How Do Cities Change When We Work from Home?
Matthew J. Delventhal
Eunjee Kwon
Andrii Parkhomenko
This version: December 4
th
, 2020.
Abstract
How would the shape of our cities change if there were a permanent increase in working
from home? We study this question using a quantitative model of the Los Angeles metropoli-
tan area featuring local agglomeration externalities and endogenous traffic congestion. We
find three important effects: (1) Jobs move to the core of the city, while residents move to the
periphery. (2) Traffic congestion eases and travel times drop. (3) Average real estate prices
fall, with declines in core locations and increases in the periphery. Workers who are able to
switch to telecommuting enjoy large welfare gains by saving commute time and moving to
more affordable neighborhoods. Workers who continue to work on-site enjoy modest welfare
gains due to lower commute times, improved access to jobs, and the fall in average real estate
prices.
Keywords: COVID-19, urban, work at home, commuting.
JEL codes: E24, J81, R31, R33, R41.
We thank Gilles Duranton, Edward Glaeser, Jeff Lin, Jorge De la Roca, and two anonymous referees, as well as
seminar participants at USC and the AREUEA virtual seminar for useful comments and discussions. The authors
acknowledge generous grant funding from METRANS-PSR (grant number 65A0674) and the USC Lusk Center for
Real Estate.
The Robert Day School of Economics and Finance, Claremont McKenna College, Claremont, CA 91711 mdel-
Department of Finance and Business Economics, Marshall School of Business, University of Southern Califor-
1
1 Introduction
The potential savings in overhead costs and commuting time from remote work are significant.
1
Technological conditions have been improving steadily for years, yet the fraction of Americans
working from home has remained small. In 2019, just 4.2% of all workers worked from home. In
2020, COVID-19 social distancing requirements forced many companies and organizations to pay
a part of the fixed cost of transition to remote work. Abundant survey evidence suggests that
many now plan to continue remote work at much higher rates even after the pandemic is over.
2
A lasting increase in working from home could have far-ranging consequences for the distribu-
tion of economic activity inside urban areas.
3
One of the critical factors driving workers’ location
choices is the need to commute between their job and their residence. Increasing the number of
telecommuters makes this trade-off moot for a significant fraction of the workforce. In this paper,
we quantify the potential impact of this change using a general equilibrium model of internal city
structure. The model features employment, residence, and real estate development choices, as
well as local agglomeration and congestion externalities, and endogenous traffic congestion across
3,846 non-rural census tracts of the Los Angeles-Long Beach combined statistical area.
We calibrate our model to match residence and employment patterns prevalent in Los Angeles
during the period 2012–2016, with an average of 3.7% of workers working from home. We then
conduct a counterfactual exercise in which we gradually increase the fraction of telecommuters
all the way to 33%, which according to Dingel and Neiman (2020), corresponds to the share of
L.A. metro area workers whose jobs could be performed mostly from home. The effects on city
structure over the long run can be broken into three categories.
First, jobs relocate to the core of the urban area, while residents move to the periphery. The
largest driver of this effect is workers who previously had to commute and can now work at home.
They tend to move farther away from the urban core to locations with more affordable houses.
This increases demand for real estate in peripheral locations and lowers demand in the core,
pushing jobs from the suburbs into more central locations.
Second, average commuting times fall, while commuting distances increase. Since fewer work-
ers commute, traffic congestion eases, which increases average speed of travel. Commuters take
advantage of this and also move farther away from their workplaces to live in locations with lower
real estate prices.
1
Mas and Pallais (2020) provide an overview of the current state of research in telecommuting. Bloom, Liang,
Roberts, and Ying (2015) present experimental evidence that telework increases employee work satisfaction without
necessarily reducing their productivity.
2
A May 2020 survey by Barrero, Bloom, and Davis (2020) finds that 16.6% of paid work days will be done
from home after the pandemic ends, compared to 5.5% in 2019. Results of a survey by Bartik, Cullen, Glaeser,
Luca, and Stanton (2020) also indicate that remote work will be much more common after the pandemic.
3
A study by Upwork in October 2020 finds that since the beginning of the pandemic 2% of survey participants
had already moved residences because of the ability to work at home and another 6% planned to do so (Ozimek,
2020).
2
Third, average real estate prices fall. As many workers move into distant suburbs, prices in
the periphery increase. However, these price increases are more than offset by the decline of
prices in the core. This decline is driven by two factors. The first is the decline in demand for
residential real estate in core locations. The second is the reduced demand for on-site office space
from workers who now telecommute. In the counterfactual where 33% of workers telecommute,
average house prices fall by nearly 6%.
In addition to these three broad trends, our quantitative model predicts considerable hetero-
geneity in outcomes that is not accounted for by the simple core-periphery continuum. Within the
core, locations with high productivity gain jobs while less productive locations lose them. At all
distances from the center, locations with better exogenous residential amenities either gain more
or lose fewer residents than less attractive equi-distant locations. Overall, the single monocentric
dimension of distance from the center only accounts for about half of all variation in predicted
outcomes.
The shift to telecommuting implies changes in the income both of workers and the owners
of real estate. On the one hand, labor productivity is pushed upward as jobs leave peripheral
areas, and employment in the most productive tracts increases. Productivity receives a further
boost from the accompanying increase in spatial agglomeration externalities. Simultaneously,
labor productivity is pushed downward because more employees work at home and teleworkers do
not contribute to agglomeration. In our quantitative exercise, these two effects offset each other
almost completely, leading to very small increases in average wages. At the same time, changes
in the spatial distribution of real estate demand and the reduced need for office space lead to
lower real estate prices and thus a reduction in the income earned by landowners and property
developers.
Our results conform fundamentally with previous theoretical findings by, for example, Safirova
(2003), Rhee (2008), and Larson and Zhao (2017). Recent work by Lennox (2020) explores the
effects of working from home in an Australian context using a quantitative spatial equilibrium
model. A related study of ours, Delventhal and Parkhomenko (2020), extends the analysis to the
entire U.S. and multiple types of telecommuters.
This paper also follows a number of recent efforts to assess the impact of urban policies and
transport infrastructure on city structure, such as those by Ahlfeldt, Redding, Sturm, and Wolf
(2015), Severen (2019), Tsivanidis (2019), Owens, Rossi-Hansberg, and Sarte (2020), and Anas
(2020). Our paper uses a similar framework to assess the impact of a change to the underlying
technology of production on urban structure.
The remainder of the paper is organized as follows. Section 2 describes the model. Section
3 provides an overview of how we calibrate the model. Section 4 describes and discusses the
counterfactual exercises. Section 5 concludes.
3
2 Model
Consider an urban area with I discrete locations, each populated by workers, firms, and
floorspace developers. The total employment of the urban area is fixed and normalized to 1.
Workers supply their labor to firms and consume residential floor space and a numeraire con-
sumption good. Workers suffer disutility from time spent in commuting between home and work,
and this time depends endogenously on aggregate traffic volume. Their choice of residence and
employment locations depends on the commuting time, wages at the place of employment, housing
costs and amenities at the place of residence, and idiosyncratic location preferences. Residential
amenities depend on agglomeration spillovers, which are increasing in the residential density of
nearby locations. Firms use labor and commercial floorspace to produce the consumption good,
which is traded costlessly inside the urban area. Firms’ total factor productivity depends on
agglomeration spillovers, which are increasing in the density of employment in nearby locations.
Developers use land and the numeraire to produce floorspace, which can be put to residential or
commercial use. The supply of floor space in each location is restricted by zoning regulations that
limit commercial development and overall density.
We introduce work from home by proposing a second type of worker–the telecommuter.
Telecommuters only come to their worksite a small fraction of workdays and thus suffer much
less disutility from commuting. On the days that they are not in the office, they do not use com-
mercial floorspace and instead produce output using “home office” floorspace in their residence
location. Working from home uses floorspace less intensively than on-site work, has a different
total factor productivity, and neither contributes to nor benefits from agglomeration spillovers.
This model is similar in many respects to Ahlfeldt, Redding, Sturm, and Wolf (2015). The
remainder of the section presents the model and Appendix B provides additional details.
2.1 Workers
2.1.1 Commuters and Telecommuters
Before choosing where to work and where to live, workers draw their commuter type. With
probability ψ ≥ 0, a worker becomes a “telecommuter.” With probability 1 − ψ, the worker
becomes a “commuter.” The two types differ in the fraction of workdays they commute to work,
θ. Commuters must come daily and therefore have θ = 1, while telecommuters have θ = θ
T
< 1.
4
2.1.2 Preferences
A worker n who resides in location i ∈ {1, ..., I}, works in location j ∈ {1, ..., I}, and has to
commute from i to j a fraction θ of time, enjoys utility
U
ijn
(θ) =
z
ijn
d
ij
(θ)
c
1 − γ
1−γ
h
γ
γ
, (1)
where z
ijn
represents an idiosyncratic preference shock for the pair of locations i and j, and d
ij
(θ)
is the disutility from commuting given by d
ij
(θ) = (1 − θ) + θe
κt
ij
. Individuals consume c units
of the final good and h units of housing. The share of housing in expenditures is given by γ,
and consumption choices are subject to the budget constraint (1 + τ)w
ij
(θ) = c + q
i
h. In this
constraint, w
ij
(θ) is the wage earned by a worker who commutes from i to j a fraction θ of days,
and q
i
is the price of residential floorspace in location i. In addition to wages, workers also earn
proportional transfers, τw
ij
(θ), which distribute income from land and the consumption good sold
to real estate developers equally among all city workers.
Idiosyncratic shocks z
ijn
are drawn from a Fr`echet distribution with c.d.f. F
z
(z) = e
−z
−
. The
indirect utility of worker n who lives in location i and works in location j is given by u
ijn
(θ) =
z
ijn
v
ij
(θ), where
v
ij
(θ) ≡
X
i
E
j
w
ij
(θ)
d
ij
(θ)q
γ
i
(2)
is the utility obtained by a worker, net of the preference shock. In the above formulation, X
i
is
the average amenity derived from living in location i, and E
j
is the average amenity derived from
working in location j.
Commuting time is a function of total vehicle miles traveled and road capacity in the city:
t
ij
= t
ij
(V MT, Cap). We assume that the capacity is fixed and the elasticity of time on each link
(i, j) with respect to total volume is a constant ε
V
. Appendix E provides more details.
2.1.3 Location Choices
Optimal choices imply that the probability that a worker with a given θ chooses to live in
location i and work in location j is
π
ij
(θ) =
(X
i
E
j
w
ij
(θ))
(d
ij
(θ)q
γ
i
)
−
P
r∈I
P
s∈I
(X
r
E
s
w
rs
(θ))
(d
rs
(θ)q
γ
r
)
−
. (3)
5
As a result, the equilibrium residential population of workers with a given θ in location i, and the
equilibrium employment in location j are given by
N
Ri
(θ) =
I
X
j=1
π
ij
(θ) and N
W j
(θ) =
I
X
i=1
π
ij
(θ). (4)
Finally, total residential population is N
Ri
= (1 − ψ)N
Ri
(1) + ψN
Ri
(θ
T
), and total employment is
N
W j
= (1 − ψ)N
W j
(1) + ψN
W j
(θ
T
).
2.2 Firms
2.2.1 Production
In each location, there is a representative firm which hires both on-site and remote labor and
produces a homogeneous consumption good which is traded costlessly across locations. The total
output of the firm in location j is Y
j
= Y
C
j
+ Y
T
j
, where Y
C
j
and Y
T
j
are the amounts produced
on-site and remotely, respectively. The on-site production function is given by
Y
C
j
= A
j
N
C
W j
α
H
C
W j
1−α
, (5)
where N
C
W j
= (1 − ψ)N
W j
(1) + θ
T
ψN
W j
(θ
T
) is the supply of on-site labor, H
C
W j
is commercial
floorspace, and α is the labor share. The remote production function is also Cobb-Douglas and it
combines workers from different locations as follows:
Y
T
j
= νA
j
X
i∈I
N
T
ij
α
T
H
T
ij
1−α
T
. (6)
In this specification, N
T
ij
= (1 − θ
T
)ψπ
ij
(θ
T
) is the supply of remote labor of telecommuters who
reside in location i and work for a firm in location j, whereas H
T
ij
is the amount of home office
space the firm rents on behalf of these workers in the place of their residence.
4
Parameter ν is the
productivity gap between on-site and remote work, common to all workers and firms. We let the
labor share in remote production, α
T
, to be different from the labor share in on-site production.
5
2.2.2 Wages
Firms take wages and floorspace prices as given, and choose the amount of on-site labor,
telecommuting labor, and floorspace that maximize their profits. Equilibrium payments for on-
4
We assume that the firm rents the floorspace that remote workers need in order to work from home, however
this specification is isomorphic to the one in which the firm only pays for labor services of a telecommuter and the
telecommuter uses his labor income to rent additional floorspace in his house.
5
One may expect that α < α
T
because telecommuters tend to work in jobs that require little floorspace. While
we do not impose this inequality in our theoretical analysis, it holds in our calibration.
6
site work at location j and remote work for a firm in location j while living in location i are,
respectively,
w
C
j
= αA
1
α
j
1 − α
q
j
1−α
α
and w
T
ij
= α
T
(νA
j
)
1
α
T
1 − α
T
q
i
1−α
T
α
T
, (7)
where q
j
is the local price of floorspace. The take-home wage of a worker with a given θ is the
weighted average of payments to his commuting labor and his telecommuting labor: w
ij
(θ) =
θw
C
j
+ (1 − θ)w
T
ij
.
6
2.3 Developers
There is a large number of perfectly competitive floorspace developers operating in each loca-
tion. Floorspace is produced using the following technology:
H
i
= K
1−η
i
(φ
i
(H
i
)L
i
)
η
, (8)
where L
i
≤ Λ
i
and K
i
are the amounts of land and the final good used to produce floorspace,
and η is the share of land in production. Λ
i
is the exogenous supply of buildable land, and
in equilibrium it is optimal for developers to use all buildable land, i.e., L
i
= Λ
i
. Function
φ
i
(H
i
) ≡ 1 −
H
i
¯
H
i
determines the local land-augmenting productivity of floorspace developers.
7
Parameter
¯
H
i
determines the density limit in tract i. When H
i
approaches
¯
H
i
, φ
i
(H
i
) approaches
zero. As a result, it becomes very costly to build due to regulatory or political barriers, such as
zoning, floor-to-area ratios, or local opposition to development.
Floorspace has three uses: commercial, residential, and home offices. Commercial floorspace
can be purchased at price q
W j
per square foot. Residential and home office floorspace is located in
the same structure (e.g., a house) and each can be bought at price q
Ri
. Developers sell floorspace
at price ¯q
i
≡ min {q
Ri
, q
W i
} to either residential or commercial users. However, the effective price
that residents or firms pay for floorspace may differ from ¯q
i
due to zoning restrictions. The wedge
between prices for residential and commercial floorspace is denoted by parameter ξ
i
> 0. If ξ
i
> 1,
regulations increase the relative cost of supplying commercial floorspace. Thus, the relationship
between residential and commercial floorspace prices is
8
q
W i
= ξ
i
q
Ri
. (9)
The demand for commercial floorspace (H
C
W j
) and home office floorspace (H
T
ij
) arises from
6
Note that the wage of a commuter does not depend on her location of residence i. However, the wage of a
telecommuter depends on his location of residence i because production uses home-office floorspace.
7
This function was also used in Favilukis, Mabille, and Van Nieuwerburgh (2019) to model density limits.
8
This equality does not need to hold if the supply of commercial or residential floorspace in a given tract is
zero. In our quantitative model, however, these corner cases do not occur.
7
profit-maximizing choices of firms. The demand for residential floorspace (H
Ri
) comes from utility-
maximizing choices of residents. Equilibrium selling price ¯q
i
equalizes the demand and the supply
of floorspace:
H
C
W j
+
X
i∈I
H
T
ij
+ H
Rj
= H
i
. (10)
Appendix B provides more details.
2.4 Externalities
Local total factor productivity and residential amenities depend on density. In particular,
the productivity in location j is determined by an exogenous component, a
j
, and an endogenous
component that is increasing in the density of on-site labor in this location, as well as every other
locations s, weighted inversely by the travel time from j to s:
A
j
= a
j
"
I
X
s=1
e
−δt
js
N
C
W s
Λ
s
#
λ
. (11)
Parameter λ > 0 measures the elasticity of productivity with respect to the density of workers,
while parameter δ accounts for the decay of spillovers from other locations. Productive external-
ities may include learning, knowledge spillovers, and networking that occur as a result of face-
to-face interactions between workers. Hence, we assume that only commuters and telecommuters
who are on-site on a given day contribute to these externalities.
Similarly, the residential amenity in location i is determined by an exogenous component, x
j
,
and an endogenous component that depends on the density of residence in every other location,
weighted inversely by the travel time to that location from i:
X
i
= x
i
"
I
X
s=1
e
−ρt
is
N
Rs
Λ
s
#
χ
. (12)
Parameter χ > 0 measures the elasticity of amenities with respect to the density of residents,
and ρ is the decay of amenity spillovers. The positive relationship between density and amenities
represents, in reduced form, the greater propensity for both public amenities, such as parks and
schools, and private amenities, such as retail shopping, to locate in proximity to greater concen-
trations of potential users and customers. All types of workers, commuters and telecommuters,
contribute equally to amenity externalities at their location of residence.
9
9
It would also be possible that telecommuters, by spending more time in the area of their residence, contribute
more to local amenities than commuters.
8
2.5 Equilibrium
An equilibrium consists of residential and workplace employment of commuters and telecom-
muters, N
Ri
(θ) and N
W j
(θ); wages of commuters and telecommuters, w
C
j
and w
T
ij
; residential and
commercial floorspace prices, q
Ri
and q
W j
; local productivities, A
j
; and local amenities, X
j
; such
that equations (4), (7), (9), (10), (11), and (12) are satisfied.
3 Data and Calibration
The Los Angeles-Long Beach Combined Statistical Area had a total population of 18.7 million
in 2018, distributed across a total land area of 88,000 square kilometers (U.S. Census Bureau,
2020). It comprises five counties (Los Angeles, Orange, Riverside, San Bernardino, and Ventura)
and 3,917 census tracts. To exclude nearly empty desert and mountain tracts with large land
areas, we exclude any tracts that are below the 2.5th percentile of both residential and employment
density. This excludes less than 1% of workers and leaves us with 3,846 tracts. We focus on the
five years between 2012 and 2016. To construct tract-level data on the residential and workplace
employment, we use the LEHD Origin-Destination Employment Statistics (LODES) data for years
2012 to 2016. Tract-level wages are constructed using the data from the American Community
Survey (ACS) and the Census Transportation Planning Products (CTPP). We also use CTPP to
estimate bilateral commuting times. Finally, prices of residential and commercial floorspace come
from the universe of transactions provided by DataQuick. Please refer to Appendix A for more
details on the data.
The baseline probability of telecommuting, ψ, is set to 0.0374. This number corresponds to the
fraction of workers who report that they primarily work from home in the 2012–2016 individual-
level data from the American Community Survey for the Los Angeles-Long Beach CSA. The
fraction of time that telecommuters spend at an on-site workplace, θ, is set to 0.114, based on a
survey questionnaire of Global Work-from-Home Experience Survey (Global Workplace Analytics,
2020).
10
The elasticity of commuting time with respect to traffic volume, ε
V
, is set to 0.2, following
Small and Verhoef (2007). In Appendix section E we discuss robustness of results to different
values of ε
V
.
We calibrate the relative TFP of telecommuters, ν, so that the wages of commuters and
telecommuters are identical in the benchmark economy.
11
The floorspace share of telecommuters,
10
The survey asked the number of days an employee worked from home per week. We classify workers as
telecommuters if they work from home three or more days per week. According to the survey, 9% of workers
work from home five days per week, 2% do this four days a week, and 3% work from home three days per week.
Based on these numbers, we calculate the fraction of time spent on-site as 1 − [0.09 × (5/5) + 0.02 × (4/5) + 0.03 ×
(3/5)]/[0.09 + 0.02 + 0.03] = 0.114.
11
Our empirical analysis finds that wages of telecommuters are higher than those of commuters, however, the
wage premium disappears once we control for age, education, industry, and occupation. It is also unclear how the
wage gap between the two types will change if many more workers start working remotely.
9
α
T
, is calibrated so that, on average, the home office of a telecommuter constitutes 20% of her
house.
12
The calibrated values of ν and α
T
are equal to 0.71 and 0.934, respectively.
We borrow values for the remaining city-wide parameters from previous studies. The share
of housing in expenditures, γ, is equal to 0.25, following Davis and Ortalo-Magn´e (2011). The
labor share in production, α, is 0.8 (Valentinyi and Herrendorf, 2008), and the land share in
construction, η, is 0.25 (Combes, Duranton, and Gobillon, 2018). Parameters that determine
the strength of agglomeration forces and decay speed for productivity and residential amenities
are borrowed from Ahlfeldt, Redding, Sturm, and Wolf (2015). In particular, we set λ = 0.071,
δ = 0.3617, χ = 1.0326, and ρ = 0.7595.
13
We also take the variance of the Fr`echet shocks and the
elasticity of utility with respect to commuting from Ahlfeldt, Redding, Sturm, and Wolf (2015),
and set = 6.6491 and κ = 0.0105.
14
Besides city-wide parameters, in order to solve the model, we also need to know vectors of
structural residuals: E, x, a, ξ, and
¯
H. The model provides equilibrium relationships that allow
us to identify these residuals from observed prices and quantities uniquely. Appendix C provides
more details on how we back out these residuals using the data.
4 Counterfactuals
The COVID-19 outbreak in early 2020 has forced many workers to work from home. While
before the epidemic around 4% of workers in Los Angeles metropolitan area worked from home,
Dingel and Neiman (2020) estimate that as many as 33% of workers in Los Angeles have jobs that
can be done remotely.
In this section, we study how a permanent reallocation from working on site to working at
home would affect the urban economy of Los Angeles. We simulate this increase by permanently
raising the probability of telecommuting, ψ. The maximum permanent increase we consider is all
the way to 0.33. We also calculate results for a range of intermediate values.
As the number of teleworkers increases, both firms and workers change their locations within
the urban area. In response, the city also experiences endogenous adjustments in the supply
of commercial and residential floorspace, as well as commuting speeds. In what follows, we
12
The average house size was 2,430 square feet in 2010, according to Muresan (2016). Home-based teleworkers
have, on average, 500 square feet larger homes than other workers (Nilles, 2000). Hence, telecommuters’ houses are
about 20% larger. This gap may reflect differences in income, location within a city, and the need for designated
workspace within a house. All of these factors are also present in our model.
13
Note that the parameter χ in our model corresponds to the product of the variance of the Fr`echet shocks and
the elasticity of residential amenities with respect to density in Ahlfeldt, Redding, Sturm, and Wolf (2015).
14
These two parameters, as well as the four parameters that determine amenity and productivity spillovers,
were estimated for the city of Berlin, and we leave the estimation for Los Angeles for future work. Nonetheless,
similar structural models with parameters estimated for other cities are characterized by similar magnitudes of
productivity agglomeration effects and spatial spillovers. At the same time, the estimates of amenity agglomeration
effects and spatial spillovers differ substantially across studies. See Berkes and Gaetani (2019),Tsivanidis (2019),
and Heblich, Redding, and Sturm (2020), among others.
10
describe the effects on the spatial allocation of workers and firms, floorspace prices, commuting
patterns, wages and land prices. Then we discuss the drivers of counterfactual changes, the role
of endogenous productivity and amenities, and welfare effects.
4.1 Spatial Reallocation
When workers are freed from the need to commute to their workplace, they tend to choose
residences farther from the urban core in locations with more affordable housing. As the share of
telecommuters rises, this drives a reallocation of residents from the core of the urban area towards
the periphery.
15
The top panel of Figure 1 maps the predicted reallocation of residents when the
fraction of telecommuters rises to 33%.
As residents decentralize, employment centralizes. There are three main factors driving this
reallocation. First, the flipside of a telecommuter being able to access jobs even if they live
far away, is that employers can access the labor of telecommuters even if they are located far
from where they live. Therefore, employment shifts from locations which are less productive
but closer to workers’ residences, toward locations closer to the core which have higher exogenous
productivity and benefit from greater productivity spillovers. Second, the reallocation of residents
increases demand for floorspace in peripheral locations and reduces it in the core, creating a
cost incentive for jobs to move in the opposite direction. Third, the fact that telecommuters
require less on-site office space further increases the cost-efficiency of firms in core locations with
high productivity but high real estate prices. The middle panel of Figure 1 maps the predicted
reallocation of jobs.
16
The net effect of these reallocations is to reduce the price of floorspace in core locations and
increase it in the periphery. The bottom panel of Figure 1 maps predicted changes in real estate
prices when the fraction of telecommuters rises to 33%.
15
Althoff, Eckert, Ganapati, and Walsh (2020) find that months after the COVID-19 pandemic saw a reallocation
of residents from the densest locations to the least dense locations in the U.S.
16
For a breakdown of residence and job changes by worker type, see Appendix D.
11
Figure 1: Changes in residence, jobs, and real estate prices
Note: Absolute change in residential density (top), job density (middle) and % change in floorspace prices (bot-
tom).
12
4.2 Commuting
A shift to telecommuting brings large benefits to those workers who do not have to come to
the office every day anymore and therefore suffer less disutility from commuting. However, those
who still have to commute benefit too, as traffic congestion drops and commuting speeds increase.
As the upper left panel of Figure 2 shows, with lighter traffic and faster speeds, the average
commuting time for those who still commute falls from 31 to 30 minutes. At the same time, the
average commute distance for commuters increases by nearly 1 km, as they relocate farther away.
This can be seen in the upper right panel. However, the total amount of kilometers traveled falls
by 29%, which suggests possible environmental benefits of the increase in telecommuting. The
magnitudes of these effects depend importantly on the elasticity of speed with respect to traffic
volume, ε
V
. Simulations for alternative values of ε
V
can be found in Appendix E.
Figure 2: Commuting, wages, and prices
Commute times Commute distances
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
22
24
26
28
30
Commuting time, minutes
All workers
Commuters
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
22
24
26
28
30
Commuting distance, km
All workers
Commuters
Wages and land prices Floorspace prices
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
-8
-6
-4
-2
0
Change, %
Wages
Land prices
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
-6
-5
-4
-3
-2
-1
0
Change, %
Residential floorspace price
Commercial floorspace price
Note: Upper left: average commuting time for all workers and commuters. Upper right: average commuting
distance. Lower left: % change in average wages and land prices. Lower right: % change in floorspace prices. All
variables plotted as a function of the share of teleworkers.
13
4.3 Wages and Floorspace Prices
When the share of telecommuters increases, two opposite forces influence average wages. On
the one hand, jobs are being reallocated to more productive locations that also benefit from
agglomeration externalities. On the other hand, a larger fraction of the workforce does not
contribute to these externalities. In our calibration, these two forces almost perfectly balance
each other. As can be seen in the lower left panel of Figure 2, a full increase in the fraction of
telecommuters to 33% leads to a 0.3% increase in average wages.
As can also be seen in the lower left panel of Figure 2, an increasing share of telecommuters is
decisive for the average price of land. Residents reallocate themselves to less expensive locations,
and firms with more telecommuters need less office space. If the fraction of telecommuters rose
to 33%, the income of landowners would fall by 8%.
The lower right panel of Figure 2 shows that the value of both types of real estate falls by
about 6%.
17
The relative decreases in residential and commercial prices depend on the fraction
of telecommuters. When the change in the amount of telecommuting is relatively small, the
decrease in residential prices is somewhat larger. After the fraction of telecommuters passes 28%,
commercial prices are hit harder.
4.4 Accounting for Counterfactual Changes
What are the main factors which drive these results? We find that a substantial part of the
variation in predicted changes is accounted for by simple measures of centrality such as distance
to cental business district. In this way, there is substantial overlap between our predictions and
the predictions that could be obtained from a uni-dimensional “monocentric city” model. We also
find that there is significant heterogeneity in predicted outcomes between tracts that are roughly
the same distance from downtown L.A. This additional heterogeneity reflects the differences in
exogenous local characteristics and transport network connections which our quantitative model
allows us to account for.
This heterogeneity is highlighted in the three panels of Figure 3. Each panel plots predicted
changes on the y-axis against the land-area weighted centrality rank of a tract on the x-axis–a
centrality rank of 0 represents the most distant tract, a centrality rank of 1 represents the tract
closest to the center of the metropolitan area.
18
In the middle panel, we see that while there is
an unambiguous prediction of job losses in the periphery, roughly equal numbers of tracts gain
and lose jobs from the 60
th
percentile and higher of centrality. In the left panel, we see that
while peripheral tracts are projected to gain residents, predictions are much more ambiguous
17
In this model, residential and commercial prices in a given location move one to one; see equation (9). However,
changes in average prices of each type may differ due to changes in the supply of each type of real estate.
18
We calculate an eigenvector centrality from the I × I matrix of inverse commuting disutilities. This measure
is highly correlated with both straight-line distance from downtown Los Angeles, and travel time from downtown
Los Angeles–the correlation is higher than 0.97 in both cases. More details can be found in Appendix F.
14
Figure 3: Quantiles of Centrality and Counterfactual Reallocations
Note: The x-axis is scaled to quantiles of the centrality measure, weighted by land area. The size of each circle is
proportional the the land area of the tract.
once the centrality is higher than the 60
th
percentile. In the right panel, we see that real estate
price changes fall systematically as we move towards the center of the city. All three panels show
substantial disparities in predicted outcomes between tracts of very similar centralities.
19
What
can account for this variation?
To help answer this question, we perform a Shapley-style decomposition of the variation in
predicted outcomes between centrality, exogenous local productivity and exogenous local employ-
ment and residential amenities.
20
We find that distance from the center can account for at most
60% of the variation in changes in floorspace prices, around 40% of the variation in changes in
employment, and 50% of the variation in changes in residence across space. Two of the key
takeaways from this exercise are that (1) locations with higher exogenous residential amenities
have bigger resident gains and smaller resident losses, all else equal; and (2) locations with higher
exogenous productivity have bigger job gains and smaller job losses, all else equal.
21
4.5 Role of Endogenous Productivities, Amenities, and Congestion
In the baseline counterfactual we assume that local productivities and amenities are endoge-
nous. We also assume that commuting speeds fall as total vehicle miles traveled goes down, and
that these increased speeds also mean that spillovers have a broader reach.
What is the role of these specification choices in driving our results? We turn each of them off
19
We can also see this variation between equidistant tracts if we return to look at Figure 1. It is perhaps most
striking in the middle panel of Figure 1, where we can see that one set of tracts which are close to downtown
experience strong gains in employment, while other tracts, equally close or even closer to downtown, lose jobs. The
bottom panel of Figure 1 shows large differences in the size of real estate price reductions between different tracts
close to downtown.
20
Full details of this decomposition are provided in Appendix F.
21
Maps of structural residuals are shown in Figure 4 in Appendix C.
15
and on in turn, and show the results in Table 1. Column (6), when all margins are turned “on,”
corresponds to the benchmark scenario. It turns out that in none of these permutations are our
main results significantly altered. Commuting times go down, floorspace prices fall, and overall
welfare goes up, all in roughly the same proportions, no matter which set of assumptions is turned
on. There are, however, some variations which illustrate the role of different model mechanisms
in shaping the results.
Table 1: Breakdown of results
Engogenous productivities: no no yes yes yes yes
Endogenous amenities: no yes no yes yes yes
Endogenous congestion: no no no no yes yes
Spillovers affected by congestion: n/a n/a n/a n/a no yes
(1) (2) (3) (4) (5) (6)
Wages of all workers, % chg 1.77 1.79 -0.39 -0.41 -0.37 0.31
Wages of commuters, % chg 2.66 2.80 0.34 0.45 0.51 1.21
Wages of telecommuters, % chg -0.13 -0.38 -1.98 -2.26 -2.26 -1.61
Residential floorspace prices, % chg -4.37 -5.03 -5.75 -6.16 -6.23 -5.63
Commercial floorspace prices, % chg -6.43 -7.39 -6.14 -6.86 -6.97 -6.41
Time spent commuting, all workers, % chg -31.42 -30.69 -31.46 -30.80 -32.23 -32.13
Time spent commuting, commuters, % chg -1.46 -0.43 -1.52 -0.57 -2.63 -2.49
Distance traveled, all workers, % chg -31.91 -30.69 -31.96 -30.85 -28.96 -28.82
Distance traveled, commuters, % chg -2.18 -0.42 -2.24 -0.65 2.06 2.27
Welfare by source, % chg
consumption 2.30 2.52 0.45 0.55 0.64 1.17
goods only 0.84 0.85 -1.26 -1.29 -1.24 -0.57
housing only 4.75 5.70 3.95 4.58 4.73 4.79
+ commuting 11.71 11.54 9.74 9.46 10.02 10.56
+ amenities 11.90 14.38 10.16 12.23 12.80 14.15
+ Fr`echet shocks 17.97 19.67 15.28 16.79 17.14 18.91
Welfare by commuter type, % chg
commuter 1.94 2.06 -0.48 -0.41 0.82 2.24
telecommuter -3.33 -1.69 -5.52 -4.06 -3.94 -2.47
Note: Columns (1)–(6) present results from specifications with different combinations of engogenous productivities,
amenities and congestion, and whether spillovers increases when traffic congestion goes down. Each column reports
results of a counterfactual experiment with an increase of the fraction of telecommuters to 0.33.
First, let us compare columns (1) and (2) of Table 1 with columns (3) and (4). If local pro-
ductivities do not adjust endogenously, wages increase. This is primarily because telecommuters
do not contribute to productivity spillovers. If these adjust, the locations which lose in-person
workers–nearly every location–see a fall in productivity. As a result, average wages fall.
Second, let us compare columns (1) and (3) with columns (2) and (4). If residential amenities
do not adjust, there is a bigger reduction in travel times and distances. This is because allowing
amenities to follow telecommuters out to the periphery increases the attractiveness of peripheral
16
locations for regular commuters, making them willing to put up with longer commutes.
Finally, let us compare columns (5) and (6) with column (4). We see that endogenous conges-
tion leads to larger reductions in time spent commuting. It also flips small reductions in distance
traveled into small increases–increased speeds allow workers to travel further while spending less
time on the road. Comparing columns (5) and (6), we can see that allowing the reach of spillovers
to increase when travel speeds go down gives a small but significant boost to wages.
4.6 Welfare
The lower half of Table 1 shows that the increase in telecommuting to 33% of workforce results
in significant welfare gains, which we measure as consumption-equivalent changes in expected
utility (see Appendix section B.3 for details). We find that reduced commuting is the single biggest
driver of welfare improvements, even when traffic congestion remains fixed at the benchmark
level. Focusing on column (6): when commuting is accounted for in addition to the 1.2% gain
from consumption, welfare gains rise by over 9 percentage points. After this, improved access
to amenities adds another 3.6 percentage points, while workers’ improved ability to fulfill their
idiosyncratic preferences contributes less than 5 percentage points.
One important driver of welfare gains for commuters is access to jobs. In large, sprawled and
congested cities, such as Los Angeles, good jobs are often inaccessible for households who live on
the periphery. To study how a shift to telecommuting impacts job access, we calculate commuter
market access for each tract as CMA
i
=
P
j
(w
j
e
−κt
ij
)
. We find that an increase in the fraction of
telecommuters improves average job access for those who keep commuting by 16%, largely thanks
to lower traffic congestion. We also find that the elasticity of floorspace prices with respect to
market access at the tract level falls, meaning that places with better access to jobs command a
lower price premium.
22
The utility of the average telecommuter is significantly higher than that of the average com-
muter, due to reduced disutility from commuting, access to lower-cost housing, and access to
better-paying jobs and amenities. As a result, the shift of workers from commuting to telecom-
muting is an important source of the welfare increases. Workers who remain commuters or telecom-
muters, see their welfare change only marginally. Commuters who continue to commute benefit
from reduced time commuting, access to lower-cost housing, and access to better-paying jobs and
amenities, and see their welfare rise by more than 2%. At the same time, telecommuters who were
already telecommuting do not benefit from the increase in their mode of work. On the contrary,
they need to compete with an increasing fraction of the workforce for residence and job sites that
were previously accessible only to them. Their welfare falls by about 2.5%.
22
Further details of these calculations as well as other results can be found in Appendix D.
17
5 Conclusion
In this paper we used a detailed quantitative model of internal city structure to study what
would happen in Los Angeles if telecommuting becomes popular over the long run. We find
substantial changes to the city structure, wages and real estate prices, and commuting patterns.
We also find that more widespread telecommuting could bring significant welfare benefits.
Our analysis necessarily omits several important channels which could dampen or amplify our
findings. First, in our model all workers are ex-ante identical and have the same chances of being
able to telecommute. In reality, the ability to telecommute is correlated with occupation, industry
and income. Accounting for this would likely have two effects. First, it would center the large
shifts in jobs and residence even more on the high-density center-city locations where the share of
skilled, telecommute-ready workers is likely to be highest. Second, there would be more downward
pressure on the average wage. This is because these center-city locations have higher than average
local productivity. If these locations lose proportionally more in-person workers, their reduction
in productivity from spillovers will be greater, as will the impact on aggregate average wages.
It is also likely that different skill levels of workers differ in their contribution to productivity
externalities.
23
If higher skilled workers are also more likely to telecommute, the effect of this
detail would be similar to the previous one: additional downward pressure on wages.
Second, we calibrated the productivity gap between commuters and telecommuters to ensure
that their average wages are the same in the benchmark economy, and assume that this parameter
remains constant in the counterfactual. We also assume that telecommuters do not contribute at
all to productivity spillovers. However, as telecommuting becomes more widespread, technological
changes might increase the relative productivity of telecommuters and allow them to contribute
more to productivity spillovers even without literal face-to-face interaction. This would put up-
ward pressure on wages, as we find in a related paper of ours, Delventhal and Parkhomenko
(2020).
Third, we do not take account of non-commuting travel.
24
If we did, we would probably
find an increase in local traffic congestion in the peripheral areas that telecommuters relocate to,
alongside the reduction in congestion along the main commuting arteries. This would mitigate
gains from moving to the periphery and lead to less decentralization overall. We also do not
distinguish between transportation modes in the model. The reduction in congestion brought by
more telecommuting could be offset if some transit users start commuting by car.
Finally, we do not allow migration in and out of the city. In practice, as some workers gain
the ability to work remotely, they may choose to leave Los Angeles and move to a different city,
or even a different country. On the other hand, telecommuters from elsewhere may move into Los
23
This is a finding of, e.g., Rossi-Hansberg, Sarte, and Schwartzman (2019).
24
Couture, Duranton, and Turner (2018) estimate that work-related trips account for only 40% of vehicle miles
traveled in the U.S.
18
Angeles to enjoy local amenities. Indeed, this is what we find in Delventhal and Parkhomenko
(2020), which expands the scope of analysis to include the entire U.S.
One more caveat is recommendable in interpreting our predictions for welfare. We model
telecommuting as a fact imposed exogenously on workers. They love it because they commute
less. Most welfare gains come through this channel. In reality, some workers may dislike remote
work. If telecommuting were a choice in which workers balance the benefits against their individual
dislike, welfare gains would almost surely be smaller than what we report.
References
Ahlfeldt, G. M., S. J. Redding, D. M. Sturm, and N. Wolf (2015): “The Economics of
Density: Evidence From the Berlin Wall,” Econometrica, 83(6), 2127–2189.
Akbar, P. A., V. Couture, G. Duranton, and A. Storeygard (2020): “Mobility and
congestion in urban India,” Working paper.
Althoff, L., F. Eckert, S. Ganapati, and C. Walsh (2020): “The City Paradox: Skilled
Services and Remote Work,” Working paper.
Anas, A. (2020): “The cost of congestion and the benefits of congestion pricing: A general
equilibrium analysis,” Transportation Research Part B, 136, 110–137.
Barrero, J. M., N. Bloom, and S. J. Davis (2020): “COVID-19 Is Also a Reallocation
Shock,” Working Paper.
Bartik, A. W., Z. B. Cullen, E. L. Glaeser, M. Luca, and C. T. Stanton (2020):
“What Jobs are Being Done at Home During the Covid-19 Crisis? Evidence from Firm-Level
Surveys,” NBER Working Paper 27422.
Bento, A., J. D. Hall, and K. Heilmann (2020): “Estimating Congestion Externalities
using Big Data,” Working Paper.
Berkes, E., and R. Gaetani (2019): “Income segregation and rise of the knowledge economy,”
Available at SSRN 3423136.
Bloom, N., J. Liang, J. Roberts, and Z. J. Ying (2015): “Does working from home work?
Evidence from a Chinese experiment,” The Quarterly Journal of Economics, 130(1), 165–218.
Combes, P.-P., G. Duranton, and L. Gobillon (2018): “The Costs of Agglomeration:
House and Land Prices in French Cities,” The Review of Economic Studies, 86(4), 1556–1589.
19
Couture, V., G. Duranton, and M. A. Turner (2018): “Speed,” The Review of Economics
and Statistics, 100(4), 725–739.
Davis, M. A., and F. Ortalo-Magn
´
e (2011): “Household expenditures, wages, rents,” Review
of Economic Dynamics, 14(2), 248 – 261.
Delventhal, M. J., and A. Parkhomenko (2020): “Spatial Implications of Telecommuting,”
Working paper.
Dingel, J. I., and B. Neiman (2020): “How Many Jobs Can be Done at Home?,” Working
Paper 26948, National Bureau of Economic Research.
Favilukis, J., P. Mabille, and S. Van Nieuwerburgh (2019): “Affordable Housing and
City Welfare,” Working Paper 25906, National Bureau of Economic Research.
Global Workplace Analytics (2020): Work From Home Experience Survey, 2020. Global
Workplace Analytics.
Heblich, S., S. J. Redding, and D. M. Sturm (2020): “The Making of the Modern Metropo-
lis: Evidence from London,” Quarterly Journal of Economics, 135(4), 2059–2133.
Larson, W., and W. Zhao (2017): “Telework: Urban form, energy consumption, and green-
house gas implications,” Economic Inquiry, 55(2), 714–735.
Lennox, J. (2020): “More working from home will change the shape and size of cities,” Centre
of Policy Studies/IMPACT Centre Working Papers g-306, Victoria University, Centre of Policy
Studies/IMPACT Centre.
Mas, A., and A. Pallais (2020): “Alternative work arrangements,” Discussion paper, National
Bureau of Economic Research.
Muresan, A. (2016): “Who Lives Largest? The Growth of Urban American Homes in the Last
100 Years,” Press Release.
Nilles, J. (2000): “Telework in the US: Telework America survey 2000,” Washington, DC:
International Telework Association and Council.
Owens, R., E. Rossi-Hansberg, and P.-D. Sarte (2020): “Rethinking Detroit,” American
Economic Journal: Economic Policy, 12(2), 258–305.
Ozimek, A. (2020): “Remote Workers on the Move,” Upwork Economist Report.
Rhee, H.-J. (2008): “Home-based telecommuting and commuting behavior,” Journal of Urban
Economics, 63(1), 198–216.
20
Rossi-Hansberg, E., P.-D. Sarte, and F. Schwartzman (2019): “Cognitive Hubs and
Spatial Redistribution,” Working Paper 26267, National Bureau of Economic Research.
Safirova, E. (2003): “Telecommuting, traffic congestion, and agglomeration: a general equilib-
rium model,” Journal of Urban Economics, 52(1), 26–52.
Severen, C. (2019): “Commuting, Labor, and Housing Market Effects of Mass Transportation:
Welfare and Identification,” Working Paper.
Shorrocks, A. (2013): “Decomposition procedures for distributional analysis: a unified frame-
work based on the Shapley value,” The Journal of Economic Inequality, 11(1), 99–126.
Small, K. A., and E. T. Verhoef (2007): The economics of urban transportation. Routledge.
Spear, B. D. (2011): NCHRP 08-36, Task 098 Improving Employment Data for Transportation
Planning. American Association of State Highway and Transportation Officials.
Tsivanidis, N. (2019): “The Aggregate and Distributional Effects of Urban Transit Infrastruc-
ture: Evidence from Bogot´a’s TransMilenio,” Working Paper.
U.S. Census Bureau (2020): American Community Survey, 2012-2016. Washington, DC: U.S.
Census Bureau.
Valentinyi, A., and B. Herrendorf (2008): “Measuring factor income shares at the sectoral
level,” Review of Economic Dynamics, 11(4), 820–835.
21
A Data Appendix
A.1 Property Price Data
Our commercial and residential property price data comes from DataQuick, which provides
the universe of property transactions and the characteristics of individual properties. The dataset
covers 2,354,535 properties over 2007–2016 in the Los Angeles-Long Beach combined statistical
area. The data provides information such as sales price, geographical coordinates, transaction
date, property use, transaction type, number of rooms, number of baths, square-footage, lot size,
year built, etc.
We categorize properties as commercial or residential based on their reported use. Examples
of residential use include “condominium”, “single family residence”, and “duplex”. Examples
of commercial use include “hotel/motel”, “restaurant”, and “office building”. Table 2 provides
descriptive statistics. Table 3 reports the number of observations in each county over the pe-
riod of 2007–2016. Note that the commercial transactions are far less frequent than residential
transactions.
We then use the transactions data to estimate hedonic tract-level residential and commercial
property indices. For a residential transaction of a property p, in tract j in year-month t, we
estimate
ln(P
pjt
) = α + βX
p
+ τ
t
+ ζ
res
j
+
pjt
, (13)
where P
pjt
is the price per square foot; X
p
contains property characteristics including property
use, transaction type, number of rooms, number of baths, lot size, and year built; and τ
t
is the
year-month fixed effect. Then the residential price index in tract j corresponds to ζ
res
g
, the tract
fixed effect.
Because commercial transactions are less numerous and more spatially concentrated, for many
Census tracts we only observe very few or no transactions in the period of interest. To overcome
this issue, we calculate commercial property indices at the Public Use Microdata Area (PUMA)-
level.
25
For a transaction of a commercial property p, in tract j of PUMA g in year-month t, we
estimate
ln(P
pgjt
) = α + βX
p
+ τ
t
+ ζ
com
g
+ υ
pgjt
, (14)
where P
pgjt
is the price per square foot; X
p
is property characteristics including property use; and
τ
t
is the year-month fixed effect. The commercial price index in PUMA g corresponds to ζ
com
g
,
which is the PUMA fixed effect. Then, to obtain tract-level commercial price indices ζ
com
j
, we
simply assign the same value of ζ
com
g
to all tracts j that belong to PUMA g.
25
PUMA is a geographic unit used by the US Census for providing statistical and demographic information.
Each PUMA contains between 100,000 and 200,000 inhabitants. There are 123 PUMAs in the Los Angeles-Long
Beach combined statistical area.
22
Table 2: Descriptive Statistics
Panel A. Residential Properties
County sqft (mean) sqft (median) sales price, (mean) sales price, (median)
Los Angeles 1752.25 1499 774734.19 389000
Orange 1969.92 1578 714043.38 495000
Riverside 2046.06 1855 489885.35 246649
San Bernardino 1759.41 1584 345662.41 200000
Ventura 1860.88 1626 569042.40 410000
Panel B. Commercial Properties
County sqft (mean) sqft (median) sales price, (mean) sales price, (median)
Los Angeles 20687.28 5203 5661399.99 1300000
Orange 16447.48 5329 3879699.73 1260000
Riverside 1329.38 1201 1813988.76 590000
San Bernardino 19486.08 3541 2472923.09 522000
Ventura 12087.09 4565 3513023.97 982500
Table 3: Number of Transactions by County and Property Type
Los Angeles Orange Riverside San Bernardino Ventura Total
Residential 909,954 330,689 557,204 363,173 105,518 2,266,538
Commercial 47,408 12,084 14,045 11,099 3,361 87,997
A.2 Wage Data
Our sources of wage data are the Census Transportation Planning Products (CTPP) and
the American Community Survey (ACS). CTPP data sets produce tabulations of the ACS data,
aggregated at the Census tract level. We use the data reported for years 2012 to 2016. We use the
variable “earnings in the past 12 months (2016 ), for the workers 16-year-old and over,” which is
based on the respondents’ workplace locations. The variable provides the estimates of the number
of people in several earning bins in each workplace tract. Table 4 provides an overview of the
number of observations in each bin for the five counties included in our study.
We calculate mean tract-level labor earnings as
w
j
=
Σ
b
nworkers
b,j
× meanw
b
Σ
b
nworkers
b,j
, (15)
where nworkers
b,j
is the number of workers in bin b in tract j, and meanw
b
is mean earnings in
bin b for the entire Los Angeles-Long Beach combined statistical area, calculated from the ACS
microdata.
Next, to control for possible effects of workers’ heterogeneity on tract-level averages, we run
the following Mincer regression,
w
j
= α + β
1
age
j
+ β
2
sexratio
j
+ Σ
r
β
2,r
race
r,j
+ Σ
i
β
3,i
ind
i,j
+ Σ
o
β
4,o
occ
o,j
+
j
, (16)
23
Table 4: Number of observations in each earnings bin, by county
Income Bin Los Angeles Orange Riverside San Bernardino Ventura
1 to 9,999 or loss 416,469 147,484 86,219 85,854 34,973
10,000 to 14,999 279,132 90,871 51,959 52,605 21,143
15,000 to 24,999 541,649 168,284 97,184 97,059 40,458
25,000 to 34,999 440,298 146,337 79,994 81,911 34,829
35,000 to 49,999 493,434 170,364 77,170 87,969 37,487
50,000 to 64,999 387,533 138,932 57,409 62,487 27,979
65,000 to 74,999 176,079 63,244 24,869 27,687 13,895
75,000 to 99,999 308,994 114,436 39,159 44,409 23,871
100,000 or more 486,179 189,108 44,925 43,158 36,346
No earnings 520 134 144 85 55
Earnings in the past 12 months (2016 ) (Workers 16 years and over), based on workplace location, Source: CTPP.
where age
j
is the average age of workers; sexratio
j
is the proportion of males to females in the
labor force; race
r,j
is the share of race r ∈ {Asian, Black, Hispanic, W hite}; ind
i,j
is the share
of jobs in industry i; occ
o,j
is share of jobs in occupation o in tract j.
26
Finally, the estimated
tract-level wage index corresponds to the sum of the estimated constant and the estimated tract
fixed effect, ˆα + ˆ
j
. Table 5 presents summary statistics for the estimated tract-level earnings.
Table 5: Descriptive statistics: the estimated tract-level earnings, by county
Obs Mean Std. Dev. Min Max
Los Angeles 2,339 61,203.81 13,589.54 21,376.82 170,987.1
Orange 582 63,455.76 11,197.14 24,120.39 113,428.8
Riverside 452 61,477.51 13,606.08 17,286.49 138,802.9
San Bernardino 369 59,823.33 12,741.2 21,101.49 132,544.9
Ventura 172 61,034.83 10,709.51 29,174.4 89,796.23
Earnings in U.S. dollars in the past 12 months (2016 ) (Workers 16 years and over), based on workplace location,
Source: CTPP.
26
We use the following industry categories: Agricultural; Armed force; Art, entertainment, recreation, ac-
commodation; Construction; Education, health, and social services; Finance, insurance, real estate; Information;
Manufacturing; Other services; Professional scientific management; Public administration, Retail. We use the fol-
lowing occupation categories: Architecture and engineering; Armed Forces; Arts, design, entertainment, sports, and
media; Building and grounds cleaning and maintenance; Business and financial operations specialists; Community
and social service; Computer and mathematical; Construction and extraction; Education, training, and library;
Farmers and farm managers; Farming, fishing, and forestry; Food preparation and serving related; Healthcare
practitioners and technicians; Healthcare support; Installation, maintenance, and repair; Legal; Life, physical, and
social science; Management; Office and administrative support; Personal care and service; Production;Protective
service; Sales and related.
24
A.3 Commuting Time Data
The CTPP database provides commuting time data for 270,436 origin-destination tract pairs
in the Los Angeles-Long Beach Combined Statistical Area for 2012-2016. There are 15,342,889
possible trajectories, and the LODES data for 2012-2016 reports positive commuting flows for
5,647,791 of them. We follow the practice recommended by Spear (2011) and use LODES data
as a measure of commuting flows and CTPP data to provide information on commute times.
Table 6: Commuting time coverage, by distance
N. of trajectories % covered by time data % w/ observed positive flows N. of commuters
< 1 km 10,105 60.8% 96.4% 239,188
< 2 km 36,205 40.5% 93.3% 410,571
< 5 km 188,047 24.4% 86.9% 1,088,797
< 10 km 649,005 15.0% 79.9% 2,248,646
< 20 km 2,099,417 8.2% 69.8% 3,995,134
< 40 km 5,549,775 4.3% 54.4% 5,508,736
< 80 km 10,752,785 2.5% 43.4% 6,515,595
All 15,342,889 1.8% 36.8% 6,935,765
Table 7: Commuting time coverage, by N. of commuters
N. of trajectories % covered by time data N. of commuters
> 100 commuters 1,778 94.4% 259,259
> 50 commuters 8,678 89.9% 723,849
> 25 commuters 27,833 82.2% 1,380,081
> 10 commuters 96,177 63.7% 2,417,561
> 5 commuters 220,555 46.5% 3,289,529
> 1 commuters 1,108,755 17.9% 5,247,370
All > 0 5,647,791 4.8% 6,935,760
Table 6 summarizes CTPP data coverage by trajectory distance. Table 7 summarizes CTPP
data coverage by trajectory and the number of commuters observed using that trajectory. These
tables show that CTPP has the greatest coverage of high-volume short-distance trajectories, just
as Spear (2011) observes and just as would be expected from a dataset based on a partial sample.
The CTPP data places commuting times into 10 bins: less than 5 minutes, 5 to 14 minutes,
15 to 19 minutes, 20 to 29 minutes, 30 to 44 minutes, 45 to 59 minutes, 60 to 74 minutes, 75 to
89 minutes, 90 or more minutes, and work from home. In order to get as accurate commute times
as possible for the set of primitive connections of the network, we drop all home-workers, who are
irrelevant for transit times. We drop workers in the top time bin, because this bin has no upper
bound and so the mean may vary substantially across trajectories. We assign mean commute
times to all the remaining bins as the mid-points between the bin bounds. We then drop all
observations which report an average commuting speed that is either less than 8 kilometers per
hour, a brisk walking pace, or more than 70 miles per hour (112.7 kilometers per hour), the
25
standard rural freeway speed limit in the United States. Finally, we calculate tractpair mean
commuting times as the average of the mean commuting times in each bin weighted by the share
of commuters on that tractpair reporting times in each bin. Table 8 provides a summary of the
overall share of commuters in each bin before and after the cleaning steps described above, and
the mean commute time assigned to each bin.
Table 8: Commuting time bins
share in raw data share in cleaned data bin mean time
< 5 min 1.6% 0.9% 5
5-14 min 19.4% 18.9% 10
15-19 min 14.0% 15.7% 17
20-29 min 19.1% 22.5% 25
30-44 min 20.5% 24.4% 37
45-59 min 8.0% 9.6% 52
60-74 min 6.1% 6.9% 67
75-89 min 0.9% 1.0% 82
> 90 min 2.8% 0 ??
work from home 7.6% 0 n/a
The previous cleaning steps eliminate observations for 36,279 trajectories, and we are left with
commuting time data for 234,157 origin-destination pairs. We then find that there are 211,521
paths for which a commuting time estimate exists for the outbound route but not the reverse. We
impute commute times for these missing return journeys, assuming that they can be completed
in the same time as the outbound trajectories. This set of connections is then almost enough
to connect all tracts–there are only a set of eight tracts that are still detached from the rest of
the network. In order to remedy this, we create a connection at the mean travel speed of 31.3
kilometers per hour between these left-out tracts and any tracts within a radius twice as large as
the hypothetical radius of tract if its land area formed a circle.
27
The final directed network contains 447,277 directed paths. We use the Dijkstra’s algorithm to
calculate the fastest path through this network for each origin-destination pair. We assume that
these calculated times represent the time require to travel from tract centroid to tract centroid.
We then add time to each trajectory to represent the time need to travel from place of residence
within tract to residence tract centroid, and from workplace tract centroid to workplace within the
tract. Naturally, these times are proportional to tract land area–larger tracts should on average
require more internal travel time. Specifically, we assume that the “internal” distance traveled
on each end of the trip is equal to the hypothetical average straight-line distance from any point
in the tract to the tract centroid, if the tract were a circle.
28
We then assume that each of these
distances is traveled at twice the overall average commuting speed in the cleaned data of 31.3
kilometers per hour. For the vast majority of tracts this adds a negligible amount to commuting
27
2 ×
p
landarea/π
28
2
3
p
landarea/π
26
time–two minutes or less. For a handful of very large tracts it adds considerable travel time–up to
half an hour. We think that this is reasonable given the time that is required to travel within these
much larger tracts. These origin-destination distribution effects are also applied to self-commute
times, so that a worker that lives and works in the same tract will still have to spend some time
travelling to their workplace–more time for larger tracts.
A.4 Summary
Table 9 gives summary statistics by tract for seven key variables: residential density; employ-
ment density; wage by workplace weighted by employees; average constant-quality price of one
square foot of residential floorspace; average constant-quality price of one square foot of com-
mercial floorspace; average commute time by residence tract; and average commute distance by
residence tract.
Table 9: Data overview
Mean Median St. Dev. Max. N. Obs.
Residents/km
2
1,621.4 1,380.3 1,376.6 15,929.3 3,846
Workers/km
2
1,285.8 578.7 3,961.0 157,995.7 3,846
Wages ( , weight by employees) 58,874 58,528 7,609 159,059 3,846
Res. price/sq ft ( , weight by residents) 369 331 347 9,349 3,846
Comm. price/sq ft ( , weight by employees) 644 581 324 6,709 3,846
Av. commute time (min, weight by residents) 28.3 26.3 6.8 96.1 3,846
Av. commute distance (km, weight by residents) 26.2 22.8 11.3 131.2 3,846
B Model Details
B.1 Floorspace Markets
B.1.1 Floorspace Supply
Land-market clearing and profit maximization imply that the equilibrium supply of floorspace
is
H
i
= φ
i
(H
i
) ((1 − η)¯q
i
)
1−η
η
L
i
. (17)
Solving this expression for H
i
and using the definition of construction efficiency φ
i
(H
i
), yields
H
i
=
((1 − η)¯q
i
)
1−η
η
L
i
1 + ((1 − η)¯q
i
)
1−η
η
L
i
/
¯
H
i
. (18)
27
B.1.2 Floorspace Demand
From equation (3), the probability that an individual who commutes a fraction θ of days works
in j, conditional on living in i, is given by
π
ij|i
(θ) =
E
j
w
ij
(θ)
d
ij
(θ)
−
I
P
s=1
E
s
w
rs
(θ)
d
is
(θ)
−
. (19)
Define ˜w
i
as the average wage earned by residents of location i. This is given by
˜w
i
≡
I
X
j=1
w
C
j
π
ij|i
(1)N
Ri
(1)
N
Ri
+ w
T
ij
π
ij|i
(θ
T
)N
Ri
(θ
T
)
N
Ri
. (20)
Therefore, the demand for residential and home-office floorspace is given by
H
Ri
=
γ(1 + τ ) ˜w
i
q
Ri
N
Ri
. (21)
The demand for home offices is
H
T i
=
1 − α
T
q
Ri
1
α
T
I
X
j=1
(νA
j
)
1
α
T
N
T
ij
. (22)
Finally, the demand for commercial floorspace is given by
H
W j
=
(1 − α)A
j
q
W j
1
α
N
C
W j
. (23)
B.2 Factor Incomes and Transfers
Since developers optimally use all land available for development, Λ
i
, equilibrium land prices
are given by
l
i
=
η
Λ
i
(q
Ri
(H
Ri
+ H
T i
) + q
W i
H
W i
) . (24)
The city-wide total land income is
I
X
i=1
l
i
Λ
i
. (25)
Income generated by land and the consumption good sold for the purposes of real estate
development is redistributed to all workers, proportionally to their incomes. The transfers increase
labor income by a fraction of τ which is equal to
τ =
P
i
(l
i
Λ
i
+ K
i
)
P
i
˜w
i
N
Ri
+
P
i
(l
i
Λ
i
+ K
i
)
. (26)
28
B.3 Welfare
The expected utility enjoyed by a resident of the city is given by
U ≡ Γ
− 1
"
I
X
r=1
I
X
s=1
X
r
E
s
(1 − ψ)
e
−κt
rs
(1 + τ)w
C
s
q
−γ
Rr
+ ψ
(1 − θ + θe
−κt
rs
)(1 + τ)w
T
rs
q
−γ
Rr
#
1
,
(27)
where Γ(·) is the gamma function. Note that the expected utility is defined before the telecom-
muting lottery and before the location preference shocks realize.
A consumption-equivalent measure of change in welfare is given by ∆. This quantity represents
the percentage amount by which the composite consumption of goods and housing, c
1−γ
h
γ
, must
change in order to make the expected utility in the benchmark economy equal to the expected
utility in the counterfactual economy. Note that in this model the composite consumption is
proportional to wages. Let “
˜
·” denote variables in the counterfactual economy. Then ∆ must
satisfy
"
I
X
r=1
I
X
s=1
X
r
E
s
(1 − ψ)
e
−κt
rs
(1 + τ)(1 + ∆)w
s
q
−γ
Rr
+ ψ
(1 − θ + θe
−κt
rs
)(1 + τ)(1 + ∆)w
T
rs
q
−γ
Rr
#
1
=
"
I
X
r=1
I
X
s=1
˜
X
r
E
s
h
(1 − ψ)
e
−κ
˜
t
rs
(1 + ˜τ ) ˜w
s
˜q
−γ
Rr
+ ψ
(1 − θ + θe
−κ
˜
t
rs
)(1 + ˜τ ) ˜w
T
rs
˜q
−γ
Rr
i
#
1
.
It follows that the change in welfare ∆ is a function of the ratio of expected utilities in the
counterfactual and the benchmark economies,
∆ =
˜
U
U
− 1. (28)
C Structural Residuals
The amounts of commuting workers and residents are related as
N
W j
(1) =
I
X
i=1
π
ij|i
(1)N
Ri
(1), (29)
Let
ˆ
E
j
≡ E
j
w
C
j
. From equations (19) and (29),
ˆ
E
j
can be defined implicitly as:
ˆ
E
j
= N
W j
(1)
I
X
i=1
e
−κt
ij
P
I
s=1
ˆ
E
s
e
−κt
is
N
Ri
(1)
!
−1
, (30)
29
where N
W j
and N
Ri
are observed tract-level employment and residential populations, and t
ij
are
observed average commuting times from tract i to tract j. Since we do not observe how many
workers telecommute in each tract and since the share of telecommuters in the data is small
(3.74% of workforce), we perform this and the following calculations assuming that all workers
commute to their jobs. A vector
ˆ
E is solved recursively using equation (30) and then the vector
of residuals E is recovered as E
j
=
ˆ
E
j
w
C
j
−
, using observed tract-level wages.
A similar procedure is applied to solve for vector X. First, let
ˆ
X
j
≡ X
j
q
−γ
Rj
.
ˆ
X
j
can be defined
implicitly as:
ˆ
X
i
= N
Ri
I
X
j=1
e
−κt
ij
P
I
r=1
ˆ
X
r
e
−κt
rj
N
W j
!
−1
. (31)
The vector
ˆ
X is solved recursively using equation (31) and then the vector of residuals X is
recovered as X
j
=
ˆ
X
j
q
γ
Rj
, using observed tract-level prices of residential floorspace. Then the
exogenous part of local amenities, x
j
, can be recovered using equation (12) and the data on local
residential population and land area.
The vector of local productivities A can be solved for using (7) and the data on wages and
commercial floorspace prices as follows:
A
j
=
w
C
j
α
!
α
q
W j
1 − α
1−α
. (32)
Then the exogenous part a
j
can be recovered using equation (11) and the data on local employment
and land area.
Since we observe commercial and residential floorspace prices for all Census tracts, we can
calculate the zoning parameter ξ
i
as
ξ
i
=
q
W i
q
Ri
. (33)
To calculate ξ
i
, we replace q
W i
and q
Ri
with tract-level quality adjusted indexes of commercial
and residential prices, ζ
com
j
and ζ
res
j
, respectively, as described in Appendix A.
Finally, in order to recover
¯
H
i
, we use market clearing conditions for land and floorspace
(L
i
= Λ
i
and equation 18). Combining them, we can recover
¯
H
i
from the following relationship:
¯
H
i
=
¯
φ ((1 − η)¯q
i
)
1−η
η
Λ
i
¯
φ ((1 − η)¯q
i
)
1−η
η
Λ
i
/H
i
− 1
, (34)
where Λ
i
is the observed land area and H
i
= H
Ri
+ H
W i
+ H
T i
is the total demand for floorspace
in tract i.
Figure 4 maps the recovered values for three key structural paramters: the exogenous com-
ponent of residential amenities, x
i
, the exogenous component of productivity, a
i
, and exogenous
employment amenities, E
i
.
30
Figure 4: Structural residuals
Note: Exogenous residential amenities (top figure), exogenous productivities (middle figure) and exogenous em-
ployment amenities (bottom figure).
31
D Additional Results of Counterfactual Experiments
D.1 Land Use
When the fraction of telecommuters rises, land use becomes more specialized. Figure 5 shows
that in the economy with more widespread telework, commercial development becomes relatively
more prevalent in core areas and less prevalent in the periphery. In addition, both types of
development become more concentrated in space. As a consequence, the numbers of primarily
residential and primarily commercial tracts increase, while the number of mixed tracts goes down
(right panel of Figure 6).
29
Figure 5: Land Use
Note: Benchmark (upper figure) and the ψ = 0.33 counterfactual (lower figure). Maps show the fraction of
commercial floorspace in each tract, varying from 0 (green) to 1 (brown). See main text for details.
29
We label a tract as commercial if the share of commercial floorspace in the tract is more than 3 times the
share of the average tract. Similarly, we label a tract as residential if the share of commercial floorspace in the
tract is less than 1/3 of the share of the average tract. All other tracts are labeled mixed.
32
Figure 6: Land Use Specialization
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
100
120
140
160
180
200
1280
1300
1320
1340
# commercial tracts (LHS)
# residential tracts (RHS)
Note: The figure shows the number of commercial and residential tracts, as a function of the share of teleworkers.
See main text for details.
D.2 Job access
In large, sprawled and congested cities, such as Los Angeles, good jobs are often inaccessible
for households who live on the periphery. To study how a shift to telecommuting impacts job
access, we calculate commuter market access for each tract as CMA
i
=
P
j
(w
j
e
−κt
ij
)
. We find
that, as the number of teleworkers grows, the average job access increases for those who keep
commuting (left panel of Figure 7). Moreover, in the counterfactual economy the elasticity of
housing prices with respect to the market access halves, meaning that places with better access
to jobs command a lower price premium (right panel of Figure 7).
Figure 7: Access to jobs
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
-5
0
5
10
15
Change, %
CMA
1 2 3 4 5 6 7
Log commuter market access
-3
-2
-1
0
1
2
Log housing prices
Benchmark, elasticity: 0.31
Counterfactual, elasticity: 0.14
Note: Left panel shows the weighted-average commuter market access for each level of ψ. Right panel shows the
relationship between CMA
i
and housing prices q
Ri
in the benchmark economy and the counterfactual economy
with ψ = 0.33.
33
Figure 8: Residence changes for continuing commuters, old telecommuters, and new telecommuters
Note: Absolute change in residential density for continuing commuters (top figure), old telecommuters (middle
figure) and new telecommuters (bottom figure). Relative to benchmark economy in counterfactual with ψ = 0.33.
35
Figure 9: Job changes for continuing commuters, old telecommuters, and new telecommuters
Note: Absolute change in job density for continuing commuters (top figure), old telecommuters (middle figure)
and new telecommuters (bottom figure). Relative to benchmark economy in counterfactual with ψ = 0.33.
36
E Elasticity of Speed to Traffic Volume
We set the elasticity of commuting speed with respect to traffic volume is ε
V
= 0.2, following
Small and Verhoef (2007). In the counterfactual economy, we calculate changes in commuting
speeds as
speed
CF
ij
− speed
BM
ij
speed
BM
ij
= −ε
V
V MT
CF
− V MT
BM
V MT
BM
,
assuming that the road capacity remains unchanged and only taking into account the change
in total vehicle miles traveled (V MT ) in the metropolitan area.
30
Then we recover commuting
times as t
CF
ij
= distance
ij
/ max{speed
CF
ij
, 65mph}. The maximum operator caps speeds at 65
mph which is the speed limit on most highways in California. Since t
ij
and V MT endogenously
depend on each other, when solving for an equilibrium in a counterfactual economy, we iterate
the model until V MT converges.
Robustness. Since the results of the counterfactual experiments described in Section 3 cru-
cially depend on changes in commuting speeds, we investigate whether our results are robust to
the value of ε
V
. While 0.2 is a standard value in the traffic modeling literature, other studies
used higher values.
31
At the same time, a low value of ε
V
ensures that many of our counterfactual
results are conservative.
To understand how sensitive our results are to the value of ε
V
, we compute the counterfactual
economy with fraction ψ = 0.33 telecommuters at different levels of ε
V
ranging from 0 to 1. Our
three main sets of results remain robust to the value of ε
V
. First, regardless of the value of ε
V
, the
economy exhibits the decentralization of residents and centralization of jobs. Second, commuters’
trips are characterized by shorter times and longer distances (Figure 10). Third, residential and
commercial floorspace prices fall for all values of ε
V
(Figure 11).
30
Note that our methodology does not allow for differential impact of changes in traffic on individual routes.
31
For example, Akbar, Couture, Duranton, and Storeygard (2020) used values of 0.2 and 0.3. Bento, Hall, and
Heilmann (2020) estimate a value of about 0.9 for peak-hour commuting in Los Angeles.
37
Figure 10: Commuting time and distance
BM 0 0.2 0.4 0.6 0.8 1
Elasticity of speed to volume
20
25
30
Commuting time, minutes
All workers
Commuters
BM 0 0.2 0.4 0.6 0.8 1
Elasticity of speed to volume
20
25
30
Commuting distance, km
All workers
Commuters
Note: Left panel displays the average commuting time for all workers and commuters in the benchmark and the
counterfactual economies at different levels of the elasticity of commuting speed with respect to traffic volume.
Right panel shows the average commuting distance.
Figure 11: House prices
0 0.2 0.4 0.6 0.8 1
Elasticity of speed to volume
-6.5
-6
-5.5
-5
-4.5
-4
Change compared to BM, %
Residential floorspace price
Commercial floorspace price
Note: The figure displays the counterfactual change in average residential and commercial floorspace prices, as a
function of the elasticity of commuting speed with respect to traffic volume.
At the same time, quantitative implications of more telecommuting for wages and welfare are
sensitive to the value of ε
V
. In our main counterfactual with ε
V
= 0.2, the average commuter
market access (CMA) increases by about 17%. However, as ε
V
approaches 1, commutes become
speedier and the average CMA increases by nearly 80% (left panel of Figure 12). In addition, the
higher the elasticity of speed, the stronger will be spatial productivity spillovers. Hence, when ε
V
goes to 1, wage gains for commuters are much larger and wage losses for telecommuters turn into
38
small gains, resulting in larger average wage increases (right panel of Figure 12).
Figure 12: Commuter market access, wages, and land prices
0 0.2 0.4 0.6 0.8 1
Elasticity of speed to volume
0
20
40
60
Change compared to BM, %
CMA
0.0374 0.1 0.15 0.2 0.25 0.3 0.33
Fraction of telecommuters (
)
-8
-6
-4
-2
0
Change, %
Wages
Land prices
Note: Left panel displays the average commuter market access for commuters, as a function of the elasticity of
commuting speed with respect to traffic volume. Right panel shows average wages and land prices.
As a result, with higher values of ε
V
, welfare gains are larger (Figure 13). In particular, as ε
V
goes to 1, commuters see their welfare increase by almost 10% (compared to 2.2% at ε
V
= 0.2),
telecommuters experience a 2% increase (compared to a 2.5% loss), and overall welfare increases
by nearly 25% (compared to 18.9%).
39
Figure 13: Welfare
0 0.2 0.4 0.6 0.8 1
-10
0
10
20
30
Change compared to BM, %
Commuters
0 0.2 0.4 0.6 0.8 1
Elasticity of speed to volume
-10
0
10
20
30
Telecommuters
Total welfare
Consumption and commuting welfare
Consumption welfare
0 0.2 0.4 0.6 0.8 1
-10
0
10
20
30
All workers
Note: Left panel shows the change in total expected welfare of commuters (“total welfare”), welfare net of pref-
erence shocks and amenities (“consumption and commuting welfare”), and welfare net of shocks, amenities, and
commuting costs (“consumption welfare”). Central and right panels report changes in welfare for telecommuters
and all workers, respectively.
F Accounting for Spatial Variation in Outcomes
Centrality. Distance from the center is a key driver of outcomes in most theoretical models
of the city. When dealing with data on real cities, it has been customary to measure this factor
simply as the straight-line distance from a “central business district” whose location is determined
by convention. Our alternative, which uses information on the city’s transportation network, is
the eigenvector centrality of each tract. We calculate it by finding the eigenvector associated
with the largest eigenvalue of the I × I matrix whose ij
th
element is given by exp{−κτ
ij
}. This
measure reflects the total strength of a given tract’s connections, taking into account not only
its direct connections, but also the connections of its connections (second order), and their (third
order) connections, and so on ad infinitum.
Interestingly, this measure picks out downtown LA as the most central location on the map. It
also turns out to be highly correlated with both straight line distance and travel time to downtown
LA (Pearson’s correlation coefficient 0.97 for each). Figure 14 shows the evolution of some key
variables along the centrality gradient.
32
Real estate prices, the density of employment, and the
32
The x-axis is scaled to quantiles of the centrality measure, weighted by land area. In other words, 0.5 on the
40
Figure 14: Quantiles of Centrality and Initial Allocations
Note: The x-axis is scaled to quantiles of the centrality measure, weighted by land area.
density of residence all increase on average the closer one gets to the center. The time required
to reach the downtown LA is also, naturally, lower near the center.
In Figure 3, we plot the changes that take place in the counterfactual exercise in the same
manner as in Figure 14. Here again we see that on average jobs move towards the center and
residents move away from it, and that there are big property price increases in the periphery. We
can also see that there is a great deal of variation that is unexplained.
Accounting for counterfactual changes. In order to have a more complete idea of what
is driving the variation in counterfactual outcomes, we expand our view to consider not only a
location’s initial centrality, but also the change in centrality between the baseline and counterfac-
tual due to changes in average speed, and the exogenous local characteristics a
i
, E
i
and x
i
. We
run a multivariate regression at the tract level, weighted by land area, of these five variables on
the log differences between counterfactual and baseline floorspace prices, employment density, and
residential density. From the estimated coefficients of these regressions we can infer the sign of
each relationship. We then use the Shapley method to decompose the coefficient of determination
(R
2
) for each regression.
33
The share assigned to each explanatory variable is a measure of its
importance in accounting for the variation across space in each counterfactual outcome.
Table 10 shows the results of this exercise for the change in floorspace prices. The negative
estimated coefficient on centrality confirms the core-periphery gradient of price changes, with
prices falling in the core and rising in the periphery. Once this is accounted for, locations whose
centrality increases due to change in speed in the counterfactual also see a more positive overall
change in prices. The negative coefficients on a
i
indicates that the relative value of real estate in
locations with high productivity falls, which is to be expected as workers on average need much
x-axis represents the single square meter of land area such that 50% of the land area in the metro area is less
central (and 50% is more central.
33
See, e.g., Shorrocks (2013).
41
Table 10: Accounting for counterfactual floorspace price changes
Coeff. Var. expl.
constant 0.274
(0.063)
centrality 0.022 32.0%
(0.015)
∆ centrality 3.918 32.1%
(0.389)
a
i
-0.270 3.4%
(0.012)
E
i
-0.018 1.8%
(0.001)
x
i
0.024 15.0%
(0.001)
Total 84.33%
less worksite floorspace than before. The positive coefficient on x
i
indicates that the premium for
locations with good natural amenities has increased in the counterfactual, driven by telecommuters
who can now choose their residence location more freely. We see that position relative to the core
drives the lion’s share of the action here: centrality and ∆ centrality together account for 64.1%
of the variation in outcomes. Overall, the factors we consider here account for about 84% of the
total variation.
Table 11: Accounting for counterfactual employment changes
Always commuter New telecommute Always telecommute All
Coeff. Var. expl. Coeff. Var. expl. Coeff. Var. expl. Coeff. Var. expl.
constant -1.292 -7.139 -0.924 -3.835
(0.199) (0.362) (0.119) (0.251)
centrality -0.138 34.5% -1.384 5.0% -0.149 32.9% -0.684 20.6%
(0.049) (0.089) (0.029) (0.062)
∆ centrality -15.619 35.0% -40.542 5.9% -10.146 33.9% -25.506 22.2%
(1.237) (2.253) (0.737) (1.561)
a
i
1.012 3.1% 2.508 18.7% 0.534 3.1% 1.596 8.3%
(0.039) (0.070) (0.023) (0.049)
E
i
0.057 1.7% 0.189 4.7% 0.022 1.8% 0.107 2.2%
(0.004) (0.008) (0.003) (0.006)
x
i
-0.041 10.0% -0.015 4.8% -0.010 7.3% -0.030 3.5%
(0.003) (0.005) (0.002) (0.003)
Total 84.36% 39.04% 78.95% 56.78%
For employment density and residential density, we further break the overall changes down
into changes in the average choices made by three groups of workers. These groups are: those
that commute both in the baseline and the counterfactual (67% of all workers), those that switch
from commuting to telecommuting (29.3%), and those that telecommute both in the baseline and
the counterfactual (3.7%). Table 11 shows the results for changes in employment density and
Table 12 shows the same for changes in resident density. Workers who continue commuting take
jobs closer to the urban core and also choose residences that are, on average, closer to the core.
42
Table 12: Accounting for counterfactual residence changes
Always commuter New telecommute Always telecommute All
Coeff. Var. expl. Coeff. Var. expl. Coeff. Var. expl. Coeff. Var. expl.
constant 0.840 -0.078 9.789 -0.148
(0.119) (0.225) (0.502) (0.141)
centrality 0.153 29.5% 0.026 35.6% 1.038 12.4% -0.045 24.9%
(0.029) (0.055) (0.123) (0.035)
∆ centrality -1.282 29.5% 15.681 36.0% 44.651 13.1% 3.514 24.9%
(0.737) (1.400) (3.122) (0.879)
a
i
0.171 3.1% -0.240 2.8% -8.139 17.2% 0.009 1.9%
(0.023) (0.044) (0.097) (0.027)
E
i
0.050 2.4% 0.036 2.6% -1.051 27.8% 0.046 5.5%
(0.003) (0.005) (0.011) (0.003)
x
i
0.004 5.8% 0.010 10.2% 0.338 7.6% -0.001 7.8%
(0.002) (0.003) (0.007) (0.002)
Total 70.33% 87.19% 78.08% 65.01%
New telecommuters, with the new-found freedom, do the opposite: they choose jobs and residence
that are, on average, farther from the core than before. Continuing telecommuters make smaller
shifts overall, taking jobs a bit closer to the core and moving their residences a bit farther from
it. Across all categories of workers there is a strong shift from commercial to residential use of
land in locations where there is a larger increase in centrality due to commuting speed changes.
There is also some heterogeneity in the way that location-specific characteristics correlate with
changes in choices for the three groups. For example, those who telecommute both in the baseline
and the counterfactual move their residences out of high-a
i
and high-E
i
tracts, presumably to
make room for the overall shift of employment into those tracts, while this pattern isn’t seen for
the other two groups.
As with changes in land prices, initial centrality and changes in centrality together account
for the lion’s share of the explained variation: 42.6% out of 56.78% total for employment changes,
and 49.8% out of 65.01% total for residence changes. The positive coefficient on a
i
for employment
changes, and its 8.3% share in the variation in outcomes, is consistent with an improvment in the
allocation of workers to high-productivity locations in the counterfactual. Overall, the included
factors account for less of the variation than in the case of floorspace prices. This is partly due
to opposing tendencies in the three different types of workers cancelling each other out.
43
