the above results suggest that low-income households are not a monolith, at least
in regard to their response in residential mobility patterns to new transit. The lowest income
households and low-income incumbent households responded in a manner consistent with the con-
clusions of (Glaeser, Kahn, and Rappaport 2008) that suggest that low income households prefer to
locate near public transit. Households in the very-poor category, on the other hand, responded in
the manner consistent with gentrication narratives in which move-out rates increase in response
to positive neighborhood shocks. We suspect this may partly be due to a smaller reliance on
public transit among these households as rising car-ownership among LA low-income households
suggests (Manville, Taylor, and Blumenberg 2018). Finally, we observe that households earning
over 50% of AMI largely did not alter their residential mobility patterns in response to transit
station opening. We suspect that higher car-ownership and higher income makes these households
less responsive to changes in access to public transit and positive amenity shocks on housing prices.
However, this insight suggests room for future research. For example, along the Expo line, we ob-
serve an increased net-out
ow of very-poor households but no increased out
ow of extremely poor
households. If extremely-poor households are able to remain in the neighborhood but very-poor
households are not then it raises the question of whether the displacement mechanism is through
prices or preferences. We leave this question to future research.
Our results do suggest that the opening of rail stations modestly aect residential patterns but
even these shifts have no discernible impact on neighborhood income distribution trends. Given that
we show how similar income distribution trends of the transit neighborhoods are when compared
to control neighborhoods and Los Angeles county-level trends, this is not very surprising. This
suggests that income distribution changes near transit neighborhoods are not unique to transit and
likely re
ect the impact of the same factors that aect income distributions in urban LA County
more widely.
The heterogeneity of impact on mobility rates across stations suggests that the eect of rail
transit is highly localized. We suspect that the way local authorities and developers choose to
develop the land and mitigate the eects of price increases near rail transit likely has more eect
on how transit stations impact local neighborhoods than the opening of rail transit itself. Prior
reports suggest that property development is not uniform across rail stations and the fact that
rail transit crosses numerous municipal boundaries inevitably rides into the issue that some cities
penalize developers with long delays or have laws such as rent control that mitigate the impact
of increased amenity qualities (Boarnet, Rodnyansky, and Boarnet 2020). Future research on rail
transit induced displacement must address the issue of local housing development patterns, city
ordinances, and household preferences. Our study does not observe housing stock, quality, or price
at a geographical level suciently ne to address the particular factors that could impact TOD
gentrication.
There is the potential that we do not see strong eects on mobility rates because the impact
of new rail-transit takes a long time to unfold; hence, not observed in the data. We view that as a
potential opening for future studies though our tests on the Red 1999 and Red 2000 stations do not
36
show much time-variation in the impact on mobility rates. The other potential argument of the
small results is that rail-station neighborhoods are dened by a radius larger than 800 meters, say
1200 meters. Although not shown, we examined summary statistics for rail station neighborhoods
dened by 1200m and the results are similar to the 800m setup.
Our results are consistent with other prior studies that document the preference of low-income
households to reside in areas with public transit options (Dragan, Ellen, and Glied 2019, Glaeser,
Kahn, and Rappaport 2008, Kahn 2007). Beside clear evidence of extremely-poor household
mobility-rate patterns skewing toward staying near transit, one reason why we may not observe
a large displacement of low-income households once rail transit opens is that poor households are
the ones that prefer or benet most from having access to rails. The persistently large share of
extremely and very poor households locating near rail is also consistent with this hypothesis. Al-
ternatively, there is the potential that in Los Angeles, the convenience or necessity for cars is so
high that access to rail transit does not oer a suciently high impact amenity value to price out
low-income households. In fact, a 2018 report on transit ridership in Los Angeles nds that transit
ridership has been falling primarily due to increased access to vehicles for low-income households
(Manville, Taylor, and Blumenberg 2018). It is also worthy to mention that in the working-paper
Aron-Dine and Bunten (2019), the authors nd displacement and gentrication eects using setups
that are similar to ours.
We wish to remind the reader that looking at mobility rates and income distribution shares may
miss important perspectives. For example, consider 100 low-income households in a neighborhood of
150 households in yeart. If 50 additional units are opened in year t+1 and are occupied by non-poor
households then the share of poor-households goes down from 100/150 to 100/200 even though the
neighborhood experienced no change in the number of poor households. Whether this is detrimental
is normative and may depend on the trade-o between perception of one's neighbors, neighborhood
amenities, and rents. Due to this, characterizing change in income distribution and observing
losses in the share of low-income households may paint an insucient picture of neighborhood
change. Similarly, we provide no insights into the impact of income changes on neighborhood
income distributions. Finally, we wish to emphasize that without observing housing prices, we
cannot make statements on household welfare. Even if an extremely-poor household decides to
remain in a rail transit neighborhood, if housing prices rise then the household must substitute
away from non-housing consumption such as food, health-care, or schooling or alternatively borrow.
A recent study on how low-income households cope with rising price shows evidence of substitution
that may be welfare-reducing (Rosen, Angst, Gregorio, and Painter 2019).
37
Chapter 2
Does size matter for land barons? The existence of local monopolies and
their eects on rents
by
Evgeny Burinskiy
Anecdotal evidence suggests some landlords collect rents that are 50% or more above cost in
low-tier rental housing markets. By contrast, their tenants often pay 50-100% of their income
towards rent. Census data conrm the possibility of high markups but only in neighborhoods
with high levels of low-income households. I investigate local market concentrations as a possible
mechanism behind landlords' ability to maintain high markups using ownership records from
the ocial Milwaukee city property assessor database. As expected, I nd near-0 Herndahl-
Hirschmann Indexes (HHIs) at the city level suggesting that the ownership structure is fairly
competitive at large geographic scales. However, HHIs suggest the existence of market con-
centrations at the neighborhood level, which, given sucient market segmentation, can lead
to higher rents. I test whether neighborhood-level concentrations have an impact on rents by
pairing ownership concentration indexes with CoStar data on rents. Instrumenting ownership
concentrations using shift-share instruments of business and wage incomes between 2007 and
2019 in the landlord's neighborhood, I nd that a 10% increase in ownership concentration leads
to a 1.3-2.1% increase in rents. This eect seems independent of discrimination and search fric-
tions. Comparison between low and high income neighborhoods suggests market concentration
dierences explain about a third of the mark-ups.
38
1 Introduction
We tend to assume a competitive ownership structure in the housing market but, at least for
rental housing, there is reason to be skeptical. US metropolitan housing markets are often spatially
stratied by socio-economic status and race which suggests the existence of sub-markets within
MSAs. As a result, it is feasible that high market concentrations in sub-markets with geographically
constrained households may lead to the ability of landlords to extracts rents. At least in the case
of Milwaukee, Matthew Desmond's anthropological work hints at the possibility of rent extraction
by citing high price to cost ratios, low vertical rent-spreads between sub-markets (Desmond 2016,
p. 151), and large rental unit portfolios in low-end Milwaukee neighborhoods (Desmond 2016,
p. 152). Spurred by this evidence, I use ownership data from the Milwaukee tax-assessor database
to assess whether there are localized market concentrations and if they lead to higher rents. Given
the persistence of market segmentation by income and race, rent extraction by large landlords may
bear signicant policy and social implications, especially on the low-income households who tend
to be rent-burdened.
While Desmond cites an instance of a high mark-up in a low-tier Milwaukee rental market, I
generalize his observation using Census data and show the problem is unique to low-income neigh-
borhoods. According to Desmond, "Sherrena collected roughly $20,000 in rent...[She] gured she
netted roughly $10,000 a month"(p. 151) which suggests a mark-up of 50% over her costs including
mortgage payments, maintenance, and property taxes. For a poignant perspective, consider that
Sherrena primarily operated properties in which her tenants were paying 80-100% of their income
towards rent. The welfare implications of lower rents for these households cannot be overstated
since any income not lost to rent can go to cover utilities, internet, food, schooling, health care,
and more. By contrast, if markups were as high as they are in higher-income neighborhoods, the
average low-income renter would have an additional $300 for monthly disposable income (Table 1).
Using ACS data, I estimate median mark-ups across the income distribution as the dierence
between a block-group's median rent and the estimated monthly mortgage payment on the median
home value
10
to show that 50% markups are not an anomaly and that the problem is unique to
neighborhoods with high shares of low-income households
11
(Table 1). Namely, when we examine
block groups in which low-income households comprise less than 10% of residents, the mean and
median mark-up is 8% and 13%, respectively, but when we examine neighborhoods in which over
30% of residents are low-income, the mean and median markups are 38-45% and 50-51%. Why are
these high markups not being competed away in low-income neighborhoods?
10
Assuming 20% down and 3% interest rate. Also includes a 2.53% property tax for Milwaukee
11
Low-income households are dened as households earning $25,000 or less which is, given Milwaukee's Area Median
Income of $45,000 is roughly 60% of AMI and is considered very-poor by HUD
39
Table 1: Estimated Mark-Ups by Neighborhood Income
Share of Low-Income Households [0.0-0.1) [0.1-0.2) [0.2-0.3) [0.3-0.5) [0.5-1.0)
stat
numb obs 87.0 139.0 113.0 144.0 42.0
med income 5496.0 4348.0 3113.0 2168.0 1381.0
med rent 911.5 891.0 842.0 801.5 709.0
med cost 792.5 692.0 476.0 392.0 346.5
mean prot margin 7.9 18.2 37.9 38.1 45.3
25th percentile prot margin -8.7 1.0 22.7 26.1 34.9
50th percentile prot margin 13.4 23.3 44.4 50.3 51.6
75th percentile prot margin 30.5 39.5 57.3 63.5 66.8
max prot margin 72.6 75.4 84.8 83.6 83.4
median renter share 24.9 29.5 35.7 39.7 42.7
mean multi-family HHI 0.0 0.0 0.0 0.0 0.1
Assessment of ownership concentrations among multi-family properties in Milwaukee, WI and
other select U.S. metropolitan areas suggest neighborhood-level market concentrations exist. Using
assessor records of residential parcels and their owners, I infer ownership shares at neighborhood
levels using shared owner addresses and compute Herndahl-Hirschmann Indexes (HHIs) accord-
ingly. High levels of ownership concentration among multi-unit properties are not wide-spread but
exist. In Milwaukee, 57 of 576 examined block groups have concentrations that can be considered
high. Moreover, Milwaukee HHIs regressed on demographic and housing data from the Ameri-
can Community Survey (ACS) show that areas with higher ownership concentrations tend to have
higher proportions of renter households, lower vacancy rates, and, higher shares of low-income
households. These correlations are consistent with a context in which landlords with large unit
portfolio concentrations could extract rents and maintain high markups over cost.
Results suggest that for Milwaukee City, a 10% increase in market concentration leads to a 1.4-
2.1% increase in rents but market concentrations cannot explain the maintenance of high markups
in low-income neighborhoods entirely. To test whether there is a relationship between rents and
market concentration, I use multi-family CoStar data on rents and unit locations. Since neither
residents nor landlords randomly sort into neighborhoods, I instrument market concentrations
using the 2007 business and wage income shares of landlord neighborhoods. Results suggest higher
ownership concentrations raise neighborhood-level rents.
The paper describes past literature and motivation in section 2, methods of estimation including
the choice of instruments in section 4, the data used for the study in section 5, estimation results
in section 6, and enumerates study limitations and future paths in section 7.
2 Motivation
Empirical studies on the role of market concentration in the determination of rents are scant
and, to my knowledge, no work examines the role of market power within metropolitan areas
derived from the housing market's spatial segmentation. Cosman and Quintero (2018) examine
the role of concentration in the housing construction industry and nd that large concentrations
exist and have been increasing over time. Moreover, this leads to greater price volatility, decreased
40
construction, and fewer unsold units. Though their analysis focuses on the impact of concentration
on business cycle dynamics, their ndings are consistent with monopoly predictions. Sommerville
(1999) examines the distribution of home-builder size and their shares across MSAs. He nds that
areas with more stringent regulation and smaller-regulatory districts tend to have less concentration
but the author does not examine any price eects. A yet unpublished study examines concentration
in multi-unit residential buildings in Manhattan, New York in a BLP framework (Watson and
Ziv 2019, Berry, Levinsohn, and Pakes 1995). They nd that monopoly power derived markups
are, on average, 20-21% of rents and that market concentrations are positively associated with rents
until building unobservables are accounted for; thus, suggesting that markups are more likely due
to building and neighborhood unobservables. However, in this case, the authors measure landlord
market power that is derived from the monopoly ownership of a particular parcel and they do not
assess the potential role of market segmentation.
Evidence of spatial sorting in the housing market is abound though primarily revolves around
socio-economic attainment and racial sorting (select studies: (Bayer, Fang, and McMillan 2014,
Glaeser, Kahn, and Rappaport 2008, Berry 1976)). As a result, if there are frictions in a household's
ability to move between spatially-segmented markets due to preferences (Bayer and McMillan 2005,
Glaeser, Kahn, and Rappaport 2008) or discrimination (Bayer, Casey, Ferreira, and McMillan 2017,
Bosch, Carnero, and Farre 2010, Ondrich, Stricker, and Yinger 1999, Yinger 1986, Courant 1978)
then a property owner does not need to corner the entire metropolitan housing market to exert
market power but rather just a small segment thereof.
However, beside preferences and discrimination there is also the potential role of information
constraints. Online search tools such as Zillow and Craigslist can help households transcend spatial
boundaries to see available products elsewhere so it is dicult to envision a household encountering
information constraints. Economic studies of how households search for rental units are scant but
existing surveys, audits, and interviews suggest that while middle and high-income or educated
households often nd housing online, low-income households have less discretion over when and
where they move and are more likely to rely on social-networks, prior landlords, and rent-signs
to nd housing (Desmond 2016, Desmond, Gershenson, and Kiviat 2015, Skobba and Goetz 2013,
Hanson and Hawley 2011, Palm and Danis 2001). To assess the generalizability of these qualitative
studies, I brie
y document low take-up of online tools among low-income households using American
Housing Survey data.
However, even if internet use across socio-economic groups were uniform, Boeing (2019) shows
that online listings themselves over-represent whiter, more educated, and wealthier neighborhoods
and under-represent neighborhoods of lower socio-economic attainment. Were low-income house-
holds to use the internet, they would not necessarily nd the housing opportunities that cater to
them and their sense of price to quality schedule in the market may still be inadequate.
These information and geographical frictions among certain household subsets imply that con-
trolling for market concentrations alone may not be sucient to capture rent-extraction. Since
these frictions do not uniformly impact all tenants, tenants more burdened by these frictions are
more likely to be charged abnormally-high rents. To capture this, I study the role of market concen-
tration on its own and its interaction with proxies that aim to capture some of the aforementioned
frictions.
41
3 Housing Search Frictions
In the 2015 AHS, a bit over 9,000 renter households who moved within the past 2 years had
their method of search recorded. I t a logistic regression on whether a household found its unit
online to the household's log-income, age, whether it is a female-headed household, whether the
head of house has a college or higher education, whether it is married, black, Asian, the number of
adults in the unit, and dummies for the CBSA of the unit. Unsurprisingly, coecients in table 2
show that younger, college-educated households are more likely to use online resources to nd their
dwelling. The positive and signicant parameter estimate of .34 on log-income in Table 2 suggests
that households with higher income are more likely to search for housing online. When evaluated
at the 2015 means, the average household (with an annual income of $36,600) has a .36 chance of
using online resources to nd housing. When I instead evaluate by lowering log-incomes by a point
(annual income of $13,500) then the chance of using online resources is only .26. This is likely an
underestimated dierence in chance since we are assuming all other household characteristics, such
as education, remain constant which is not likely to hold.
Moreover, the use of internet to nd housing among low-income households did not increase
substantially between 2015 and 2017. Table 3 tabulates the share of households that used a par-
ticular technology to nd housing by income quantile. In 2015, 15% and 22% of households in the
bottom 2 income quantiles used the internet to nd housing while 18% and 26%, respectively, in
2017. Given the shares, the increase in usage is almost entirely due to a switch away from use of
vacancy signs and paper publications such as newspapers.
One immediate concern with the results below is that the question asked from the tenants
pertains to how they found the unit in which they currently reside, not which technologies they
used during the search process. For example, it is possible, though not necessarily feasible, that
100% of poor households used the internet to look for housing but, for whatever reason, abandoned
online search and instead turned to alternative means to nd housing. This would mean the numbers
discussed above underestimate low-income household exposure to information on the online rental
market and overestimate the constraints on its search. I am not aware of any surveys that eschew
this conundrum.
Table 2: Income and online search in AHS
2015 2017
logincome 0.339***(0.000) 0.311***(0.000)
HHAGE -0.034***(0.000) -0.028***(0.000)
femaleHH -0.037(0.434) 0.093*(0.059)
collegeOrMore 0.672***(0.000) 0.662***(0.000)
marriedHH 0.064(0.258) 0.009(0.876)
Black -0.01(0.865) -0.047(0.453)
Asian -0.157*(0.087) -0.051(0.552)
NUMADULTS -0.142***(0.000) -0.139***(0.000)
obs 9365 8109
pseudo R 0.11 0.09
t-statistics in parentheses
*p<.1; **p<.05; ***p<.01
42
Table 3: Income and method of apartment search (National)
SEARCHFAM SEARCHLIST SEARCHNET SEARCHOTH SEARCHPUB SEARCHREA SEARCHSIGN
income quintile
2015 1 0.38 0.06 0.16 0.18 0.07 0.03 0.12
2 0.35 0.07 0.24 0.14 0.04 0.03 0.13
3 0.31 0.06 0.31 0.12 0.03 0.04 0.14
4 0.28 0.07 0.36 0.10 0.03 0.06 0.11
5 0.22 0.06 0.42 0.09 0.02 0.10 0.08
2017 1 0.38 0.06 0.18 0.20 0.04 0.04 0.10
2 0.32 0.07 0.28 0.15 0.03 0.04 0.10
3 0.27 0.07 0.34 0.12 0.03 0.06 0.11
4 0.24 0.08 0.39 0.10 0.02 0.08 0.08
5 0.22 0.07 0.43 0.09 0.02 0.11 0.07
Table 4: Income and method of apartment search (Milwaukee - AHS)
SEARCHFAM SEARCHLIST SEARCHNET SEARCHOTH SEARCHPUB SEARCHREA SEARCHSIGN
income quintile
2015 1 0.34 0.10 0.24 0.09 0.10 0.03 0.11
2 0.34 0.09 0.26 0.13 0.06 0.01 0.12
3 0.28 0.06 0.25 0.19 0.10 0.00 0.13
4 0.25 0.03 0.42 0.10 0.03 0.02 0.15
5 0.23 0.07 0.50 0.02 0.01 0.03 0.15
If landlords are aware of selection into search technology by income, then they can lter prospec-
tive tenants based on the form of advertisement they choose. A landlord catering to low-income
households will probably not advertise online since, as the above table suggests, low-income house-
holds won't be looking there anyway. Boeing (2019) shows that Craigslist listings overrepre-
sent units in whiter, wealthier, and more educated neighborhoods. Eectively, this constrains
a searcher's choice set. For example, instead of being aware of tens or hundreds of units available
at a particular price as one would through an online search, the option set found through social-
networks, print, and signs will likely oer much fewer options and, on top of that, lower quality
information regarding the unit itself.
Although indirect, the above suggests the possibility of information constraints among some
households which when paired with mobility frictions across sub-markets or geographies may allow
landlords to rent extraction or, otherwise, maintain market power.
4 Method
4.1 Herndahl-Hirschmann Index
I use the Herndahl-Herschmann Index (HHI) to gauge market concentration. I construct
HHIs based on assessor property records for select cities and counties that make their master
property data available. The data contain the location, land-use, zoning, ownership, and structure
characteristics for all properties in a jurisdiction. I exclude properties that do not have a residential
zoning or land use designation and use the owner's address and number of units on the property
to estimate HHIs. Namely, let n
ig
be the number of units owned by owner i at geography g. If
N
g
=
P
i
n
ig
is the total number of units in a given location, then market share is s
ig
=
n
ig
Ng
and
the HHI at location g is equation 13.
HHI
g
=
X
i
s
2
ig
(13)
43
An HHI that is close to 0 suggests a competitive market while an HHI of 1 implies a complete
monopoly. According to the US Justice Department, markets with a concentration of over .2
are considered to be highly concentrated
12
. The Herndahl-Hirschman Index is sensitive to the
denition of market since varying the base can signicantly alter the implication of HHI. I compute
HHIs at two levels of geography: city and Census block groups. It is commonly argued that the
housing market is competitive and, surely, when we examine the HHI at the city level, it should
be very close to 0. However, Desmond and literature on low-income households suggest that the
low-income housing market may be highly localized and span less than the entire city due to market
segmentation and mobility frictions so I examine HHIs at a neighborhood scale of block groups.
4.2 Rents, Market Concentration, and Instruments
To assess whether neighborhoods with higher concentrations of ownership exhibit higher rents
ceteris paribus, I regress market concentrations on log-rents in hedonic model 14 and its instru-
mented rendition. Let r
ig
be the rent of unit i in neighborhood g, X
i
the vector of structure
characteristics, Z
g
be the neighborhood characteristics such the distance of unit i to the CBD or
nearest employment center and share of renters, H
i
a vector of occupant characteristics, f(HHI)
g
the measure of neighborhood g's market concentration, and "
ig
be the white noise. X
i
contains
the number of bathrooms, number of bedrooms, number of stories, and categorical variables for
building type (ie 2-unit apartments, single-family detached, etc), year built, and others. CoStar
data do not have occupant-related observation so only when I use IUF American Housing Survey
data does H
i
contain the occupant's length of tenure at the current residence, age, sex, race, and
other characteristics.
The variable of interest is . If > 0 then neighborhoods with higher concentrations exhibit
higher rents, if = 0 then they don't, and if< 0 then most likely is endogenized by the quality
of the unit though there is scope for economies of scale. will also be interacted with either search
or demographic variables to see whether search frictions and market segmentation by race increase
its impact.
r
ig
= +X
i
+
Z
g
+H
i
+f(HHI)
g
+"
ig
(14)
Empirically, the highest diculty is discerning between the eect of power market and quality
on rents. Given sorting and geographic segmentation in the housing market, if neighborhoods in
which market power is concentrated coincide with unobserved neighborhood and unit qualities then
in a hedonic model the coecient on market concentration will likely be bias. In Milwaukee, high
market concentrations geographically tend to coincide with low socio-economic demographics which
also tend to coincide with low neighborhood amenities and low unit quality. As a result, an OLS
is likely to return a downward biased estimate.
To mitigate endogeneity issues, I use an instrument inspired by the ? working paper in which
the authors use 2007 business income from households earning over $100,000 as an instrument that
predicts the presence of institutional investors in the post-Recession period of 2009-2014 but that is
uncorrelated with factors that determine housing prices. Accordingly, they show their instrument
does not predict housing prices or construction and shows parallel trends for its lowest and highest
values. The authors suggest that the Quantitative Easing (QE) programs unveiled during the
nancial crisis sharply decreased returns from safe assets that, in turn, led sophisticated investors
12
https://www.justice.gov/atr/herfindahl-hirschman-index
44
to invest in the housing market. Although the program was national, there were regional dierences
in how investors reacted (?).
In that spirit, I instrument market concentrations using Bartik-like 2007 business and wage
income shares of the property holders. Property assessor data enable me to see the address at
which the property is registered and; hence, match the IRS income data of the owner's address to
each property. Equipped with this link, for each Census Block Group, I sum the property owners'
2007 business and wage incomes and divide it by the total income to get the shares. The instrument
I use is the sum of 2007 owners' business and wage shares
13
. I only use business and wage incomes
because, as I detail below, dividend incomes do not predict ownership concentrations and other
income types are not available in 2007 income data.
I also construct shift-share instruments within the MSA using wage and business income sources
as an alternative to 2007 business incomes. The shift-share instruments are constructed per equation
15 where y
t
g
is the total income for a particular income type in neighborhood g. Y
g
is the total
income in neighborhood g, and n
t
is the national change in that income type between 2007 and
2015. Heterogeneity in the instrument stems from the fact that in 2007, the initial income shares
varied across neighborhoods.
z
g
=
X
t2fbusiness;wageg
n
t
y
t
g
Y
g
(15)
However, since all shares are being multiplied by a constant, the component of the instruments
is identical to their 2007 shares in a regression setting (ie corr(fz
g
g
g
;f
y
business
g
+y
wage
g
Yg
g
g
) = 1). As a
result, shift-share instruments yield the same results as the sum of 2007 business and wage income
shares.
4.3 Proxy Frictions
To proxy information frictions, I use the predicted probability of a household using news-papers
and signs to nd housing. These are estimated using a simple Probit model with household-
level characteristics such as income, age, education, and race serving as covariates. Although
search methods are only available for recent movers, I use the parameter estimates to extrapolate
probabilities to households for whom search technology was not recorded.
5 Data
To estimate ownership shares, I use data on parcel ownership from city and county property
databases. The data contain details on ownership, zoning, land-use, number of units, and assessed
value on every parcel within a jurisdiction. The study primarily leans on the City of Milwaukee
though also demonstrates the existence of local monopolies in Los Angeles County, and Austin
City, and Wake County. For Milwaukee, to compute concentration indexes, I restrict observations
to those that have a land use code ranging from single-family to commercial real-estate
14
and to
properties with at least 2 residential units. When estimating relationship between rents and market
13
Note, the ideal instrument would perhaps entail the 2007 income shares of 2007 property owners. Data limitations
on historical property ownership prevent me from assembling such an instrument
14
Observations with land use gp code between 1-4.
45
concentrations using 2015 American Housing Survey data, I use property record data downloaded
in 2018 while when using 2021 CoStar data, I use property record data downloaded in 2021.
Although the data contain the universe of ownership knowledge, data limitations guarantee
that ownership shares will be underestimated with nigh certainty. The biggest factor is the lack of
transparency behind LLC ownership though I mitigate this slightly by using the address at which
ownership is registered since dierent LLCs by the same set of owners are often registered at the
same address. The other issue behind underestimation stems from data entry errors that, due to
the size of data, are impossible to systematically correct.
I compute concentration indexes for non-single family residential properties at the Census block
group level. To identify non-single-family units, I exclude properties with single-family land-use
designations and block groups with less than 2 properties. The distinction is vital since 70%
of Milwaukee consists of single-family homes which are most likely owner-occupied; hence, not
relevant to the rental market. As seen in the left-most boxplot in Figure 1, indexes computed
with single-family homes will dilute the relevant concentration index. Looking at the other box
plots, the majority of Milwaukee's 576 block groups with more than 1 property are solidly below
the .2 threshold with only 57 block groups having concentration indexes over .2. From a data
perspective, this suggests that the high and low concentration groups are not balanced which may
give more weight to low-concentration groups in the model estimates. One way to reconcile this
and many confounding issues would be to compare highly concentrated areas to adjacent areas.
However, as discussed below, this method is precluded by the small sample of units available for
each neighborhood.
Figure 1: Milwaukee City Market Concentration Indexes
A potential source of error stems from the fact that ownership data for Milwaukee City were
pulled in autumn of 2018 while matched AHS data are from 2015. Unless ownership did not change
between 2015 and 2018, the mismatch in time frames may introduce measurement error into the
study. For example, if neighborhood A was highly concentrated in 2015 but ownership changes
46
diluted concentration levels by 2018 then impact estimates will be downward bias. Of course,
the case holds in the other direction, too. Due to this, for the study, I assume that ownership
concentrations did not change between 2015 and 2018 when estimating using American Housing
Survey data. .
To examine search frictions, I use data from the 2015 AHS. For households who moved within 2
years of the survey, the AHS collects a nationally-representative sample on how households found
the unit they currently inhabit. Households can choose one of 7 dierent search technologies
including 1) family, friends, etc 2) rent sign, 3) newspaper 4), real-estate agent, 5) internet, 6)
apartment rental agency listing ,and 7) other. The 2015 AHS has just over 9,000 observations on
search technology.
To examine the relationship between market concentration and rents, I primarily lean on 2021
CoStar data for Milwaukee. These data contain the number of units, the average eective rent per
unit, number of units, and other property characteristic data. They also contain details such as
the latest tax expense per unit and operating costs per unit along with details such as the distance
to the nearest transit hub.
I also place restrictions on CoStar data to exclude observations that may not generalize to the
wider Milwaukee market. For this, I exclude all observations that are labeled either aordable or
senior housing. I also exclude all observations with more than 40 units which is around the 80%
quantile among all CoStar observations. Finally, as in American Housing Survey estimations, I
exclude all block groups with an HHI of 1 because such neighborhoods tend to be outliers and I
only include block groups with at least 2 properties.
47
Table 5: Sample Means with Co-Star Data
full sample restricted sample
HHImf 0.10 0.09
prop LI 0.28 0.27
Direct Vacant Space 37.71 8.02
Number Of Stories 2.84 2.40
dist2cbd 6.47 6.54
Number Of Units 30.06 11.97
Year Built 1949.98 1943.54
Star Rating 2.24 2.14
Typical Floor Size 9459.08 4675.96
Taxes Per SF 1.60 1.66
operExpense 1503.79 1546.09
propBlack 0.34 0.35
propRent 0.69 0.69
median value 161786.66 152799.43
propVacant 0.11 0.11
medLastTransYr 6.46 6.36
population density 11674.38 11941.92
propVacant 0.11 0.11
marriedShare 0.22 0.22
collegeOrMoreShare 0.31 0.30
median year structure built 1230.17 1200.72
Closest Transit Stop Dist (mi) 10.67 10.71
median age 32.77 32.04
average family income 83067.16 79389.84
To estimate my instruments, I use 2007 and 2018 IRS income data that tabulate total income
by income-type for each Zip code in the U.S. Total Zip code incomes are matched to the assessor
data using the parcel owner's Zip code address and then aggregated to the property's block group
by taking the sum of all owners' incomes within each block group. Business, wage, and dividend
income shares are computed using the aggregated totals for each block group. National changes in
48
income shares are computed as the percent dierence between 2007 and 2015 income shares across
all Zip codes
15
. Income shifts are computed on households with incomes greater than $100,000
16
though results do not categorically dier when all incomes are used instead.
Instrument relevance can be observed in Table 6. All considered instrument variations are
statistically signicant in predicting multi-family market concentrations, except in case of dividend
incomes, and F-stats are all over 10.
Table 6: Relevance Tests
bi2ti 07 comp instr 15 z15 bi2ti 07f comp instr 15f z15f
business share 07 -40.229***(12.226) -81.703***(13.088)
z15 -581.59***(176.752)
z18 -397.615***(63.695)
business instr 15 -1567.62***(216.929) -1789.961***(213.872)
wage instr 15 -39.402*(20.563) -72.222***(20.733)
div instr 15 1909.035(3574.982) -5065.167(3577.16)
other cov no no no yes yes yes
rsqrd 0.0171 0.0995 0.0171 0.1116 0.168 0.1116
nobs 565 565 565 530 530 530
Fstat 10.827(0.001) 21.767(0.0) 10.827(0.001) 12.074(0.0) 14.348(0.0) 12.074(0.0)
condNum 270.522 82817.5 3908.82 81341 2.29687e+07 395784
Standard errors are in parentheses. ***<.01, **<.05, *<.1
Other covariates include: median home value, number of multi-family properties interacted with number of multi-family units, and the share of
households who rent.
Note, ? suggest that QE predicts market concentration because it led high-income households
to shift their investments toward real estate which, at least in case of Milwaukee, does not seem
to be the case. The negative sign on business income in Table 6 suggests that highly-concentrated
neighborhoods are actually ones that had less business investment, at least in 2007. This is true
regardless whether income type shares are computed using all households or high-income households
(ie earning over $100,000) as in ?. This does not necessarily invalidate the instrument for two
reasons. First, lack of investment in 2007 suggests that property in the area was likely inexpensive
so it could lead bargain investors to buy up property in under-invested areas. Second, because
Milwaukee is not a growing metropolitan area, established investors may overlook it and; thus,
make it accessible to new investors. A brief glance at Figure 2 shows that areas with higher
concentrations tend to have more recent transactions which suggests concentrations are related
to post-Recession investment activity. Moreover, the fact that business incomes are negatively
correlated with concentrations while wages are positively correlated suggests that, at least in case
of Milwaukee, landlords who amass property holdings are likely non-institutional players. This
coincides with evidence in Evicted, in which Desmond suggests established landlords and institutions
are reluctant more reluctant to enter the low-income rental market.
15
To get annual national income shares, I sum each income type across all Zip codes available in the data and
divide it by the total income across all Zip codes. In case of 2007 data, business incomes are 'a00900', wages are
'a00200', and total income is 'a00100'. These elds change slightly across the years so in 2018, the comparable elds
are 'a00900' for business income, 'a00200' for wages, and 'a02650' for total income.
16
In the data, this is 'agi stub' 5.
49
Figure 2: Higher HHI, more recent transactions
To test instrument exogeneity, I regress the three sets of instruments on variables that are typi-
cally indicative of neighborhood quality including log average family income, the share of neighbor-
hood households who are poor, and the share of neighborhood households who are black. As seen
in table 7, 2007 business income shares predict both share of black households and neighborhood
incomes though not poverty levels. The 2015 shift-share business and wage income instruments
do not predict poverty levels or family incomes though they do predict black household shares.
Dividend income shares predict all of the considered neighborhood-quality variables so I exclude
dividend income shares in the construction of instruments. It is worth noting that although the
components of the instrument, ie shift-share business and wage incomes, do not predict incomes
and poverty levels, the instrument, z15 in the table, does correlate with these variables. It is not
entirely certain why the sum of two uncorrelated variables correlate with quality outcomes.
Table 7: Exogeneity Tests
Dependent variable
avg family income avg family income avg family income propBlack propBlack propBlack prop LI prop LI prop LI
business share 07 11.341***(3.965) 22.762***(3.604) -0.356(1.395)
business instr 15 -47.073(64.713) 119.344**(58.005) -31.128(23.396)
wage instr 15 3.655(6.265) -19.248***(5.623) 3.924*(2.268)
div instr 15 3136.892***(1084.792) -784.203(970.174) 978.859**(391.312)
z15 163.95***(57.323) 329.07***(52.099) -5.153(20.164)
other cov yes yes yes yes yes yes yes yes yes
rsqrd 0.6294 0.6294 0.6556 0.4497 0.5 0.4497 0.5509 0.5568 0.5509
nobs 529 529 529 530 530 530 530 530 530
Fstat 150.467(0.0) 150.467(0.0) 126.61(0.0) 73.051(0.0) 67.117(0.0) 73.051(0.0) 109.132(0.0) 84.064(0.0) 109.132(0.0)
condNum 1.17245e+06 81114.5 2.30146e+07 81341 2.29687e+07 1.17572e+06 81341 2.29687e+07 1.17572e+06
Standard errors are in parentheses. ***<.01, **<.05, *<.1
Other covariates include: median home value, number of multi-family properties interacted with number of multi-family units, and the share of
households who rent.
It is noteworthy to mention that the predominant driving force behind signicant results are
2007 wage shares both on their own and as an element in the shift-share instrument which may
give cause for concern. If wage and business shares are simply a proxy of tenant incomes then the
shares mirror tenant socio-economic levels which tend to be highly correlated with neighborhood
and unit quality. As a result, the income shares predict precisely what they are intended to avoid.
In the ideal case, both business and wage shares are independent of local socio-economic levels and,
are instead, only predictive of landlords' investment decisions.
50
Examination of the Zip codes that are represented in income shares suggests that income shares
predominantly represent the neighborhoods of the investors, not the tenants. Over 50% of multi-
family properties are registered to a Zip code that is dierent from its situs suggesting that income
shares are not from the property's tenants. Moreover, as seen in table 8, the share of non-local
landlords increases with the market concentration decile and the highest concentration decile also
has the highest share of non-local owners. This implies that wage and business income shares re
ect
the landlords' neighborhoods and investment capacity rather than the tenants' socio-economic
standing
17
.
Table 8: Outside ownership shares
non-local ownerr share
MF HHI decile
(0.00377, 0.0111] 0.42
(0.0111, 0.0154] 0.49
(0.0154, 0.0194] 0.51
(0.0194, 0.0245] 0.53
(0.0245, 0.0335] 0.50
(0.0335, 0.0494] 0.54
(0.0494, 0.073] 0.57
(0.073, 0.117] 0.56
(0.117, 0.206] 0.57
(0.206, 1.0] 0.59
6 Results
6.1 Rent extraction
Under most specications, estimates using CoStar data indicate higher neighborhood-level mar-
ket concentrations raise rents though results are only signicant at the 5% levels and are sensitive
to sample selection (Table 10). Coecient estimates suggest that a 10% increase in market con-
centration leads to an 1.5 - 2.1% increase in rents within the neighborhood. Although not reported
here due to result reporting restrictions, these coecient estimates are virtually identical to those
using American Housing Survey data.
The estimation procedure with OLS suers from the downward bias discussed in the section
4. Per table 9, the coecient estimate on the impact of log market concentrations on log rents
decreases from .05 to 0 as more controls are introduced. When the full suite of covariates and
restrictions are included, Column 4 suggests that higher concentrations decrease rents though,
statistically, the coecient is not distinguishable from 0. Because these estimates suggest a pretty
clear endogeneity issue, I primarily lean on the instrumented coecient estimates.
17
As a robustness check, future renditions of this work will also t the models using income shares that exclude
owners whose Zip codes coincide with the location of the property.
51
Table 9: OLS Results
18
Dependent variable:
log(rent)
(1) (2) (3) (4)
log(HHImf) 0.050

0.040

0.021 0.002
(0.014) (0.017) (0.015) (0.019)
Controls No No Some Yes
Sample restriction No Yes Yes Yes
Observations 987 626 571 479
R
2
0.013 0.009 0.322 0.560
Adjusted R
2
0.012 0.008 0.304 0.524
Note:

p<0.1;

p<0.05;

p<0.01
Variable details: HHImf is the multi-family Herndahl-Hirschmann Index at the Census block group level. Full suite of controls is here.
IV results in table 10 suggest that estimates are sensitive to the choice of instrument. In column
1, the instrument on multi-family HHIs is the shift-share dierence between 2007 and 2018 wage
and business incomes in owner neighborhoods and suggests an increase of 2.1% in rents for every
10% increase in HHI. This estimate is similar to the one in column 2 in which the instrument is
simply the 2018 business and wage shares of property owners. Interestingly, when I use the sum of
business and wage shift-shares instead of the individual components, as in column 3, the coecient
estimates blows up and becomes non-signicant.
Results on the interaction terms between market concentration indexes and proxies for dis-
crimination or information frictions do not suggest those channels aect the impact of market
concentration (Columns 4 and 5 in table 10). Estimates in both columns are statistically indistin-
guishable from 0 and, categorically, results are robust to covariate choice and sample restrictions.
Nonetheless, when the categorical variable for black households is interacted with market concen-
trations, we do observe that the coecient on market concentration is near 0 while the coecient on
the interaction term between black households and market concentration indexes is similar to non-
interacted estimates. This suggests that black households in Milwaukee may be more susceptible
to rent spikes induced by high ownership concentrations though more study is needed.
18
Full control set includes: Style, Building Class, property has features, near university, Direct Vacant Space,
Number Of Stories, dist to CBD, Number of Units, Year Built, Star Rating, Typical Floor Size, Taxes Per SF,
operating expense, prop low income, prop rent, median value, prop vacant, units per prop, median year of last
transaction in neighborhood, population density, prop vacant, married share, college or more share, median years
structure built, Construction Material, Closest Transit Stop Dist in Miles, median age, average family income,prop
black, constant. All dollar and distance variables were either log or square-root transformed.
52
Table 10: IV Results
19
Dependent variable:
log(rent)
'18 wage & biz instr '18 biz & wage share z18 biz & wage share biz & wage share
(1) (2) (3) (4) (5)
log(HHImf) 0.210

0.244

0.681 0.154

0.198

(0.106) (0.100) (0.813) (0.068) (0.085)
log(HHImf):highLI 0.049
(0.112)
propBlack:log(HHImf) 0.027
(0.133)
highLI 0.255
(0.282)
propBlack 0.076 0.080 0.268 0.030 0.023
(0.102) (0.103) (0.383) (0.093) (0.468)
Observations 479 479 479 479 479
R
2
0.433 0.382 0.787 0.482 0.457
Adjusted R
2
0.387 0.336 0.920 0.442 0.411
Note:

p<0.1;

p<0.05;

p<0.01
Variable details: highLI is a dummy variable for Census block groups in which over 30% of households make under $25,000. propBlack is the
proportion of black households in the Census block group. HHImf is the multi-family Herndahl-Hirschmann Index at the Census block group
level.
6.2 Impact estimate
To see whether the estimates can explain the high mark-ups in neighborhoods with high poverty
levels, I compare the market concentration levels between neighborhoods with the lowest (<10%)
and highest (>30%) share of low-income households
20
. Figure 3 shows that all quantiles of multi-
family HHIs are higher in neighborhoods with the highest shares of low-income households. Taking
the dierence in log HHI quantiles between these neighborhoods and applying the impact estimate
of .21 (Table 10) shows that relative to low-share neighborhoods, rents in the high-share neighbor-
hoods are up to 20% higher and an average 12.5% higher due to higher neighborhood-level market
concentrations. Given that the mark-up dierence between the low and high-share neighborhoods
is nearly 35 percentage points, this calculation suggests that higher market concentrations explain
only a third of the mark-up dierences.
19
Full control set includes: Style, Building Class, property has features, near university, Direct Vacant Space,
Number Of Stories, dist to CBD, Number of Units, Year Built, Star Rating, Typical Floor Size, Taxes Per SF,
operating expense, prop low income, prop rent, median value, prop vacant, units per prop, median year of last
transaction in neighborhood, population density, prop vacant, married share, college or more share, median years
structure built, Construction Material, Closest Transit Stop Dist in Miles, median age, average family income,prop
black, constant. All dollar and distance variables were either log or square-root transformed.
20
Recall from table 1, low-income is dened as households earning less than $25,000.
53
Figure 3: Estimated Impact on Lowest Income Neighborhoods
7 Discussion
This study nds that neighborhood-level market concentrations raise rents and the estimated
impact is able to explain about a third of landlords' ability to maintain 50% markups in low-income
neighborhoods. This suggests other channels may be at play. One channel could be high barriers
to entry. Although property prices in low-income areas are inexpensive, if property owners or
landlords are reluctant to enter the low-income rental market due to lack of experience, knowledge,
or the hassles that come with operating low-income properties then existing landlords can maintain
market power without the need to be geographically concentrated. Alternatively, because property
values in low-income rental markets are low and oer little to no appreciation, investors buying into
these markets will likely seek cash
ows over property investment income. As a result, landlords
will charge rents that make them indierent between current cash-
ows in low-value properties and
the net present value of higher-value properties. A third potential cause of the cause markups is
risk. If landlords are unsure whether they will receive rent from their tenants, they may charge a
premium on the rent that can oset the expected losses. Future work will try to distinguish the
impact of these channels.
7.1 Milwaukee is not unique
The problem of high market concentrations is not unique to Milwaukee and other cities may
be better suited to this sort of study. As can be seen in gures 4a, 4b, 4c, high neighborhood-level
ownership concentrations exist throughout U.S. metros. Austin has the highest levels of localized
54
market concentrations of the areas in comparison. In fact, among multi-family properties, the
median neighborhood HHI is over the .2 threshold. As a result, Austin City may be a more
appropriate study area as it has a more balanced ownership concentration distribution which solves
many of the biases that are incurred from data limitations in Milwaukee.
(a) Market Concentration in Austin (b) Wake County Market Concentration
(c) Los Angeles County Market Concentration
7.2 Policy Implications
Although estimates suggest that neighborhood-level ownership concentrations do not explain
the entire ability of landlords to maintain high mark-ups, they still suggest that tenants in neigh-
borhoods with the highest low-income shares could pay on average 12.5% or $88 less in monthly
rent were ownership more competitive. Given a median income and rent of $1,381 and $709, respec-
tively (table 1), such a rent saving could increase disposable incomes by 13%. The highly-localized
nature of these phenomena and the potentially ample alternative product choice that could be avail-
able to low-income households implies that achieving these savings for low-income households may
require only minor interventions such as assisting low-income households with apartment search
and, otherwise, providing them information regarding unit availability outside of their immediate
neighborhood. This is particularly salient in cities such as Milwaukee where the median rent does
not vary drastically across neighborhoods.
55
Chapter 3
Can't buy, will rent. The role of ownership barriers in determining rent
distributions by
Evgeny Burinskiy
Housing prices are often analyzed either within owner or renter markets while the relationship
between the two gets neglected. I show large variation in size-adjusted rent spreads across
MSAs whereby cities with lower barriers to home-ownership have lower rent spreads while cities
with high ownership barriers have high rent spreads. I hypothesize that this is partly due to
who is able to buy a house. Namely, in Milwaukee, a household with a moderate income can
easily buy a home so a landlord renting to this household is competing with the ownership
market. Conversely, San Francisco's high housing prices mean even high-income households
must rent; therefore, landlords catering to these households enjoy more pricing power. Relative
to Milwaukee, San Francisco's rent spread will be higher. I test this hypothesis using rent data
from the 2015 AHS and mortgage denial rates that are instrumented by Big 4 Banks' 2008
shares. Results suggest that inaccessibility to home-ownership raises rents relative to incomes
and that such barriers can explain the rent-spread dierences between MSAs.
56
1 Introduction
In Evicted, Matthew Desmond reports that a two-bedroom unit in one of Milwaukee's low-
income Zip codes with a median 2019 household income of $24,557 rents for $550 per month while
in a nearby Wauwatosa middle-class neighborhood with a median household income of $82,000, a
two-bedroom unit fetches for $750 or only 33% more even though the income dierence is nearly
230% (Desmond 2016, pg151)
21
. By comparison, the average household income in San Francisco's
94103 Zip code and corresponding rent are $128,009 and $1,386 while in Zip 94105, the income and
rents are $213,987 and $3,014. This equates to a 67% increase in income but a more than doubling
in the average rent
22
. Are there rent-to-income and rent-spread dierences between the MSAs and,
if so, why?
There is clearly high variability in rent distributions between MSAs. In gure 5, the distribu-
tion of 2 bedroom rents among recent movers in Milwaukee is markedly narrower than the rent
distribution of San Francisco
23
which suggests that Milwaukee has a narrowed rent spread. In an
alternative perspective, I regress log incomes on log rents after controlling for unit size and type to
show that Milwaukee's rent-to-income slope is about half of San Francisco's and is the 6th lowest
among the 25 compared MSAs (table 11). Although the actual slope estimates are sensitive to
sample specication
24
, Desmond's implied slope is :0034 = 200=$58; 000 which is similar to the
unrestricted slope estimate of .26 in table 26.
21
2019 household median income for Zip code 53206 comes from https://censusreporter.org/profiles/
86000US53206-53206/ while the same statistic for Wauwatosa stems from https://www.census.gov/quickfacts/
wauwatosacitywisconsin
22
Income data come from https://www.incomebyzipcode.com/california/94103 and rent data stem from https:
//www.city-data.com/zips/94103.html both of which derive their corresponding statistics from the U.S. Census
23
Rents are from the 2015 American Housing Survey for households who have lived in their residence for less than
5 years.
24
For details, please see Appendix Section 7
57
Figure 5: Histogram of demeaned 2-bedroom rents
Table 11: Rent to income slopes for 25 MSAs
logincome InvMills nObs R^2 adj condNum
Atlanta-Sandy Springs-Roswell, GA 0.1349 (0.000) 0.2010 (0.004) 623 0.32 327
Memphis, TN-MS-AR 0.1527 (0.000) 0.0898 (0.206) 652 0.3 256
Houston-The Woodlands-Sugar Land, TX 0.1720 (0.000) 0.3150 (0.000) 702 0.42 460
Dallas-Fort Worth-Arlington, TX 0.1925 (0.000) -0.0207 (0.756) 798 0.37 624
Phoenix-Mesa-Scottsdale, AZ 0.1960 (0.000) 0.0701 (0.290) 642 0.48 691
Milwaukee-Waukesha-West Allis, WI 0.1992 (0.000) 0.1011 (0.121) 614 0.4 279
Raleigh, NC 0.2056 (0.000) 0.0917 (0.196) 640 0.42 445
Portland-Vancouver-Hillsboro, OR-WA 0.2265 (0.000) 0.1459 (0.038) 615 0.44 293
Riverside-San Bernardino-Ontario, CA 0.2292 (0.000) 0.1751 (0.038) 624 0.36 485
Cleveland-Elyria, OH 0.2309 (0.000) 0.2541 (0.006) 497 0.33 476
Detroit-Warren-Dearborn, MI 0.2312 (0.000) 0.3571 (0.000) 454 0.45 324
Cincinnati, OH-KY-IN 0.2409 (0.000) -0.0601 (0.467) 483 0.42 282
Denver-Aurora-Lakewood, CO 0.2530 (0.000) -0.1741 (0.060) 648 0.45 401
Miami-Fort Lauderdale-West Palm Beach, FL 0.2590 (0.000) 0.2101 (0.026) 623 0.43 954
Washington-Arlington-Alexandria, DC-VA-MD-WV 0.2917 (0.000) 0.3993 (0.003) 671 0.31 464
Kansas City, MO-KS 0.2947 (0.000) 0.0615 (0.374) 565 0.48 382
Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 0.3083 (0.000) 0.0224 (0.858) 447 0.36 844
New Orleans-Metairie, LA 0.3088 (0.000) 0.2358 (0.016) 605 0.39 342
Los Angeles-Long Beach-Anaheim, CA 0.3202 (0.000) 0.1644 (0.136) 846 0.39 733
Boston-Cambridge-Newton, MA-NH 0.3790 (0.000) 0.4671 (0.001) 492 0.46 584
San Francisco-Oakland-Hayward, CA 0.4006 (0.000) 0.0789 (0.445) 718 0.45 583
Seattle-Tacoma-Bellevue, WA 0.4250 (0.000) -0.0048 (0.950) 715 0.54 330
Chicago-Naperville-Elgin, IL-IN-WI 0.4608 (0.000) -0.0637 (0.534) 552 0.46 597
Pittsburgh, PA 0.5406 (0.000) -0.1857 (0.151) 369 0.44 202
New York-Newark-Jersey City, NY-NJ-PA 0.5802 (0.000) -0.1640 (0.199) 563 0.52 279
p-value is in parentheses
It is also evident that the slopes are related to incomes within the MSAs since cities with
higher incomes seem to have higher slopes. This is veried in gure 6a in which we observe that
the median slope across renter incomes seems to be positively correlated with 4th quintile renter
incomes. Similarly, regressing supply elasticity estimates from Saiz (2010) on the slopes in table 11
suggests that supply elasticities explain 25% of the variation in these slopes and that the coecient
is signicant at the 10% level despite the mere 25 observations. These two factors suggest that lack
58
of access to housing among higher-income renters may explain dierences in rent-to-income slopes
across MSAs. As a result, cities with higher rent-slopes should also have more renters and the plot
of slopes from table 11 against renter shares in gure 6b bears this prediction out.
(a) Median rent dierences vs renter income (b) Rent slope against renter share
I propose rent spreads relative to income are higher in higher-cost MSAs because the barriers to
home ownership are higher. In MSAs where the barriers to becoming an owner are small, landlords
renting to households at or near the margin of a possible tenure transition have limited pricing
power because they are competing with home sellers. In MSAs with higher barriers to ownership,
landlords do not compete with home-sellers because it takes longer for renters to accumulate the
requisite wealth to buy a house. As a result, relative to low-barrier MSAs, rents among future
home-owners in high-barrier MSAs will be higher; hence, rent spreads will also be higher relative
to income. Using 2015 American Housing Survey data along with mortgage denial rates, I show
that higher barriers to ownership increase rent spreads relative to incomes. Moreover, barriers to
ownership seem to explain all of the dierences in rent-to-income slopes across MSAs.
This paper is as follows. Section 2 lays out prior literature, section 3 proposes a simple model
to motivate the empirical method, section 4 describes the hedonic and IV approaches, section 5
describes the used data, section 6 details the results, and section 7 provides discussion points.
2 Related Literature
There is a long literature on price variations between MSAs. An often cited culprit of high
prices and price variability between MSAs are local land-use regulations that articially constrain
housing supply and raise land prices (Glaeser and Ward 2009, Ihlanfeldt 2007, Green, Malpezzi,
and Mayo 2005, Quigley and Raphael 2005, Glaeser, Gyourko, and Saks 2005, Edward L. Glaeser
2005, Glaeser and Gyourko 2003, Malpezzi 1996, Katz and Rosen 1987). Natural geographical
constraints can also raise prices because they make building more dicult and because many such
features are sought-after amenities (Davido 2014, Saiz 2010).
The other side of the equation is housing demand and the role of urban agglomeration. The
concentration of high-skilled workers in particular cities raises worker productivity, skills, salaries,
and, in turn, housing prices (Moretti 2012, Berry and Glaeser 2005, Diamonds 2016).
59
Although the above literature can explain the large dierences in averages between metropolitan
areas, the models are mute when it comes to explaining dierences in the distribution of housing
prices. Additionally, the above models and empirical backing explain dierences in either rents
or housing prices but do not provide a framework under which the two can interact. This paper
intends to bridge that gap.
This work most closely follows Gete and Reher (2018) who show that a negative shock to
mortgage access following the Great Recession increased rents in the 2011-2014 period as households
moved from the ownership to the renter markets. The authors exploit the heterogenous decrease
in mortgage availability across MSAs due to increased regulatory requirements on the Big 4 banks
after the Great Recession to show that the increase in households who switch from shopping for
homes to renting increased local rents in the short run. To my knowledge, this is the rst paper to
explicitly tie constraints in the ownership market to the renter market.
This work borrows two insights from Gete and Reher (2018), mortgage denial rates and Big
4's share of deposits as an instrument. Although the authors use mortgage denial rates specically
to measure contractions in credit, I use mortgage denial rates as a proxy for barriers to home-
ownership. Since mortgage denial rates are likely endogenous, the authors follow the work of Chen,
Hanson, and Stein (2017),Goodman (2017) and, D'Acunto and Rossi (2017), to instrument denial
rates using the 2008 share of bank deposits held in each MSA by Big 4 banks. They argue and show
that this is a valid instrument for post-recession mortgage contractions as these banks were subject
to greater scrutiny and regulation from the national government which curtailed their ability to
originate mortgages. Although the legislation was federal, the non-uniform distribution of banks
across metropolitan areas yields heterogenous impact on local mortgage markets.
The contribution of this paper is twofold. First, it proposes and documents a framework for
explaining dierences between metropolitan rent distributions. Second, it provides a framework
under which the impact of credit shocks can heterogenously impact rents within metropolitan
areas.
3 Theory
Suppose that a unit's rent in a particular rental market is determined by simple linear model
(equation 16) where r
i
is the unit's rent, X
i
is a vector of unit characteristics, Z
i
is a vector of
location characteristics,H
i
is a vector of household characteristics, and"
i
is white noise error term.
r
i
= +X
i
+Z
i
+H
i
+"
i
(16)
Since households sort into units by income(Diamonds 2016, Guerrieri, Hartley, and Hurst 2013,
Bajari and Kahn 2005, Bayer, McMillan, and Rueben 2004), by holding the unit's and household's
size constant, we can approximate its quality by a household's income. Namely, given the occupant's
incomey
i
, I propose that+X
i
+Z
i
+H
i
D(y
i
) whereD(y
i
) is the household's demand for
rental housing as a function of its income. To nish up, let householdi's demand for rental housing
take on the form of equation 17 where P (home-ownershipjH
i
)(y
i
) is the probability of household i
becoming a homeowner as a function of its income conditional on household composition. I assume
that
@
@y
P (home-ownershipjH
i
)(y))> 0.
D(y
i
jH
i
) =y
i
(1P (home-ownershipjH
i
)) (17)
60
Plugging D(y
i
) into the rent equation and taking a derivative with respect to income suggests
that the slope is partly determined by households' ability to become an owner. If we decrease
P (home-ownership) across all incomes in an MSA, then the rent-to-income slope increases and vice
versa.
@r
i
@y
i
=[(1P (home-ownershipjH
i
)(y
i
)y
i
@
@y
i
P (home-ownershipjH
i
)(y
i
))] (18)
Given the rental market is segmented by income and the probability of ownership is not uniform
across incomes, the impact of home-ownership barriers will not be uniform across the rental market.
For households at the bottom of the renter income distribution with a low probability of home-
ownership, dierences in home-ownership barriers across MSAs should have little to no impact on
their rents. Higher income households with higher ownership-probabilities should see larger changes
in their rents across MSAs with dierent levels of ownership barriers.
4 Methodology
The goal is to estimate the rent-to-income slopes for MSAs and test whether access to the
ownership impacts that slope. Since the goal is not to extract or infer preferences and willingness
to pay for amenities but rather to see the structure of rent spreads relative to income within MSAs,
I lean on a simple hedonic framework for the slope estimates. To test whether ownership-access
impacts rent spreads, per the work of Gete and Reher (2018), I instrument measures of ownership-
accessibility because such metrics are likely to be confounded by the fact that neither mortgage
suppliers nor households randomly sort across MSAs.
4.1 Income-to-rent slopes
I estimate rent-to-income slopes after controlling for unit type and size with equation 19. Let
r
ig
be the rent of unit i, X
i
the vector of structure characteristics, H
i
a vector of occupant char-
acteristics, y
i
the household's income, HC
i
the Heckman correction for a household's inclusion in
the renter sample, and "
ig
be the white noise. X
i
contains the number of bathrooms, number
of bedrooms, number of stories, and categorical variables for building type (ie 2-unit apartments,
single-family detached, etc), year built, and features such as having a porch or garage. H
i
con-
tains the occupant's length of tenure at the current residence, the number of children, the number
of adults, and the household's marital status. Note, as mentioned in section 3, income replaces
quality-related unit and neighborhood characteristics; hence, the sparse and primarily size-related
control variables. I estimate equation 19 individually for each MSA with log rents and log incomes
and present the results in Table 11.
logr
ig
= +X
i
+H
i
+ logy
i
+HC
i
+"
i
(19)
4.2 Access to Ownership
To estimate the impact of home ownership barriers on rent spreads, I ideally would interact
y
i
with (1P (home-ownershipjH
i
)) as in equation 20. However, given the current diculty of
estimating MSA-level probabilities of home-ownership, I instead opt to proxy average dierences
61
inP (home-ownership) across MSAs using mortgage denial rates DR
g
as in equation 21 where
1
is
the parameter of interest.
logr
ig
= +X
i
+H
i
+ logy
i
(1P (home-ownershipjH
i
)) +HI
i
+"
i
(20)
logr
ig
= +X
i
+H
i
+
0
logy
i
+
1
y
i
DR
g
+HC
i
+"
ig
(21)
If the theory is correct then access to the ownership market should not matter as much for low-
income households who are not fortunate enough to ever own property. I verify this by estimating
equation 21 only on renter households whose incomes are in the top and bottom 25th percentiles
of renters. The lower quarter should either be insignicant or smaller in magnitude than the
among the upper 25th percentile.
Like Gete and Reher (2018), I use mortgage denial rates as a proxy for access to the ownership
market. My use of mortgage denial rates is an inverse measure of the diculty households in a
particular MSA have in purchasing a home. However, since denial rates are not only indicative of
ownership-opportunities but also unobserved features such as household sorting and their ability
to get credit, estimates using OLS are likely biased.
Following Reher and Gete, I instrument denial rates by the 2008 share of deposits for the Big
4 banks of Bank of America, Wells Fargo, Citi Group, and JP Morgan. The exogeneity of the
instruments stems from the fact that the ability of the Big 4 to originate mortgages was curtailed
nationally so it is independent of local credit shocks and household abilities to borrow. The non-
uniform distribution of banks across metropolitan areas yields the heterogenous impact on local
mortgage markets that allows the authors and me to use it as an instrument for local access to
mortgages.
As a robustness check on denial rates, I use a downpayment to income ratio as an alternative
measure of barrier to ownership access. To get the downpayment value, I assume everyone puts
down 20% on their mortgage so I divide the loan amount reported in HDMA by .8 to get the
total loan amount and then multiply that value by .2 to get the downpayment amount. To obtain
the ratio, I divide the estimated downpayment amount by the applicant's income as reported in
HDMA. The average 2015 downpayment to income ratio for each MSA is simply the average of the
individual ratios within each MSA. Higher ratios suggest that saving up for a downpayment is more
dicult or takes longer for the average household and; therefore, it is more dicult to purchase
a home. Although not shown, results are nigh identical if I were to instead divide the estimated
downpayment by the 50th percentile of renter incomes as derived from AHS data.
5 Data
Mortgage denial rates at the MSA level are computed from 2015 Home Mortgage Disclosure
Data (HDMA) data per the method used in Gete and Reher (2018). Namely, the mortgage denial
rate is the share of loans that were denied by the nancial institution among conventional home-
purchase loans for 1-4 family owner-occupied homes. There is however a caveat that comes with
these data.
Some large metropolitan areas such as San Francisco and New York are geocoded to CBSA
62
geographies that dier slightly from the geocoding of the American Housing Survey
25
. Since micro-
data are not available, it is not possible to re-code denial rates to dierent geographies so to
match them to AHS. To get around this, I re-assigned the big metropolitan observations to their
corresponding AHS area. For example, Los Angeles is geocoded to Los Angeles, Long Beach, and
Glendale CBSA whereas in the AHS, Los Angeles is coded to the Los Angeles, Long Beach, Anaheim
CBSA. This introduces error into the denial rates but that is also something instrumentation may
help mitigate.
Additionally, for the Miami-Fort Lauderdale-Port St. Lucie, FL CBSA, the estimated mortgage
denial rate in 2015 was over 24% which is strangely high relative to other CBSAs (table 12). As a
result, I exclude Miami from the study as it will likely bias coecient estimates on denials.
The 2008 bank deposit shares of the Big 4 for each MSA are computed using FDIC's Share of
Deposits (SOD) data. Unlike mortgage rates, these data were geocoded to CBSA areas that are
consistent with the American Housing Survey data.
Since Reher and Gete use 2008 deposit shares to instrument 2011-2014 denial rates while I
instrument 2015 rates, I check the rst stage to make sure that the 2008 deposit shares are still
a relevant instrument. In table 14, we see that the share of deposits for the Big 4 banks is still a
signicant factor in denial rates. In both columns, denial rates are signicant and, in both cases,
the F-statistic is markedly over 10.
The down payment to income ratios are computed from 2015 HDMA Data and 2015 AHS data.
The average down payment is not available in the data so I infer it for each MSA. I average loan
amount

L
MSA
, and assuming that on average households put down 20%, I compute the average
down payment amount as dp
MSA
=

L
MSA
:2
:8
. To get the average down payment to renter income
ratio, I compute the average renter income y
MSA
with AHS data and divide dp
MSA
by y
MSA
. In
essence, since the down-payment calculation does not add any new information to loan amount, this
metric should be equivalent to diving the average loan amount by renter income ratios. However,
I nd the down-payment to income ratio more interpretable in context of saving. Similarly, it
is possible to compute monthly-mortgage payments using the loan amounts but, given uniform
assumptions on mortgages, the estimated impact would not dier from the dp
MSA
= y
MSA
results.
Housing cost and household data come from the national and metropolitan samples of the 2015
American Housing Survey that are matched to mortgage and deposit data at the MSA level.
Summary statistics for numerical covariates can be observed in Tables 12 and 13. Table 12
documents summary statistics for the entire sample after ltering out observations with $0 monthly
rent and $0 annual incomes. However, there are still a few odd observations. For example, the
sample contains units with $4 rents, a mortgage denial rate of 24.6%, and a minimum annual income
of $4. To mitigate some of these odd sample features, I restrict the AHS sample to working-age
households aged between 21 and 65, rents and incomes to be between their respective 2.5th and
97.5th percentiles within the MSA, and I exclude the MSA with the 24.6% denial rate
26
.
25
The full list of large metropolitan areas who CBSAs had to be adjusted are: Los Angeles, New York, Philadelphia,
Detroit, Chicago, Washington, Seattle, Dallas, San Francisco, Boston, and Miami.
26
Removing Miami from the sample halves impact estimates suggesting that Miami had an outsized impact on the
estimation of equations 19 and 21
63
Table 12: Variable Summary Statistics
Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max
RENT 20,784 1,041.170 868.576 4 600 1,300 10,600
household age 20,784 44.827 16.553 16 31 56 85
number of adults 20,784 1.739 0.877 0 1 2 11
BEDROOMS 20,784 2.046 0.982 0 1 3 5
BATHROOMS 20,784 1.838 1.133 1 1 3 9
household income 20,784 52,832.720 70,651.610 4 18,000 68,000 2,709,000
tenure length 20,784 4.441 6.297 0 1 5 71
married household 20,784 0.274 0.446 0 0 1 1
income below rent 20,784 0.088 0.284 0 0 0 1
share with incomes greater than 10th percentile 20,784 0.615 0.086 0.484 0.546 0.692 0.823
share with incomes greater than 25th percentile 20,784 0.359 0.053 0.260 0.318 0.405 0.456
number of kids 20,784 0.541 1.004 0 0 1 9
denial rate 20,784 0.125 0.033 0.089 0.102 0.142 0.246
mean loan amount 20,784 358.991 208.510 169.674 227.891 430.309 1,057.967
mean income 20,784 149.641 55.081 88.110 107.007 170.304 312.614
mean loan amount per income 20,784 2.789 0.613 1.961 2.246 3.192 4.302
mean mortgage to income ratio 20,784 0.697 0.153 0.490 0.562 0.798 1.075
big 4 deposit share 20,784 0.236 0.167 0.000 0.096 0.369 0.639
supply elasticity 20,784 1.392 0.690 0.595 0.811 1.765 3.191
log income x denial rate 20,784 1.296 0.357 0.190 1.056 1.443 3.229
estimated mean downpayment 20,784 89.748 52.127 42.418 56.973 107.577 264.492
20,784 2.328 1.048 1.257 1.670 2.280 4.599
2015 Metropolitan and National AHS observations for renters, including Saiz' supply elasticities, share of Big4 deposits, and mortgage denial
rates at MSA level.
Table 13: Variable Summary Statistics for Estimation Sub-sample
Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max
RENT 15,229 983.062 526.817 30.000 650.000 1,200.000 5,000.000
household age 15,229 40.182 11.589 22.000 30.000 50.000 64.000
number of adults 15,229 1.782 0.868 1.000 1.000 2.000 9.000
BEDROOMS 15,229 2.097 0.964 0.000 1.000 3.000 5.000
BATHROOMS 15,229 1.841 1.113 1.000 1.000 3.000 9.000
household income 15,229 51,238.490 38,448.360 760.000 22,800.000 70,000.000 302,800.000
tenure length 15,229 3.822 5.239 0.000 1.000 5.000 60.000
married household 15,229 0.292 0.455 0.000 0.000 1.000 1.000
income below rent 15,229 0.048 0.213 0.000 0.000 0.000 1.000
share with incomes greater than 10th percentile 15,229 0.611 0.083 0.484 0.546 0.662 0.823
share with incomes greater than 25th percentile 15,229 0.358 0.053 0.260 0.318 0.402 0.456
number of kids 15,229 0.632 1.057 0.000 0.000 1.000 9.000
denial rate 15,229 0.120 0.021 0.089 0.102 0.135 0.164
mean loan amount 15,229 353.688 210.969 169.674 227.891 430.309 1,057.967
mean income 15,229 147.556 55.413 88.110 106.543 168.351 312.614
mean loan amount per income 15,229 2.781 0.622 1.961 2.246 3.216 4.302
big 4 deposit share 15,229 0.241 0.170 0.000 0.086 0.369 0.639
supply elasticity 15,229 1.441 0.686 0.627 0.858 2.109 3.191
log income x denial rate 15,229 1.264 0.239 0.630 1.072 1.419 2.003
est mean downpay 15,229 88.422 52.742 42.418 56.973 107.577 264.492
estimated mean downpayment 15,229 2.258 1.029 1.257 1.667 2.241 4.599
The conditions for this sub-sample are: age is in (21,65), rent is between 2.5th and 97.5th percentile in the MSA, denial rate is less than 20%.
64
Table 14: First Stage on denials
Dependent variable:
denial rate denial rate
(1) (2)
big4share 0.028

(0.001) 0.029

(0.001)
logincome 0.002

(0.0002)
lengthTenure 0.0003

(0.00003)
sqrt(BEDROOMS) 0.001 (0.001)
BATHROOMS 0.001

(0.0002)
incomeBelowRent 0.001 (0.001)
porchYes 0.004

(0.0004)
STORIES 0.0002 (0.0002)
NUMADULTS 0.001

(0.0002)
HHAGE 0.0001

(0.00002)
numkids 0.001

(0.0002)
married 0.001 (0.0004)
Constant 0.113

(0.0003) 0.133

(0.003)
Observations 15,229 15,229
R
2
0.051 0.100
Adjusted R
2
0.051 0.098
Residual Std. Error 0.021 (df = 15227) 0.020 (df = 15198)
F Statistic 813.590

(df = 1; 15227) 56.297

(df = 30; 15198)
Note:

p<0.1;

p<0.05;

p<0.01
6 Results
6.1 Role of Access to Ownership Markets
Estimates suggest that higher barriers of access to ownership raise rent spreads relative to
incomes (Table 15). The OLS estimate in column 1 implies that for a 1 percentage point (pp)
increase in denial rates, size-adjusted MSA rents increase by 1%. In column 3, instead of estimating
denials on their own, I interact denials with log-incomes while also estimating log-incomes separately
,but not denials. The interacted term on denials and log-income suggests that going from the
interaction term's mean of 1.3 to its 3rd quartile of 1.4 (table 13 ) increases rents relative to income
by about 2%. These estimates align with the suggestion of table 11 that housing markets with
higher ownership barriers also have higher rent spreads relative to income. When I include supply
elasticities into the model, the impact coecient in column 2 is signicant and negative when
estimated with OLS. This suggests that the estimates coming from OLS are likely endogenized and
biased.
IV estimates on the eect of MSA denial rates on rents and rent-spreads suggest that OLS
underestimates the eect by a factor of 20. The coecient in column 4 suggests that a 1pp increase
in denial rate increases MSA rents by 16%. The coecient on the interaction between log incomes
and denial rates suggests that if we raise the interacted term from its median to its 3rd quartile,
rents relative to incomes to go up by 50%. In table 11, Kansas is the median MSA with a rent-to-
income slope of .12 while Washington, DC is the 3rd quartile with its slope of .2. The dierence
is about 40% suggesting that the access to ownership estimate in column 5 can explain the entire
dierence in rent-to-income slopes. Unlike with OLS, including supply elasticities slightly increases
impact estimates in the IV specication.
In Table 15, we see other covariates behave as they should. For example, households with longer
tenures get a discount on their rents and units with more bedrooms and baths warrant higher rents.
There are occasional coecients, such as for units with kitchens and occasionally for porches, that
are the opposite of their anticipated sign but the coecients are not signicant in such cases.
65
Table 15: Estimate results with denial rates
Dependent variable:
log(RENT)
OLS instrumental
variable
(1) (2) (3) (4) (5) (6) (7)
denial rate 1.043

0.448

20.820

22.460

(0.176) (0.179) (1.063) (1.581)
(denial rate)xlog(income) 0.101

(0.017)
log(income)x(denials rate) 2.000

2.153

(0.103) (0.152)
log income 0.391

0.373

0.379

0.419

0.425

0.175

0.163

(0.005) (0.005) (0.006) (0.007) (0.008) (0.013) (0.017)
supply elasticity 0.166

0.047

0.046

(0.006) (0.017) (0.017)
tenure length 0.007

0.009

0.007

0.014

0.014

0.013

0.013

(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
sqrt(BEDROOMS) 0.034

0.039

0.034

0.090

0.092

0.089

0.091

(0.015) (0.014) (0.015) (0.020) (0.021) (0.020) (0.021)
BATHROOMS 0.061

0.068

0.061

0.072

0.070

0.073

0.072

(0.004) (0.004) (0.004) (0.006) (0.006) (0.006) (0.006)
income below rent 0.927

0.876

0.928

0.868

0.878

0.888

0.900

(0.020) (0.020) (0.020) (0.028) (0.029) (0.028) (0.029)
with porch 0.005 0.026

0.005 0.095

0.095

0.099

0.099

(0.009) (0.008) (0.009) (0.013) (0.013) (0.013) (0.013)
STORIES 0.026

0.023

0.027

0.052

0.054

0.054

0.056

(0.004) (0.004) (0.004) (0.005) (0.006) (0.005) (0.006)
number of adults 0.023

0.013

0.023

0.004 0.003 0.004 0.003
(0.005) (0.005) (0.005) (0.007) (0.007) (0.007) (0.007)
household age 0.002

0.002

0.002

0.003

0.003

0.003

0.003

(0.0003) (0.0003) (0.0003) (0.0005) (0.0005) (0.0005) (0.0005)
number of kids 0.009

0.005 0.009

0.008 0.008 0.008 0.008
(0.004) (0.004) (0.004) (0.005) (0.006) (0.006) (0.006)
married household 0.060

0.061

0.060

0.071

0.071

0.072

0.072

(0.009) (0.009) (0.009) (0.013) (0.013) (0.013) (0.013)
Constant 1.961

2.573

2.088

0.653

0.987

1.874

1.742

(0.070) (0.071) (0.066) (0.167) (0.263) (0.090) (0.109)
Observations 15,229 15,229 15,229 15,229 15,229 15,229 15,229
R
2
0.379 0.411 0.379 0.139 0.226 0.149 0.235
Adjusted R
2
0.378 0.410 0.378 0.142 0.229 0.152 0.238
Note:

p<0.1;

p<0.05;

p<0.01
6.2 Top and bottom results
As suggested in section 3, access to home-ownership should primarily impact rents for households
who are on the margin of becoming home owners and should not aect households who are not
fortunate enough to ever own a home. To this extent, I re-estimate the above equations but on
subsamples of the top and bottom 25% of renter incomes along with the middle 50%.
Results in table 16 qualitatively agree with the above hypothesis. In columns 1 and 2, we see
the OLS results for observations in the bottom and top 25% of renter incomes. The coecient in
Column 1 is positive but not signicant suggesting that barriers to ownership do not aect rents
among the bottom quarter of renter incomes. Column 2's estimate is larger than the comparable
result in table 15 in column 1, as expected, and suggests that a 1pp increase in mortgage denial
rates increases MSA rents by 2%. The IV results in columns 3 and 5 qualitatively suggest a similar
66
result though the coecient dierence between the top and bottom 25% of earners is not as large.
Among renter households with incomes in the bottom 25%, 1pp increase in denials increases MSA
rents by 10% but the same increase in denial rate among the top 25% increases MSA rents by 16%.
As expected, the IV result for the middle 50% of renters in column 4 suggests they are impacted
almost as much as the top 25% and more than the bottom 25%.
Table 16: Results on top and bottom 25% of renter income subsample
Dependent variable:
log(RENT)
OLS instrumental
variable
(1) (2) (3) (4) (5)
denial rate 0.316 2.028

9.274

14.106

15.583

(0.472) (0.311) (2.302) (1.356) (1.735)
log(income) 0.544

0.675

0.614

0.537

0.777

(0.020) (0.022) (0.030) (0.019) (0.035)
supply elasticity 0.048

0.036

0.016
(0.026) (0.015) (0.022)
tenure length 0.006

0.010

0.008

0.012

0.012

(0.002) (0.001) (0.002) (0.001) (0.002)
sqrt(BEDROOMS) 0.062

0.107

0.029 0.078

0.175

(0.034) (0.030) (0.036) (0.022) (0.037)
BATHROOMS 0.074

0.064

0.057

0.062

0.062

(0.012) (0.008) (0.013) (0.007) (0.009)
income below rent 1.080

1.082

0.859

(0.030) (0.033) (0.090)
with porch 0.049

0.027 0.038

0.068

0.065

(0.021) (0.017) (0.022) (0.013) (0.022)
with garage 0.135

0.134

0.125

(0.026) (0.013) (0.018)
STORIES 0.001 0.060

0.011 0.045

0.088

(0.009) (0.007) (0.010) (0.006) (0.009)
number of adults 0.057

0.013 0.046

0.011 0.028

(0.016) (0.008) (0.017) (0.008) (0.010)
household age 0.005

0.001 0.006

0.001

0.002

(0.001) (0.001) (0.001) (0.001) (0.001)
number of kids 0.013 0.002 0.015 0.004 0.011
(0.010) (0.008) (0.010) (0.006) (0.010)
married household 0.143

0.028

0.149

0.067

0.034

(0.032) (0.015) (0.034) (0.014) (0.018)
Constant 0.981

1.697

0.728 1.154

4.549

(0.221) (0.269) (0.533) (0.358) (0.594)
Observations 2,810 3,975 2,810 8,444 3,975
R
2
0.415 0.312 0.347 0.010 0.006
Adjusted R
2
0.409 0.307 0.339 0.014 0.014
Note:

p<0.1;

p<0.05;

p<0.01
6.3 Robustness Check
Estimates using the downpayment to annual income ratio as a measure of ownership barriers
tell the same story as mortgage denial rates, albeit at smaller magnitudes (Table 17). For example,
the OLS estimate in column 1 suggests that going from the median to the 75th percentile of the
downpayment to income ratio raises MSA-wide rents by 13%. The same but instrumented coecient
in column 3 suggests a 20% increase in MSA-wide rents. When we examine the interaction term
67
between log-incomes and the mortgage-to-income ratio, movement from the median to the 3rd
quartile implies an increase of 16% in rent-to-income slopes in OLS and 23% when instrumented.
With denial rates, the same dierence between the median and 3rd quartiles of the interaction term
implies a dierence of 50%.
The lower impact estimates of the downpayment to income ratio is expected given that log-
incomes and the incomes from which the ratio is computed are highly correlated. Removing log
income from the control variables when estimating with the downpayment-to-income ratio increases
coecient estimates by almost 100% in both OLS and instrumented versions whereas performing
the same check using denial rates results in much lower coecient dierences. This suggests that
the downpayment to income ratio is not a great metric of ownership barriers in this context due to
its correlation with MSA-level log income means.
Table 17: Estimate results with downpayment to income ratio
Dependent variable:
log(RENT)
OLS instrumental
variable
(1) (2) (3) (4)
mean mortgage to income ratio 0.976

1.434

(0.026) (0.068)
(mean mortgage to income ratio) x log(income) 0.104

(0.002)
(mean mortgage to income ratio) x log(income) 0.135

(0.006)
log(income) 0.336

0.262

0.319

0.224

(0.005) (0.006) (0.006) (0.009)
supply elasticity 0.062

0.015

0.014
(0.006) (0.009) (0.009)
tenure length 0.011

0.011

0.012

0.012

(0.001) (0.001) (0.001) (0.001)
sqrt(BEDROOMS) 0.089

0.092

0.112

0.111

(0.014) (0.014) (0.014) (0.014)
BATHROOMS 0.061

0.058

0.057

0.057

(0.004) (0.004) (0.004) (0.004)
income below rent 0.792

0.796

0.753

0.753

(0.019) (0.019) (0.020) (0.020)
with porch 0.038

0.032

0.044

0.043

(0.008) (0.008) (0.008) (0.008)
STORIES 0.030

0.031

0.033

0.032

(0.003) (0.003) (0.004) (0.004)
number of adults 0.002 0.0004 0.009

0.009

(0.005) (0.005) (0.005) (0.005)
household age 0.002

0.002

0.003

0.003

(0.0003) (0.0003) (0.0003) (0.0003)
number of kids 0.003 0.001 0.006 0.005
(0.004) (0.004) (0.004) (0.004)
married household 0.059

0.058

0.058

0.058

(0.009) (0.009) (0.009) (0.009)
Constant 2.082

2.682

1.885

2.888

(0.064) (0.063) (0.070) (0.066)
Observations 15,229 15,229 15,229 15,229
R
2
0.460 0.458 0.449 0.451
Adjusted R
2
0.459 0.457 0.448 0.450
Note:

p<0.1;

p<0.05;

p<0.01
68
7 Discussion and Conclusion
Results suggest high barriers to home-ownership increase rents - particularly for higher-income
renters. As a consequence, metropolitan areas with higher barriers also have wider rent spreads
relative to renter incomes. Although suggestive, there is more work ahead.
The dierence between low and high income renters is not as large as expected which suggests
there may be factors beside ownership-access at play. The current framework assumes that mortgage
rates only impact would-be home owners which neglects the fact that landlords also buy property. In
metros with high ownership barriers, the extra costs landlords must pay to acquire their properties
is almost inevitably passed down to their tenants. This is one reason why we may see a non-trivial
impact on rents among low-income renters. Another potential reason is that a renter in the bottom
25th percentile of San Francisco is not renting the same quality of unit as a renter situated in the
same portion of the income distribution in Milwaukee. Better weather, walkability, local art scenes,
and other unobserved factors in San Francisco may all raise the quality of local amenities and hence
raise prices relative to incomes. Finally, there is the element of endogenous gentrication whereby
high income households who cannot nd a home in their desired neighborhood instead move to a
nearby lower-income neighborhood (Guerrieri, Hartley, and Hurst 2013). This would suggest that
a lack of ownership opportunities in sought-after neighborhoods can over
ow into lower-income
renter markets.
The current model does not account for important demand factors as job growth, weather,
and population growth. To more credibly identify the impact of ownership barriers on the income
distribution, such demand factors need to be accounted for.
There are also other factors that may widen rent spreads in star cities such as agglomeration. We
know that returns to agglomeration are higher for higher-skilled workers, which, beside endogenizing
amenities, should also increase their willingness to pay to live in a particular MSA and; thus, raise
the rents they are willing to pay relative to income (Bacolod, la Roca, and Ferreyra 2020, Roca
and Puga 2017, Diamonds 2016). This would could show up in data on rents relative to incomes
whereby comparing households in the same part of the income distribution across MSAs would
suggest that in star cities, these households pay a higher share of their income toward rent. Future
work may aim to incorporate these features into the current framework.
Policy implications of these ndings primarily revolve around the fact that changes in the mag-
nitude of ownership barriers will primarily aect higher-income renters. In that sense, regulations
aimed to ease access to ownership will have no little bearing on rents and tenant welfare among
low-income renters. This is especially true if rents are sticky upward. Namely, suppose the credit
environment makes it less expensive for a landlord to buy an existing rental property. If rents are
downwardly sticky, then it is unlikely incumbent tenants will see their monthly rents drop.
69
References
Aron-Dine, S., and D. M. Bunten (2019): \Yes, Gentrication Displaces People," unpublished.
Atkinson, R. (2000): \Measuring gentrication and displacement in Greater London," Urban
Studies, 37, 149{165.
Bacolod,M.,J.D.laRoca, andM.M.Ferreyra (2020): \In Search of Better Opportunities:
Sorting and Agglomeration Eects among Young College Graduates in Colombia," .
Bajari, P., and M. Kahn (2005): \Estimating Housing Demand With an Application to Ex-
plaining Racial Segregation in Cities," Journal of Business & Economic Statistics, 23, 20{33.
Baker, D. M., and B. Lee (2017): \How does light rail transit impact gentrication? Evidence
from Fourtnee US Urbanized Areas," Journal of Planning Education and Research.
Bardaka, E., M.S.Delgado, andR.Florax (2018): \Causal identication of transit-induced
gentrication and spatial spillover eects: The case of the Denver light rail," Journal of Transport
Geography, 71, 15{31.
Barton, M., and J. Gibbons (2018): \A stop too far: How does public transportation concen-
tration in
uence median household income?," Urban Studies, 54.
Bayer, P., M. Casey, F. Ferreira, and R. McMillan (2017): \Racial and ethnic price
dierentials in the housing market," Journal of Urban Economics, 102, 91{105.
Bayer, P., H. Fang, and R. McMillan (2014): \Separate when equal? Racial inequality and
residential segregation," Journal of Urban Economics, 82, 32{48.
Bayer, P., and R. McMillan (2005): \Racial Sorting and Neighborhood Quality," Working
Paper 11813; National Bureau of Economic Research.
Bayer, P., R. McMillan, and K. Rueben (2004): \Residential Segregation in General Equi-
librium," Working Papers 885, Economic Growth Center.
Bereitschaft, B. (2020): \Gentrication and the evolution of behavior within America's urban
cores, 2000-2015," Journal of Transport Geography.
Berry, B. J. (1976): \Ghetto Expansion and single-family housing prices: Chicago, 1968{1972,"
Journal of Urban Economics, 3, 397{423.
Berry, C. R., and E. L. Glaeser (2005): \The Divergence of Human Capital Levels Across
Cities," Working Paper 11617, National Bureau of Economic Research.
Berry, S., J. Levinsohn, and A. Pakes (1995): \Automobile Prices in Market Equilibrium,"
Econometrica, 63, 841{890.
Boarnet, M. G., R. W. Bostic, S. R. Evgeny Burinskiy, and A. Prohofsky (2018):
\Gentrication Near Rail Transit Areas: A Micro-Data Analysis of Moves into Los Angeles
Metro Rail Station Areas," Discussion paper, National Center for Sustainable Transportation.
70
Boarnet, M. G., R. W. Bostic, S. Rodnyansky, A. Prohofsky, A. Eisenlohr, and
H. Jamme (2018): \The eects of rail stations on neighborhood displacement in Los Angeles
County, 1993-2013," Discussion paper, California Community Foundation.
Boarnet, M. G., S. Rodnyansky, and R. W. Boarnet (2020): \The Evolution of the Los
Angeles Metro Rail Station Neighborhoods: Moves, Rents, and Permits.," Discussion paper,
California Community Foundation.
Boeing, G. (2019): \Representation and the Digital Reproduction of Urban Inequality," Environ-
mental Planning A: Economy and Space.
Bosch, M., M. A. Carnero, and L. Farre (2010): \Information and Discrimination in the
Rental Housing Market: Evidence From the Field Experiment," regional science and urban
economics, 40, 11{19.
Chapple,K. (2009): \Mapping susceptibility to gentrication: The early warning toolkit," Discus-
sion paper, Berkeley, CA. University of California-Berkeley, Center for Community Innovation.
Chen, B. S., S. G. Hanson, and J. C. Stein (2017): \The Decline of Big-Bank Lending to
Small Business: Dynamic Impacts on Local Credit and Labor Markets," Working Paper 23843,
National Bureau of Economic Research.
Cosman,J., andL.Quintero (2018): \Market Concentration in Home Building," Johns Hopkins
Carey Business School Research Paper Series.
Courant, P. N. (1978): \Racial Prejudice in a Search Model of the Urban Housing Market,"
Journal of Urban Economics, 5, 329{345.
D'Acunto, F., and A. G. Rossi (2017): \Ditching the Middle Class with Consumer Protection
Regulation," Working paper, National Bureau of Economic Research.
Davidoff, T. (2014): \Supply constraints are not valid instrumental variables for home prices
because they are correlated with many demand factors," Critical Finance Review, 5, 177{206.
Deka, D. (2017): \Benchmarking gentrication near commuter rail stations in New Jersey," Urban
Studies.
Desmond, M. (2016): Evicted. Penguin Random House LLC.
Desmond, M., C. Gershenson, and B. Kiviat (2015): \Forced Relocation and Residential
Instability among Urban Renters," Social Service Review.
Diamonds, R. (2016): \The Determinants and Welfare Implications of US Workers' Diverging
Location Choices by Skill: 1980-2000," American Economic Review, 106, 479{524.
Ding, L., J. Hwang, and E. Divringi (2016): \Gentrication and residential mobility in
Philadelphia.," Regional science and urban economics, 61, 38{51.
Dong, H. (2017): \Rail-transit-induced gentrication and the aordability paradox of TOD,"
Journal of Transport Geography, 63, 1{10.
71
Dragan, K., I.Ellen, andS.A.Glied (2019): \Gentrication And The Health Of Low-Income
Children In New York City," Health Aairs, 38, 1425{1432.
Edward L. Glaeser, Joseph Gyuorko, R. S. (2005): \Why is Manhattan So Expensive?
Regulation and the Rise in Housing Prices," Journal of Law and Economics, 48.
Freeman, L. (2005): \Displacement or Succession? Residential Mobility in Gentrifying Neighbor-
hoods," Urban Aairs Review, 40, 463{491.
Freeman, L., A. Cassola, and T. Cai (2015): \Displacement and gentrication in England
and Wales: A quasi-experimental approach," Urban Studies.
FTB (2006): \Tax Gap Plan: A Strategic Approach to Reducing California's Tax Gap," Discussion
paper, California Franchise Tax Board (FTB).
(2013): \California 540: 2013 Personal Income Tax Booklet," .
GamperRabindran, S., and C. Timmins (2011): \Does cleanup of hazardous waste sites raise
housing values? Evidence of spatially localized benets," The American Economic Review,
101(3), 620{624.
Gete, P., and M. Reher (2018): \Mortgage Supply and Housing Rents," Review of Financial
Studies, 31, 4884{4911.
Glaeser,E., andB.Ward (2009): \The causes and consequences of land use regulation: evidence
from Greater Boston," Journal of Urban Economics, 65, 265{278.
Glaeser, E. L., and J. Gyourko (2003): \The Impact of Building Restrictions on Housing
Aordability," Economic Policy Review, pp. 21{39.
Glaeser, E. L., J. Gyourko, and R. E. Saks (2005): \Why Have Housing Prices Gone Up?,"
American Economic Review, 95(2), 329{333.
Glaeser, E. L., M. E. Kahn, and J. Rappaport (2008): \Why do the poor live in cities? The
role of public transportation," Journal of Urban Economics, 63, 1{24.
Goodman, L. (2017): \Quantifying the Tightness of Mortgage Credit and Assessing Policy Ac-
tions," Discussion paper, Urban Institute.
Green, R. K., S. Malpezzi, and S. K. Mayo (2005): \Metropolitan-Specic Estimates of
the Price Elasticity of Supply of Housing and Their Sources," American Economic Review, 95,
334{339.
GrubeCavers, A., and Z. Patterson (2015): \Urban rapid rail transit and gentrication in
Canadian urban centres: A survival analysis approach," Urban Studies.
Guerrieri, V., D. Hartley, and E. Hurst (2013): \Endogenous Gentrication and Housing
Price Dynamics," Journal of Public Economics, pp. 45{60.
Hanson, A., and Z. Hawley (2011): \Do landlords discriminate in the rental housing market?
Evidence from an internet eld experiment in US cities," journal of urban economics, 70, 99{114.
72
Heilmann, K. (2018): \Transit access and neighborhood segregation. Evidence from the Dallas
light rail system," Regional Science and Urban Economics, 73, 237{250.
Hess, C. L. (2020): \Light-Rail Investment in Seattle Gentrication Pressures and Trends in
Neighborhoods," Urban Aairs Review, 56, 154{187.
Ihlanfeldt, K. (2007): \The eect of land-use regulation on housing and land prices," Journal
of Urban Economics, 61, 420{435.
IRS (2014): \EITC Participation Rate by States," .
Kahn, M. (2007): \Gentrication Trends in New Transit-Oriented Communities: Evidence from
14 Cities That Expanded and Built Rail Transit Systems," Real Estate Economics, 35, 155{182.
Katz,L., andK.T.Rosen (1987): \The Interjurisdictional Eects of Growth Control on Housing
Prices," Journal of Law and Economics, 30.
Lee, K. O. (2014): \Why do renters stay in or leave certain neighborhoods? The role of neigh-
borhood characteristics, housing tenure transitions, and race," Journal of Regional Science, 54,
755{787.
Malpezzi, S. (1996): \Housing Prices, Externalities, and Regulation in U.S. Metropolitan Areas,"
Journal of HOusing Research, 7.
Manville, M., B. D. Taylor, and E. Blumenberg (2018): \Falling Transit Ridership: Cali-
fornia and Southern California," Discussion paper, UCLA Institute of Transportation Studies.
Martin, I. W., and K. Beck (2018): \Gentrication, Property Tax Limitation, and Displace-
ment," Urban Aairs Review, 54, 33{73.
McKinnish, T., R. Walsh, and K. White (2010): \Who Gentries Low-Income Neighbor-
hoods?," Journal of Urban Economics, 67, 180{193.
Metro LA (2009): \Long Range Transportation Plan, 2009," Discussion paper, Los Angeles
County and Metropolitan Transportation Authority (L.A. Metro).
Moretti, E. (2012): The New Geography of Jobs. Houghton Miin Harcourt.
Nilsoon, I., and E. Delmelle (2020): \New rail transit stations and the out-migration of low-
income residents," .
Nilsson, I., and E. Delmelle (2018): \Transit investments and neighborhood change: On the
likelihood of change," Journal of Transport Geography, 66, 167{179.
Ondrich, J., A. Stricker, and J. Yinger (1999): \Do Landlords Discriminate? The Incidence
and Causes of Racial Discrimination in Rental Housing Markets," Journal of Housing Economics,
8, 185{204.
Ong, P., M. Zuk, C. Pech, and K. Chapple (2017): Development of an o-model displacement
assessment methodologypp. 148{153. University of California.
73
Palm, R., and M. A. Danis (2001): \Residential Mobility: The Impacts of Web-Based Informa-
tion on the Search Process and Spatial Housing Choice Patterns," Urban Geography, 22, 641{655.
Pathak, R., C. K. Wyczalkowski, and X. Huang (2017): \Public transit access and the
changing spatial distribution of poverty," Regional Science and Urban Economics, pp. 198{212.
Quigley, J. M., and S. Raphael (2005): \Regulation and the High Cost of Housing in Califor-
nia," American Economics Association Paper and Proceedings, pp. 323{328.
Rayle, L. (2015): Housing Policy Debate25, 531{548.
Roca, J. D. L., and D. Puga (2017): \Learning by Working in Big Cities," .
Rosen, J., S. Angst, S. D. Gregorio, and G. Painter (2019): \How do renters cope with
unaordability?," Discussion paper, USC Sol Price School of Public Policy.
Saiz, A. (2010): \The Geographic Determinants of Housing Supply," The Quarterly Journal of
Economics, pp. 1253{1296.
SCAG (2012): \2012-2035 Regional Transportation Plan/Sustainable Communities Strategy," Dis-
cussion paper, Southern California Association of Governments (SCAG).
Skobba, K., and E. G. Goetz (2013): \Mobility Decisions of Very Low-Income Households,"
Cityscape: A Journal of Policy Development and Research, 15.
Sommerville, T. (1999): \The Industrial Organization of Housing Supply: Market Activity, Land
Supply, and the Size of Homebuilder Firms," Real Estate Economics, 27, 669{694.
Wang,K., andM.Woo (2017): \The relationship between transit rich neighborhoods and transit
ridership," Applied Geography, pp. 183{196.
Watson, C. L., and O. Ziv (2019): \Is the rent too high? Land Ownership and Monopoly
Power," unpublished.
Yinger, J. (1986): \Measuring Racial Discrimination with Fair Housing Audits: Caught in the
Act," American Economic Association, 76, 881{893.
Zuk, M., B. A. C. K. G. K. L. S. (2018): \Gentrication, Displacement, and the Role of Public
Investment," Journal of Planning Literature.
Zuk, M., A. Bierbuaum, K. Chapple, K. Gorska, A. L. Sideris, P. Ong, and T. Thomas
(2015): \Gentrication, Displacement and the Role of Public Investment: A Literature Review,"
Discussion paper.
Zuk,M., andK.Chapple (2015): \Urban Displacement Project. Mapping Neighborhood Change
in Los Angeles County, 2015," .
74
Appendix
Supplementary Tables and Figures
Context
Figure 7: Map of Los Angeles Metro Rail Lines (2013)
75
Table 18: Observable characteristic medians by treatment group
adjIncome tpdobyy depndnts lestat
treated control treated control treated control treated control treated
year
1994 18100.0 16000.0 NaN NaN 1 1 2 2
1995 18126.0 16452.0 NaN NaN 1 1 2 2
1996 18566.0 16729.0 1958.0 1963.0 1 1 2 2
1997 18942.0 17229.0 1962.0 1963.0 1 1 2 2
1998 19894.0 18017.0 1964.0 1964.0 1 1 2 2
1999 20080.0 18430.0 1964.0 1965.0 1 1 2 2
2000 20236.0 18461.0 1961.0 1962.0 0 0 2 2
2001 20101.0 18383.0 1962.0 1963.0 0 0 2 2
2002 20145.0 18390.0 1963.0 1963.0 0 0 2 2
2003 20198.0 18569.0 1963.0 1964.0 0 0 2 2
2004 20102.0 18446.0 1964.0 1965.0 0 0 2 2
2005 20000.0 18339.0 1965.0 1966.0 0 0 2 2
2006 19906.0 18386.0 1966.0 1967.0 0 0 2 2
2007 19761.0 18290.0 1967.0 1968.0 0 0 2 2
2008 19426.0 18140.0 1968.0 1969.0 0 0 2 2
2009 18627.0 17535.0 1968.0 1969.0 0 0 2 2
2010 18541.0 17462.0 1970.0 1971.0 0 0 2 2
2011 18387.0 17402.0 1970.0 1971.0 0 0 2 2
2012 18664.0 17570.0 1971.0 1971.0 0 0 2 2
2013 18655.0 17636.0 1972.0 1973.0 0 0 2 2
2014 19852.0 18846.0 1972.0 1972.0 0 0 2 2
76
Table 19: annual AMI Table
-100ami 0ami 30ami 50ami 80ami 100ami 200ami 500ami
year
1993 -42300 0 12690.0 21150.0 33840.0 42300 84600 211500
1994 -45200 0 13560.0 22600.0 36160.0 45200 90400 226000
1995 -45200 0 13560.0 22600.0 36160.0 45200 90400 226000
1996 -46900 0 14070.0 23450.0 37520.0 46900 93800 234500
1997 -47800 0 14340.0 23900.0 38240.0 47800 95600 239000
1998 -49800 0 14940.0 24900.0 39840.0 49800 99600 249000
1999 -51300 0 15390.0 25650.0 41040.0 51300 102600 256500
2000 -52100 0 15630.0 26050.0 41680.0 52100 104200 260500
2001 -54500 0 16350.0 27250.0 43600.0 54500 109000 272500
2002 -55100 0 16530.0 27550.0 44080.0 55100 110200 275500
2003 -50300 0 15090.0 25150.0 40240.0 50300 100600 251500
2004 -54200 0 16260.0 27100.0 43360.0 54200 108400 271000
2005 -54450 0 16335.0 27225.0 43560.0 54450 108900 272250
2006 -56200 0 16860.0 28100.0 44960.0 56200 112400 281000
2007 -56500 0 16950.0 28250.0 45200.0 56500 113000 282500
2008 -59800 0 17940.0 29900.0 47840.0 59800 119600 299000
2009 -62100 0 18630.0 31050.0 49680.0 62100 124200 310500
2010 -63000 0 18900.0 31500.0 50400.0 63000 126000 315000
2011 -64000 0 19200.0 32000.0 51200.0 64000 128000 320000
2012 -64800 0 19440.0 32400.0 51840.0 64800 129600 324000
2013 -61900 0 18570.0 30950.0 49520.0 61900 123800 309500
2014 -60600 0 18180.0 30300.0 48480.0 60600 121200 303000
2015 -63000 0 18900.0 31500.0 50400.0 63000 126000 315000
77
Table 20: Annual number of households in treated basin
num treated HHs median number of HHs mean number of HHs station count
year
1994 34808 669.0 669.0 52
1995 44708 835.0 860.0 52
1996 48252 890.5 928.0 52
1997 52006 922.0 1000.0 52
1998 55583 989.5 1069.0 52
1999 59914 1091.5 1152.0 52
2000 65835 1197.0 1291.0 51
2001 67655 1205.5 1301.0 52
2002 67818 1196.5 1304.0 52
2003 65865 1171.5 1267.0 52
2004 69622 1240.5 1339.0 52
2005 72513 1277.0 1394.0 52
2006 76266 1338.5 1467.0 52
2007 81090 1411.5 1559.0 52
2008 79873 1415.5 1536.0 52
2009 80156 1446.5 1541.0 52
2010 81198 1437.5 1562.0 52
2011 82893 1419.0 1594.0 52
2012 84255 1460.5 1620.0 52
2013 86422 1490.5 1662.0 52
2014 89009 1524.0 1712.0 52
78
Mobility rates
Figure 8: Mean mobility rates by income
Whiskers are 1 standard deviation above and below mean
79
Parallel Tests
Table 21: Parallel trend test estimates
dep var movedOut movedOutOut movedIn movedInIn
ind var lag1:treated lag2:treated lag1:treated lag2:treated lag1:treated lag2:treated lag1:treated lag2:treated
model line-year AMI
base Expo 30 0.003 (0.0079) -0.002 (0.0035) 0.001 (0.0062) -0.002 (0.0040) 0.002 (0.0057) 0.015** (0.0049) 0.001 (0.0057) 0.012* (0.0054)
50 0.002 (0.0056) 0.006 (0.0071) 0.006 (0.0053) 0.005 (0.0057) 0.009 (0.0070) 0.006 (0.0038) 0.006 (0.0056) 0.005 (0.0029)
80 0.009 (0.0071) 0.006 (0.0087) 0.008 (0.0070) 0.003 (0.0076) -0.001 (0.0055) 0.007 (0.0071) -0.005 (0.0061) 0.004 (0.0066)
200 0.006 (0.0070) 0.011 (0.0065) 0.001 (0.0052) 0.005 (0.0073) 0.008 (0.0066) 0.003 (0.0062) 0.003 (0.0056) 0.002 (0.0068)
500 -0.02** (0.0074) -0.022 (0.0179) -0.014* (0.0056) -0.02 (0.0180) 0.013 (0.0107) 0.008 (0.0216) 0.008 (0.0108) 0.015 (0.0212)
Gold2003 30 -0.01 (0.0058) -0.006 (0.0063) -0.005 (0.0046) -0.004 (0.0054) 0.01** (0.0036) -0.006 (0.0042) 0.013*** (0.0031) -0.008** (0.0031)
50 -0.001 (0.0108) -0.007 (0.0064) 0.001 (0.0089) -0.004 (0.0059) -0.004 (0.0053) -0.004 (0.0096) -0.0 (0.0051) -0.003 (0.0070)
80 -0.011 (0.0112) 0.001 (0.0076) -0.011 (0.0101) -0.0 (0.0085) 0.016 (0.0108) 0.004 (0.0081) 0.008 (0.0087) 0.002 (0.0061)
200 -0.001 (0.0074) -0.01 (0.0068) 0.001 (0.0075) -0.008 (0.0063) 0.003 (0.0067) 0.007 (0.0062) 0.003 (0.0064) 0.006 (0.0053)
500 0.01 (0.0168) -0.006 (0.0126) 0.003 (0.0146) -0.006 (0.0109) 0.015 (0.0127) -0.007 (0.0152) 0.013 (0.0117) -0.01 (0.0145)
Gold2009 30 -0.001 (0.0030) -0.0 (0.0037) -0.005 (0.0035) 0.003 (0.0038) 0.002 (0.0062) -0.002 (0.0074) 0.003 (0.0028) -0.002 (0.0059)
50 -0.013*** (0.0039) -0.001 (0.0052) -0.012*** (0.0037) 0.001 (0.0046) -0.005* (0.0018) 0.002 (0.0047) -0.004** (0.0017) 0.001 (0.0028)
80 0.006 (0.0066) 0.003 (0.0035) 0.005 (0.0053) 0.002 (0.0042) 0.007 (0.0047) -0.008 (0.0050) 0.009*** (0.0028) -0.006 (0.0063)
200 -0.01 (0.0082) -0.003 (0.0106) -0.01 (0.0075) -0.007 (0.0103) -0.003 (0.0099) -0.0 (0.0072) -0.007 (0.0100) -0.001 (0.0061)
500 -0.135* (0.0585) 0.058 (0.0756) -0.133* (0.0549) 0.059 (0.0750) -0.036 (0.0245) -0.094 (0.0508) -0.032 (0.0236) -0.051 (0.0477)
Red1999 30 -0.014 (0.0122) -0.004 (0.0115) -0.01 (0.0136) -0.003 (0.0055) 0.009 (0.0152) 0.007 (0.0099) 0.009 (0.0119) 0.004 (0.0083)
50 0.008 (0.0189) 0.007 (0.0186) 0.008 (0.0161) 0.013 (0.0178) -0.007 (0.0140) -0.007 (0.0162) -0.004 (0.0086) 0.001 (0.0151)
80 0.003 (0.0218) -0.015 (0.0255) 0.007 (0.0123) -0.001 (0.0167) -0.02 (0.0132) -0.01 (0.0211) -0.01 (0.0069) -0.009 (0.0163)
200 -0.009 (0.0291) -0.014 (0.0275) 0.007 (0.0185) 0.006 (0.0207) -0.009 (0.0078) 0.012 (0.0121) 0.0 (0.0081) 0.017 (0.0087)
500 -0.037 (0.0194) -0.002 (0.0200) -0.027 (0.0169) -0.02 (0.0151) -0.01 (0.0286) 0.006 (0.0178) -0.018 (0.0147) -0.017 (0.0217)
Red2000 30 -0.043* (0.0175) -0.033 (0.0236) -0.022 (0.0128) -0.021 (0.0238) -0.016 (0.0109) -0.005 (0.0217) -0.01 (0.0087) -0.004 (0.0175)
50 -0.013 (0.0140) 0.018** (0.0062) -0.012 (0.0143) 0.014*** (0.0034) -0.022 (0.0180) -0.03 (0.0293) -0.019 (0.0203) -0.02 (0.0299)
80 -0.021*** (0.0046) 0.015*** (0.0035) -0.014 (0.0098) 0.023* (0.0101) -0.009 (0.0070) -0.004 (0.0216) -0.005 (0.0050) 0.001 (0.0164)
200 0.011*** (0.0032) 0.011 (0.0061) 0.009 (0.0092) 0.012*** (0.0016) 0.012 (0.0143) 0.0 (0.0053) 0.014 (0.0139) -0.001 (0.0071)
500 0.003 (0.0400) -0.026 (0.0234) -0.004 (0.0418) -0.019 (0.0187) -0.02 (0.0179) 0.02 (0.0119) -0.022 (0.0193) 0.016 (0.0088)
incumbents Expo 30 0.003 (0.0079) -0.002 (0.0035) 0.001 (0.0062) -0.002 (0.0040) 0.002 (0.0057) 0.015** (0.0049) 0.001 (0.0057) 0.012* (0.0054)
50 0.002 (0.0056) 0.006 (0.0071) 0.006 (0.0053) 0.005 (0.0057) 0.009 (0.0070) 0.006 (0.0038) 0.006 (0.0056) 0.005 (0.0029)
80 0.009 (0.0071) 0.006 (0.0087) 0.008 (0.0070) 0.003 (0.0076) -0.001 (0.0054) 0.007 (0.0071) -0.005 (0.0061) 0.004 (0.0065)
200 0.006 (0.0070) 0.011 (0.0065) 0.001 (0.0052) 0.005 (0.0073) 0.008 (0.0066) 0.003 (0.0062) 0.003 (0.0055) 0.002 (0.0068)
500 -0.02** (0.0074) -0.022 (0.0179) -0.014* (0.0056) -0.02 (0.0180) 0.013 (0.0107) 0.008 (0.0219) 0.008 (0.0108) 0.015 (0.0214)
Gold2003 30 -0.01 (0.0059) -0.006 (0.0063) -0.006 (0.0047) -0.004 (0.0054) 0.01** (0.0036) -0.006 (0.0041) 0.013*** (0.0031) -0.008** (0.0030)
50 -0.001 (0.0108) -0.008 (0.0065) 0.001 (0.0089) -0.004 (0.0059) -0.004 (0.0052) -0.004 (0.0096) -0.0 (0.0051) -0.003 (0.0070)
80 -0.011 (0.0111) 0.002 (0.0076) -0.011 (0.0100) 0.0 (0.0084) 0.016 (0.0108) 0.004 (0.0082) 0.008 (0.0086) 0.002 (0.0062)
200 -0.001 (0.0074) -0.01 (0.0067) 0.001 (0.0076) -0.008 (0.0063) 0.003 (0.0067) 0.008 (0.0061) 0.003 (0.0064) 0.007 (0.0053)
500 0.01 (0.0167) -0.007 (0.0126) 0.003 (0.0145) -0.006 (0.0108) 0.015 (0.0124) -0.007 (0.0150) 0.013 (0.0114) -0.01 (0.0144)
Gold2009 30 -0.001 (0.0030) -0.0 (0.0037) -0.005 (0.0035) 0.003 (0.0038) 0.002 (0.0062) -0.002 (0.0075) 0.003 (0.0028) -0.002 (0.0059)
50 -0.013*** (0.0039) -0.001 (0.0052) -0.013*** (0.0037) 0.001 (0.0046) -0.005* (0.0018) 0.002 (0.0047) -0.004** (0.0016) 0.001 (0.0029)
80 0.006 (0.0066) 0.003 (0.0035) 0.005 (0.0053) 0.002 (0.0042) 0.007 (0.0047) -0.008 (0.0050) 0.009*** (0.0028) -0.006 (0.0063)
200 -0.01 (0.0082) -0.003 (0.0107) -0.01 (0.0075) -0.007 (0.0103) -0.002 (0.0098) 0.0 (0.0073) -0.007 (0.0099) -0.001 (0.0062)
500 -0.135* (0.0584) 0.059 (0.0755) -0.133* (0.0547) 0.059 (0.0749) -0.038 (0.0215) -0.096 (0.0523) -0.034 (0.0213) -0.051 (0.0491)
Red1999 30 -0.014 (0.0121) -0.004 (0.0113) -0.01 (0.0137) -0.003 (0.0054) 0.009 (0.0150) 0.007 (0.0099) 0.009 (0.0117) 0.003 (0.0083)
50 0.008 (0.0189) 0.007 (0.0189) 0.008 (0.0161) 0.013 (0.0181) -0.008 (0.0140) -0.008 (0.0163) -0.004 (0.0086) 0.0 (0.0153)
80 0.002 (0.0221) -0.014 (0.0258) 0.007 (0.0124) -0.001 (0.0168) -0.02 (0.0132) -0.01 (0.0213) -0.01 (0.0072) -0.01 (0.0164)
200 -0.009 (0.0292) -0.014 (0.0274) 0.007 (0.0186) 0.006 (0.0207) -0.009 (0.0079) 0.012 (0.0120) -0.0 (0.0082) 0.017 (0.0087)
500 -0.036* (0.0173) -0.0 (0.0211) -0.026 (0.0146) -0.017 (0.0163) -0.009 (0.0275) 0.004 (0.0176) -0.018 (0.0137) -0.019 (0.0219)
Red2000 30 -0.043* (0.0174) -0.033 (0.0237) -0.022 (0.0127) -0.021 (0.0237) -0.016 (0.0106) -0.004 (0.0217) -0.01 (0.0081) -0.003 (0.0171)
50 -0.013 (0.0144) 0.018** (0.0063) -0.012 (0.0148) 0.014*** (0.0034) -0.023 (0.0187) -0.029 (0.0294) -0.019 (0.0209) -0.02 (0.0300)
80 -0.021*** (0.0046) 0.015*** (0.0043) -0.014 (0.0102) 0.023* (0.0100) -0.01 (0.0068) -0.004 (0.0216) -0.005 (0.0050) 0.002 (0.0165)
200 0.011*** (0.0034) 0.011* (0.0057) 0.01 (0.0093) 0.013*** (0.0012) 0.012 (0.0143) 0.0 (0.0049) 0.014 (0.0137) -0.001 (0.0068)
500 0.001 (0.0417) -0.026 (0.0236) -0.006 (0.0432) -0.02 (0.0192) -0.017 (0.0191) 0.021* (0.0103) -0.019 (0.0202) 0.018* (0.0087)
longobs obsExpo 30 0.002 (0.0071) -0.001 (0.0040) -0.001 (0.0055) -0.0 (0.0048) 0.003 (0.0063) 0.016*** (0.0050) 0.001 (0.0061) 0.014** (0.0053)
50 0.002 (0.0064) 0.006 (0.0074) 0.006 (0.0058) 0.005 (0.0067) 0.009 (0.0073) 0.008* (0.0038) 0.005 (0.0056) 0.006* (0.0028)
80 0.012 (0.0067) 0.004 (0.0080) 0.01 (0.0075) 0.003 (0.0072) -0.002 (0.0052) 0.006 (0.0069) -0.005 (0.0062) 0.004 (0.0061)
200 0.005 (0.0063) 0.01 (0.0056) 0.002 (0.0056) 0.006 (0.0058) 0.007 (0.0064) 0.004 (0.0056) 0.004 (0.0057) 0.003 (0.0061)
500 -0.01 (0.0098) -0.019 (0.0176) -0.004 (0.0068) -0.017 (0.0175) 0.015 (0.0110) 0.007 (0.0227) 0.01 (0.0096) 0.014 (0.0224)
obsGold2003 30 -0.005 (0.0060) -0.006 (0.0073) -0.003 (0.0050) -0.004 (0.0060) 0.011* (0.0050) -0.008 (0.0052) 0.013** (0.0045) -0.009* (0.0039)
50 0.001 (0.0123) -0.007 (0.0064) 0.003 (0.0096) -0.004 (0.0052) -0.007 (0.0042) -0.002 (0.0076) -0.004 (0.0051) -0.003 (0.0060)
80 -0.015 (0.0126) 0.002 (0.0091) -0.015 (0.0116) 0.001 (0.0102) 0.015 (0.0108) 0.005 (0.0079) 0.006 (0.0088) 0.004 (0.0062)
200 0.0 (0.0080) -0.009 (0.0069) 0.002 (0.0076) -0.007 (0.0064) 0.001 (0.0064) 0.01 (0.0058) 0.001 (0.0060) 0.007 (0.0051)
500 0.011 (0.0194) -0.012 (0.0147) 0.002 (0.0168) -0.013 (0.0119) 0.012 (0.0119) -0.008 (0.0162) 0.011 (0.0109) -0.01 (0.0152)
obsGold2009 30 -0.001 (0.0037) 0.002 (0.0068) -0.006 (0.0032) 0.004 (0.0053) 0.002 (0.0074) -0.004 (0.0065) 0.003 (0.0032) -0.003 (0.0044)
50 -0.013*** (0.0040) -0.004 (0.0054) -0.012** (0.0044) -0.001 (0.0048) -0.007*** (0.0009) -0.0 (0.0048) -0.006*** (0.0013) -0.002 (0.0031)
80 0.005 (0.0069) 0.002 (0.0031) 0.004 (0.0051) 0.001 (0.0039) 0.003 (0.0050) -0.007 (0.0056) 0.007* (0.0034) -0.005 (0.0064)
200 -0.009 (0.0069) -0.003 (0.0104) -0.009 (0.0068) -0.007 (0.0102) -0.002 (0.0088) -0.001 (0.0073) -0.005 (0.0091) -0.002 (0.0062)
500 -0.181*** (0.0547) 0.095 (0.0681) -0.178*** (0.0543) 0.098 (0.0665) -0.035 (0.0244) -0.109* (0.0495) -0.032 (0.0237) -0.061 (0.0472)
obsRed1999 30 -0.007 (0.0170) 0.008 (0.0130) -0.006 (0.0184) 0.008 (0.0083) 0.006 (0.0141) -0.003 (0.0041) 0.003 (0.0133) -0.003 (0.0078)
50 0.001 (0.0229) 0.008 (0.0185) -0.002 (0.0188) 0.011 (0.0162) -0.001 (0.0192) 0.005 (0.0168) -0.0 (0.0138) 0.008 (0.0154)
80 -0.006 (0.0199) -0.02 (0.0220) -0.001 (0.0107) -0.01 (0.0130) -0.028* (0.0138) -0.018 (0.0176) -0.014** (0.0049) -0.015 (0.0127)
200 -0.014 (0.0341) -0.017 (0.0294) 0.002 (0.0214) 0.001 (0.0231) -0.007 (0.0105) 0.011 (0.0133) 0.001 (0.0103) 0.017 (0.0108)
500 -0.039 (0.0199) -0.001 (0.0261) -0.028 (0.0183) -0.02 (0.0204) -0.012 (0.0306) -0.006 (0.0187) -0.02 (0.0158) -0.023 (0.0272)
obsRed2000 30 -0.034 (0.0208) -0.017 (0.0220) -0.016 (0.0152) -0.01 (0.0232) -0.016 (0.0094) -0.002 (0.0249) -0.011 (0.0073) -0.005 (0.0138)
50 -0.014 (0.0148) 0.007 (0.0169) -0.016 (0.0143) 0.003 (0.0134) -0.023 (0.0149) -0.026** (0.0095) -0.02 (0.0167) -0.02* (0.0102)
80 -0.018* (0.0073) 0.014 (0.0126) -0.009 (0.0152) 0.029 (0.0216) -0.023 (0.0127) -0.006 (0.0228) -0.012 (0.0080) 0.0 (0.0186)
200 0.013** (0.0043) 0.015** (0.0057) 0.011 (0.0104) 0.016*** (0.0039) 0.016 (0.0128) 0.002 (0.0026) 0.017 (0.0131) 0.0 (0.0060)
500 0.008 (0.0386) -0.024 (0.0340) -0.001 (0.0406) -0.017 (0.0274) -0.019 (0.0117) 0.023 (0.0138) -0.021 (0.0132) 0.019 (0.0105)
superlong Expo 30 -0.003 (0.0068) 0.004 (0.0082) -0.001 (0.0059) 0.004 (0.0079) -0.001 (0.0083) 0.01 (0.0067) 0.003 (0.0077) 0.011 (0.0076)
50 -0.002 (0.0069) 0.001 (0.0063) -0.0 (0.0064) 0.0 (0.0068) 0.005 (0.0067) 0.004 (0.0074) 0.007 (0.0053) 0.001 (0.0051)
80 0.013 (0.0103) 0.004 (0.0103) 0.011 (0.0106) 0.002 (0.0094) -0.005 (0.0075) 0.0 (0.0067) -0.008 (0.0074) -0.003 (0.0046)
200 -0.001 (0.0077) 0.013*** (0.0038) -0.002 (0.0070) 0.009** (0.0034) 0.003 (0.0087) 0.002 (0.0056) -0.002 (0.0070) 0.002 (0.0046)
500 -0.015* (0.0073) -0.005 (0.0184) -0.016* (0.0077) -0.003 (0.0175) 0.007 (0.0119) 0.014 (0.0155) 0.006 (0.0082) 0.022 (0.0147)
Gold2003 30 -0.008 (0.0104) -0.014* (0.0065) -0.006 (0.0087) -0.01* (0.0048) -0.001 (0.0089) -0.001 (0.0068) -0.001 (0.0090) -0.006 (0.0068)
50 0.0 (0.0165) 0.003 (0.0066) 0.004 (0.0137) 0.002 (0.0064) -0.01 (0.0087) -0.007 (0.0108) -0.007 (0.0066) -0.004 (0.0079)
80 -0.009 (0.0117) -0.003 (0.0138) -0.009 (0.0101) -0.008 (0.0146) 0.016 (0.0118) 0.0 (0.0065) 0.007 (0.0104) -0.002 (0.0061)
200 -0.003 (0.0080) -0.008 (0.0056) -0.002 (0.0084) -0.004 (0.0055) -0.006 (0.0046) 0.003 (0.0065) -0.006 (0.0055) 0.002 (0.0059)
500 0.007 (0.0210) -0.008 (0.0205) 0.004 (0.0205) -0.009 (0.0177) 0.001 (0.0132) -0.015 (0.0180) 0.001 (0.0115) -0.018 (0.0176)
Gold2009 30 -0.006 (0.0042) 0.003 (0.0052) -0.007* (0.0034) 0.005 (0.0049) 0.006 (0.0109) -0.002 (0.0070) 0.007 (0.0052) -0.001 (0.0045)
50 -0.016*** (0.0031) -0.003 (0.0049) -0.014*** (0.0040) 0.001 (0.0035) -0.004 (0.0050) -0.001 (0.0051) -0.004 (0.0046) -0.003 (0.0042)
80 0.005 (0.0061) -0.002 (0.0045) 0.005 (0.0056) -0.003 (0.0050) 0.003 (0.0046) -0.005 (0.0065) 0.005 (0.0030) -0.002 (0.0067)
200 -0.011 (0.0106) -0.001 (0.0100) -0.009 (0.0103) -0.006 (0.0099) 0.002 (0.0122) -0.003 (0.0057) -0.004 (0.0124) -0.004 (0.0051)
500 -0.128 (0.0895) 0.114 (0.0993) -0.129 (0.0935) 0.113 (0.0970) -0.034 (0.0316) -0.101*** (0.0300) -0.032 (0.0329) -0.046 (0.0526)
Red1999 30 -0.008 (0.0173) 0.012 (0.0097) -0.006 (0.0148) 0.008 (0.0107) 0.007 (0.0072) 0.009 (0.0082) 0.0 (0.0066) 0.013 (0.0127)
50 0.018 (0.0134) -0.002 (0.0171) 0.008 (0.0091) -0.001 (0.0144) 0.005 (0.0131) 0.007 (0.0108) -0.003 (0.0077) 0.01 (0.0108)
80 0.005 (0.0107) -0.009 (0.0087) 0.004 (0.0112) -0.01 (0.0081) -0.007 (0.0070) -0.002 (0.0116) -0.006 (0.0054) -0.011 (0.0082)
200 0.009 (0.0119) 0.003 (0.0115) 0.009 (0.0115) 0.001 (0.0107) 0.0 (0.0081) 0.007 (0.0059) 0.003 (0.0062) 0.004 (0.0040)
500 -0.019 (0.0225) 0.019 (0.0212) -0.003 (0.0183) -0.002 (0.0131) -0.012 (0.0189) -0.015 (0.0248) -0.013 (0.0156) -0.021 (0.0296)
Red2000 30 0.003 (0.0133) -0.019 (0.0393) 0.014 (0.0184) -0.009 (0.0412) -0.018 (0.0230) 0.028 (0.0255) -0.017 (0.0242) 0.018 (0.0190)
50 0.032 (0.0224) 0.051*** (0.0125) 0.017 (0.0209) 0.036** (0.0128) 0.001 (0.0298) 0.006 (0.0101) -0.005 (0.0275) 0.008 (0.0087)
80 -0.03*** (0.0024) 0.008 (0.0182) -0.032*** (0.0032) 0.011 (0.0182) -0.047 (0.0244) -0.005 (0.0192) -0.035*** (0.0076) -0.003 (0.0096)
200 0.009 (0.0255) 0.011 (0.0208) 0.005 (0.0248) 0.006 (0.0199) -0.001 (0.0086) -0.022 (0.0173) 0.002 (0.0060) -0.021 (0.0118)
500 0.031 (0.0248) -0.007 (0.0181) 0.018 (0.0303) -0.007 (0.0261) -0.011 (0.0297) 0.031** (0.0107) -0.016 (0.0295) 0.024*** (0.0038)
standard error in parentheses (robust errors clustered at station level) *p<.05; **p<.01;
***p<.001
80
Table 22: Parallel trend test estimates by income
movedIn movedInIn movedOut movedOutOut
lineYear lag1 30 -0.001 (0.0039) -0.002 (0.0035) -0.001 (0.0058) -0.001 (0.0048)
50 0.005 (0.0052) 0.003 (0.0044) 0.001 (0.0033) 0.004 (0.0030)
80 -0.002 (0.0049) -0.004 (0.0055) 0.004 (0.0053) 0.004 (0.0054)
200 0.006 (0.0036) 0.003 (0.0038) 0.0 (0.0052) -0.002 (0.0038)
500 0.003 (0.0105) 0.001 (0.0093) -0.009 (0.0066) -0.003 (0.0072)
lag1treatedRed1999 30 -0.001 (0.0056) 0.001 (0.0060) -0.013 (0.0075) -0.009 (0.0068)
50 -0.012 (0.0150) -0.009 (0.0089) -0.01 (0.0110) -0.01 (0.0075)
80 -0.006 (0.0080) 0.003 (0.0065) -0.004 (0.0084) 0.003 (0.0103)
200 -0.01 (0.0063) -0.004 (0.0066) -0.014 (0.0091) -0.001 (0.0061)
500 -0.013 (0.0185) -0.012 (0.0172) -0.021 (0.0290) -0.016 (0.0285)
lag1treatedRed2000 30 -0.009 (0.0054) -0.004 (0.0048) -0.01 (0.0071) -0.002 (0.0071)
50 -0.005 (0.0076) -0.005 (0.0073) -0.006 (0.0054) -0.01* (0.0042)
80 -0.007 (0.0083) -0.003 (0.0107) -0.018** (0.0065) -0.011 (0.0084)
200 0.005 (0.0093) 0.01 (0.0080) 0.006 (0.0057) 0.005 (0.0055)
500 -0.006 (0.0132) -0.003 (0.0120) 0.014 (0.0227) -0.001 (0.0287)
lag1treatedGold2003 30 0.007 (0.0053) 0.01* (0.0045) 0.001 (0.0065) 0.0 (0.0052)
50 -0.004 (0.0072) -0.0 (0.0062) -0.006 (0.0078) -0.008 (0.0067)
80 0.012 (0.0071) 0.009 (0.0070) -0.017* (0.0086) -0.015 (0.0083)
200 -0.008 (0.0052) -0.006 (0.0056) -0.002 (0.0072) 0.001 (0.0061)
500 0.01 (0.0158) 0.009 (0.0141) 0.01 (0.0149) -0.003 (0.0159)
lag1treatedGold2009 30 -0.004 (0.0053) -0.003 (0.0042) 0.0 (0.0067) -0.001 (0.0056)
50 -0.011 (0.0060) -0.006 (0.0051) -0.014** (0.0044) -0.014*** (0.0039)
80 0.002 (0.0059) 0.006 (0.0067) -0.002 (0.0071) -0.002 (0.0064)
200 -0.009 (0.0063) -0.005 (0.0057) -0.017* (0.0078) -0.009 (0.0058)
500 -0.048* (0.0232) -0.043* (0.0208) -0.042 (0.0274) -0.041 (0.0261)
standard error in parentheses (robust errors clustered at station level) *p<.05; **p<.01;
***p<.001
81
Table 23: Parallel trend test estimates on incumbent households
movedIn movedInIn movedOut movedOutOut
incumbents lineYear lag1 30 0.005 (0.0052) 0.003 (0.0050) 0.004 (0.0061) 0.004 (0.0059)
50 0.003 (0.0060) 0.001 (0.0051) 0.001 (0.0040) 0.004 (0.0037)
80 0.001 (0.0075) -0.004 (0.0072) 0.006 (0.0078) 0.004 (0.0080)
200 0.005 (0.0057) 0.0 (0.0056) 0.003 (0.0065) 0.003 (0.0060)
500 0.006 (0.0059) 0.004 (0.0053) -0.018* (0.0086) -0.014 (0.0081)
lag1treatedRed1999 30 -0.008 (0.0070) -0.005 (0.0065) -0.019* (0.0084) -0.013 (0.0080)
50 -0.008 (0.0157) -0.011 (0.0108) -0.005 (0.0115) -0.009 (0.0070)
80 -0.005 (0.0104) 0.005 (0.0088) 0.002 (0.0100) 0.003 (0.0110)
200 -0.009 (0.0097) -0.002 (0.0096) -0.009 (0.0099) -0.004 (0.0082)
500 -0.015 (0.0187) -0.015 (0.0162) -0.025 (0.0363) -0.009 (0.0371)
lag1treatedRed2000 30 -0.009 (0.0064) -0.004 (0.0056) -0.006 (0.0092) -0.003 (0.0089)
50 0.004 (0.0092) 0.004 (0.0071) -0.008 (0.0067) -0.011* (0.0050)
80 -0.014 (0.0114) -0.01 (0.0130) -0.022* (0.0104) -0.017 (0.0115)
200 0.006 (0.0116) 0.01 (0.0112) -0.001 (0.0064) -0.003 (0.0078)
500 -0.015 (0.0102) -0.008 (0.0091) 0.011 (0.0105) -0.003 (0.0144)
lag1treatedGold2003 30 0.005 (0.0059) 0.009 (0.0056) -0.005 (0.0065) -0.004 (0.0053)
50 -0.003 (0.0078) -0.001 (0.0059) -0.003 (0.0083) -0.006 (0.0062)
80 0.0 (0.0087) 0.002 (0.0081) -0.011 (0.0101) -0.01 (0.0101)
200 -0.005 (0.0068) -0.001 (0.0067) -0.006 (0.0079) -0.004 (0.0072)
500 -0.003 (0.0123) -0.0 (0.0118) 0.017 (0.0158) 0.007 (0.0149)
lag1treatedGold2009 30 -0.003 (0.0067) -0.002 (0.0061) -0.001 (0.0074) -0.004 (0.0068)
50 -0.007 (0.0071) -0.005 (0.0058) -0.011** (0.0041) -0.015*** (0.0044)
80 -0.002 (0.0083) 0.006 (0.0084) -0.005 (0.0080) -0.001 (0.0077)
200 -0.007 (0.0072) -0.001 (0.0070) -0.01 (0.0076) -0.009 (0.0068)
500 -0.029 (0.0238) -0.036* (0.0143) -0.02 (0.0279) -0.014 (0.0262)
standard error in parentheses (robust errors clustered at station level) *p<.05; **p<.01;
***p<.001
82
Income Distributions
Figure 9: Annual LTM income distribution in dense parts of LA County
83
Figure 10: Annual LTM Income Distribution by Treatment Group
84
Mobility rates
Figure 11: Station-to-station mobility rates
In and Out Mobility Denition
When labeling the year in which a household moved out or in, we have two options. First, if a
ler moved then she can be marked as moved-out in yeart1 and moved-in in yeart. Alternatively,
we can mark the ler's moved-in and moved-out time in the same year though then the choice of
whether it should bet1 or yeart arises. Of course, if a ler is observed in locationA in yeart1
and location B in year t then her residence could have changed anytime between April t 1 and
April
27
t. Presumably, if a ler moved then she moved directly from place A to place B so it does
not make much sense to say that she moved out in year t 1 and moved-in in yeart. On the other
hand, we do not know in which year exactly she moved. Moreover, we also do not know whether
her new residenceB was the residence to which she moved fromA since she could have moved from
A to residence C or D or any number of intermediate locations prior to settling in B where taxes
were led. All of this introduces error into our mobility-estimates. We know from experimentation
that if we choose the rst option where movers are marked as moved-out in yeart1 and moved-in
in year t then we get move-out and move-in trends that exhibit similar but shifted trends. For
example, if there is a spike in move-out rates in year t 1 then we'll see a similar spike in move-in
rates in year t. Although it is possible the spikes occurred in separate years if vacated housing
units are not lled immediately, because the housing market in Los Angeles is tight, we believe it
is more likely that we observe turnover increases so that if move-out rates spike then move-in rates
to spike too.
Suppose now that when comparing year pairst 1 andt, we label households as moved-out(in)
in yeart. Suppose a station opened in year 2000 and that there is a signicant impact on mobility
rates in the year that the station opens. When comparing households in year 1999-2000, we can
27
Note, tax season technically ends in April but we observe some lers who le years after or even early so April
is by no means binding.
85
label households as moved-out(in) either in year 1999 or 2000
28
. If a signicant share of lers
moved in 2000 as a result of station-opening but we label the year of move-out(in) in 1999 then if
we put in a placebo one year prior to the station's opening date then the placebo test on parallel
trends will falsely fail. On the other hand, if we label households as moved-out(in) in year 2000
then even if some households who moved in year 1999 are captured by the year 2000, the placebo
test will not fail falsely. Moreover, if there are any households who moved-out(in) in year 1999 due
to anticipatory eects then that eect will be captured in the coecient on station-opening among
households near rail-transit. This suggests that we ought to mark households as moved-out(in) in
year t. However, all of this misses a crucial issue in the data structure.
When a household moves away from a rail station in year t, it disappears from the dataset.
Similarly, if a household moves into a station area from somewhere beyond the control and treatment
basins then it is not present in the data in year t 1. This has two practical implications. First,
when assigning move-in and move-out actions to year t then by construction, the two trends are
identical which defeats the purpose of the analysis. The second issue is that if a household moves out
of all station areas then we see the household in yeart 1 as a mover but sincet is no longer in the
dataset, there is no variable that shows that this household moved out. As a result, all households
that move out of stations areas are mislabeled as non-movers which severely undercounts households
that move-out.
For all of the above reasons, we dene move-out as the last year in which a household is seen
at a rail station provided it was in the dataset in years t and t 1 while move-in is the rst year
in which we observe a household at a station area.
Data Details
To associate coordinates with 9-digit zip codes, and, thus, link them to LA Metro rail station
locations, we use the Geolytics database on 9-digit zip code centroids for years 2000, 2002, 2004,
2007, 2009, 2011, 2012, 2013, and 2016. If coordinates are not available for a particular year, we
forward and backward ll coordinates from years for which coordinates were available. Although
this may introduce error into zip locations, we are not too worried because fewer than 1% of all
California 9-digit zip codes moved more than 100 meters between 2000 and 2012 (Boarnet, Bostic,
Rodnyansky, Prohofsky, Eisenlohr, and Jamme 2018). We measure the straight-line distance of
every 9-digit zip code to every LA Metro rail station and associate the closest rail station within
a particular distance band to each zip. During the matching process, we encounter the issue that
some coordinates for 9-digit zip codes change across the years. Since we associate lers to stations
via zip codes, a change in a zip's coordinates can associate a household to dierent stations without
any changes in the household's zip or residence. Unfortunately, we do not know whether zip location
changes are deliberate or erroneous. Overall, we have 230,000 9-digit zip codes within 2km of a rail
station that are related to a station in every year. Of these, 4,754 are associated with 2 stations
across the years and 62 with more than 3 stations. We keep all 9-digit zips with 2 or less station
associations or 229,566 codes. Additionally, there are about 950 zip codes that are within .5 miles
of a rail station in some years but 900-1300m in others which means they fall into both control and
treated areas. Because we do not know why a particular zip code's location changed between some
28
We can decompose the set of households who moved-out(in) between years t 1 and t into 4 groups: at1, at,
bt1 and bt where a us the share of movers impacted by rail transit opening and b is the share that are not. Don't
know if this is a worthy pursuit, this can discuss lagged and auto-correlational issues
86
years, these zip codes can result in the contamination of the treatment groups and falsely label
some households as movers so we exclude them.
Comparison with Census data
Since this is a novel dataset and not everyone les taxes, we test whether it is representative of
the underlying population. Tax compliance and ling rates are high in the United States (IRS 2010,
IRS 2012, IRS 2017). Mandatory federal requirements toward the end of our study period in 2013
for income tax ling were incomes of $20,000 and above for married households and $10,000 and
above for single households (IRS 2013). In California, single persons with annual incomes above
$12,562 are required to le state income tax while married persons with annual incomes above
$25,125 must le (FTB 2013). The FTB estimates that 89 percent of California residents who
were required to le taxes did le as of the mid 2000s (FTB 2006)
29
. Persons with incomes below
the ling requirement still have incentives to le, for example, to be eligible for the federal earned
income tax credit (EITC). The IRS estimates that in 2014, 76 percent of California residents who
were federal EITC-eligible led for federal (IRS 2014). These gures suggest that our data should
be close to the universe of income-earning households.
To verify, we compare our data against population data from the American Community Survey
(ACS) and U.S. Census. For each Census Block Group located within .5 miles of an L.A. Metro rail
station, we pull the 1990, 2000, and 2010 Census estimates and supplement the sample with the
5-year 2011-2015 ACS population estimates. In Table 25, we compare years 2000 and 2010 directly
with the FTB household counts while for 1990 and 2011-2015 periods we compare to the 1994-1999
and 2011-2014 FTB average annual counts respectively. Furthermore, we compare two types of
gures. Under Total, we compare the number of FTB households in our station areas against the
ACS/Census total population in these stations. Under Mean, we compare the average number of
FTB households in our station areas against the average number in ACS/Census data. As seen in
column percent of Census (Total), our household counts are approximately 54-72% of the Census
estimates. When we compare the average FTB station household counts to those of the Census,
we get a very similar result in column percent of Census (median). In general, we expect FTB
counts to underestimate the actual count of households but there are some additional variations
that may help explain the discrepancy. Not all observations in all FTB years were geocoded at
some zip code level and that tends to vary from 97-100% for the majority of our sample (Table
24). Of these observations, the share that are geocoded to the 9-digit zip code level range in the
45-55%
30
. Within this sample, we can track 70-80% of households across years t 1 and t. These
losses result in a wide range of usable sub-sample from a low of 20% in the early 1990s to high
mid-40% in the post-2000 data. As seen in Table 24, the share of lers with 9-digit zip codes
increases rapidly from 39% to 47% between 1994 and 1997. Similarly, the share of households who
le in consecutive years also increases drastically from 63% in 1994 to 76% in 1999. As a result, the
growth in our population looks explosive with an almost 100% increase in the 7 years between 1994
and 2001. Although evidence from the Census (Table 25) does suggest a growth in population near
LA Metro rail stations, it is highly unlikely that the population grew as quickly as that suggested
by our data. More generally, according to Table 24, prior to 1996, less than 30% of our data could
be used for analysis while 1996 onwards, 30-40% of households meet the 9-digit geocoded zip code
29
The California Franchise Tax Board last assessed ling tax rate in 2006 but they take steps every year to ensure
compliance with tax laws.
30
Earlier years (1993-1996) in our sample have much lower shares geocoded to any and the 9-digit zip code level
87
and 2-consecutive year requirements to be included in our sample. To achieve some homogeneity,
for residential mobility estimates we only estimate on years in which more than 30% of the data
were available for use in our analysis which excludes 1993-1996
31
and leaves us with observation
years 1997-2015.
Table 24: Share of lers included
Number of lers geocoded share share geocoded to 9-digit zip level share of lers in both years usable share of lers
1993 5634125 0.972 0.181 NaN NaN
1994 5670780 0.978 0.362 0.440 0.156
1995 5774398 0.982 0.440 0.633 0.274
1996 5772398 0.988 0.467 0.686 0.317
1997 5929430 0.986 0.487 0.713 0.342
1998 6083822 0.985 0.499 0.725 0.356
1999 6223989 0.990 0.513 0.746 0.379
2000 6354230 0.992 0.520 0.757 0.390
2001 6507393 0.993 0.528 0.766 0.402
2002 6550678 0.997 0.525 0.777 0.407
2003 6480527 0.998 0.521 0.765 0.398
2004 6738562 0.998 0.525 0.767 0.402
2005 6856892 0.998 0.521 0.772 0.401
2006 6981286 0.998 0.526 0.778 0.408
2007 7252045 0.998 0.536 0.786 0.420
2008 7095101 0.999 0.530 0.786 0.416
2009 7112825 0.999 0.533 0.785 0.418
2010 7125754 0.999 0.533 0.786 0.419
2011 7478334 0.999 0.549 0.786 0.431
2012 7451093 0.999 0.555 0.780 0.432
2013 7315914 0.999 0.549 0.790 0.433
2014 7663030 0.999 0.558 0.794 0.443
2015 7645201 0.999 0.524 NaN NaN
Table 25: FTB Sub-sample and Census data
Total Mean
census FTB percent of Census census FTB percent of Census
year
1990census 55839 37875 0.678 1095 728 0.66
2000census 91471 65829 0.720 1794 1266 0.71
2010census 142849 77102 0.540 2801 1483 0.53
acs5 2015 145666 79764 0.548 2856 1547 0.54
*FTB data are averages of 1994-1999 and 2011-2014 periods.
Noise and rate volatility
Mobility rates for some income groups will inherently suggest higher entropy than others. For
example, extremely poor households (0-30%) occupy the largest share of the income distribution
every year across the sample suggesting that the base of our mobility rate estimates are higher
and; therefore, more precise. On the other side, small number of households in the 200-500% of
AMI group occupy so even an additional 10 or 20 moving households can drastically shift mobility
rates. Consequently, we expect mobility rates for poor households to have the highest precision and
mobility rates for our highest income category to be noisiest or most volatile. Moreover, although
the income categories 0-200% of AMI are similar in size, the smaller size of the 200-500% may also
impact our statistical power since deviations from the average are more likely to be detected for
poor households than non-poor households.
31
Note there is likely substantial variation in the share of usable data. Dense, urban parts of LA County likely
contain higher share of households geocoded to 9-digit zip code levels but we eschew these complications in our
analysis.
88
In and Out Appendix
Data Details
In order to construct HHIs, I use the publicly available data sets on parcel ownership across
select jurisdictions in the US. The only data set not publicly available is the Los Angeles County
assessor database which was acquired via a Freedom of Information Act request submitted by
Los Angeles Times. The data contain all commercial, manufacturing, and residential properties
in their jurisdictions at the time the data were pulled. In case of the Milwaukee study, the data
were pulled in 2018. Since I am only concerned about residential units, I exclude all observations
that are not zoned residential or designated as residential in their land use. Neither zoning nor
land-use designations are not always correct and they sometimes con
ict (ie multi-family property
on an airport). As a result, I also use presence of residential units as additional criteria to select
residential properties. Although all hand-checked properties were in fact residential, data errors
and inconsistencies leave the possibility of having non-residential properties in the data.
Another ubiquitous issue arises from parcels with single-family land-uses or zoning having more
than one unit and vice versa so unit counts are not guaranteed to be correct. In case of Milwaukee,
at least two hand-checked unit counts were corrected after online investigation. However, unit
counts seem to be in the ballpark with other counts. In case of Milwaukee, my data contain
150,510 residential addresses representing 257,453 units. According to the Census, there were an
estimated 230,784 households in Milwaukee and 260,581
32
housing units which suggests the Tax
Assessor unit count is fairly close to the universe of housing units. According to the Census
33
, Los
Angeles County had 3.6 million housing units and 3.3 million households in 2018 and 2013-2017
respectively but my assessor data suggest there are only 3.1 million housing units once I lter out
properties with non-residential land-use codes. Although it is both strange that there are half a
million more households than units
34
and that the assessor estimate is lower, I assume that the LA
County assessor data are representative of the County housing market.
To compute multi-family ownership shares, I use properties with non-single family land use
designations or properties with more than 1 unit. In case of Milwaukee, there is a
ag for owner-
occupied units yet 10% of these parcels also have more than 1 unit suggesting that
ag may fail to
accurately distinguish between renter and owner-occupied properties.
Due to idiosyncrasies in the owner name and address records, my Herndahl-Hirschman In-
dexes are almost certainly under-estimated. Though all jurisdictions have this feature, I will only
exemplify Milwaukee. According to the data, there are 114,703 unique owners in the data and
108,111 unique owner addresses. Clearly, mistakes in names and addresses exist. For example,
DOORS DWELLINGS LLC owns two separate properties but the other property is recorded under
DOOR4S DWELLINGS LLC. Minor mistakes in addresses such as W7026 COTTTONVILLE DR
and W7026 COTTONVILLE DR also exist. It is not possible to correct all of these mistakes so
inevitably some commonly owned properties will show up as separately owned properties in my
data. Thus, it is with near certainty that my data undercount the number of commonly owned
properties. The other concern is that many properties are owned by LLCs so even if I can identify
properties owned commonly by an LLC, it is impossible for me to tell whether multiple LLCs have
the same owner. For example, ZYCH LIVING TRUST D8-24-10 and ZYCH TRUST DATED
32
https://censusreporter.org/profiles/16000US5553000-milwaukee-wi/
33
https://www.census.gov/quickfacts/losangelescountycalifornia
34
Los Angeles County has one of the nation's highest doubling-up rates
89
5-4-96 are almost certainly owned by the same set of people but I cannot know this unambiguously.
As such, I must contend with the fact that my HHIs underestimate market power.
Slope Sensitivity Appendix
Slope sensitivity to sample
Tables 26 and 27 show the estimates of equation 19 on Milwaukee using dierent sample subsets.
The coecient on income varies substantially depending on whether I trim the top and bottom
rents. For example, level regressions tted on sub-samples that did not restrict rents suggest that
the rent-to-income slope is $27 for every $10,000 increase in household incomes whereas the trimmed
samples imply a $14 and $19 slope for the 10% and 5% trims respectively. Even with the median
regressions, coecient estimates using a trimmed sub-sample show smaller slope estimates. This
implies that coecient estimates are highly sensitive to the inclusion of the lowest-cost units. The
log-log variant implies similar pattern in the coecient's variation, suggesting an income to rent
elasticity that ranges from .056% to .18% depending on whether the lowest segment is included or
not.
Table 26: Level-level rent regressions
no sample restrictions working age 10% rents trimmed 5% rents trimmed no top chop applied to all
delta 0.0024 (0.000) 0.0025 (0.000) 0.0016 (0.000) 0.0025 (0.000) 0.0025 (0.000) 0.0026 (0.000)
invMills 265.7359 (0.000) 134.9378 (0.006) 89.0268 (0.008) 112.8812 (0.018) 112.8812 (0.018) 101.0731 (0.035)
R^2 adj 0.34 0.37 0.36 0.38 0.38 0.39
nObs 894 710 655 692 692 673
Table 27: Log-log rent regressions
no sample restrictions working age 10% rents trimmed 5% rents trimmed no top chop applied to all
delta 0.3184 (0.000) 0.2872 (0.000) 0.1169 (0.000) 0.1877 (0.000) 0.1877 (0.000) 0.1851 (0.000)
invMills 0.0931 (0.294) 0.0349 (0.728) 0.0897 (0.053) 0.1071 (0.073) 0.1071 (0.073) 0.1053 (0.077)
R^2 adj 0.28 0.24 0.34 0.37 0.37 0.38
nObs 894 710 655 692 692 673
90

Abstract (if available)

Abstract This dissertation is a collection of three empirical essays in Urban and Housing Economics that explore the factors behind residential security and housing rents among low-income renters. The first chapter examines the impact that the introduction of rail-transit has on households' residential mobility and, in turn, the neighborhood income distribution. In essence, it answers whether transit-induced neighborhood shocks cause neighborhoods to gentrify through an influx of high income households and an outflow of low-income residents. The chapter uses 20 years of micro-data on household income and place of residence provided by the California Franchise Tax Board to assess residential mobility flows before and after a transit station opens in Los Angeles County using a Difference-in-Difference approach. The study finds substantial variation in impact across low-income households and transit lines. Residential patterns among extremely low-income households suggest a preference for locating near public transit opportunities while patterns among very low-income households suggest that positive neighborhood shocks may reduce their ability to stay in transit neighborhoods. Residential flows among low and higher-income households do not seem to be impacted by the opening of transit stations. Although the study finds that station-opening causes some changes in residential mobility patterns, these shifts are largely drowned out by the same macro-factors that drive distributional shifts across urban Los Angeles County. The second chapter proposes and empirically supports the theory that the ability of households to become homeowners has an impact on rent distributions. The study uses American Housing Survey and HDMA mortgage data along with an instrumented variable approach to show that markets in which households have a more difficult time purchasing a house have substantially wider rent distributions than markets in which access to ownership is easier. Moreover, the impact is not uniform across renter households as high-income renters are impacted substantially more than low-income renters. This suggests that policies that decrease barriers to ownership primarily benefit higher-income renters. The third chapter shows the existence of high markups in Milwaukee's low-income rental market and tests whether high ownership concentrations enable landlords to maintain the markups. The study pairs ownership records from Milwaukee's tax assessor database with rent and CoStar multi-family housing data to implement an instrumented variable approach that tests the impact of ownership concentrations on rents. Estimates show that higher ownership concentrations raise rents and that differences in ownership concentrations between high and low income neighborhoods explain about a third of landlords' ability to maintain high markups.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

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

PDF

Three essays on housing demographics: depressed housing access amid crisis of housing shortage

PDF

Household mobility and neighborhood impacts

PDF

Reshaping Los Angeles: housing affordability and neighborhood change

PDF

Housing market crisis and underwater homeowners

PDF

Environmental justice in real estate, public services, and policy

PDF

The interactions between housing and business

PDF

Three essays on aging, wealth, and housing tenure transitions

PDF

Urban air pollution and environmental justice: three essays

PDF

A better method for measuring housing affordability and the role that affordability played in the mobility outcomes of Latino-immigrants following the Great Recession

PDF

Crowded with potential: housing and social mobility strategies among China's educated migrants

PDF

Emissions markets, power markets and market power: a study of the interactions between contemporary emissions markets and deregulated electricity markets

PDF

Active travel, outdoor leisure, and neighborhood environment: path analysis, Los Angeles County

PDF

The role of the timing of school changes and school quality in the impact of student mobility: evidence from Clark County, Nevada

PDF

Healthy mobility: untangling the relationships between the built environment, travel behavior, and environmental health

PDF

Essays on the economics of cities

PDF

Risks, returns, and regulations in real estate markets

PDF

The economic and political impacts of U.S. federal carbon emissions trading policy across households, sectors and states

PDF

Spatial and temporal expenditure-pricing equity of rail transit fare policies

PDF

Student mobility in policy and poverty context: two essays from Washington

Asset Metadata

Creator Burinskiy, Evgeny (author)

Core Title Insights into residential mobility and pricing of rental housing: the role of gentrification, home-ownership barriers, and market concentrations in low-income household welfare

School School of Policy, Planning and Development

Degree Doctor of Philosophy

Degree Program Urban Planning and Development

Degree Conferral Date 2021-08

Publication Date 08/04/2021

Defense Date 04/20/2021

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

Tag Economics,Housing,housing market,OAI-PMH Harvest,Urban,urban economics

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Green, Richard K. (committee chair)

Creator Email gene.burinsky@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC15709329

Unique identifier UC15709329

Legacy Identifier etd-BurinskiyE-9996

Document Type Dissertation

Format application/pdf (imt)

Rights Burinskiy, Evgeny

Type texts

Source 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. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

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