Robots and Jobs (Acemoglu 2019)

Manuscript Click here to
access/download;Manuscript;robots_and_jobs_May_15_2019_jp
Copyright The University of Chicago 2019. Preprint (not copyedited or formatted). Please use DOI when citing or quoting. DOI: 10.1086/705716
Robots and Jobs: Evidence from US Labor

Markets.∗
Daron Acemoglu Pascual Restrepo

MIT Boston University
June 28, 2019
Abstract
As robots and automation technologies take over tasks performed by labor, there is in-
creasing concern about the future of jobs and wages. We analyze the effect of the increase
in industrial robot usage between 1990 and 2007 on US labor markets. Using a model in
which robots compete against workers, we show that advances in robotics technology may
reduce employment and wages. The local labor market impacts of robots can be estimated
by regressing the changes in employment and wages on exposure to robots in each local
labor market—defined from industry-level advances in robotics technology and the local dis-
tribution of employment across industries. Using this approach, we estimate robust negative
effects of robots on employment and wages across commuting zones. We bolster this evidence
by showing that the commuting zones most exposed to robots in the post-1990 era do not
exhibit any differential trends before 1990. The impact of robots is distinct from other types
of capital and technologies. According to our estimates, one more robot per thousand work-
ers reduces the employment to population ratio by about 0.2 percentage points and wages
by 0.42 percent.
Keywords: automation, industrial robots, employment, jobs, labor, wages.
JEL Classification: J23, J24.
∗
We thank David Autor, James Bessen, Lorenzo Caliendo, Amy Finkelstein, Matthew Gentzkow, Michael
Greenstone, Mikell Groover, Steven Miller, Brendan Price, John Van Reenen, three anonymous referees and
participants at various seminars and conferences for comments and suggestions; Giovanna Marcolongo, Mikel
Petri and Joonas Tuhkuri for outstanding research assistance; and the Institute for Digital Economics and the
Toulouse Network of Information Technology for financial support.
Journal of Political Economy

Downloaded from www.journals.uchicago.edu by Nottingham Trent University on 08/15/19. For personal use only.
1 Introduction
In 1929 John Maynard Keynes famously predicted that the rapid spread of automation tech-
nologies would bring “technological unemployment” (Keynes, 1930). Wassily Leontief prophe-
sied similar problems for workers, writing “Labor will become less and less important. . . More
and more workers will be replaced by machines. I do not see that new industries can employ
everybody who wants a job” (quoted in Curtis, 1983, p. 8). Though these predictions have not
come to pass, there is renewed concern that advances in robotics and artificial intelligence will
lead to massive job losses (e.g., Brynjolfsson and McAfee, 2014; Ford, 2015). There is mounting
evidence that the automation of a range of low- and medium-skill occupations has contributed
to wage inequality and employment polarization (e.g., Autor, Levy and Murnane, 2003; Goos
and Manning, 2007; Michaels, Natraj and Van Reenen, 2014). These concerns notwithstanding,
we have little systematic evidence on the equilibrium impact of automation technologies, and
especially of robots, on employment and wages.1
In this paper we estimate the equilibrium impact of a leading automation technology, in-
dustrial robots, on local US labor markets. The International Federation of Robotics—IFR for
short—defines an industrial robot as “an automatically controlled, reprogrammable, and mul-
tipurpose [machine]” (IFR, 2014). That is, industrial robots are fully autonomous machines
that do not need a human operator and can be programmed to perform several manual tasks
such as welding, painting, assembly, handling materials and packaging. Textile looms, elevators,
cranes or transportation bands are not robots since they have a unique purpose, cannot be re-
programmed to perform other tasks and/or require a human operator. This definition excludes
other types of equipment and enables an internationally and temporally comparable measure-
ment of a class of technologies—industrial robots—that are capable of replacing human labor
in a range of tasks.
Robotics technology advanced significantly in the 1990s and 2000s, leading to a fourfold
rise in the stock of (industrial) robots in the United States and Western Europe between 1993
and 2007. As Figure 1 shows, in the United States the increase amounted to one new robot
per thousand workers, and in Western Europe to 1.6 new robots per thousand workers. The
automotive industry employs 38 percent of existing robots, followed by the electronics industry
(15 percent), plastics and chemicals (10 percent) and metal products (7 percent).
Our empirical approach is based on a model where robots and workers compete in the
production of different tasks. Our model builds on Acemoglu and Autor (2011), Acemoglu and
Restrepo (2018a) and Zeira (1998), but extends these frameworks so that the share of tasks
performed by robots varies across sectors and there is trade between labor markets specializing
in different industries. Greater robot adoption negatively affects wages and employment owing
1
Frey and Osborne (2013), World Development Report (2016) and McKinsey (2017) estimate which types of
jobs are susceptible to automation based on various technological projections. Such approaches are not infor-
mative about the equilibrium impact of automation since they do not take into account how other sectors and
occupations will respond to these changes. See also Arntz, Gregory and Zierahn (2016) on other problems with
these methodologies.
1
to a displacement effect (as they directly displace workers from tasks they were previously
performing), but there is also a positive productivity effect (as other industries and/or tasks
increase their demand for labor). Our framework clarifies that, because of the displacement
effect, robots can have very different implications for labor demand than capital deepening
or factor-augmenting technologies. It further shows that the effects of robots on employment
and wages can be estimated by regressing the change in these variables on exposure to robots.
Exposure to robots is a Bartik-style measure (Bartik, 1991), constructed from the interaction
between baseline industry shares in a local labor market and technological possibilities for the
introduction of robots across industries.
We first document that there is considerable variation in robot adoption across industries
and show that the same industries are rapidly adopting robots in both the United States and
Europe. We further show that, at the industry level, there is no strong positive correlation
between robot adoption and any of the other major trends affecting US local labor markets such
as import competition from China, competition from Mexico, offshoring, the decline of routine
tasks, investments in IT capital and overall capital deepening. Moreover, consistent with theory,
robot adoption at the industry level is associated with lower labor share and employment, and
greater value added and labor productivity.
After presenting industry-level correlations, we investigate the equilibrium impact of robots
in local labor markets, proxied by commuting zones in the United States.2 We construct our
measure of exposure to robots using data from the IFR on the increase in robot usage across
19 industries (roughly at the two-digit level outside manufacturing and at the three-digit level
within manufacturing) and their baseline employment shares from the Census before the onset
of recent robotic advances. In order to focus on the component of investment in robots driven
by technological advances, we exploit adoption trends in European economies that are ahead
of the United States in robotics. Our identifying assumption is that commuting zones housing
industries where there are greater advances in robotics technology are not differentially affected
by other labor market shocks or trends—a presumption we investigate from a number of angles.3
Using this strategy, we estimate a negative relationship between a commuting zone’s exposure
to robots and its post-1990 labor market outcomes. Our estimates imply that between 1990
2
Not all equilibrium responses take place within commuting zones—the most important other responses are
trade with other local labor markets, which we model explicitly below; migration, which we investigate empirically;
and the response of technology and new tasks to changes in factor prices emphasized in Acemoglu and Restrepo
(2018a). All the same, recent research suggests that much of the adjustment to shocks, both in the short run
and the medium run, takes place locally (e.g., Acemoglu, Autor and Lyle, 2005, Moretti, 2011, Autor, Dorn and
Hanson, 2013).
3
We show in Acemoglu and Restrepo (2018c) that greater robot adoption in these countries is largely a
consequence of their more rapid demographic change than in the US.
This empirical strategy is similar to that used by Autor, Dorn and Hanson (2013) and Bloom, Draca and
Van Reenen (2015) to estimate the effects of Chinese imports. Though not a panacea for all sources of omitted
variable bias, this strategy allows us to filter out variation in robot adoption coming from idiosyncratic US factors
(e.g., US-specific declines or worsening labor relations in some industries). This strategy would be compromised
if changes in robot usage in other advanced economies are correlated with adverse shocks to US industries. For
instance, there may be common shocks affecting the same industries across advanced economies, such as other
technological changes or import competition, and these shocks could induce the same industries everywhere to
adopt robots. We show later that these confounders are not responsible for our results.
2
and 2007 the increase in the stock of robots (approximately one additional robot per thousand
workers from 1993 to 2007) reduced the average employment to population ratio in a commuting
zone by 0.39 percentage points and average wages by 0.77 percent (relative to a commuting zone
with no exposure to robots). These numbers are sizable but not implausible. For example, they
imply that one more robot in a commuting zone reduces employment by about 6 workers; this
estimate includes both direct and indirect effects, the latter caused by the decline in the demand
for nontradables as a result of reduced employment and wages in the local economy.
To understand the aggregate implications of these estimates, we need to make additional as-
sumptions about how different commuting zones interact. Greater use of robots in a commuting
zone generates benefits for the rest of the US economy by reducing the prices of tradable goods
produced using robots and by creating shared capital gains. Our model enables us to quantify
these positive spillovers across commuting zones and leads to smaller but still uniformly negative
aggregate effects. With our preferred specification, our estimates imply that one more robot per
thousand workers reduces the aggregate employment to population ratio by about 0.2 percent-
age points and wages by about 0.42 percent (compared to its larger local effects, 0.39 percentage
points and 0.77 percent, respectively). Or equivalently, one new robot reduces employment by
about 3.3 workers.
We verify that our measure of exposure to robots is unrelated to past trends in employment
and wages from 1970 to 1990, a period that preceded the onset of rapid advances in robotics
technology. Several robustness checks bolster our interpretation. First, our results are robust
to including differential trends by various baseline characteristics, linear commuting zone trends
and controls for other changes affecting demand or productivity in various industries. Second,
we show that the automotive industry, which is the most robot-intensive sector, is not driving
our results. Third, consistent with our theoretical emphasis that robots (and more generally,
automation technologies) have very different labor market effects than other types of machinery
and overall capital deepening, we find no similar negative impact from capital, other measures
of IT or overall productivity increases.
The employment effects of robots are most pronounced in manufacturing, and in particular
in industries most exposed to robots. They are also concentrated in routine manual, blue collar,
assembly and related occupations. Consistent with the presence of spillovers on nontradables,
we estimate negative effects on construction and retail and personal services as well.
Besides the papers that we have already mentioned, our work is related to the empirical
literature on the effects of technology on wage inequality (Katz and Murphy, 1992), employment
polarization (Autor, Levy and Murnane, 2003; Goos and Manning, 2007; Autor and Dorn, 2013;
Michaels, Natraj and Van Reenen, 2014), aggregate employment (Autor, Dorn and Hanson, 2015;
Gregory, Salomons and Zierahn, 2016), the demand for labor across cities (Beaudry, Doms and
Lewis, 2006) and firms’ organization and demand for workers with different skills (Caroli and
Van Reenen, 2001, Bartel, Ichniowski and Shaw, 2007, and Acemoglu et al., 2007).
Most closely related to our work is the pioneering paper by Graetz and Michaels (2018).
Focusing on the variation in robot usage across industries in different countries, Graetz and
3
Michaels estimate that industrial robots increase productivity and wages, but reduce the em-
ployment of low-skill workers. Although we rely on the same IFR data, we utilize a different
empirical strategy, which enables us to go beyond cross-country, cross-industry comparisons,
to exploit plausibly exogenous changes in the spread of robots and to estimate the equilibrium
impact of robots on local labor markets.
The rest of the paper is organized as follows. Section 2 presents a simple model of the effects
of robots on employment and wages. Section 3 introduces our data and sources. Section 4
documents the correlation between robot adoption at the industry level and employment, the
labor share and value added. Section 5 presents our main empirical results and various robust-
ness checks. This section also looks at the differential effects of robots on workers in different
industries, occupations and skill groups. Section 6 presents our IV estimates and evaluates the
local and aggregate implications of the spread of robotics technology in the US. Section 7 con-
cludes. The Appendix, which is available online, presents proofs, additional theoretical results
and robustness checks.
2 Robots, Employment and Wages: A Model

This section presents a model building on Acemoglu and Restrepo (2018a) to exposit the po-
tential effects of robots on employment and wages, and derives our estimating equations. To
develop intuition, we start with a model without trade between commuting zones.
2.1 The Effects of Robots in Autarky Equilibrium

The economy consists of |C| commuting zones. Each commuting zone c ∈ C has preferences
defined over an aggregate of the output of |I| industries, given by
! σ
X 1 σ−1 σ−1
σ σ
Yc = νi Yci , (1)
i∈I
where σ > 0 denotes the elasticity of substitution across goods produced by different industries
and the νi ’s are share parameters that designate the importance of industry i in the consumption
P
aggregate (with i∈I νi = 1).
In the autarky equilibrium a commuting zone consumes its own production of each good,
denoted by Xci . Hence, for all i ∈ I and c ∈ C, we have Yci = Xci . We choose the consumption
aggregate in each commuting zone as numeraire (with price normalized to 1) and denote the
price of the output of industry i in commuting zone c by PciX .
Each industry produces output by combining capital with a continuum of tasks indexed by
s ∈ [0, 1], each of which can be produced using industrial robots or human labor. We use xci (s)
to denote the quantity of task s utilized in the production of Xci . These tasks must be combined
4
in fixed proportions so that
Xci = α−α (1 − α)−(1−α) Aci [ min {xci (s)}]α Kci

1−α
, (2)
s∈[0,1]
where Kci denotes the non-robot capital used in industry i, 1 − α is its share in the production
process, Aci is the productivity of industry i, and the term α−α (1 − α)−(1−α) is a convenient
normalization. Differences in the Aci ’s will translate into different industrial compositions of
employment across commuting zones.
Industrial robots replace workers in some of the tasks they were previously performing.
Specifically, in industry i tasks [0, θi ] are technologically automated and can be performed by
robots. We assume that all commuting zones have access to the same technology, i.e., the same
θi in industry i. Denoting the productivity of labor by γL and the productivity of robots by
γM > 0, we have (
γM Mci (s) + γL Lci (s) if s ≤ θi
xci (s) =
γL Lci (s) if s > θi ,
where Lci (s) and Mci (s) are, respectively, the numbers of workers and robots used in task s.
Because tasks above θi have not yet been technologically automated, they must be performed
by labor.
In each commuting zone c labor is supplied by a representative household with preferences
Cc1−ψ − 1 B
− L1+ε ,
1−ψ 1+ε c
where Cc denotes this household’s consumption and Lc is its labor supply. Its budget constraint
is Cc ≤ Wc Lc + Πc , where Πc is non-labor (capital and profit) income. In this specification, ψ
determines the income elasticity of labor supply, and ε is the inverse of the Frisch elasticity of
labor supply.
Robots are produced using investment (in units of the final good), denoted by Ic , with the
1
production function Mc = D(1 + η)Ic1+η , and have a rental price of RcM . This formulation,
with η > 0, allows the supply of robot services to a commuting zone to be upward sloping.
This is reasonable in the medium term, since about two thirds of the costs of robots are for
services supplied by local, specialized robot integrators that install, program and maintain this
equipment (Leigh and Kraft, 2018). Finally, in the autarky model we take the supply of capital
in commuting zone c to be fixed at Kc , and denote its price by RcK .
An equilibrium is a tuple of prices {Wc , RcM , RcK }c∈C and quantities {Cc , Yc , Ic , Lc , Mc }c∈C
such that in all commuting zones, firms maximize profits, households maximize their utility, and
the markets for capital, labor, robots and final goods clear:
XZ XZ X
Lci (s) =Lc , Mci (s) =Mc , Kci (s) =Kc , Cc =Yc − Ic .
i∈I [0,1] i∈I [0,1] i∈I
We prove in the Appendix that an equilibrium exists and is unique.
5
To analyze the equilibrium impact of robots, let us first define cost savings from using robots
in commuting zone c as
γL RcM
πc = 1 − .
γM W c
Robots will not be adopted when πc < 0; in what follows we focus on the case where πc > 0 in
all commuting zones. The next proposition characterizes the partial equilibrium impact of an
advance in automation/robotics technology for industry i, denoted by dθi .
Proposition 1 Suppose that πc > 0. Then

dθi 1 1
d ln Lci = − + d ln Yc − σ + − 1 d ln PciX , (3)
1 − θi α α
where Lci denotes the employment in industry i in commuting zone c.
Like all other results in this section, a proof of this proposition is presented in the Appendix.
Equation (3) highlights three different forces shaping labor demand of industry i, repre-
sented by Lci . First, there is a negative displacement effect: an increase in θi leads to the use
of robots in tasks otherwise performed by labor, displacing workers employed in these tasks.
This displacement effect always reduces the labor share in the industry undergoing automation
and may also reduce its overall labor demand.4 Labor demand does not necessarily decline
following automation because of the countervailing (positive) productivity effect, which is repre-
sented by the second term. Intuitively, automation lowers the cost of production (thus increases
“productivity”), and via this channel raises the demand for labor in non-automated tasks in
all industries. Finally, there is a composition effect, represented by the third term: industries
undergoing automation expand at the expense of others, and this raises the demand for labor
coming from their non-automated tasks.
We can aggregate the industry-level implications of Proposition 1 to derive the effects of
robots on local labor demand as follows:
X
X dθi 1 1
d ln Lc = − `ci + d ln Yc − σ + − 1 (`ci − χci )d ln PciX , (4)
1 − θi α α
i∈I i∈I
where `ci is industry i’s share in total employment in commuting zone c, while χci is this
industry’s share of value added in the local economy. The first two terms are direct analogues
of the displacement and productivity effects in (3). The third term shows that the impact of
the composition effect for labor demand depends on whether automation is reallocating output
towards sectors that are more labor intensive than average (those for which `ci > χci ). This
composition effect disappears when all industries have the same labor share.
4
The negative impact on theh labor
n share can beoiseen by computing the total task production in industry i
α
as Xci = Aci α−α (1 − α)−(1−α) min γMθMi
ci γL Lci
, 1−θi
K 1−α
ci , which shows that an increase in θi always makes
production less labor intensive (see Acemoglu and Restrepo, 2018a, 2019). In the Appendix we establish that
a sufficient condition for the displacement effect to dominate the other forces and reduce (relative) industry
1−π sL
employment is σ < 1 + απ csLic , where sL ic is the industry’s labor share in production tasks. This condition is
c ic
easily satisfied for plausible parameter values.
6
Equation (4) provides a partial equilibrium characterization of how the demand for labor
changes following automation. The next proposition links changes in prices and total output to
automation technologies and thus derives the full equilibrium impact of automation.
Proposition 2 Suppose that πc > 0 for all c ∈ C and θi = 0 for all i ∈ I. Then
X dθi γL
d ln Lc =[−ζ disp + ζ prod πc − ζc,L
inc
ψ] · `ci , (5)
1 − θi γ M
i∈I
X dθi γL
d ln Wc =[−ζ disp ε + ζ prod επc + ζc,W
inc
ψ] · `ci , (6)
1 − θi γ M
i∈I
where ζ disp = (1−α+η)/Λ, ζ prod = (1+η)/Λ, ζc,L

inc = απ /Λ , ζ inc = α(π −(1−π )(1−α+η))/Λ,
c c c,W c c
γL
and Λ = γM (1 − α + αψ + ε) > 0.
The assumption that θi = 0 for all i simplifies the relevant expressions by removing the
composition effect. The economic effects are similar when this assumption is relaxed as shown
in the Appendix.5
Proposition 2 establishes that the response of both employment and wages to automation
P dθi γL
is shaped by the term i∈I `ci 1−θ i γM
, which is the basis of our exposure to robots measure.
In addition, the coefficient on this variable in both equations comprises three distinct terms.
The first term, −ζ disp , represents the displacement effect. The second, ζ prod , represents the
productivity effect, generated by cost savings, πc . When cost savings from automation are
limited, automation decreases employment and wages. Conversely, when πc is large, automation
increases them. Finally, the third term in both equations incorporates the negative income effect
of automation on labor supply.
The impacts of robots highlighted in Proposition 2 are very different from the effects of overall
capital deepening (an increase in the supply of capital, Kc ), or from technological changes that
increase the productivity of robots, γM , or industry productivity, Aci . Capital deepening, greater
productivity of robots and increases in industry productivity do not displace workers from the
tasks they are performing and always raise wages and employment.6 This observation clarifies
that the displacement effect created by automation is responsible for its potentially negative
impact on labor demand.
2.2 The Effects of Robots When Commuting Zones Trade

The autarky model transparently illustrates the displacement and productivity effects of au-
tomation, but ignores how its economic consequences may spill over across local labor markets.
5
Composition effects arise when the term `ci − χci (or equivalently the labor share) is correlated with the
introduction of robots across industries. The correlation between the labor share of an industry in 1992 and
subsequent robots usage is 0.1 across all industries and -0.04 within manufacturing, suggesting a minor role for
composition effects.
6
As we show in Acemoglu and Restrepo (2018b, 2019), labor-augmenting technological changes have very
different effects from automation as well unless the elasticity of substitution between labor and machines is
implausibly low (in particular lower than the share of machines in value added). Here we took this elasticity to
be zero for simplicity, since this simplification does not impact any of the implications we are focusing on.
7
Trade in goods and services change the sensitivity of employment and wages to robot adoption
and their aggregate implications. We now incorporate automation/robots into a simple model
of trade between commuting zones building on Armington (1969) and Anderson (1979). Specif-
ically, we modify our model in two ways. First, we assume that the representative household’s
utility depends on a tradable good, Cc , and a nontradable (service) good, Sc :
(Ccφ Sc1−φ )1−ψ − 1 B

− L1+ε . (7)
1−ψ 1+ε c
This specification implies that a constant share φ ∈ (0, 1) of spending goes to the tradable good.
We assume that this nontradable good is produced with labor, that is, Sc = LSc , and we denote
the price of the nontradable good in commuting zone c by Pc . The remaining labor, Lc − LSc , is
used in the production of tradable goods.
The second modification is to assume that the tradable good is produced as in (1), but now
with inputs sourced from all commuting zones so that
! λ
λ−1 λ−1
X 1 λ
λ
Yci = υsi Xsci (for all c and i), (8)
s∈C
where λ is the elasticity of substitution between varieties sourced from different commuting
zones, and the share parameters, the υsi ’s, indicate the desirability of varieties from different
sources. We assume that there are no trade costs, so that the price of the tradable good is
equalized across commuting zones, and we choose it as the numeraire. Denoting the amount
of good i exported from commuting zone c to destination d by Xcdi (including d = c), market
clearing imposes
X
Xci = Xcdi (for all c and i).
d∈C
We also assume that the initial stock of capital of the economy, K, is perfectly mobile
across commuting zones, and we modify the budget constraint of households to Cc + Pc Sc ≤
Wc Lc +χΠ Π
c Π, where Π is the national non-labor income and a share χc of this income is allocated
to commuting zone c (with c∈C χΠ
P
c = 1). The main result of this section is presented in the
next proposition, which parallels Proposition 2.
Proposition 3 Suppose that πc = π0 for all c ∈ C and θi = 0 for all i ∈ I. Then

X dθi γL price
d ln Lc =[−ζ̄ disp φ + ζ̄ prod φπ0 − ζ̄Linc ψ] `ci + ζ̄LY d ln Y + ζ̄LΠ d ln Π + ζ̄cL (9)
1 − θi γ M
i∈I
X dθi γL price
d ln Wc =[−ζ̄ disp ε + ζ̄ prod επ0 + ζ̄W
inc
ψ] `ci Y
+ ζ̄W Π
d ln Y + ζ̄W d ln Π + ζ̄cW , (10)
1 − θi γ M
i∈I
where the ζ̄’s are functions of the underlying parameters.
This proposition assumes that πc is the same across commuting zones as well as θi = 0 for
8
all i; we provide a more general version with similar implications in the Appendix.
As before, the ζ̄’s summarize the local impact of robots on employment and wages. Trade
between commuting zones implies that productivity gains and price changes in one area will be
shared with others. The productivity spillovers, generated by the change in national income
d ln Y , are captured by the ζ̄ Y terms, while spillovers from changes in prices are summarized
by the ζ̄ price terms. Finally, the ζ̄ Π terms represent the income effects and the demand for
nontradables resulting from non-labor income, d ln Π. These general equilibrium effects are
not functions of exposure to robots in the own commuting zone, and thus we obtain the same
reduced-form relationship between robots and local labor demand as in the autarky model. The
aggregate effects of robots, however, depend on the extent of trade across commuting zones
because of the additional spillover terms and because the ζ̄’s in this proposition differ from their
autarky counterparts in Proposition 2. We take these differences into account in our quantitative
evaluation.
2.3 Empirical Specification

Propositions 2 and 3 summarize the effects of advances in the robotics technology on local
employment and wages. The key equations, (9) and (10), show that the equilibrium impact of
robots depends on the same object, which we will call a commuting zone’s US exposure to robots,
X
US exposure to robotsc = `ci · AP Ri , (11)
i∈I
where recall that `ci is the baseline employment share of industry i in commuting zone c, and
dθi γL dMi dYi Mi

AP Ri = = − (12)
1 − θi γ M Li Yi Li
is the (US) adjusted penetration of robots in industry i. Exposure to robots is thus a Bartik-
style measure combining industry-level variation in the usage of robots and baseline employment
shares. Our model implies a specific form for this relationship, including an adjustment for the
overall expansion of each industry’s output, given by the last term in (12).
With this measure of exposure to robots, we can estimate
d ln Lc =βL · US exposure to robotsc + L W

c ; d ln Wc =βW · US exposure to robotsc + c , (13)
regardless of whether or not there is trade between commuting zones, though the coefficients
βL and βW have different interpretations in these two cases. In these equations, L W
c and c
represent other factors affecting labor supply and demand, and in our empirical work, we model
them as functions of various baseline characteristics and observed economic changes.
The models in equation (13) can be estimated using OLS with the US exposure to robots
variable computed from US data on the adjusted penetration of robots. Yet there are two related
reasons why the US exposure to robots could be correlated with the error terms, L W
c and c ,
9
leading to biased estimates. First, some industries may be adopting robots in response to other
changes that they are undergoing, which could directly impact their labor demand. Second, any
shock to labor demand in a commuting zone affects the decisions of local businesses, including
robot adoption.7
Ideally, we only want to use changes in robot penetration driven by exogenous improvements
in technology, dθi . To identify the component of robot penetration driven by changes in tech-
nology, we instrument the US exposure to robots using an analogous measure constructed from
the penetration of robots in European countries that are ahead of the United States in robotics.
Namely, we construct
X
Exposure to robotsc = `ci · AP Ri , (14)
i∈I
where AP Ri is the adjusted penetration of robots computed from European countries. We

describe and motivate this choice in greater detail in the next section.
3 Data
In this section we describe our main data sources.
3.1 Robots
Our main data consist of counts of the stock of robots by industry, country and year from the
IFR. The IFR data are based on yearly surveys of robot suppliers and cover 50 countries from
1993 to 2014, corresponding to about 90 percent of the industrial robots market. However, the
stock of industrial robots by industry going back to the 1990s is only available for Denmark,
Finland, France, Germany, Italy, Norway, Spain, Sweden, and the United Kingdom, which
together account for 41 percent of the world industrial robot market.8 Outside of manufacturing,
we have consistent data for the use of robots in six broad industries: agriculture, forestry
and fishing; mining; utilities; construction; education, research and development; and services.
Within manufacturing, we have consistent data on the use of robots for 13 more disaggregated
industries: food and beverages; textiles (including apparel); wood and furniture; paper and
printing; plastics and chemicals; minerals; basic metals; metal products; industrial machinery;
electronics; automotive; shipbuilding and aerospace; and miscellaneous manufacturing (e.g.,
production of jewelry and toys). We use this industry classification throughout and refer to it
as the “IFR industries”.
Figure 1 and Table A1 in the Appendix depict the evolution of robot stocks for different
groups of European countries and for the United States. In Figure 1 we show separately the
evolution of the stock of robots for Germany, the average for Denmark, Finland, France, Italy
7
An example for the first concern would be the automotive industry adopting more robots in the United States
because of greater wage push from its unions. An example for the second would be a local recession in Detroit,
MI that impacts the automotive industry that has a large footprint there.
8
Though the IFR also reports data by industry for Japan, these data underwent a major reclassification. We
follow the recommendations of the IFR and exclude Japan from our analysis.
10
and Sweden, the average for Norway, Spain and the UK, and the United States. The trends
for Denmark, Finland, France, Germany, Italy and Sweden are particularly interesting, because
these countries are technologically more advanced than the United States in robotics.9 US robot
usage starts near 0.4 robots per thousand workers in the early 1990s, increases to 0.7 in 2000, and
then rises rapidly to 1.4 robots per thousand workers in the late 2000s; this evolution closely
tracks the average of Denmark, Finland, France, Italy and Sweden, but its level is about 20
percent lower.10
The IFR data have some noteworthy shortcomings. First, not all robots are classified into
one of the 19 IFR industries. About 30 percent of robots are unclassified, and this fraction
has declined throughout our sample. We allocate these unclassified robots to industries in the
same proportions as in the classified data. Second, although the IFR reports data on the total
stock of industrial robots in the United States from 1993 onwards, it does not provide industry
breakdowns until 2004. This does not affect the exposure to robots measure computed from
European data, and in Section 6.1 we describe how we use US data in our IV strategy. Finally,
the IFR only reports the overall stock of robots for North America. Though this aggregation
introduces noise in our measures of US exposure to robots, this is not a major concern, since
the United States accounts for more than 90 percent of the North American market and our IV
procedure purges this type of measurement error from the US exposure to robots.11
We combine the IFR data with employment counts and output by country and industry from
the EUKLEMS dataset (see Jägger, 2016),12 which allows us to construct adjusted penetration
9
These countries have more robots than the United States at the beginning of the sample, in 1993, and have
invested more in robots since. They also have greater “robot exports” (measured as exports of intermediates
related to robotics from the Comtrade dataset, see Acemoglu and Restrepo, 2018c, for details). For example,
robot exports per worker are three to four times as large in Italy, France and Denmark as in the United States,
and more than six times as large in Germany, Finland and Sweden. Norway and the UK are behind the United
States in all of these metrics. Spain has adopted robots rapidly in the automotive industry since 1993, but is
behind or comparable to the United States in other sectors, and its robot exports are at the same level as the
United States.
10
Acemoglu and Restrepo (2018c) show that demographic factors account for a large fraction of this cross-
country variation and for why European countries are ahead of the United States in robotics. The relative
shortage of middle-aged (production) workers in countries aging rapidly, such as Germany, France, Italy, Japan
and South Korea, encourages the development and adoption of robotics technology, which is then exported to
other countries, including the United States, experiencing less rapid demographic change.
11
Robots in different sectors have similar capabilities and prices. Industrial robots belong to one of a handful
of standardized types—articulated robots, SCARA (selective compliance assembly robot arm) robots, Cartesian
robots and parallel robots. Consistent with this, robot prices are fairly similar across sectors (ranging from about
$44,000 per robot to about $88,000), and our results in Table 5 suggest that the quantitative effects of robots
in different sectors are similar. We investigate the role of robot prices further in Tables A24 and A25 in the
Appendix.
12
To obtain comparable data, we first use information on hours worked to obtain a count of US equivalent
workers by industry in 1990. We then compute the number of robots by industry, country and year divided
by US equivalent workers in 1990. Because the data for Norway are missing from the EUKLEMS, we use the
distribution of employment in the remaining Scandinavian countries in our sample (Denmark, Finland, and
Sweden) to impute the Norwegian distribution. In addition, we were able to match most of the industries used
in the EUKLEMS dataset to the 19 IFR industries. One exception is wood and furniture, since employment
in furniture products is pooled with miscellaneous manufacturing. To address this issue, we allocate 40% of
the employment in miscellaneous manufacturing to the wood and furniture sector based on the proportions of
employment in the US in these detailed industries (obtained from the NBER-CES dataset described below).
Finally, because the IFR data for Denmark are not classified by industry before 1996, we construct estimates for
1993-1995 by deflating the 1996 stocks by industry using the total growth in its stock of robots.
11
of robots, AP Ri and AP Ri , for different time periods. Following equation (12), our baseline
measure of the adjusted penetration of robots between two dates, t0 and t1 , is given by
j j
" #
1 X Mi,t 1
− Mi,t 0 j Mi,t0
AP Ri,(t0 ,t1 ) = − gi,(t 0 ,t1 )
(15)
5
j∈EU RO5 Lji,1990 Lji,1990
j
where Mi,t is the number of robots in industry i in country j at time t (from the IFR data),
j
gi,(t 0 ,t1 )
is the growth rate of output of industry i in country j between t0 and t1 (also from the
EUKLEMS), and Lji,1990 is the baseline employment level in industry i and country j (from the
EUKLEMS).13 In our long-differences models, we take t0 = 1993 and t1 = 2007, though we also
present models where we focus on other periods.
For our baseline measure we use the average penetration in EU RO5, comprising Denmark,
Finland, France, Italy, and Sweden, that is, countries ahead of the United States in robotics
excluding Germany. Focusing on countries that are ahead of the United States helps us isolate
the source of variation coming from global technological advances (rather than idiosyncratic
US factors). We exclude Germany from our baseline measure because, as Figure 1 shows, it
is so far ahead of the other countries that its adoption trends may be less relevant for US
patterns than the trends in EU RO5. The Appendix presents versions of our main results for
different constructions of the AP Ri measure, including a specification where we use all European
countries, one where we use both Germany and the EU RO5, one where we use the observed
j
increase in robot density without the gi,(t 0 ,t1 )
Mi,t0 /Lji,1990 term, and a complementary measure
where we include an adjustment for variation in the average price of a robot in an industry.
We also measure the US adjusted penetration of robots as
US − M US
Mi,t
US 1 i,t0 US Mi,t0
AP Ri,(t 0 ,t1 )
= − gi,(t 0 ,t1 )
. (16)
LU S
i,1990 LU S
i,1990
Given the coverage of the IFR data for US industries, this variable only goes back to t0 = 2004.
3.2 Industry Data

To explore the industry-level correlates of robot adoption, we use data on US industry employ-
ment, wage bill, value added and labor share. The employment and wage bill data come from
the County Business Patterns (CBP). We supplement the CBP with the NBER-CES dataset,
which covers the manufacturing sector and reports data on employment and wage bills for all
workers and for production workers (see Acemoglu et al., 2016). We also use data on value added
and labor share from the Bureau of Economic Analysis input-output tables (BEA), and on IT
capital and the overall capital stock from the Bureau of Labor Statistics (BLS). These data
are available for a detailed set of industries, which we then aggregate to the 19 IFR industries.
j
13
Because there were few robots in 1993, the adjustment term gi,(t 0 ,t1 )
Mi,t0 /Lji,1990 is not quantitatively im-
portant; 96 percent of the variation in the adjusted penetration of robots across industries between 1993 and 2007
j j
is driven by the increase in robot density, the term (Mi,t1
− Mi,t0
)/Lji,1990 in equation (15). The exception is the
electronics industry, which had a high stock of robots in 1993 and experienced rapid growth thereafter.
12
Industry-level imports from China and Mexico and exports from Germany, Japan and South
Korea are computed from Comtrade data (following Acemoglu et al., 2016). Finally, we use
the share of tasks in an occupation that can be offshored (“task offshorability” from Autor and
Dorn, 2013) and the share of imported intermediates as a proxy for offshoring (from Feenstra
and Hanson, 1999, and Wright, 2014).
3.3 Commuting Zone Data and Exposure to Robots

In our main analysis, we focus on the 722 commuting zones covering the US continental territory
(Tolbert and Sizer, 1996). Following equations (11) and (14), we measure US exposure to robots
in a commuting zone as
X
US exposure to robotsc,(t0 ,t1 ) = `1990
ci · AP Ri,(t0 ,t1 ) , (17)
i∈I
where `1990
ci is industry i’s share in the total employment of commuting zone c and AP Ri is as
defined in (16). Exposure to robots is defined analogously, exploiting variation in industry-level
adoption of robots in the EU RO5 countries,
X
Exposure to robotsc,(t0 ,t1 ) = `1970
ci · AP Ri,(t0 ,t1 ) , (18)
i∈I
where AP Ri,(t0 ,t1 ) is given in (15). We now use the 1970 employment shares, `1970
ci , as the baseline
in order to focus on historical, persistent differences in the industrial specialization of commuting
zones that predate robotics technology. This choice avoids any mechanical correlation due
to robot adoption before the 1990s or mean reversion associated with temporary changes in
industry employment in the 1980s. It is also worth noting that even when we consider changes
in subperiods (e.g., in our models with stacked differences), we keep the baseline employment
shares constant to avoid endogenous and serially correlated changes in our exposure variable.
We use the public use data from the 1970, 1990 and 2000 Censuses and the American Com-
munity Survey (see Ruggles et al., 2010) to construct measures of population, employment,
employment by industry and occupation, and demographics for each commuting zone. To in-
crease sample size, we follow Autor, Dorn and Hanson (2013) and measure the 2007 outcomes
using the ACS for 2006-2008. Similarly, we measure the 2014 outcomes from the ACS for 2012-
2016. We also use the Census and ACS to compute the average hourly and weekly wages within
250 demographic × commuting zone cells, which corrects for changes in the observed character-
istics of employed workers. Our demographic cells are defined by gender, education (less than
high school, high school degree, some college, college and professional degrees, and masters and
doctoral degrees), ten-year age bins (16-25, 25-35, 36-45, 46-55, 56-65 and 65+) and race. All
top-coded wage income observations are set equal to 1.5 times the value of the top code and
winsorized at 2 dollars per hour as in Acemoglu and Autor (2011). We additionally use county-
level data (which we again aggregate to the commuting zone level) on employment counts from
the County Business Patterns for 1990, 2000 and 2007; wage and non-wage income from the
13
BEA; and wage income and migration flows from the IRS.
To control for potentially confounding changes in trade patterns and other technological
changes, we rely on data on exposure to Chinese imports from Autor, Dorn and Hanson (2013),
and data on the fraction of employment in a commuting zone in routine occupations (as defined
in Autor and Dorn, 2013). To distinguish the effects of robots from the effects of capital ac-
cumulation, investments in IT and other technologies raising productivity, we construct Bartik
measures of increases in capital stock, IT capital and value added across the 19 IFR industries.
Finally, we use data compiled by Leigh and Kraft (2018), who scraped the web to obtain the
location and employment of robot integrators—companies that install, program and maintain
robots for different industrial applications. Using these data, we construct estimates of robot
integrator activity in each commuting zone.
4 Industry Correlations
We start by documenting industry trends. Figure 2 depicts the relationship between AP Ri,(1993,2007)
(computed from EU RO5) and AP Ri,(2004,2007) (computed from the US data and scaled to a 14-
year equivalent change). Both variables are expressed in terms of robots per thousand workers.
Consistent with the notion that US industry trends in robotics are driven by technological im-
provements, there is a strong correlation between adoption of robots in the EU RO5 countries
and in the United States (see also Table A2 in the Appendix). The figure also reveals significant
heterogeneity across industries. While some industries, such as automotive, plastics and chem-
icals and metal products, exhibit increases in robot penetration of more than 7.5 robots per
thousand workers, others like paper and printing, textiles and wood and furniture experienced
modest increases both in Europe and the US.
In the rest of this section, we focus on the variation in AP Ri , which we interpret as a proxy
for improvements in robotics technologies available to US firms. Table A3 in the Appendix
documents that improvements in robotics do not mimic other industry-level trends. Industries
that are adopting more robots are not the same industries affected by Chinese or Mexican import
competition or offshoring, nor are they the industries experiencing rapid growth in total capital
or IT capital. Nor are they the ones with a high fraction of routine jobs. Within manufacturing,
the correlation between our measure of adjusted penetration of robots, AP Ri , and the change in
imports from China is −0.39 (the overall correlation is 0.15). The correlation of AP Ri with the
share of routine tasks is −0.24 within manufacturing and −0.01 overall. The correlations with the
change in imports from Mexico, with task offshorability and with offshoring of intermediates are,
respectively, −0.03, −0.41 and −0.17 within manufacturing (and 0.31, −0.26 and 0.19 overall).
The correlation with the increase in capital is 0.22 within manufacturing (and −0.37 overall),
and the correlation with the increase in IT capital is 0.23 within manufacturing (and −0.17
overall). This strengthens our presumption that the use of industrial robots is a technological
phenomenon that is largely unrelated to other industry trends.14
14
This interpretation is bolstered by the close association between AP Ri and Graetz and Michaels’s (2018)
14
Our model shows that under plausible conditions, industries adopting robots reduce their
labor demand. Table 1 reports regressions of various industry-level measures of labor demand on
AP Ri for different time periods. Panel A focuses on the wage bill and Panel B looks at employ-
ment. The first four columns present long-differences specifications where we regress the change
in log wage bill from 1993 to 2007 on our baseline measure of adjusted robot penetration for the
same period, AP Ri,(1993,2007) . In column 1 of Panel A we show the relationship between AP Ri
and log wage bill, which is negative, indicating that industries experiencing greater penetration
of robots have also seen significant (relative) declines in labor demand. To control for other
industry trends over this time period, column 2 includes the change in imports from China and
dummies for manufacturing and light manufacturing—the latter consists of the textile industry
and the paper, publishing and printing industry. These two light manufacturing industries have
been on a steep downward trend for reasons unrelated to robots (mostly due to offshoring and
the rise of digital media). Controlling for the light manufacturing dummy ensures that the
estimates in column 2 are not driven by the comparison of these declining industries to other
manufacturing industries. Including these three controls reduces the magnitude of the coefficient
on AP Ri , but also makes it more precisely estimated (-0.923, standard error = 0.419). This
estimate implies that an increase of one robot per thousand workers in our AP Ri measure is
associated with a 0.92 percent relative decline in the wage bill. This implies that the average
increase in the stock of robots in manufacturing, 7 robots per thousand workers, is associated
with a 6.3 percent decline in the wage bill. Columns 3 and 4 show similar patterns for the wage
bill of all workers and production workers within manufacturing using the NBER-CES dataset.
Columns 5-9 present stacked-differences models for two subperiods of seven years, 1993-2000
and 2000-2007, with analogues of our AP Ri variable computed for each subperiod (in this case we
have two observations per industry). These models are appealing because they focus on within-
industry changes and exploit the timing of robot adoption. For instance, robot penetration
in the automotive industry accelerated in the 2000s, whereas it decelerated in shipbuilding
and aerospace during the 2000s. We now see a more precisely estimated relationship than the
one shown in columns 1-4. For example, the equivalent of the estimate in column 2 is -1.096
(standard error = 0.235), which implies that one more robot per thousand workers (in AP Ri )
is associated with over a 1 percent decline in labor demand. Stacked-differences models also
enable us to include linear industry trends, thus controlling more flexibly for the possibility that
industries have been on differential trends for other reasons (and in particular controlling for
declining industries). Although specifications controlling for industry trends are demanding, in
column 7 we estimate a similar negative relationship between robot adoption and labor demand.
Finally, columns 8 and 9 show similar patterns for the wage bill of all workers and production
workers within manufacturing using the NBER-CES dataset. Panel B shows analogous results
for employment and Figure 3 illustrates the relationship between AP Ri and log wage bill and
log employment from column 8 visually.
replaceability index, which measures the fraction of occupations in an industry involving tasks that can be
automated using industrial robots. See Figure A1 in the Appendix.
15
In the Appendix we present a series of robustness checks for these industry correlations.
Figure A2 in the Appendix verifies that there are no significant pre-trends correlated with the
adjusted penetration of robots for log wage bill and log employment (for all workers and for
production workers). Tables A4 and A5 confirm that the patterns shown in Table 1 are similar
when we use different constructions for the AP Ri variable and when we focus on more recent time
periods. Finally, Table A6 shows that the results are also similar when, rather than including the
light manufacturing dummy, we directly control for industry value added or the factors affecting
value added trends. In particular, in Panel A of this table we control for the change in industry
value added between 1992 and 2007 (from the BEA input-output tables) and in Panel B we
instrument for the change in value added using intermediate imports in supplier industries. The
estimates are broadly similar to, but larger than, the estimates in Table 1, presumably because
controlling for value-added isolates the displacement effect. In Panel C we control for differences
in task offshorability, which is one of the factors leading to the rapid decline in production and
value added in the light manufacturing industries, while in Panel D we include a dummy for
industries adopting robots. The results are again similar.
We also use the BEA data to estimate the relationship between robots and industry labor
share and value added between 1992 and 2007. Column 10 in Panel A of Table 1 shows that,
consistent with robots increasing productivity, value added is increasing in industries adopting
more robots—even though employment is contracting.15 This result suggests that, as in our
theory, industries adopting robots are becoming not just more productive but also less labor
intensive; this is confirmed by our estimate in column 10 in Panel B, which shows a large decline
in the labor share. This estimate implies that one more robot per thousand workers is associated
with a 0.8 percentage point decline in the labor share between 1992 and 2007.
Although we view the industry correlations mostly as descriptive, they establish that indus-
tries where robotics technology has made greater advances have experienced expanding output
and declining labor demand, employment and labor share. We next turn to the implications of
robots for employment and wages in local labor markets.
5 The Effect of Robots across Committing Zones

In this section we describe our measure of exposure to robots and document its variation.
We then present reduced-form results for employment and wages, investigate their robustness
and explore the heterogeneous effects of robots across industries, occupations, gender and skill
groups. We present IV estimates and discuss their quantitative implications in the next section.
15
Within manufacturing, the industries that adopted the greatest number of robots (in the US and in EU RO5)—
automotive, plastics and chemicals and metal products—experienced the fastest growth in value added between
1992 and 2007, ranging between 2 percent and 4 percent per year. In contrast, light manufacturing industries—
textiles and paper and printing—did not adopt many robots and experienced absolute declines in value added.
In Table A7 in the Appendix, we also document the significant positive effect of robots on labor productivity,
which confirms one of the main findings of Graetz and Michaels (2018) from cross-industry, cross-country data.
Because of data availability, we focus on long differences for value added, labor productivity and labor share.
16
5.1 Exposure to Robots and Robot-Related Activities

We focus on the exposure measure defined in equation (18) and constructed from European data
on robot penetration by industry. We use this variable as an instrument to uncover the effects
of the spread of robots on US labor markets.
Panel A of Figure 4 depicts the geographic distribution of exposure to robots between 1993
and 2007. In many parts of the United States there is only a small increase of about 0.27-
0.67 robots per thousand workers. In others, including parts of Kentucky, Louisiana, Missouri,
Tennessee, Texas, Virginia and West Virginia, our measure of exposure ranges between 2 and 5
robots per thousand workers. More strikingly, in some parts of the Rust Belt and Texas, robot
penetration increases by 5-10 per thousand workers. Figure 2 highlighted that there is greater
penetration of robots in the automotive industry than in other sectors (both in the United States
and Europe). Panel B of Figure 4 verifies that even after this industry is left out, there is still
considerable geographic variation in exposure to robots.
Are commuting zones with a high exposure to robots adopting more industrial robots as our
model predicts? Though data on robot adoption at the commuting zone level are not available,
in Figure 5 we provide evidence of greater robot-related activities in exposed commuting zones
using the data on integrators from Leigh and Kraft (2018). The figure shows the residual plot of
a regression of log of one plus the number of integrators in a commuting zone against exposure to
robots (as in most figures that follow, we partial out the covariates from our main specification
in column 4 of Table 2, which we describe below). The dashed line corresponds to the regression
relationship after the top one percent of commuting zones with highest exposure to robots are
excluded.16 In both cases we see a positive association between exposure to robots and the
number of integrators in a commuting zone. Table A8 shows that this relationship is robust to
alternative specifications and to different ways of measuring robot integrator activity.
5.2 Reduced-form Results for Employment and Wages

Table A9 in the Appendix provides a first look at how commuting zones with high and low
exposure to robots differ in terms of their labor market characteristics. Columns 2-5 present the
mean for various outcomes and covariates by quartiles of exposure to robots, while columns 6
and 7 show the correlations between these variables and exposure to robots. Three patterns are
notable. First, only three covariates show significant differences between high and low exposure
commuting zones. These are the shares of manufacturing employment and light manufacturing
employment, the female share of manufacturing employment, and we control for these variables
in our base specification. Second, across commuting zones at different quartiles of exposure to
robots, there are only very small differences in our two main labor market variable: hourly wages
in 1990 and private employment to population ratio in 1990 (which focuses on salaried workers
in the private sector and thus excludes public employment and self-employment). Finally and
16
These are Alpena, MI; Defiance, OH; Detroit, MI; Houghton Lake, MI; Lansing, MI; Lorain, OH; Mount
Pleasant, MI; Saginaw, MI; Sault Ste. Marie, MI; Wilmington, DE.
17
most notably, from 1990 to 2007, more exposed commuting zones experienced more negative
labor market trends.
To explore these patterns in detail, we estimate reduced-form specifications similar to equa-
tion (13). We regress changes in our main labor market outcomes on exposure to robots. Our
identifying assumption is that there are no differential shocks or trends affecting labor markets
with greater exposure to robots (based on baseline industry composition and European adoption
trends) relative to those with less exposure. We discuss threats to the validity of this identifying
assumption in Section 5.4.
Table 2 presents results for a long-differences specification for 1990-2007, where we regress
changes in employment and wage measures between 1990 and 2007 on the exposure to robots
variable for the same period. We end our sample in 2007 to avoid the potentially confounding
effects of the Great Recession and present results for a longer time window as well as for more
recent periods in the Appendix.17 The table focuses on our main outcome variables: the (private)
employment to population ratio in Panel A and log hourly wages in Panel B.18 Our baseline
specifications are weighted by population in 1990 and report standard errors that are robust
against arbitrary heteroskedasticity and spatial correlation within US states in parentheses.
Column 1 presents a parsimonious specification that only includes Census division dummies
as covariates. In Panel A we see a strong negative relationship between exposure to robots and
employment changes in a commuting zone with a coefficient of -0.44 (standard error = 0.09).
This estimate implies that an increase of one robot per thousand workers in exposure to robots
is associated with a relative decline in the (private) employment to population ratio of 0.44
percentage points.19
In column 2 we control for demographic characteristics in 1990, specifically, log population;
the share of females in the population; the share of population above 65 years; the shares
of population with no college, some college, college or professional degrees, and masters and
doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians in the population.
Since our regression specification is in changes, these controls allow for differential trends by
baseline demographic characteristics. Their inclusion slightly reduces our estimate of the impact
of exposure to robots on employment to population ratio to -0.41.
In column 3 we control for the baseline shares of employment in manufacturing and light
manufacturing and the female share of manufacturing employment. These controls allow for
differential trends by the baseline industrial structure of a commuting zone and ensure that our
17
To match the time window over which we measure the adjusted penetration of robots, we rescale the outcomes
to a 14-year equivalent change. In particular, for each variable, we define long differences as (y2007 − y2000 ) + 0.7 ×
(y2000 − y1990 ).
18
Equation (13) has change in log employment on the left-hand side. We estimate this relationship in Table
A15 in the Appendix, but opt for the employment to population ratio as our baseline because it is the standard
specification in the literature.
19
A difference in exposure of one robot per thousand workers between 1993 and 2007 (from EU RO5) is approx-
imately the increase in US exposure to robots over the same time period. It also corresponds to the interquartile
range of this variable (between Pittsburgh, PA at the 75th percentile and Omaha, NE at the 25th percentile).
The difference in exposure between the 1st percentile (West Palm Beach, FL) and the 99th percentile (Detroit,
MI) is much larger—about 9 robots per thousand workers.
18
exposure variable does not proxy for other trends affecting manufacturing employment. These
controls have a minor effect on our coefficient of interest, which now stands at -0.43, and is more
precisely estimated with a standard error of 0.06.
In column 4 we control for other changes that have impacted labor market outcomes during
our period of analysis: imports from China between 1990 and 2007 and the decline of routine
occupations proxied by their baseline shares in employment (the coefficient estimates for these
controls are shown in Table A12). Consistent with the lack of correlation between these measures
and exposure to robots, shown in Table A9, these controls have no impact on our estimates.
The point estimate remains at -0.45 (standard error = 0.06).20
Figure 6 provides a residual regression plot for our specification from column 4 in Panel A,
with the regression estimate shown with the solid line. The figure highlights that there are several
commuting zones with very large exposure to robots. Column 5 estimates the specification in
column 4 after excluding the top one percent of commuting zones with the highest exposure
(which are the ones listed in footnote 16) and demonstrates that these high exposure commuting
zones are not driving our negative estimates. The coefficient estimate in Panel A increases to
-0.57 (standard error = 0.14) and is shown with the dashed line in Figure 6.
Finally, column 6 shows that the results are similar in unweighted regressions. In the same
specification as in column 4, we now estimate a coefficient of -0.51 (standard error = 0.12).
Panel B presents results for log hourly wages. Because wages are only available for employed
workers and our evidence in Panel A suggests that employment declines in more exposed com-
muting zones, we present estimates adjusted for changes in the composition of wage earners.
Specifically, we use the change in the average log wage between 1990 and 2007 for each of the 250
demographic cells in a commuting zone as our left-hand side variable.21 Because we now have
multiple observations for each commuting zone—one for each demographic group—we weight
each observation by the size of the demographic group in the commuting zone in 1990. The
estimates show that greater exposure to robots reduces wages. In column 4, when we control
for our baseline covariates, the coefficient estimate is -0.88 (standard error = 0.13). This implies
that a one robot increase in our exposure measure (per thousand workers) leads to 0.87 percent
lower hourly wages. Figure 6 provides a residual regression plot for this specification.
We next turn to stacked-differences models where we exploit variation over two periods: 1990-
2000 and 2000-2007 (the former converted to a seven-year equivalent change for consistency). In
this case our standard errors are robust against heteroskedasticity and within-state serial and
20
As shown in Table A9 and discussed above, the shares of employment in manufacturing and light manufactur-
ing and the female share of manufacturing employment differ between high and low exposure commuting zones.
Table A10 shows that these variables are significant predictors of exposure to robots, exposure to robots in the
automotive industry (which we use in Table 5 below) and exposure to robots when we simultaneously control for
all covariates on the right-hand side. We continue to estimate negative and significant effects on employment and
wages when we do not control for light manufacturing and/or the female share of manufacturing employment,
though the estimates for employment are about 30 percent smaller in specifications that do not control for light
manufacturing (see Table A11). This is presumably because, as discussed above, the decline in employment in
light manufacturing industries is negatively correlated with exposure to robots.
21
For each demographic group g in commuting zone c, we compute the long difference of average log hourly
wages ∆ ln Wcg as explained in footnote 17. We then regress ∆ ln Wcg on the exposure measure for commuting
zone c and control for a full set of cell fixed effects.
19
spatial correlation. Table 3 presents our findings; Panel A is for employment and Panel B for
log hourly wages. The first six columns have the same structure as Table 2. In both panels,
the estimates are more negative than before and remain precisely estimated. For example, in
column 4 of Panel A the coefficient of interest is -0.55 (standard error = 0.05), while in Panel B
the estimate for log hourly wages increases to -1.4 (standard error = 0.18). The bottom panels
of Figure 6 illustrate these stacked-differences estimates, separately marking observations from
the two periods and showing that the negative relationship is present in both periods.
The stacked-differences model focuses on the differential changes in exposure to robots be-
tween these two time periods and enables us to control for linear commuting zone trends. Al-
though this specification is demanding and exploits a different source of variation than long
differences, we estimate a similar negative impact of exposure to robots on both employment
and wages. In column 7, for example, the estimates for employment and wages are, respec-
tively, -0.5 (standard error = 0.08) and -1.6 (standard error = 0.27). These findings bolster our
confidence that exposed commuting zones are not simply on a differential trend unrelated to
advances in robotics technology.
We also obtain similar patterns for the entire period 1990-2014 as well as for 2000-2007 and
2000-2014 (see Table A13).
5.3 Other Labor Market Outcomes

In the Appendix, we investigate the effects of exposure to robots on a range of other labor
market outcomes. Table A14 shows robust negative effects on employment in manufacturing,
which is relevant since robots are mostly adopted in manufacturing and substitute directly for
production workers in this industry.
Tables A15 and A16 look at alternative measures of employment and wages. These include
the (private) employment to population ratio including self-employment; the total employment
to population ratio including public employment and self-employment; employment counts from
the CBP divided by population; log employment; log weekly wages; log yearly wages; and log
wage bill from the CBP. The results are broadly similar to our baseline estimates in both long-
differences and stacked-differences specifications for all of these measures.
Table A15 also explores the implications for nonemployment by looking at the participation
and unemployment margins. We estimate a positive impact of exposure to robots on the non-
participation and unemployment rates. Quantitatively, our estimates imply that about three
quarters of the additional nonemployed drop out of the labor force, while a quarter remain
unemployed. In line with the rise in non-participation, in Table A17 we also estimate increased
use of SSA retirement and disability benefits and other government transfers.
Table A18 explores the response of migration. Some of our estimates show a negative impact
on population and net migration (computed from the IRS data), though these effects are neither
consistent across specifications nor precisely estimated. Quantitatively, they imply that the
migration responses are about one fourth of the size of the employment responses.22 Consistent
22
For example, using our stacked-differences specification from column 4 of Table 3 and comparing this to the
20
with the spillovers on the nontradable sector we report in Section 5.7 and the lack of powerful
migration responses, Table A18 also documents a decline in house prices and rents in exposed
commuting zones.
Finally, we use data from the BEA and the IRS to estimate the effects of robots on wage and
non-wage income separately. Our estimates in Table A19 show precise and large negative effects
on wage income, and no significant impact on non-wage income. This last result is consistent
with the notion that owners of robot integrators and firms introducing robotics technology are
not necessarily located in exposed commuting zones.
5.4 Threats to Validity

There are two main threats to the identifying assumption behind our estimates. First, the
industries that have been adopting more robots over the last two decades (in the United States
and Europe) could have been on a downward trend because of declining demand, international
competition, other technological changes or worsening labor relations. Second, the commuting
zones that house the industries adopting more robots may be affected by other negative shocks.
In either case, our estimates might confound the impact of robots with these pre-existing industry
and commuting zone trends. (A third possible threat, that exposure to robots measure is
correlated with other concurrent technological changes, is discussed in the next subsection).
Our analysis in Section 4, which demonstrates that the penetration of robots in EU RO5
is not correlated with industry pre-trends or with other major sources of changes in labor
demand, is reassuring for the first threat. Moreover, the fact that value added has expanded in
industries with the greatest penetration of robots suggests that our measure is not correlated
with negative demand shocks to industries. Regarding the second threat, our stacked-differences
analysis, which controlled for commuting zone trends, already established that linear commuting
zone trends do not explain our estimates. These results notwithstanding, we next investigate
these issues directly by checking for pre-trends (which could result from either concern) and by
controlling for other industry and ommuting zone trends.
Panel A of Table 4 shows that there are no significant pre-trends. Specifically, we estimate
the relationship between exposure to robots and the changes in the employment to population
ratio (columns 1-4) and in log hourly wages (columns 5-8) between 1970 and 1990. Our base
specification, in columns 2 and 6, shows that there is no quantitatively or statistically significant
association between exposure to robots and pre-1990 changes in employment or wages. The
picture is similar when we exclude highly-exposed commuting zones in columns 3 and 7, and
when we report unweighted specifications in columns 4 and 8. These results are summarized
in Figure 7, which presents residual plots for employment and wages from the specifications
in columns 2, 3, 6 and 7. In Table A20 in the Appendix we confirm that there are also no
pre-trends in other key labor market variables for which we have pre-1990 data—in particular,
equivalent specification in column 4 of Table A18 in the Appendix (Panel D), we see that an increase of one robot
per thousand workers in our exposure measure reduces total employment by 2 percent, of which 0.5 percent is
explained by the decline in population and 1.5 percent is explained by the decline in the employment to population
ratio.
21
manufacturing employment, the employment rate including self-employment and the public
sector, the non-participation rate, the unemployment rate and log weekly wages.
Panel B of Table 4 investigates potentially confounding industry trends. Specifically, we
present estimates that control for predicted employment declines based on a Bartik measure of
past industry trends, “exposure to industry trends”. To construct this measure, we use the 19
IFR industries and interact their national log employment decline between 1970 and 1990 with
their baseline employment share in 1970 in each commuting zone. Though the coefficient over
the exposure to industry trends variable is negative and significant in a few specifications, the
point estimates for exposure to robots are unaffected. These results support our interpretation
that the exposure to robots variable is not proxying for declining industries.
Finally, Panel C goes one step further and directly controls for the 1970-1990 change in
the employment to population ratio or hourly wages on the right-hand side of our baseline
specifications. This control has no effect on our parameter estimates from Table 2.
A different concern is that our estimates may have been driven or unduly influenced by the
automotive industry, which adopted more robots than any other sector between 1993 and 2007
and may be impacted by other economic trends.23
To address this concern, in Table 5 we decompose our measure of exposure to robots into
two parts; one exploiting the penetration of robots in the automotive industry and the other
exploiting the penetration of robots in all other industries. We include both of these measures
on the right-hand side of our employment and wage regressions. The table presents both long-
differences (columns 1-3) and stacked-differences (columns 4-6) specifications. Panel A is for
employment, while Panel B is for wages. In both panels we find that the effects of exposure
to robots in the automotive industry are similar to the effects of exposure to robots in other
sectors. In none of these models do we reject the hypothesis that the coefficients of these two
variables are equal. These results are reassuring for two distinct reasons. First, they indicate
that our results are not solely driven by the automotive industry. Second, they also suggest that
the effects of robots in different sectors are broadly similar.
5.5 Robots, Capital and Other Technologies

Our model demonstrates that capital deepening and technological changes that do not automate
tasks previously performed by labor do not generate a displacement effect and should thus have
very different impacts on labor demand. We now investigate the effects of capital deepening,
increases in IT capital and growth in value added on employment and wages. We want to
understand whether the effects of these variables differ from the effects of robots and also verify
that controlling for these trends does not change our main estimates.
23
Indeed, the share of employment in the automotive industry explains 67 percent of the cross-commuting zone
variation in exposure to robots, and Table A21 in the Appendix shows that the automotive industry has the
highest Rotemberg weight, which ranges from 50 to 90 percent in the specifications presented in Table 2 (see
Goldsmith-Pinkham, Sorkin and Swift, 2018). In our stacked-differences specifications in Table 3 this industry
also receives a large weight, but only during the 2000-2007 period when its robot penetration accelerated. These
high Rotemberg weights indicate that our reduced-form estimates may be sensitive to other shocks affecting local
labor markets specializing in the automotive industry during this period.
22
Table 6 presents the results from this exercise. We again report both long-differences and
stacked-differences specifications. The first four columns are for employment while the next four
are for wages. In Panel A we control for “exposure to capital”, which is a Bartik measure of
the increase in (log) industry capital stock. In Panels B and C we control for “exposure to IT
capital” and “exposure to industry value added” (which are Bartik measures of increases in log
industry IT capital and in log industry value added, respectively).
Including these variables has little effect on our estimates of the impact of robots. Moreover,
and in line with our theoretical emphasis that automation is conceptually different from capital
deepening and other types of technological changes that increase value added, these variables
are, if anything, positively correlated with changes in employment and wages (see Table A23 in
the Appendix for broadly similar results with other measures of computer technology). These
results confirm our expectation that, due to the displacement effect, industrial robots should have
a very different impact on labor demand than other (non-automation) technologies. They also
bolster the case that our estimates are not capturing the effects of other concurrent technological
changes (since these tend to have different impacts from robots and, as already shown in Table
A3 in the Appendix, are uncorrelated with exposure to robots).24
5.6 Other Robustness Checks

The Appendix reports a range of additional robustness checks. First, Tables A24 and A25 show
that the exact construction of exposure to robots does not affect our results. We report esti-
mates where this measure is computed from the average of all European countries and from the
average of EU RO5 plus Germany, and a specification where we use the 1990, rather than the
1970, employment distribution. In addition, we present estimates using an exposure measure
using raw penetration of robots (rather than our theoretically grounded measure based on ad-
justed penetration of robots) and from a specification where we weight each industry’s adjusted
penetration of robots by the average cost of robots in that industry. In all cases, the reduced-
form estimates using these alternative measures of exposure are negative and significant (and
the IV estimates reported in column 7 of Tables A24 and A25 are also similar to the estimates
in Table 7 presented in the next section).
Table A26 explores the role of outliers. Our results are robust when we exclude Detroit
(the commuting zone with the highest exposure to robots); when we exclude observations with
residuals above or below 1.95 standard deviations; when we estimate Li’s (1985) robust regression
which downweights influential observations; and when we estimate median regressions.
Table A27 shows that our results are also robust to controlling for a full set of state fixed
effects, to allowing for mean-reverting dynamics in employment and wages by including the
24
Another related threat to our IV strategy is that, as noted in footnote 3, international competition may affect
robot adoption decisions both in the United States and Europe. Relatedly, investments in robots in the countries
that are most advanced in robotics technology, Germany, Japan and South Korea, may increase international
competition for US industries. Table A22 in the Appendix investigates these issues and shows that our estimates
of the effects of robots are similar when we control for exposure to imports from Mexico, task offshoring and
exports from Germany, Japan and South Korea.
23
baseline value of the dependent variable, and to controlling for contemporaneous changes in all
of our baseline demographic variables. Table A28 reports the results from a two-step LASSO
specification with a large number of covariates (following Belloni et al., 2014), and establishes
that those included in our main specifications are very similar to those identified by the LASSO
procedure, and the two-step estimates are very close to our baseline estimates.
5.7 Effects by Industry, Occupation, Gender and Skill

This subsection investigates how exposure to robots has impacted employment in different in-
dustries and occupations as well as the employment and wages of different workers.
The top panel in Figure 8 presents estimates of the effects of exposure to robots on employ-
ment in different industries. We present point estimates and confidence intervals for two long-
differences specifications analogous to columns 4 and 5 of Table 2 and one stacked-differences
specification corresponding to column 4 of Table 3 (Figure A4 in the Appendix presents es-
timates from analogous unweighted specifications). Figure 8 shows that the effects of robots
concentrate in manufacturing (also shown in Table A14), and especially in heavily-robotized
industries, which include automotive, plastics and chemicals, metal products, basic metals, elec-
tronics and food and beverages. There are no significant effects on the remaining manufacturing
industries. Consistent with the indirect effects on nontradables discussed in Section 2.2, we find
negative impacts on construction and retail and personal services. The only two sectors that
show positive effects in some specifications are agriculture and education, health care and the
public sector (although these estimates are neither precise nor robust).
The bottom panel in Figure 8 presents analogous results for employment by occupation.
In line with our expectations, the negative employment effects of robots are mostly in routine
manual occupations, and in particular in blue collar occupations such as machinists, assemblers,
material handlers and welders. Workers in these occupations engage in tasks that are being au-
tomated by industrial robots, so it is natural for them to experience the bulk of the displacement
effect created by robots. We do not estimate positive employment effects in other occupations,
suggesting that, at least locally, the productivity gains from using industrial robots have not
resulted in an expansion of employment in non-automated tasks.
Figure 9 in the main text and Table A29 in the Appendix investigate the employment and
wage effects by gender. We estimate negative impacts for both men and women. The effects
are larger for men. For example, with our baseline specification in long differences, reported in
Figure 9 and in column 1 of Table A29, the impact of exposure to robots on the employment
to population ratio of men is -0.57, while for women it is -0.34. Table A29 further shows that
the decline for male employment is concentrated in manufacturing, while the decline in female
employment is more pronounced in nonmanufacturing.
Figure 9 summarizes the effects of robots on employment and wages for workers in different
education groups. We present estimates for all workers and estimates for men and for women
separately. We see negative employment and wage effects for both men and women with less than
high school, a high school degree, some college, a college or professional degree, and masters
24
or doctoral degrees. We find it surprising that there is no positive effect on workers with
postgraduate degrees. One interpretation is that this result reflects reduced demand for these
workers from the nontradable sector. A complementary explanation is that, in contrast to
other computer-assisted technologies, industrial robots are not directly complementing high-skill
workers. Figure 10 investigates the impact of exposure to robots on the wage distribution by
estimating quantile regressions (using our baseline specification from column 4 of Table 2). For
all workers, we estimate negative and significant effects below the 35th percentile of the wage
distribution. When we focus on workers with no college degree, the negative and significant
effects extend all the way up to the 85th percentile; while for workers with a college degree or
more, the negative effects concentrate below the 15th percentile. These results confirm that the
negative wage effects of robots are mostly at the bottom and the middle of the distribution.
6 IV Estimates and Local and Aggregate Implications

In this section, we report our IV estimates and explore their quantitative implications for local
and aggregate changes in employment and wages.
6.1 IV Estimates
We use the US exposure to robots to compute two-stage least squares (2SLS) estimates of β L
and β W in equation (13). Figure 2 already showed the close association between the industry-
level spread of robots in the United States (AP Ri,(2004,2007) ) and in Europe (AP Ri,(1993,2007) ).
Figure A3 in the Appendix depicts our first-stage relationship by plotting the US exposure to
robots at the commuting zone level computed using AP Ri,(2004,2007) against exposure to robots
computed using AP Ri,(1993,2007) .
Table 7 reports our IV estimates for the long-differences specifications analogous to those in
Table 2 and also reports the corresponding first-stage coefficients and F-statistics (Table A33
in the Appendix reports the corresponding OLS estimates). Because we use the IV estimates
to quantify the aggregate implications of robot adoption, we focus on the population-weighted
specifications. Columns 1-3 are for the (private) employment to population ratio and columns
4-6 are for log hourly wages. The first column in each block presents a parsimonious specification
that only control for Census division dummies. The second column in each block additionally
controls for demographic characteristics and industry composition of commuting zones. Finally,
the third column in each block adds controls for imports from China and the decline of routine
jobs.
Panel A covers our baseline period, 1990-2007, where the 2004-2007 US data are rescaled to
a 14-year equivalent change. Panel B also focuses on 1990-2007, but constructs the US exposure
to robots by imputing US industry data using the aggregate US change between 1993 and 2004
rather than rescaling the 2004-2007 industry data. Panel C is for 1990-2014, while Panels D and
E focus on 2000-2007 and 2000-2014, time windows that closely overlap with US robots data.
In all cases, the 2SLS estimates are negative and precise for both employment and wages. Our
25
base estimates in columns 3 and 6 in Panel A, which we use in our quantitative evaluation in
the next subsection, are -0.39 (standard error = 0.09) for employment and -0.77 (standard error
= 0.15) for log hourly wages.25
The estimates in Table 7 present standard errors that are robust against heteroskedasticity
and within-state spatial correlation. As pointed out by Borusyak et al. (2018), these standard
errors do not take into account potential correlations across commuting zones resulting from
other industry shocks. Table A30 reproduces our IV estimates with standard errors computed
following Borusyak et al.’s (2018) procedure to account for such correlation (Tables A31 and
A32 present these adjusted standard errors for our main long-differences and stacked-differences
reduced-form models). We do not find systematic differences between our baseline standard
errors and these potentially more conservative standard errors, presumably because our spec-
ifications already control for the most important industry shocks affecting US labor markets,
such as import competition from China, trends in overall manufacturing, light manufacturing
and the decline of routine jobs.
6.2 Magnitudes
The IV estimates in the previous subsection quantify the impact of one additional robot per
thousand workers on employment and wages in a commuting zone relative to other areas. Our
estimates in columns 3 and 6 in Panel A, for example, imply that the adoption of one additional
robot per thousand workers in a commuting zone reduces its employment to population ratio by
0.39 percentage points (roughly a 1 percent decline) and hourly wages by 0.77 percent relative
to other commuting zones. These numbers suggest that one more robot reduces employment by
about 6 workers in the affected commuting zone relative to others.26 These are sizable magni-
tudes, but not implausible since they include both the direct effects of robots on employment
and wages and the spillover effects on nontradables resulting from the decline in local demand
following the loss of employment and wage income in the area we saw in Section 5.7.27
The more challenging question is how much (and whether) employment and wages in the ag-
gregate decline in response to the adoption of industrial robots. As emphasized in Proposition 3,
when commuting zones interact through trade and capital markets, our local IV estimates do not
25
The IV estimate for employment is 40 percent smaller and the estimate for wages is 40 percent larger for
1990-2014 in Panel C. This might reflect the fact that as wages have continued to adjust in the affected commuting
zones, some of the initial employment response may have been reversed.
26
The increase of one more robot per thousand workers between 1993 and 2007 is equivalent to an increase
of 0.6 robots per thousand people or a total increase of 120,000 robots. Consequently, our estimates imply that
this increase led to a 0.39 percentage points lower (private) employment to population ratio, which is equivalent
to one robot reducing employment by 6 (≈ 0.0039/(0.6/1000)) workers. Equivalently, the increase of 120,000 in
the stock of robots during this period is predicted to have reduced employment by 756,000 jobs. We obtain a
reduction in employment of 720,000 jobs (or about 4 jobs per robot) if we use the estimate for 1990-2014 from
Panel C together with the larger increase of 180,000 in the stock of robots over this longer time period.
27
These magnitudes can be compared to the local effects from exposure to imports from China. Using the
stacked-differences estimates from Table A12 in the Appendix, which correspond to the specification used in
Autor, Dorn and Hanson (2013), the implied magnitude from the national rise in Chinese imports for a commuting
zone with the average exposure to Chinese imports is a decline of about 1 percentage point in the employment to
population ratio—2.5 times the 0.39 percentage points decline due to the industrial robots.
26
directly translate into aggregate effects because robot adoption in one commuting zone reduces
the costs of goods consumed in other areas and generates capital gains shared by households in
those areas. To explore these aggregate implications, we need to make further assumptions on
cross-commuting zone spillovers (and this suggests greater caution in interpreting these aggre-
gate estimates than the local effects discussed in the previous paragraph).
We first assume that Proposition 3 provides a reliable approximation to these cross-commuting
zone interactions and use our regression evidence and external information to discipline the key
parameters of the model. Specifically, we use equations (9) and (10) in Proposition 3, which
provide expressions for βL and βW in terms of the underlying parameters of the model as well
as the labor share. We then use information on the labor share and the parameters σ, λ, α, π0
and γM /γL to solve for the values of the inverse of the Frish elasticity of labor supply, ε, and
the inverse of the elasticity of the robot supply, η, that are consistent with our IV estimates, βbL
and βbW . With these parameter estimates we compute the aggregate implications of robots from
our model. Proposition A7 in the Appendix provides formulas linking the aggregate effects of
robots on employment and wages to the parameters η and ε.
Our parameter choices are as follows: (1) σ = 1 as the elasticity of substitution between
different industries (see Oberfield and Raval, 2014); (2) λ = 5 for the elasticity of substitution
between traded varieties, which follows the trade literature (e.g., Simonovska and Waugh, 2014;
and Head and Meyer, 2015); (3) sL = 0.9916 as the baseline share of labor in task production,
which is implied by the number of robots in US industries in 1993;28 (4) α = 0.67, which together
with the estimate for sL yields an initial labor share of approximately 2/3 in all commuting
zones; (5) φ = 0.25, which matches the 18 percent share of employment in the tradable or
manufacturing sector; (6) π0 = 0.3 which, in line with the evidence surveyed in BCG (2015),
implies that robot adoption reduces costs by about 30 percent; (7) γM /γL = 3, which implies
that in automated tasks a robot performs, on average, the work of three workers;29 and (8)
ψ = 0.02, which is consistent with a marginal propensity to consume leisure of 10 percent (see
Imbens, Rubin and Sacerdote, 2001).
Given these parameter values, equations (9) and (10) yield η = 0.79 and ε = 0.17. The
estimate for η implies a fairly inelastic supply of robots to the local economy, which limits the
productivity gains from robot adoption and instead generates greater rents for robot integrators
and producers. If we suppose that local robot services are provided by combining (elastically
supplied) robotics equipment and (mostly inelastically) supplied services of robot integrators,
28
In Propositions 2 and 3 we simplified the exposition by focusing on the case where θ0 = 0, which implied a
share of labor in task production equal to sL = 1. Our more general expressions in Proposition A2 and A6 clarify
the role of sL .
29
This number is consistent with both the conclusions of the early engineering studies on the capabili-
ties of industrial robots (see Groover et al., 1986) and more recent estimates. For example, Ford (2015,
p. 2) reports that robots can move six times as many boxes as humans in the same time period. Robots
also appear to be three times as productive in welding as humans (https://www.sciencechannel.com/tv-
shows/worlds-biggest-shipbuilders/?utm source=scigo&utm medium=redirect), six times as productive in plac-
ing bricks (https://www.cnbc.com/2018/02/17/construction-robotics-bricklaying-robot-five-times-faster-than-
human.html), and nine times as productive as humans in smart phone assembly (https://futurism.com/2-
production-soars-for-chinese-factory-who-replaced-90-of-employees-with-robots).
27
this estimate is equivalent to the share of the inelastic component of integrate services, η/(1+η),
being approximately 0.44. Since the share of local robot integrators in total costs is about
0.75 (Leigh and Kraft, 2018), this number implies that about two thirds of their services are
inelastically supplied to the local economy. The estimate for ε, on the other hand, implies an
elastic response of labor supply. This estimate is in line with the “macro” Frisch elasticities that
are consistent with the observed relationship between employment and wages in the aggregate
(see Table 1 in Chetty et al., 2011).
Using these parameter estimates we compute the aggregate effects of robots. One more robot
per thousand workers is predicted to reduce aggregate wages by 0.42 percent and the aggregate
employment to population ratio by 0.2 percentage points (or 400,000 jobs); equivalently, one more
robot reduces employment by 3.3 workers.30 With these parameter values, about two thirds of
the decline in the demand for labor in an exposed commuting zone is driven by the contraction
of the nontradable sector. This estimate is consistent with the magnitudes of the decline in
employment in nontradables such as construction, retail and personal services documented in
Section 5.7 (where manufacturing accounts for 0.16 of the 0.45 decline in the employment to
population ratio in response to one more robot per thousand workers, with the rest of the
decline accounted for by nontradables). Our model also implies a 0.33 percent increase in the
productivity of the tradable sector, a sizable capital gain of 1.87 percent and a 136 percent
increase in industrial robot utilization. This last estimate closely matches the 139 percent
increase in the stock of robots observed during this period.
Table A34 in the Appendix considers variations in the values of the key parameters, σ, λ,
π0 , γM /γL and ψ, and shows that both the implied values of η and ε and the resulting aggregate
effects are not sensitive to reasonable variations.
7 Concluding Remarks
The spread of robots, artificial intelligence and other automation technologies has raised concerns
about the future of jobs and wages. Nevertheless, there has been relatively little work on the
equilibrium effects of new automation technologies, and in particular of robots. In this paper, we
investigate the effects of industrial robots on US local labor markets. Robots, and automation
technologies more generally, displace workers from tasks they were previously performing and
should thus have very different labor market effects than overall capital deepening and other
types of technological changes (such as factor-augmenting ones). This is what we find in our
empirical work.
We focus on the variation in robot adoption originating from the technological frontier,
proxied by trends in other economies that are more advanced than the US in robotics technology
30
The aggregate effects are broadly similar if we use the IV estimate for 1990-2014 from Panel C of Table 7. In
this case, the same procedure leads to η = 1 and ε = 0.39, which imply that one additional robot per thousand
workers reduces employment by 0.15 percentage points and hourly wages by 0.67 percent. Then the increase of
180,000 robots during this period is estimated to reduce aggregate employment by 420,000 jobs and hourly wages
by 1 percent.
28
(which are ahead of the US partly because their demographic trajectories have generated greater
demand for robots; see Acemoglu and Restrepo, 2018c). This strategy enables us to purge
confounding changes across US industries from exposure to robots. Using this methodology and
approximating local labor markets with commuting zones, we estimate large and robust negative
effects of robots on employment and wages. We show that commuting zones most affected by
robots in the post-1990 era were on similar trends to other labor markets before 1990, and
that the impact of robots is distinct from and only weakly correlated with the prevalence of
routine jobs, the effects of imports from China, imports from Mexico, offshoring, IT capital and
capital deepening. Moreover, consistent with our theoretical emphasis, robots are estimated to
have very different effects from IT technologies and overall capital deepening. Our estimates
imply that each additional robot per thousand workers reduces local employment to population
ratio by 0.39 percentage points and wages by about 0.77 percent. Because adopting robots
creates benefits for other commuting zones resulting via trade linkages, the implied aggregate
effects are somewhat smaller—one additional robot per thousand workers reduces the aggregate
employment to population ratio by 0.2 percentage points and aggregate wages by 0.42 percent.
There are relatively few robots in the US economy, so the number of jobs lost due to robots
has been limited thus far (a 0.2 percentage point decline in the employment to population ratio
or about 400,000 jobs). However, if robotics technology proceeds as expected by experts over the
next two decades (e.g., Brynjolfsson and McAfee, 2014, especially pp. 27-32, and Ford, 2015),
the future aggregate implications of robots could be larger. For example, BCG (2015) offers two
scenarios for the next decade. In their aggressive scenario, the world stock of robots will quadru-
ple by 2025. This corresponds to 5.25 more robots per thousand workers in the United States,
and with our estimates would lead to a 1 percentage point lower employment to population ratio
and 2 percentage points lower wage growth between 2015 and 2025. Their more conservative
scenario involves a less than threefold increase in the stock of robots and would correspondingly
have a more modest impact (a reduction of 0.6 percentage point decline in the employment
to population ratio and 1 percent lower wage growth). Crucially, however, any extrapolation
about the future effects of robots should acknowledge not only the usual uncertainty associated
with such exercises but also the possibility that some of the general equilibrium effects working
through technology might emerge only slowly (Acemoglu and Restrepo, 2018a) and that the
response of employment and wages may be different once robots become sufficiently widespread.
We view our paper as a first step in exploring the labor market implications of different types
of technologies. Our conceptual framework highlights that, in contrast to the prevailing pre-
sumption in economic discussions, automation and non-automation technologies have distinct
impacts and different waves of automation technologies may have different consequences de-
pending on the balance between displacement and productivity effects (Acemoglu and Restrepo,
2018d, 2019). The next decade is likely to witness major advances in AI, machine learning, com-
munication technologies and new manufacturing technologies, including augmented reality and
modular design. Whether these technologies will increase labor demand, employment and wages
is an open and important question that needs to be investigated using a number of approaches.
29
References
Acemoglu, Daron, David Autor and David Lyle (2004) “Women, War, and Wages:
The Effect of Female Labor Supply on the Wage Structure at Midcentury,” Journal of Political
Economy, 112(3): 497–551.
Acemoglu, Daron and David Autor (2011) “Skills, tasks and technologies: Implications
for employment and earnings,” Handbook of Labor Economics, 4: 1043–1171.
Acemoglu, Daron , Philippe Aghion, Claire Lelarge, John Van Reenen and Fab-
rizio Zilibotti (2007) “Technology, Information and the Decentralization of the Firm,” The
Quarterly Journal of Economics 122(4): 1759–1799.
Acemoglu, Daron and Pascual Restrepo (2018a) “The Race Between Machine and
Man: Implications of Technology for Growth, Factor Shares and Employment” American Eco-
nomic Review, 108(6): 1488–1542.
Acemoglu, Daron and Pascual Restrepo (2018b) “Modeling Automation” NBER WP
No. 24321.
Acemoglu, Daron and Pascual Restrepo (2018c) “Demographics and Automation”
NBER WP No. 24421.
Acemoglu, Daron and Pascual Restrepo (2018d) “Artificial Intelligence, Automation
and Work” NBER Working Paper No. 24196.
Acemoglu, Daron and Pascual Restrepo (2019) “Automation and New Tasks: How
Technology Displaces and Reinstates Labor” Journal of Economic Perspectives, 33(2): XXX-
XXX.
Anderson, James (1979) “A Theoretical Foundation for the Gravity Equation,” Anerican
Economic Review, 69(1): 106–160.
Arntz, Melanie, Terry Gregory, and Ulrich Zierahn (2016) “The Risk of Automation
for Jobs in OECD Countries,” OECD Social, Employment and Migration Papers No. 189.
Armington, Paul S. (1969) “A Theory of Demand for Products Distinguished by Place
of Production,” Staff Papers International Monetary Fund, 16(1): 159–178.
Autor, David H. and David Dorn (2013) “The Growth of Low-Skill Service Jobs and
the Polarization of the U.S. Labor Market,” American Economic Review, 103(5): 1553–97.
Autor, David H., David Dorn, and Gordon H. Hanson (2013) “The China Syn-
drome: Local Labor Market Effects of Import Competition in the United States.” American
Economic Review 103(6): 2121–68
Autor, David H., David Dorn, & Gordon H. Hanson (2015) “Untangling Trade and
Technology: Evidence from Local Labor Markets,” Economic Journal, 125(584): 621-646.
Autor, David H., Frank Levy and Richard J. Murnane (2003) “The Skill Content of
Recent Technological Change: An Empirical Exploration,” The Quarterly Journal of Economics,
118(4): 1279–1333.
Bartel, Ann, Casey Ichniowski, and Kathryn Shaw “How Does Information Technol-
ogy Affect Productivity? Plant-Level Comparisons of Product Innovation, Process Improvement,
30
and Worker Skills,” The Quarterly Journal of Economics, 122(4): 1721–1758.

Bartik, Timothy (1991) Who Benefits from State and Local Economic Development Poli-
cies? W.E. Upjohn Institute.
Beaudry, Paul , Mark Doms, and Ethan Lewis (2006) “Endogenous Skill Bias in
Technology Adoption: City-Level Evidence from the IT Revolution,” NBER WP No. 12521.
Belloni, Alexandre, Victor Chernozhukov, and Christian Hansen (2014) “High-
Dimensional Methods and Inference on Structural and Treatment Effects,” Journal of Economic
Perspectives, 28(2): 29–50.
Bloom, Nicholas, Mirko Draca, and John Van Reenen (2016) “Trade Induced Tech-
nical Change? The Impact of Chinese Imports on Innovation, IT and Productivity,” The Review
of Economic Studies, 83(1): 87–117.
Boston Consulting Group (2015) “The Robotics Revolution: The Next Great Leap in
Manufacturing.”
Brynjolfsson, Erik and Andrew McAfee (2014) The Second Machine Age: Work,
Progress, and Prosperity in a Time of Brilliant Technologies, W. W. Norton & Company.
Caroli, Eve, and John Van Reenen (2001) “Skill Biased Organizational Change? Ev-
idence from a Panel of British and French Establishments.” Quarterly Journal of Economics,
116(4): 1449–92.
Chetty, Raj, Adam Guren, Day Manoli, and Andrea Weber (2011) “Are Micro
and Macro Labor Supply Elasticities Consistent? A Review of Evidence on the Intensive and
Extensive Margins.” American Economic Review 101(3): 471–475.
Curtis, Charlotte (1983) “Machines vs Workers,” The New York Times—February 8,
Page 8, Print version.
Doms, Mark and Ethan Lewis (2006) “Labor Supply and Personal Computer Adop-
tion,” Federal Reserve Bank of San Francisco Working Paper 2006-18.
Goos, Maarten, and Alan Manning (2007) “Lousy and Lovely Jobs: The Rising Po-
larization of Work in Britain,” The Review of Economics and Statistics, 89(1): 118-133.
Gordon, Robert (2016) The Rise and Fall of American Growth, Princeton University
Press, Princeton New Jersey.
Graetz, Georg and Guy Michaels (2018) “Robots at Work,” Review of Economics and
Statistics, forthcoming.
Feenstra, Robert C., and Gordon H. Hanson (1999) “The Impact of Outsourcing
and High-Technology Capital on Wages: Estimates for the United States, 1979?1990.” The
Quarterly Journal of Economics 114(3): 907–940.
Ford, Martin (2015) The Rise of the Robots, Basic Books, New York.
Frey, Carl B. and Michael A. Osborne (2013) “The Future of Employment: How
Susceptible are Jobs to Computerisation?” Mimeo. Oxford Martin School.
Goldsmith-Pinkham, Paul, Isaac Sorkin and Henry Swift (2018) “Bartik Instru-
ments: What, When, Why, and How,” NBER WP No. 24408.
Groover, Mikell, Mitchell Weiss, Roger N. Nagel, and Nicholas G. Odrey (1986)
31
Industrial Robotics: Technology, Programming and Applications, Mcgraw-Hill Inc.

Gregory, Terry, Anna Salomons, and Ulrich Zierahn (2016) “Racing With or Against
the Machine? Evidence from Europe,” ZEW - Centre for European Economic Research Discus-
sion Paper No. 16-053
Head, Keith and Thierry Mayer (2014) “Gravity Equations: Workhorse, Toolkit, and
Cookbook.” Handbook of International Economics 4: 131.
Imbens, Guido W., Donald B. Rubin, and Bruce I. Sacerdote (2001) “Estimating
the Effect of Unearned Income on Labor Earnings, Savings, and Consumption: Evidence from
a Survey of Lottery Players,” American Economic Review 91(4): 778–794.
International Federation of Robotics (2014) Wold Robotics: Industrial Robots.
Jägger, Kirsten (2016) “EU KLEMS Growth and Productivity Accounts 2016 release -
Description of Methodology and General Notes.”
Katz, Lawrence F., and Kevin M. Murphy (1992) “Changes in Relative Wages, 1963-
1987: Supply and Demand Factors.” The Quarterly Journal of Economics, 107(1): 35–78.
Keynes, John Maynard (1930) “Economic Possibilities for our Grandchildren,” Chapter
in Essays in Persuasion.
Leigh, Nancey Green, and Benjamin R. Kraft (2018) “Emerging Robotic Regions in
the United States: Insights for Regional Economic Evolution.” Regional Studies 52(6): 804–815.
Li, Guoying (1985) “Robust Regression,” in Exploring Data Tables, Trends, and Shapes,
pp. 281–343.
McKinsey Global Institute (2017) “A Future that Works: Automation, Employment
and Productivity,” Online report.
Michaels, Guy, Ashwini Natraj and John Van Reenen (2014) “Has ICT Polarized
Skill Demand? Evidence from Eleven Countries over Twenty-Five Years,”Review of Economics
and Statistics, 96(1): 60–77.
Moretti, Enrico (2011) “Local Labor Markets,” Handbook of Labor Economics, Elsevier.
Oberfield, Ezra and Devesh Raval (2014) “Micro Data and Macro Technology,” NBER
WP No. 20452.
Ruggles, Steven, Matthew Sobek, Trent Alexander, Catherine A. Fitch, Ronald
Goeken, Patricia Kelly Hall, Miriam King, and Chad Ronnander (2010) “Integrated
Public Use Microdata Series: Version 3.0.” Minnesota Population Center.
Simonovska, Ina, and Michael E. Waugh (2014) “The elasticity of trade: Estimates
and evidence” Journal of international Economics 92(1): 34–50.
Tolbert, Charles M., and Molly Sizer (1996) “US Commuting Zones and Labor Market
Areas: A 1990 Update.” Economic Research Service Staff Paper 9614.
World Bank (2016) World Development Report 2016: Digital Dividends.
Wright, Greg C. (2014) “Revisiting the Employment Impact of Offshoring,” European
Economic Review 66:63–83.
Zeira, Joseph (1998) “Workers, Machines, and Economic Growth,” Quarterly Journal of
Economics, 113(4): 1091–1117.
32
Figure 1: Industrial robots in the United States and Europe.

Industrial robots per thousand workers in the United States and Europe.
Figure 2: Adjusted penetration of robots in the US and EU RO5 by industry.

Plot of the adjusted penetration of robots between 1993 and 2007 (AP Ri ) and the adjusted
penetration of robots in the United States between 2004 and 2007 (AP Ri rescaled to a 14-year
equivalent change). Adjusted penetration of robots is given in number of robots per thousand
workers in the industry. The solid line corresponds to the 45o line. Marker size indicates the
baseline US employment in the industry.
33
34
Figure 3: The relationship between robots and labor demand across industries.
The figure presents residual plots of the relationship between adjusted penetration of robots (AP Ri ) and the change in log wage bill (left
panel) and the change in log employment (right panel) from stacked-differences models, with data for 1993-2000 (in light gray) and 2000-2007
(in dark gray). The solid line shows the coefficient estimates from column 8 of Panel A (left panel) and column 8 of Panel B (right panel) of
Table 1. The covariates from these models are partialled out. The dashed line is for a regression which in addition excludes the automotive
industry. Marker size indicates the baseline US employment in the industry.
Figure 4: Geographic distribution of exposure to robots 1993-2007

The top panel shows the distribution of exposure to robots. The bottom panel shows the
distribution of exposure to robots outside of the automotive industry.
35
Figure 5: Exposure to robots and the location of robot integrators.

The figure presents the relationship between exposure to robots for 1993-2007 and the log of
one plus the number of robot integrators in a commuting zone. The covariates from column 4 of
Table 2 are partialled out. Data on the location of robot integrators are from Leigh and Kraft
(2018). The solid line corresponds to a regression with commuting zone population in 1990 as
weights. The dashed line is for a regression which in addition excludes the top one percent of
commuting zones with the highest exposure to robots. Marker size indicates the 1990 population
in the commuting zone.
36
37
Figure 6: The effects of robots on employment and wages.
The top panels of the figure present the long-differences relationship between exposure to robots and changes in the employment to population
ratio 1990-2007 (top left panel) and log hourly wages 1990-2007 (top right panel). The covariates from column 4 of Table 2 are partialled
out. The bottom panels of the figure present stacked-differences relationships between exposure to robots and changes in the employment
to population ratio (bottom left panel) and log hourly wages (bottom right panel) for 1990-2000 and 2000-2007. In the bottom panels,
the observations for 1990-2000 are shown in light gray and the observations for 2000-2007 are shown in dark gray, and the covariates from
column 4 of Table 3 are partialled out. In all panels, the solid line shows the coefficient estimate from a regression with commuting zone
population in 1990 as weights. The dashed line is for a regression which in addition excludes the top one percent of commuting zones with
the highest exposure to robots. Marker size indicates the 1990 population in the commuting zone.
Figure 7: Pre-trends in employment and wages.

The figure presents the relationship between exposure to robots for 1993-2007 and the change in
the employment to population ratio 1970-1990 (top panel) and the change in log hourly wages
1970-1990 (bottom panel). The covariates from column 2 of Table 4 are partialled out. The solid
line shows the coefficient estimate from a regression with commuting zone population in 1990
as weights. The dashed line is for a regression which in addition excludes the top one percent
of commuting zones with highest exposure to robots. Marker size indicates the 1970 population
in the commuting zone.
38
39
Figure 8: The effects of robots on industries and occupations.
The figure presents estimates of the effects of exposure to robots on changes in industry employment to population ratios (top panel) and
changes in occupation employment to population ratios (bottom panel). The capped lines provide 95% confidence intervals. The first
set of estimates are from long-differences specifications as in column 4 of Table 2. The second set of estimates are from long-differences
specifications as in column 5 of Table 2 (where we remove the top one percent of commuting zones with the highest exposure to robots).
The third set of estimates are from stacked-differences specifications as in column 4 of Table 3.
Figure 9: Effects of robots on employment and wages by education and gender.

The figure presents estimates of the effects of exposure to robots on changes in the employment
to population ratio (top panel) and changes in log hourly wages (bottom panel) for all workers
and for men and women with different education levels (less than high school; high school degree;
some college; college or professional degree; and masters or doctoral degree). The capped lines
provide 95% confidence intervals.
40
Figure 10: The effects of robots on the wage distribution.

The figure presents estimates of the effects of exposure to robots on quantiles of the wage
distribution for a specification equivalent to column 4 in Table 2 following the methodology of
Chetverikov, Larsen and Palmer (2016). Estimates for the 5th, 10th, . . . , and 95th quantiles
together with their 95% confidence intervals are shown. The different panels are for all workers;
for workers with no college degree and workers with a college degree or more separately; and for
men and women separately.
41
Table 1: Robots, labor demand, labor share and value added: industry-level results
Long
Long differences 1993-2007 Stacked differences 1993-2000 and 2000-2007 differences
1992-2007
CBP (all industries) NBER-CES (within manufacturing) CBP (all industries) NBER-CES (within manufacturing) BEA-IO
Production Production
All workers All workers All workers All workers All workers All workers All workers All workers
workers workers
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A. Change in log wage bill Value added
Adjusted penetration -2.718 -0.923 -0.816 -0.993 -2.510 -1.096 -1.492 -1.037 -1.150 0.128
of robots, AP Ri (0.732) (0.419) (0.378) (0.324) (0.673) (0.235) (0.481) (0.177) (0.205) (0.061)
Observations 19 19 13 13 38 38 38 26 26 19
R-squared 0.19 0.91 0.84 0.91 0.53 0.90 0.95 0.87 0.91 0.72
Panel B. Change in log employment Labor share
Adjusted penetration -1.967 -0.754 -0.831 -0.991 -1.904 -0.883 -1.325 -0.921 -1.016 -0.797
of robots, AP Ri (0.654) (0.347) (0.339) (0.261) (0.609) (0.129) (0.329) (0.114) (0.152) (0.281)
Observations 19 19 13 13 38 38 38 26 26 19
42
R-squared 0.12 0.90 0.87 0.92 0.30 0.86 0.93 0.89 0.93 0.37
Covariates:
Time period dummies X X X X X
Broad industry
X X X X X X X X
dummies
Chinese imports X X X X X X X X

Industry fixed effects X
The table presents estimates of the relationship between adjusted penetration of robots and wage bill, employment, value added and labor share across US industries. Columns
1-4 present long-differences estimates for changes in log wage bill for 1993-2007 (Panel A) and log employment for 1993-2007 (Panel B). Columns 5-9 present stacked-differences
estimates for changes in log wage bill for 1993-2000 and 2000-2007 (Panel A) and log employment for 1993-2000 and 2000-2007 (Panel B). Column 10 presents long-differences
estimates for changes in log value added for 1992-2007 (Panel A) and labor share for 1992-2007 (Panel B). Changes in log value added are annualized and given in percent change per
year. Changes in labor share are in percentage points. Data sources and time periods are reported at the top of the table and the set of covariates is reported at the bottom. Column
1 does not include any covariates, and column 5 only includes time period dummies. Columns 2-4, 6-9 and 10 control for dummies for manufacturing and light manufacturing
(paper and printing, and textiles) and exposure to Chinese imports by industry from Acemoglu et al. (2016). Column 7 includes a full set of industry fixed effects. The regressions
in columns 1-9 are weighted by baseline industry employment in 1993 and the regressions in column 10 are weighted by baseline value added by industry in 1992. Standard errors
that are robust against heteroskedasticity and serial correlation at the industry level are in parentheses.
Table 2: The effects of robots on employment and wages: long differences
Long differences 1990-2007

Excludes
zones with the
Weighted by population Unweighted
highest
exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in the employment to population ratio, 1990-2007
Exposure to robots -0.445 -0.414 -0.434 -0.448 -0.572 -0.516
(0.094) (0.076) (0.057) (0.059) (0.138) (0.118)
Observations 722 722 722 722 712 722
R-squared 0.27 0.46 0.66 0.67 0.66 0.62
Panel B. Change in log hourly wages, 1990-2007
(0.163) (0.126) (0.134) (0.132) (0.274) (0.205)
Observations 87100 87100 87100 87100 85776 87100
R-squared 0.32 0.33 0.33 0.33 0.33 0.08
Covariates:
Census divisions X X X X X X
Demographics X X X X X
Industry shares X X X X
Trade, routine jobs X X X
The table presents estimates of the effects of exposure to robots on employment and wages. Panel A presents long-differences
estimates for changes in the employment to population ratio, 1990-2007. Panel B presents long-differences estimates for
changes in log hourly wages, 1990-2007. The specifications in Panel B are estimated at the demographic cell × commuting
zone level, where demographic cells are defined by age, gender, education and race. Columns 1-5 present regressions
weighted by population in 1990. Column 5 presents results excluding the top one percent of commuting zones with the
highest exposure to robots. Column 6 presents unweighted regressions. The covariates included in each model are reported
at the bottom of the table. Column 1 only includes Census division dummies. Column 2 adds demographic characteristics
of commuting zones in 1990 (log population; the share of females; the share of population above 65 years; the shares
of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the
shares of Whites, Blacks, Hispanics and Asians). Column 3 adds the shares of employment in manufacturing and light
manufacturing, and the female share of manufacturing employment in 1990. Columns 4-6 add exposure to Chinese imports
and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation
within states are in parentheses.
43
Table 3: The effects of robots on employment and wages: stacked differences
Stacked differences 1990-2000 and 2000-2007

Commuting zone trends
Excludes
zones with the Weighted by
Weighted by population Unweighted Unweighted
highest population
exposure
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Change in the employment to population ratio, 1990-2000 and 2000-2007
Exposure to robots -0.625 -0.591 -0.525 -0.551 -0.702 -0.743 -0.508 -1.007
(0.092) (0.076) (0.059) (0.052) (0.150) (0.092) (0.079) (0.116)
Observations 1444 1444 1444 1444 1424 1444 1444 1444
R-squared 0.24 0.32 0.39 0.41 0.40 0.39 0.71 0.44
Panel B. Change in log hourly wages, 1990-2000 and 2000-2007
(0.211) (0.199) (0.191) (0.182) (0.551) (0.295) (0.271) (0.407)
44
Observations 183606 183606 183606 183606 180818 183606 183606 183606
R-squared 0.28 0.28 0.28 0.29 0.27 0.09 0.32 0.10
Covariates:
Time period dummies X X X X X X X X
Census divisions X X X X X X X X
Demographics X X X X X X X
Industry shares

X X X X X X
Trade, routine jobs X X X X X
Commuting zone trends X X
The table presents estimates of the effects of exposure to robots on employment and wages. Panel A presents stacked-differences estimates for changes in the employment to
population ratio, 1990-2000 and 2000-2007. Panel B presents stacked-differences estimates for changes in log hourly wages, 1990-2000 and 2000-2007. The specifications in Panel
B are estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Columns 1-5 and 7 present regressions
weighted by population in 1990. Column 5 presents results excluding the top one percent of commuting zones with the highest exposure to robots. Columns 6 and 8 present
unweighted regressions. The covariates included in each model are reported at the bottom of the table. Column 1 only includes Census division dummies and time period dummies.
Column 2 adds demographic characteristics of commuting zones (log population; the share of females; the share of population above 65 years; the shares of population with no
college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians). Column 3 adds the shares
of employment in manufacturing and light manufacturing, and the female share of manufacturing employment. Columns 4-6 add exposure to Chinese imports and the share
of employment in routine jobs. In addition, columns 7 and 8 include a full set of commuting zone fixed effects. Standard errors that are robust against heteroskedasticity and
correlation within states are in parentheses.
Table 4: The effects of robots on employment and wages: testing and controlling for pre-trends
Change in the employment to population ratio Change in log hourly wages

Excludes zones Excludes zones
Weighted by population with highest Unweighted Weighted by population with highest Unweighted
exposure exposure
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Past change in labor market outcomes, 1970-1990
Exposure to robots in -0.014 0.006 -0.158 -0.217 0.106 0.145 0.566 -0.097
subsequent period, 1993-2007 (0.077) (0.070) (0.174) (0.151) (0.242) (0.243) (0.453) (0.359)
Observations 722 722 712 722 59230 59230 58402 59230
R-squared 0.41 0.43 0.44 0.32 0.50 0.50 0.49 0.30
Panel B. Estimates controlling for exposure to industry trends between 1970 and 1990
(0.046) (0.049) (0.138) (0.100) (0.136) (0.135) (0.273) (0.226)
Exposure to industry trends -0.041 -0.041 -0.037 -0.063 -0.065 -0.067 -0.077 -0.115
1970-1990 (0.019) (0.018) (0.019) (0.019) (0.034) (0.034) (0.034) (0.034)
Observations 722 722 712 722 87100 87100 85776 87100
R-squared 0.67 0.67 0.66 0.63 0.33 0.33 0.33 0.08
Panel C. Estimates controlling for the change in the dependent variable between 1970 and 1990
45
(0.057) (0.060) (0.132) (0.123) (0.144) (0.142) (0.291) (0.211)
Observations 722 722 712 722 55988 55988 55182 55988
R-squared 0.66 0.67 0.66 0.62 0.41 0.41 0.40 0.16
Covariates:
Demographics and industry

X X X X X X X X
shares
Trade, routine jobs X X X X X X
The table presents estimates of the effects of exposure to robots on past employment and wages, as well as estimates on contemporary changes in employment and wages controlling
for past changes in these outcomes. Panel A presents estimates for changes in the employment to population ratio (columns 1-4) and log hourly wages (columns 5-8) between 1970
and 1990. For comparison with our main results, these outcomes are scaled to a 14-year equivalent change. Panel B presents long-differences estimates for changes in the employment
to population ratio (columns 1-4) and log hourly wages (columns 5-8) between 1990 and 2007 controlling for a measure of exposure to industries that were in decline between 1970
and 1990 (see the main text for details on this variable). Panel C presents long-differences estimates for changes in the employment to population ratio (columns 1-4) and in log
hourly wages (columns 5-8) between 1990 and 2007 controlling for the change in the dependent variable between 1970 and 1990. The specifications for log hourly wages are estimated
at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Columns 1-3 and 5-7 present regressions weighted by
population in 1990. Columns 3 and 7 present results excluding the top one percent of commuting zones with highest exposure to robots. Columns 4 and 8 present unweighted
regressions. The covariates included in each model are reported at the bottom of the table. Columns 1 and 5 include Census division dummies, demographic characteristics of
commuting zones (log population; the share of females; the share of population above 65 years; the shares of population with no college, some college, college and professional
degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, and the
female share of manufacturing employment. In Panel A, the baseline covariates are from 1970, and in Panels B and C they are from 1990. Columns 2-4 and 6-8 add exposure to
Chinese imports and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
Table 5: The effects of robots on employment and wages: the role of the auto-
motive industry
Long differences 1990-2007 Stacked differences 1990-2000 and 2000-2007

Weighted by population Unweighted Weighted by population Unweighted
(1) (2) (3) (4) (5) (6)
Panel A. Change in the employment to population ratio
Exposure to robots in -0.429 -0.459 -0.571 -0.620 -0.566 -0.771
automotive industry (0.078) (0.065) (0.153) (0.088) (0.054) (0.127)
other industries (0.210) (0.117) (0.133) (0.212) (0.181) (0.161)
Test for equality of
0.69 0.42 0.47 0.87 0.54 0.72
coefficients (p-value)
Observations 722 722 722 1444 1444 1444
R-squared 0.27 0.67 0.62 0.24 0.41 0.39
Panel B. Change in log hourly wages
automotive industry (0.105) (0.118) (0.225) (0.162) (0.144) (0.282)
other industries (0.515) (0.331) (0.377) (0.710) (0.741) (0.595)
Test for equality of
0.80 0.52 0.93 0.73 0.63 0.81
coefficients (p-value)
Observations 87100 87100 87100 183606 183606 183606
R-squared 0.32 0.33 0.08 0.28 0.29 0.09
Covariates:
Time period dummies X X X
Baseline covariates X X X X
The table presents estimates of the effects of exposure to robots separately for the automotive industry and other indus-
tries. The p-value for a test of equality between the coefficients of exposure to robots in the automotive industry and in
other industries is reported below the estimates. Columns 1-3 present long-differences estimates for the 1990-2007 period.
Columns 4-6 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Panel A presents results for the
employment to population ratio. Panel B presents results for log hourly wages. The specifications in Panel B are estimated
at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race.
Columns 1-2 and 4-5 present regressions weighted by population in 1990. Columns 3 and 6 present unweighted regressions.
The covariates in each model are reported at the bottom of the table. Columns 1 and 4 only includes Census division
dummies (and time period dummies in the stacked-differences specifications). Columns 2-3 and 5-6 add demographic char-
acteristics of commuting zones (log population; the share of females; the share of population above 65 years; the shares of
population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the shares
of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the female
share of manufacturing employment, exposure to Chinese imports and the share of employment in routine jobs. Standard
errors that are robust against heteroskedasticity and correlation within states are in parentheses.
46
Table 6: The effects of robots controlling for capital deepening, IT capital and value added

Long differences Stacked differences Long differences Stacked differences
Weighted Unweighted Weighted Unweighted Weighted Unweighted Weighted Unweighted
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Estimates controlling for exposure to capital
(0.056) (0.114) (0.057) (0.083) (0.137) (0.213) (0.154) (0.258)
0.025 0.027 0.065 0.111 0.116 0.137 0.200 0.221
Exposure to capital
(0.020) (0.021) (0.036) (0.028) (0.033) (0.031) (0.099) (0.073)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.62 0.41 0.41 0.33 0.08 0.29 0.09
Panel B. Estimates controlling for exposure to IT capital
(0.058) (0.116) (0.053) (0.092) (0.135) (0.199) (0.202) (0.298)
0.006 0.009 0.054 0.014 0.067 0.066 0.123 0.089
Exposure to IT capital
(0.009) (0.010) (0.032) (0.013) (0.017) (0.016) (0.050) (0.038)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.62 0.42 0.39 0.34 0.08 0.29 0.09
Panel C. Estimates controlling for exposure to industry value added
47
(0.061) (0.113) (0.051) (0.087) (0.133) (0.201) (0.183) (0.292)
Exposure to industry value 0.060 0.064 0.082 0.079 0.034 0.033 0.097 0.062
added (0.013) (0.015) (0.024) (0.018) (0.021) (0.017) (0.054) (0.025)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.69 0.64 0.42 0.41 0.33 0.08 0.29 0.09
Covariates:

Time period dummies X X X X
Census divisions and baseline
X X X X X X X X
covariates
The table presents estimates of the effects of exposure to robots, capital accumulation, IT capital and value-added on employment and wages. Columns 1-2 and 5-6 present
long-differences estimates for 1990-2007. Columns 3-4 and 7-8 present stacked-differences estimates for 1990-2000 and 2000-2007. Columns 1-4 present results for employment to
population ratio. Columns 5-8 present results for log hourly wages. The specifications in columns 5-8 for log hourly wages are estimated at the demographic cell × commuting
zone level, where demographic cells are defined by age, gender, education and race. Odd-numbered columns present regressions weighted by population in 1990. Even-numbered
columns present unweighted regressions. In Panel A we control for a measure of exposure to capital, constructed by interacting the increase in log capital in each industry with
its baseline employment share in the commuting zone. In Panel B we control for a measure of exposure to IT capital, constructed by interacting the increase in log IT capital in
each industry with its baseline employment share in the commuting zone. In Panel C we control for a measure of exposure to industry value added, constructed by interacting the
increase in log value added in each industry (between 1992 and 2007 in all models) with its baseline employment share in the commuting zone. All models include Census division
dummies (and time period dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the share of females; the share of
population above 65 years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites,
Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure to Chinese imports
and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
Table 7: The effects of robots on employment and wages: IV estimates
Change in Change in
employment to population ratio log hourly wages
(1) (2) (3) (4) (5) (6)
Panel A. Long-differences, 1990-2007
US exposure to robots -0.375 -0.377 -0.388 -1.022 -0.762 -0.768
(0.119) (0.088) (0.091) (0.225) (0.147) (0.148)
Observations 722 722 722 87100 87100 87100
First-stage coefficient 1.19 1.15 1.15 1.19 1.15 1.15
First-stage F-statistic 38.16 32.19 33.62 39.81 33.17 34.35
Panel B. Alternative imputation of US data, 1990-2007
(0.123) (0.092) (0.094) (0.234) (0.153) (0.153)
Observations 722 722 722 87100 87100 87100
Panel C. Long-differences, 1990-2014
(0.126) (0.077) (0.080) (0.169) (0.165) (0.170)
Observations 722 722 722 90341 90341 90341
Panel D. Long-differences, 2000-2007
(0.137) (0.067) (0.067) (0.227) (0.172) (0.176)
Observations 722 722 722 99319 99319 99319
Panel E. Long-differences, 2000-2014
(0.190) (0.065) (0.067) (0.156) (0.183) (0.191)
Observations 722 722 722 106375 106375 106375
Covariates:
Division dummies X X X X X X
Demographics and
X X X X
industry shares
Trade, routine jobs X X
The table presents IV estimates of the effects of exposure to robots on employment and wages for different time periods.
Panels A and B present results for 1990-2007. Panel C presents results for 1990-2014. Panel D presents results for 2000-2007.
Panel E presents results for 2000-2014. In all models, we instrument the US exposure to robots using exposure to robots
from EU RO5. In Panels A, C, D and E we rescale the US exposure to robots to match the time period used. In Panel B
we use an alternative imputation strategy for US exposure to robots described in the text. Columns 1-3 present results for
the employment to population ratio. Columns 4-6 present results for log hourly wages. The specifications for log hourly
wages are estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender,
education and race. All IV estimates are from regressions weighted by population in 1990. The covariates included in each
model are reported at the bottom of the table. Columns 1 and 4 only include Census division dummies. Columns 2 and 5
add demographic characteristics of commuting zones (log population; the share of females; the share of population above
65 years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral
degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light
manufacturing, and the female share of manufacturing employment. Columns 3 and 6 add exposure to Chinese imports
and the share of employment in routine jobs. We also report the first-stage coefficients and its F-statistic in all models.
Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
48
Online Appendix (Not for Publication)

In this part of the Appendix, we provide proofs and generalizations of results in the text.
A1 Autarky Equilibrium
We start with the characterization of the autarky equilibrium, then establish existence and
uniqueness, and prove Proposition 1 and a more general version of Proposition 2.
Characterization of Equilibrium
The first-order condition for the representative household in commuting zone c is
Wc = BCcψ Lεc . (A1)
Market clearing implies
Cc = Yc − Ic (A2)
= Yc − D−1−η (1 + η)−1−η Mc1+η ,
1
where the second line follows by inverting the production function for robots, Mc = D(1+η)Ic1+η ,
introduced in the text. Combining this with (A1), we obtain
ψ
Wc = B Yc − D−1−η (1 + η)−1−η Mc1+η Lεc . (A3)
From the production function for robots we also have
RcM = D−1−η (1 + η)−η Mcη . (A4)
Recall as well that in the autarky model the supply of capital in commuting zone c is taken to
be exogenously given at Kc > 0.
Under the assumption that πc > 0 in all commuting zones, tasks below θi will be produced
RcM Wc
with robots at a cost γM and tasks above θi will be produced with labor at a cost γL . Hence,
the marginal cost—and thus the price—of industry i is
K α
1 RM Wc 1−α
PciX = θi + (1 − θi ) RcK . (A5)
Aci γM γL
Next, define the share of labor in production tasks in industry i as
Wc Lci (1 − θi ) W
γL
c
sL
ci = = . (A6)
αPciX Xci K
RM
θ i γM + (1 − θi ) W c
γL
Here note that a fraction 1 − α of the total costs of the sector are paid to capital (given the
A-1
Cobb-Douglas technology in (2)), and sL

ci is the share of labor in the remainder, and thus the
share of labor in the value added of industry i is simply αsL
ci .
Because the final good in each commuting zone is taken as numeraire, we also have the
following ideal price index condition,
X 1−σ
1= νi PciX . (A7)
i∈I
Now combining (A5) and (A6), we can express the wage bill in commuting zone c as
X
Wc Lc = Wc Lci
i∈I
X
= αsL X
ci Pci Xci .
i∈I
−σ
From equation (1), the demand for industry i in commuting zone c is Xci = νi PciX Yc , and
substituting for this, the previous expression can be rewritten as
X 1−σ
W c Lc = αsL X
ci νi Pci Yc . (A8)
i∈I
Similarly, the demand for robots can be expressed as

X 1−σ
RcM Mc = α(1 − sL X
ci )νi Pci Yc , (A9)
i∈I
and the demand for capital is

RcK Kc = (1 − α)Yc . (A10)
Definition 1 An equilibrium of the autarky model is given by a set of factor prices {Wc , RcM , RcK },
factor supplies {Lc , Mc }, and level of output Yc for each c ∈ C such that:
• factor supplies satisfy equations (A3) and (A4);
• factor prices satisfy the ideal price index condition, equation (A7);
• factor markets clear, that is, equations (A8), (A9), and (A10) hold.
Proposition A1 An equilibrium of the autarky model exists and is unique.
Proof: The existence of equilibrium can be proved using a standard fixed point argument (as
in the proof of Proposition A5 below). Here we provide a proof exploiting the second welfare
theorem that establishes existence and uniqueness more directly.
Because the autarky equilibrium is a competitive equilibrium in an economy with a repre-
sentative household, from the second welfare theorem any autarky equilibrium is a solution to
A-2
the maximization of the utility of the representative household subject to the technology and
feasibility constraints. This problem can be written as
Cc1−ψ − 1 B
max − L1+ε
{Xci ,Lci ,Mci ,Kci }i∈I ,Lc ,Mc ,Kc ,Yc ,Cc ,Ic 1 − ψ 1+ε c
! σ
X 1 σ−1 σ−1
σ σ
subject to: Yc = νi Xci .
i∈I
α
−α −(1−α) γM Mci γL Lci 1−α
Xci = α (1 − α) Aci min , Kci
θi 1 − θi
1
Mc = D(1 + η)Ic1+η
X
Mc = Mci
i∈I
X
Lc = Lci
i∈I
X
Kc = Kci
i∈I
Yc = Ic + Cc .
The objective function is continuous and strictly concave and the constraint set is convex and
compact. This ensures that this maximization problem has a unique solution, which gives us
the unique equilibrium of the autarky model.
Proposition A1 can be generalized to the case where πc is negative in some commuting zones,
in which case not all technologically automated tasks will be produced with robots as in the
general framework considered in Acemoglu and Restrepo (2018a).
Proofs of Propositions from Section 2.1

Proof of Proposition 1: Since the labor share in industry i is αsL
ci and that the value added
1−σ
of this industry is νi PciX Yc , we have
1−σ
Wc Lci = sL X
ci ανi Pci Yc .
Using the formulas for sL X

ci and Pci in equations (A5) and (A6), we obtain
(1 − θi ) W
γL
c
1−σ
Wc Lci = K ανi PciX Yc
RM
θ i γM + (1 − θi ) W
γL
c
(1 − θi ) γWL 1−α
1−σ
= 1 RcK α
ανi PciX Yc
(Aci PciX ) α
1−α
(1 − θi ) γWL

Yc α 1−σ
= 1 (1 − α) ανi PciX Yc ,
(Aci PciX ) α Kc
A-3
Yc
where in the last line we used the fact that RcK = (1 − α) Kc
(equation (A10)). Simplifying this
expression yields
1−α
α(1 − α) α νi 1
1−σ− α 1 α−1
Lci = (1 − θi ) 1 PciX Ycα Kc α . (A11)
γL Aciα
Taking logs on both sides and differentiating yields (3).
In footnote 4, we provided a sufficient condition for (relative) labor demand in industry i to
1/α
decrease following automation, 1/α+σ−1 > πc sL
ic . To prove this claim, differentiate equation (A11)
and note that from equation (3) that d ln Lcj = d ln Yc for industries that are not undergoing
any automation. Thus, d ln Lci < d ln Lcj if

dθi 1
> σ + − 1 d ln PciX
1 − θi α

1 dθi
= σ + − 1 αsLic πc .
α 1 − θi
Rearranging this expression yields the desired condition.

We now state and prove a generalization of Proposition 2.
Proposition A2 Suppose that πc > 0 and θi = θ0 ≥ 0. Then

 
ψ X dθi γL
d ln Lc = −ζcdisp + ζcprod πc − ζcL,inc α(1−sL
· `ci , (A12)
1− c) 1 − θi γ M
1+η i∈I
 
ψ X dθi γL
d ln Wc = −ζcdisp ε + ζcprod επc + ζcW,inc α(1−sL
· `ci , (A13)
1− c) 1 − θi γ M
1+η i∈I
where ζcdisp = (1 − α + η)/Λc , ζcprod = sL L,inc

c (1 + η)/Λc , ζc = α(sL 2 W,inc
c ) πc /Λc , ζc = αsL
c (πc +
πc (η − α)(1 − sL L
c ) − (1 − sc πc )(1 − α + η))/Λc , and
 
γL  ψ
Λc = 1 − α + η(1 − sL L 2
c ) + α(sc ) α(1−s
L
L ) + εsc > 0.
γM 1− c
1+η
Proof: Differentiating (A2) and rearranging, we obtain
1 ιc
d ln Cc = d ln Yc − (1 + η)d ln Mc ,
1 − ιc 1 − ιc
where ιc = Ic /Yc is the share of robot investment in aggregate output of commuting zone c.
α(1−sL
c)
Using the definition of sL
c and equation (A4), this share is equal to ιc = 1+η . Taking logs
and differentiating equation (A1) and combining with the previous expression, we obtain
ψ
d ln Yc − α(1 − sL

d ln Wc = α(1−sL c )d ln Mc + εd ln Lc . (A14)
c)
1− 1+η
A-4
Equation (A4) then yields

d ln RcM = ηd ln Mc . (A15)
Differentiating the expression for PciX and rearranging gives
Wc RcM
γL − γM
d ln PciX =α(1 − sL M
ci )d ln Rc + αsL
ci d ln Wc + (1 − α)d ln RcK −α
PciX
dθi
=α(1 − sL M L K L
ci )d ln Rc + αsci d ln Wc + (1 − α)d ln Rc − απc sci
1 − θi
L `ci dθi
=α(1 − sL M
ci )d ln Rc + αsL K
ci d ln Wc + (1 − α)d ln Rc − απc sc .
χci 1 − θi
Next taking logs and differentiating equation (A7) and combining it with the previous ex-
pression, we obtain
X dθi
α(1 − sL M L K L
c )d ln Rc + αsc d ln Wc + (1 − α)d ln Rc = αsc πc `ci , (A16)
1 − θi
i∈I
where sL
c denotes the average labor share of production tasks in commuting zone c. When
θi = θ0 , we have sL L
ci = sc , and
(1 − θ0 ) W
γL
c
sL
c = .
RK
θ 0 γM
M
+ (1 − θ0 ) W
γL
c
Taking logs and differentiating equation (A8) and combining it with the previous expression, we
obtain
X dθi
d ln Wc + d ln Lc = − (1 − sL
c πc ) `ci + (1 − sL M
c )(d ln Wc − d ln Rc ) + d ln Yc (A17)
1 − θi
i∈I
X
+ (1 − σ) `ci d ln PciX .
i∈I
Note next that
X X
(1 − σ) `ci d ln PciX = (1 − σ) χci d ln PciX = 0,
i∈I i∈I
where recall that χci is the share of industry i in value added in commuting zone c, and the first
equality follows because αsL L
ci = αsc implies that `ci = χci , and the second equality follows from
the ideal price index condition in equation (A7). Using this expression, (A17) can be further
simplified to
X dθi
d ln Wc + d ln Lc = −(1 − sL
c πc ) `ci + (1 − sL M
c )(d ln Wc − d ln Rc ) + d ln Yc . (A18)
1 − θi
i∈I
Using similar steps, we obtain a simplified expression for the demand for robots in equation
A-5
(A9) as
sL
c
X dθi
d ln RcM + d ln Mc = L
(1 − sL
c π c ) `ci − sL M
c (d ln Wc − d ln Rc ) + d ln Yc . (A19)
1 − sc 1 − θi
i∈I
Finally, the demand for capital in equation (A10) implies
d ln RcK = d ln Yc . (A20)
Equations (A14), (A15), (A16), (A18), (A19) and (A20) define six linear equations in six
unknowns, and yield a unique solution. The solution gives the formulas for d ln Lc and d ln Wc
in the proposition.
A similar result holds when we relax the assumption that θi = θ0 ≥ 0, except that because
now sL
ci varies across industries, there will be additional residual terms in the expressions (A12)
and (A13).
Proof of Proposition 2: Our main result then follows immediately as a corollary of Propo-
sition A2 by setting θ0 = 0, which implies that sL
c = 1.
Extension: Workers and Capitalists

We now extend Proposition A2 to account for the possibility that non-labor income generated
by automation does not all accrue to workers and instead may go to “capitalists” who do not
supply labor. Let us thus modify the budget constraint of the household, which supplies all
labor to the commuting zone, to
CcL ≤ Wc Lc + ωc Πc ,
where ωc ∈ [0, 1] denotes the share of non-labor income owned by the household or by “workers”,
and CcL denotes their consumption of the final good. Capitalists consume the remaining resources
Yc − Ic − CcL , which ensures market clearing.
We obtain a similar set of equilibrium equations as before, but now, the labor supply in
equation (A14) becomes
ψ
(1 − (1 − ωc )δc ) d ln Yc − α(1 − sL

d ln Wc = α(1−sL c )d ln Mc + εd ln Lc
c)
1− 1+η
+ ψ(1 − ωc )δc (d ln Wc + d ln Lc ), (A21)
where δc denotes the share of wage income in workers total income. This equation shows that,
when ωc < 1, the income effects created by capital gains are dampened. With this modification,
we obtain the following generalization of Proposition A2:
A-6
Proposition A3 Suppose that πc > 0 and θi = θ0 . Then

h i X dθi γL
d ln Lc = −ζcdisp + ζcprod πc − ζcL,inc ψ + ζcL,ω ψ(1 − ωc )δc · `ci ,
1 − θi γ M
i∈I
h i X dθi γL
d ln Wc = −ζcdisp ε + ζcprod επc + ζcW,inc ψ − ζcW,ω ψ(1 − ωc )δc · `ci ,
1 − θ i γM
i∈I
where the ζ’s are the unique solution to the system of equations given by (A15), (A16), (A18),
(A19), (A20), and (A21).
Proof: The proof is analogous to that of Proposition A2 with equation (A21) replacing equation
(A14).
Extension: Services
In the model with trade between commuting zones, preferences are defined over a tradable
“industry” good and a nontradable “service” good as shown in equation (7). In this section, we
show that incorporating the same structure of preferences into the autarky model leads to very
similar expressions, but this exercise will provide a better benchmark for comparison with the
trade model.
Namely, we adopt the same preferences as in (7), and continue to assume that Cc = Yc − Ic
and Sc = LSc . This implies that the price of the service good is equal to the wage, Wc . As in
the trade model, now a fraction φ of income will be spent on Cc and a fraction 1 − φ of it on Sc
so that
1 1−φ
Sc = (Yc − Ic ).
Wc φ
We continue to take Yc as the numeraire, which implies that the consumer price index, which
incorporates the cost of the nontradable good PcC , is
PcC = (1 − φ)1−φ φφ Wc1−φ . (A22)
Using this expression, we obtain the optimal labor supply for the representative household as
(1−φ)ψ
(1 − φ)
Wcφ+(1−φ)ψ = (1 − φ) (1−φ) φ
φ B (Yc − Ic )ψ Lεc .
φ
Naturally, when φ = 1 we recover equation (A3). Taking logs, differentiating and rearranging
this expression, we obtain
ψ

(φ + (1 − φ)ψ)d ln Wc = α(1−sL c )d ln Mc + εd ln Lc . (A23)
c)
1− 1+η
A-7
Following similar steps to those in the proof of Proposition A2, we obtain

X 1−σ
Wc Lc = αsL X
ci νi Pci Yc + Wc LSc
i∈I
X 1−σ 1−φ
= αsL X
ci νi Pci Yc + (Yc − Ic ),
φ
i∈I
and therefore,
X dθi
d ln Wc + d ln Lc = − (1 − sL
c πc ) `ci + %(1 − sL M
c )(d ln Wc − d ln Rc )
1 − θi
i∈I
1−%

+ %d ln Yc + α(1−sL c )d ln Mc , (A24)
c)
1− 1+η
φα
where % = α+φα denotes the share of employment in the tradable sector, and `ci is now defined
as the share of total employment (including the nontradable sector) in industry i. Likewise, the
demand for robots can be obtained as
sL
c
X dθi
d ln RcM + d ln Mc = L
(1 − s L
c πc ) `ci − sL M
c (d ln Wc − d ln Rc ) + d ln Yc . (A25)
%(1 − sc ) 1 − θi
i∈I
Once again from the ideal price index condition, we have
1 L X dθi
α(1 − sL M L K
c )d ln Rc + αsc d ln Wc + (1 − α)d ln Rc = αsc πc `ci . (A26)
% 1 − θi
i∈I
The supply of robots (equation (A4)) and the demand for capital (equation (A10)) remain
unchanged. Combining these equations, we obtain a generalization of Proposition A2.
Proposition A4 Suppose that πc > 0 and θi = θ0 ≥ 0. Then

h i X dθi γL
d ln Lc = −ζcdisp φ + ζcprod φπc − ζc,L
inc
φψ · `ci ,
1 − θi γ M
i∈I
h i X dθi γL
d ln Wc = −ζcdisp ε + ζcprod επc + ζc,W
inc
ψ · `ci ,
1 − θ i γM
i∈I
where the ζ’s are given by the unique solution to the system of equations given by (A15), (A20),
(A23), (A24), (A25) and (A26).
When θ0 = 0 as assumed in the main text, we have ζcdisp = ζ disp = (1 + η − α)/Λ, ζcprod =
ζ prod = (1 + η + (1 − φ)(1 − α))/Λ, ζc,L
inc = (1 − φ + φα)π /Λ, ζ inc = (1 − φ + φα)(π − (1 −
c c,W c
γL
πc (1 − α + η))/Λ, and Λ = γM φ ((1 − α)φ + ψ(1 − φ + φα) + ε).
Proof: The proof of this proposition follows by noting that now equations (A15), (A20), (A23),
(A24), (A25) and (A26) can be uniquely solved for d ln Lc , d ln Wc , d ln Mc , d ln RcM , d ln Yc , and
d ln RcK , and yield the above expressions for d ln Lc and d ln Wc .
A-8
A2 Trade Equilibrium
We next study the equilibrium of the model where there is trade between commuting zones.
First recall that in this case there is an exogenously given supply of capital for the entire
economy, K. Next, we turn to the supply of labor. First, note that given the preferences in
(7), a fraction φ of income will be spent on the tradable good Cc and a fraction 1 − φ on the
nontradable good, which implies that
1
(1 − φ) Wc Lc + χΠ

Sc = cΠ ,
Wc
where χΠ
c is the share of capital gains owned by households in commuting zone c. Because we
took the tradable good Yc as the numeraire, the consumer price index is again given by (A22),
and the labor supply now satisfies
ψ
Wcφ+(1−φ)ψ = (1 − φ)1−φ(1+ψ) φφ B Wc Lc + χΠ
cΠ Lεc . (A27)
The supply of robots continues to be given as in equation (A4).

From equation (8), the price of the (tradable) good of industry i in every commuting zone
is given by
! 1
1−λ
X 1−λ
X
PiY = υoi Poi . (A28)
o∈C
Since the price of the tradable good aggregate is chosen as the numeraire, the ideal price
index condition now becomes
X 1−σ
1= νi PiY . (A29)
i∈I
With similar steps to our analysis in the autarky model, the demand for labor takes the form
X
Wc Lc = Wc Lci + Wc LSc
i∈I
X
αsL X Π

= ci Pci Xci + (1 − φ) Wc Lc + χc Π .
i∈I
A-9
Also, equations (1) and (8) imply that

X
Xci = Xcdi
d∈C
X λ −λ
= υci Ydi PiY PciX
d∈C
X λ−σ −λ
= υci νi Yd PiY PciX
d∈C
λ−σ −λ X
=υci νi PiY PciX Yd
d∈C
λ−σ X −λ
=υci νi PiY Pci Y.
Using this formula for Xci , we obtain a simplified expression for labor demand in commuting
zone c as
X λ−σ 1−λ
αsL Y
PciX Y + (1 − φ) Wc Lc + χΠ

Wc Lc = ci υci νi Pi cΠ . (A30)
i∈I
Similarly, the demand for robots is

X λ−σ 1−λ
RcM Mc = α(1 − sL
ci )υci νi Pi
Y
PciX Y, (A31)
i∈I
and the demand for capital is given by
RK K = (1 − α)Y. (A32)
Finally, the national capital gains are given by

X X
Π=Y − Wc Lc − D−1−η (1 + η)−1−η Mc1+η . (A33)
c∈C c∈C
Definition 2 An equilibrium of the trade model is given by a set of factor prices {Wc , RcM }c∈C ,
factor supplies {Lc , Mc }c∈C , and national aggregates Y , RK , and Π such that:
• factors supplies are given by (A4) and (A30);
• factor prices satisfy the ideal price index condition, equations (A28) and (A29);
• factor markets clear, that is, equations (A30), (A31), and (A32) hold;
• capital gains are given by equation (A33).
Proposition A5 An equilibrium of the trade model exists.
Proof (sketch): Existence follows from a standard fixed point argument. Our economy con-
sists of |C| (representative) households, |C| nontradable goods, |C| × |I| tradable intermediates,
|C| nontradable robot inputs, |C| types of labor inputs, and |C| final goods. The production
A-10
possibilities sets of all of these goods, which use labor, capital and a subset of the other goods,
are convex, and consumer preferences, defined over the |C| final goods and labor supply, are
continuous and strictly concave, and in fact, are also homothetic. Existence of equilibrium then
follows by constructing the vector of the product of excess demands and prices, verifying com-
pactness, and then applying Brouwer’s fixed point theorem. Moreover, the first welfare theorem
applies and shows that equilibrium is Pareto optimal.
We next state and prove a generalized version of Proposition 3 in the text. To simplify the
expressions, we also impose the following initial allocation of non-labor income across commuting
zones χΠ P Wc Lc
c = c0 ∈C Wc0 Lc0
, which ensures that the ratio of labor to non-labor income across
commuting zones is constant. Note that this is only imposed for the baseline allocation.
Proposition A6 Suppose that the initial allocation of non-labor income satisfies χΠ P Wc Lc

c = c0 ∈C Wc0 Lc0
,
πc = π0 > 0 and θi = θ0 . Then
X dθi γL Y
d ln Lc =[−ζ̄ disp φ + ζ̄ prod φπ0 − ζ̄Linc φψ] · `ci + ζ¯L d ln Y + ζ̄LΠ d ln Π + ζ̄LG Gc,U S ,
1 − θi γ M
i∈I
X dθi γL
d ln Wc =[−ζ̄ disp ε + ζ̄ prod επ0 + ζ̄W incψ] · `ci Y
+ ζ̄W Π
d ln Y + ζ̄W G
d ln Π + ζ̄W Gc,U S ,
1 − θi γ M
i∈I
where the ζ̄’s are the unique solution to the system of equations given by Equations (A15), (A36),
(A37), (A38) and (A39), and
X
1−λ
X X Poi
αsL d ln Wo + α(1 − sL )d ln RoM .

Gc,U S = (λ − σ) `ci υoi (A34)
i∈I o∈C
PiY
Moreover, when θ0 = 0, we have ζ̄ disp = (1 + η + (λ − 1)αη)/Λ, ζ̄ prod = (1 − α + η + (λ −

1)αη + σα)/Λ, ζ̄Linc = α(π0 (σ − 1)α − (1 − π0 )(1 + η + (λ − 1)αη))/Λ, ζ̄W inc = (1 − φ + φα)(π0 (σ −
1)α − (1 − π0 )(1 + η + (λ − 1)αη)/Λ, and
(λ − 1)α2

2
Λ=φ φ+ε 1+ + ψ(1 − φ + (λ − 1)α ) > 0.
1 − φ + φα
Proof: First, note that when πc = π0 and θi = θ0 , we have sL L

ci = s for all i and c.
Next the change in household income in a commuting zone c is given by

W c Lc Wc Lc
d ln Wc Lc + χΠ

cΠ = (d ln W c + d ln Lc ) + 1 − d ln Π
W c L c + χΠcΠ Wc Lc + χΠ cΠ
P P
c0 ∈C Wc0 Lc0 c0 ∈C Wc0 Lc0
=P (d ln Wc + d ln Lc ) + 1 − P d ln Π
c0 ∈C Wc0 Lc0 + Π c0 ∈C Wc0 Lc0 + Π
αsL αsL

= 1−φ+φ (d ln Wc + d ln Lc ) + φ − φ d ln Π,
1−ι 1−ι
A-11
P
where ι = c∈C Ic /Y . Collecting terms, this expression can be rewritten as
d ln Wc Lc + χΠ L L

c Π = ω (d ln Wc + d ln Lc ) + (1 − ω )d ln Π, (A35)
L
where ω L = 1 − φ + φ αs
1−ι is the overall labor share in the economy.
Differentiating and rearranging equation (A27) and combining it with (A35), we obtain the
following expression for labor demand in commuting zone c,
(φ + (1 − φ)ψ)d ln Wc = ψω L (d ln Wc + d ln Lc ) + ψ 1 − ω L d ln Π + εd ln Lc .

(A36)
The supply of robots continues to be given by equation (A15).

Moreover, again from equation (A30), we can express labor demand in commuting zone c as
d ln Wc + d ln Lc = + %d ln Y + (1 − %) ω L (d ln Wc + d ln Lc ) + 1 − ω L d ln Π

X dθi
− (1 − sL π0 ) `ci + %(1 − sL )(d ln Wc − d ln RcM )
1 − θi
i∈I
X X
+ (1 − λ) `ci d ln PciX + (λ − σ) `ci d ln PiY .
i∈I i∈I
where % denotes the baseline share of employment in the tradable sector.

To simplify this expression, note that equation (A28) implies
X
1−λ
X Poi
d ln PiY = υoi X
d ln Poi
o∈C
PiY
X
1−λ
X Poi L L dθi
= υoi αs d ln Wo + α(1 − s )d ln RoM K
+ (1 − α)d ln R − αs π0 L
o∈C
PiY 1 − θi
X
1−λ
dθi
K
X
L Poi
αsL d ln Wo + α(1 − sL )d ln RoM .

=(1 − α)d ln R − αs π0 + υoi
1 − θi
o∈C
PiY
Using this expression for d ln PiY , we can further simplify our labor demand expression as
d ln Wc + d ln Lc =%d ln Y + (1 − %) ω L (d ln Wc + d ln Lc ) + 1 − ω L d ln Π

X dθi
− (1 − sL π0 + (1 − σ)αsL π0 ) `ci + %(1 − sL )(d ln Wc − d ln RcM )
1 − θi
i∈I
+ %(1 − λ) αs d ln Wc + α(1 − sL )d ln RcM + %(1 − σ)(1 − α)d ln RK

L

+ Gc,U S , (A37)
where Gc,U S is given in equation (A34).

Similarly, the expression for the demand for robots in equation (A31) can be rearranged to
A-12
obtain
d ln RcM + d ln Mc =d ln Y
1 (1 − sL π0 )sL
X
L dθi
+ L
+ (σ − 1)αs π0 `ci − sL (d ln Wc − d ln RcM )
% (1 − s ) 1 − θi
i∈I
+ (1 − λ) αs d ln Wc + α(1 − s )d ln RcM + (1 − σ)(1 − α)d ln RK

L L

1
+ Gc,U S , (A38)
%
while the demand for capital, equation (A10), implies
d ln RK = d ln Y. (A39)
Equations (A15), (A36), (A37), (A38) and (A39) define a system of five linear equations and
five unknowns, d ln Lc , d ln Wc , d ln RcM , d ln Mc , and d ln RK . Solving this system of equations
yields the formulas for d ln Lc and d ln Wc given in the proposition. Moreover, when θ0 = 0,
these simplified to the expressions given above.
The next proposition shows how aggregate effects of robots can be computed in the economy
with trade. We simplify the analysis by focusing on the case we use in our quantitative exercise
where πc = π0 and θi = θ0 .
Proposition A7 Suppose that the initial allocation of non-labor income satisfies χΠ P Wc Lc

c = c0 ∈C Wc0 Lc0
,
W W
P P
πc = π0 > 0, and θi = θ0 ≥ 0. Let d ln L = c∈C χc d ln Lc and d ln W = c∈C χc d ln Wc de-
note the average change in employment and wages across commuting zones, where χW
c denotes
the share of the national wage bill paid in commuting zone c. Then
h
disp prod inc
i X X dθi γL
d ln L = −ζ φ + ζ φπ − ζL φψ · χW
c `ci ,
c
1 − θi γM
i∈I
h i X X dθi γL
d ln W = −ζ disp ε + ζ prod επ − ζW
inc
ψ · χW
c `ci ,
c
1 − θi γ M
i∈I
where the ζ’s coincide with those given in Proposition A4 for πc = π0 , and sL L
c = s . In
particular, when in addition θ0 = 0, we have ζ disp = (1 + η − α)/Λ, ζ prod = (1 + η + (1 −
φ)(1 − α))/Λ, ζLinc = (1 − φ + φα)π0 /Λ, ζW
inc = (1 − φ + φα)(π − (1 − π (1 − α + η))/Λ, and
0 0
γL
Λ= γM φ ((1 − α)φ + ψ(1 − φ + φα) + ε)
Proof: Let LTc denote total employment in the tradable sector. First, note that we can rewrite
A-13
χW
c as
Wc Lc
χW
c =P
s∈C Ws Ls
Wc LTc
=P T
s∈C Ws Ls
αsL i∈I Xci PciX
P
= L
P αs Y X
Xci Pci
= i∈I ,
Y
That is, because all commuting zones have the same factor intensity, χW
c is equal to the share
of output generated by commuting zone c within the tradable sector (recall that Y denotes the
aggregate output of the tradable sector).
Using the fact that Π + c Wc Lc = φ1 (Y − I), we obtain
P

L L
1 ιc
ω (d ln W + d ln L) + 1 − ω d ln Π = d ln Y − (1 + η)d ln M (A40)
1 − ιc 1 − ιc
Differentiating, rearranging and summing equation (A27) across commuting zones yields an
expression for average wages in the United States,
(φ + (1 − φ)ψ)d ln W = ψω L (d ln W + d ln L) + ψ 1 − ω L d ln Π + εd ln L,

α(1−sL )
which Can be simplified by substituting from equation (A40) and using the fact that ι = 1+η :
ψ
d ln Y − α(1 − sL )d ln M + εd ln L.

(φ + (1 − φ)ψ)d ln W = α(1−sL )
(A41)
1− 1+η
Adding up equation (A15) across commuting zones yields
d ln RM = ηd ln M, (A42)
where d ln RM = χW M χW
P P
c∈C c d ln Rc and d ln M = c∈C c d ln Mc .
Now differentiating and rearranging (A30), and summing over commuting zones, we obtain
an aggregated version of the labor demand equation demand in equation,
d ln W + d ln L =%d ln Y + (1 − %) ω L (d ln W + d ln L) + 1 − ω L d ln Π

X X dθi
− (1 − sL π0 ) χW
c `ci + (1 − sL )(d ln W − d ln RM )
1 − θi
c∈C i∈I
XX XX
+ (1 − λ) χc `ci d ln PciX + (λ − σ)
W
χW Y
c `ci d ln Pi .
c∈C i∈I c∈C i∈I
A-14
This equation simplifies to
1−%
d ln Y − α(1 − sL )d ln M

d ln W + d ln L =%d ln Y + α(1−sL )
1− 1+η
X X dθi
− (1 − sL π0 ) χW
c `ci + %(1 − sL )(d ln W − d ln RM ) (A43)
1 − θi
c∈C i∈I
χW X χW Y
P P P P
by using equation (A40) and noting that c∈C i∈I c `ci d ln Pci = 0 and c∈C i∈I c `ci d ln Pi =
0. These last two observations follow from the price index in equation (A28) and the ideal price
index condition in equation (A29). In particular, for the former, note that
XX XX
χW Y
c `ci d ln Pi =% χW
c χci d ln Pi
Y
c∈C i∈I i∈I c∈C

X
P
XX j∈I Xcj Pcj Xci PciX
=% P X
d ln PiY
Y X
j∈I cj cjP
i∈I c∈C
X X Xci P X
=% ci
d ln PiY
Y
i∈I c∈C
X Yi P Y
=% i
d ln PiY
Y
i∈I
=0,
which follows from equation (A29) (recall that χci is the share of industry i in value added in
commuting zone c). In this derivation, we used Yi to denote the total output of industry i,
so that Yi PiY = c Xci PciX , and we also used `ci = %χci , which follows from the fact that all
P
tradable sectors have the same labor intensity.

Likewise,
XX XX
χW X
c `ci d ln Pci =% χW X
c χci d ln Pci
c∈C i∈I i∈I c∈C
X
P
XX j∈I Xcj Pcj Xci PciX
=% P X
d ln PciX
Y X
j∈I cj cjP
i∈I c∈C
X X Xci P X
=% ci
d ln PciX
Y
i∈I c∈C
X Yi P Y X Xci P X
=% i ci
d ln PciX
i∈I
Y
c∈C
Yi PiY
X Yi P Y
=% i
d ln PiY
Y
i∈I
=0,
where we have used the price index in equation (A28) and the ideal price index condition in
A-15
equation (A29).
Following the same steps, we obtain aggregate robot demand from equation (A31) as
d ln RM + d ln M =d ln Y
1 (1 − sL π0 )sL X W X dθi
+ χc `ci − sL (d ln W − d ln RM ) (A44)
% (1 − sL ) 1 − θi
c∈C i∈I
χW X
P P
Finally, the fact that i∈I c∈C c χci d ln Pci = 0 also implies
X X X X dθi
χW
c χci (αsL d ln Wc + α(1 − sL )d ln RcM + (1 − α)d ln RK ) = χW
c χci αsL π0 .
1 − θi
c∈C i∈I c∈C i∈I
Using the fact that `ci = %χci and the definition of d ln W and d ln RM , we can rewrite this
equation as
1 L X WX dθi
αsL d ln W + α(1 − sL )d ln RM + (1 − α)d ln RK = αs π0 χc `ci . (A45)
% 1 − θi
c∈C i∈I
Finally, taking logs and differentiating the demand for capital in equation (A32), we have
d ln RK = d ln Y (A46)
Equations (A41), (A42), (A43), (A44), (A45) and (A46) define a system of six linear equa-
tions in six unknowns. Solving this system of equations yields the formulas for d ln L and d ln W
in the proposition.
A3 Details of the Quantitative Exercise

Propositions A6 and A7 show how to compute the local and aggregate effects of robot adoption
in terms of the parameters of our model, the share of labor in production tasks, sL , and the
share of (non-robot) capital, 1 − α. Here, we provide some of the details omitted from the text
of how we choose parameter values to perform our quantitative exercise.
1. First, recall that γM /γL = 3 as explained in the text.
2. Let us next turn to the share of labor in production tasks, sL . Our model implies that,
among industries using robots, the baseline ratio of robots per thousand workers in the
US is
Mi θ0 γ L
1000 = 1000 .
Li 1 − θ0 γ M
In 1993, the US had around four robots per thousand workers in industries using robots
θ0
(which are almost entirely in manufacturing). Since γM /γL = 3, this implies 1−θ0 = 0.012.
A-16
We can then compute the labor share in production tasks as
1−θ0 1
θ0 1−π0
sL = 1−θ0 1
= 0.9916.
θ0 1−π0 + 1
This implies that in 1993, labor accounted for 99.16 percent of the value added in tasks
that can be automated using industrial robots, and robots accounted for the remaining
0.84 percent.
3. Because the overall labor share in the economy is αsL , the previous observation implies
α = 0.67 to match the labor share of 66.6 percent.
4. We equate tradables with manufacturing, which gives a share of employment in tradables

αφ
of % = 0.18. Using the fact that % = 1−φ+αφ , we obtain φ = 0.25.
5. We then choose the income elasticity of labor supply ψ to match empirical estimates of the
propensity to consume leisure out of additional income. In particular, our labor supply
equation implies
dLc ψ Wc Lc ψ
Wc =− = − ωL,
dCc ε Cc ε
Wc Lc
where we used the fact that, in our model, Cc = ω L —where ω L is the share of labor in
total value added.
Imbens, Rubin and Sacerdote (2001) estimate that the propensity to consume leisure out
of one additional dollar is about 0.1, which implies that
ψ L
0.1 = ω . (A47)
ε
6. Finally, we choose the value for ε (and thus for ψ from equation (A47)) as follows. Let
βL = −ζ disp φ + ζ prod φπ0 − ζ L,inc φψ and βW = −ζ disp ε + ζ prod επ0 + ζ W,inc ψ. By definition,
P dθi γL P dθi γL
d ln Lc = βL · i∈I `ci 1−θ γ
i M
and d ln W c = βW · i∈I `ci 1−θi γM solve the system of
equations given by (A15), (A36), (A37), (A38) and (A39) when d ln Y = d ln Π = Gc,U S =
0. Next, substituting the formulas for d ln Lc and d ln Wc into (A36), we obtain the equation
(φ + (1 − φ)ψ)βW = ψω L (βW + βL ) + εβL . (A48)
Solving equations (A47) and (A48) simultaneously and using our IV estimates β̂L and β̂W
yields
φβW
ε= 0.1(1−φ)
= 0.17
0.1(β̂W + β̂L ) − ωL
β̂W + β̂L
0.1 φβ̂W
ψ= L
= 0.02.
ω 0.1(β̂W + β̂L ) − 0.1(1−φ) β̂W + β̂L
ωL
A-17
Given the values of ψ and ε , η is chosen to match our IV estimate β̂L (or β̂W ). This yields
η = 0.79.
A4 Additional Figures and Tables
Figure A1: Adjusted robot penetration and share of replaceable jobs.

Plot of the adjusted penetration of robots between 1993 and 2007 (AP Ri ) and the share of
replaceable jobs by industry in 1990. The data on replaceable jobs are from Graetz and Michaels
(2018). Marker size indicates the baseline US employment in the industry.
A-18
A-19
Figure A2: Industry pre-trends.
The figure presents residual plots of the relationship between the adjusted penetration of robots for 1993-2007 (AP Ri ) and the 1970-1990
change in log wage bill (top left panel), log wage bill for production workers (top right panel), log employment (bottom left panel), and log
employment for production workers (bottom right panel). The solid lines correspond to regression models analogous to those in columns
8-9 of Panels A and B of Table 1. The covariates from these models are partialled out. The coefficients for these models and their standard
errors are reported next to each plot. The dashed line is for a regression which in addition excludes the automotive industry. Marker size
indicates the baseline US employment in the industry.
Figure A3: First-stage relationship for commuting zones.

The figure presents the relationship between exposure to robots for 1993-2007 (from EU RO5)
and US exposure to robots for 2004-2007 (rescaled to a 14-year equivalent change). The co-
variates from columns 3 and 6 of Table 7 are partialled out. The solid line corresponds to a
regression with commuting zone population in 1990 as weights. The dashed line is for a regres-
sion which in addition excludes the top one percent of commuting zones with highest exposure
to robots. Marker size indicates the 1990 population in the commuting zone.
A-20
A-21
Figure A4: The effects of robots on industries and occupations, additional estimates.
The figure presents estimates of exposure to robots on the change in industry employment to population ratios (top panel) and the change in
occupation employment to population ratios (bottom panel). The capped lines provide 95% confidence intervals. The first set of estimates
are from long-differences specifications as in column 6 of Table 2. The second set of estimates are from stacked-differences specifications
as in column 5 of Table 2 (where we remove the top one percent of commuting zones with highest exposure to robots). The third set of
estimates are from stacked-differences specifications as in column 6 of Table 3 (unweighted).
Table A1: Summary statistics: industry data
Baseline
Robots per thousand workers, EU RO5 Robots per thousand workers, US
employment, US
IFR industry 1993 2000 2007 2014 2004 2007 2014 (thousands)
Manufacturing:
Automotive 19.96 37.87 69.30 76.70 69.01 85.72 117.72 1110.78
Plastics and Chemicals 3.25 15.23 26.07 22.93 5.12 6.95 9.91 2205.11
Metal Products 8.51 13.70 21.81 21.24 4.60 5.84 8.29 1689.15
Industrial Machinery 3.99 4.67 7.41 11.74 1.32 1.67 2.37 1540.93
Food and Beverages 0.47 1.80 5.75 10.83 2.91 3.92 6.17 1862.25
Basic Metals 1.17 3.95 7.05 10.31 3.98 5.05 7.17 712.02
Electronics 2.88 5.94 9.85 7.11 5.71 8.66 13.11 2868.12
Miscellaneous Manufacturing 3.12 3.84 3.40 6.48 1.40 1.96 13.81 690.29
Minerals 0.77 2.01 3.78 4.99 0.12 0.23 0.67 557.74
A-22
Wood and Furniture 0.66 2.04 4.64 4.83 0.01 0.01 0.14 1047.85
Shipbuilding and Aerospace 0.83 4.56 4.05 3.09 0.05 0.12 0.54 1110.54
Textiles 0.33 1.03 1.27 1.53 0.00 0.01 0.05 1848.24
Paper and Printing 0.27 0.46 0.95 1.36 0.00 0.00 0.11 2467.12
Nonmanufacturing:

Mining 0.32 2.00 3.16 2.14 0.00 0.01 0.06 763.19
Education and Research 0.04 0.21 0.35 0.40 0.01 0.01 0.06 12636.45
Agriculture 0.00 0.00 0.16 0.24 0.00 0.00 0.04 2551.58
Utilities 0.00 0.02 0.02 0.18 0.00 0.00 0.03 745.44
Construction 0.00 0.03 0.07 0.14 0.00 0.01 0.02 7107.59
Services 0.00 0.00 0.00 0.00 0.00 0.00 0.00 84775.62
The table shows the evolution of the stock of robots per thousand workers by industry. The number of robots is from the IFR and the number of workers in each industry is from
EUKLEMS.
Table A2: Adjusted penetration of robots in the US and in Europe
Dependent variable: Dependent variable:

Adjusted penetration of robots, US Penetration of robots, US
Weighted by employment Unweighted Weighted by employment Unweighted
(1) (2) (3) (4) (5) (6)
Panel A. 1993-2007
Adjusted penetration of 1.475 1.608 1.960 1.443 1.513 1.844
robots, AP Ri (0.746) (0.864) (0.723) (0.673) (0.787) (0.654)
Observations 19 19 19 19 19 19
R-squared 0.63 0.64 0.73 0.66 0.66 0.74
Panel B. 2000-2007
robots, AP Ri (0.283) (0.268) (0.189) (0.236) (0.230) (0.160)
R-squared 0.83 0.85 0.89 0.85 0.86 0.90
Panel C. 2004-2007
robots, AP Ri (0.486) (0.560) (0.469) (0.433) (0.511) (0.422)
R-squared 0.63 0.64 0.72 0.65 0.65 0.74
Panel D. 2000-2014
robots, AP Ri (0.242) (0.294) (0.205) (0.298) (0.348) (0.241)
R-squared 0.82 0.82 0.83 0.82 0.82 0.85
Covariates:
Manufacturing dummy X X X X
The table presents estimates of the relationship between adjusted penetration of robots, AP Ri from EU RO5, and the
penetration of robots in the US, AP Ri . Columns 1-3 present estimates using the adjusted penetration of robots in the
US as outcome. Columns 4-6 present estimates using the penetration of robots in the US as outcome (unadjusted for
changes in output). Each panel presents results for a different time period; when necessary we rescale the penetration of
robots in the US to match its length. Column 1 and 4 do not include any covariates. Columns 2-3 and 5-6 control for a
dummy for manufacturing. The regressions in columns 1-2 and 2-3 are weighted by baseline industry employment in 1993,
and the regressions in columns 3 and 6 are unweighted. Standard errors that are robust against heteroskedasticity are in
parentheses.
A-23
Table A3: Industry-level correlations between adjusted penetration of robots

and industry characteristics.
Correlation with adjusted

penetration of robots
Within Across all Industry with
manufacturing industries largest value
Measures of trade competition:

Chinese trade competition 1990-2007 -0.39 0.15 Textiles
Measures of offshoring:
Mexican imports to the US 1990-2007 -0.03 0.31 Electronics
Share offshorable jobs in 1990 -0.41 -0.26 Textiles
Intermediate goods imports 1990-2007 -0.17 0.19 Electronics
Capital and other investments:

Percent increase in capital stock
0.22 -0.37 Construction
1990-2007
Percent increase in IT capital stock
0.23 -0.17 Construction
1990-2007
Other industry characteristics:

Share routine jobs in 1990 -0.24 -0.01 Paper and printing
Female share of employment in 1990 -0.30 -0.40 Textiles
Share replaceable jobs in 1990 0.41 0.64 Automotive
The table presents the correlation between adjusted penetration of robots for 1993 and 2007 (AP Ri,1993,2007 ) and several
industry covariates. We present the correlation within manufacturing (across 13 industries) and the overall correlation
(across 19 industries), with employment in 1993 as weights. The share of offshorable jobs, routine jobs, female workers, and
replaceable jobs are all computed using the distribution of employment by industry in the 1990 Census. See the text for
detailed definitions and sources of data.
A-24
Table A4: Industry-level results: different constructions of the adjusted penetration of robots, AP Ri
Long differences, 1993-2007 Stacked differences, 1993-2000 and 2000-2007 Long differences, 1992-2007
CBP (all industries) NBER-CES (within manufacturing) CBP (all industries) NBER-CES (within manufacturing) BEA-IO (all industries)
Wage bill Wage bill Wage bill Wage bill Wage bill Value added Labor share
workers bill workers bill
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Panel A. Baseline construction of the instrument
Adjusted penetration of robots, -0.923 -0.816 -0.993 -1.096 -1.492 -1.037 -1.150 0.128 -0.797
AP Ri (0.419) (0.378) (0.324) (0.235) (0.481) (0.177) (0.205) (0.061) (0.281)
Observations 19 13 13 38 38 26 26 19 19
R-squared 0.91 0.84 0.91 0.90 0.95 0.87 0.91 0.72 0.37
Panel B. Including penetration of robots in Germany
AP Ri (0.273) (0.246) (0.184) (0.158) (0.414) (0.094) (0.136) (0.062) (0.267)
Observations 19 13 13 38 38 26 26 19 19
R-squared 0.91 0.84 0.91 0.90 0.95 0.86 0.91 0.70 0.34
Panel C. Using penetration of robots among nine European countries with data
AP Ri (0.240) (0.218) (0.161) (0.149) (0.450) (0.094) (0.176) (0.057) (0.259)
Observations 19 13 13 38 38 26 26 19 19
R-squared 0.91 0.84 0.91 0.90 0.95 0.86 0.91 0.68 0.32
Panel D. Using penetration of robots without adjustment term
AP Ri (0.257) (0.235) (0.174) (0.177) (1.038) (0.132) (0.289) (0.057) (0.249)
A-25
Observations 19 13 13 38 38 26 26 19 19
R-squared 0.91 0.84 0.91 0.90 0.95 0.85 0.90 0.71 0.35
Panel E. Adjusting penetration of robots using robot prices
AP Ri (0.379) (0.330) (0.249) (0.224) (0.568) (0.137) (0.164) (0.082) (0.364)
Observations 19 13 13 38 38 26 26 19 19
R-squared 0.92 0.86 0.92 0.91 0.95 0.88 0.92 0.70 0.36

Covariates:
Time period dummies X X X X
Industry shares X X X X X X X X X
Chinese imports X X X X X X X X X
Industry dummies X
The table presents estimates of the relationship between adjusted penetration of robots and changes in log wage bill, log value added and labor share across US industries. Each
panel presents results using a different construction of the measure for adjusted penetration of robots. Panel A presents results for our baseline construction. Panel B presents
results including the penetration of robots in Germany. Panel C presents results using the data for nine European countries (Germany, Denmark, Finland, France, Italy, Sweden,
Norway, Spain, and the UK) to construct the adjusted penetration of robots. Panel D presents results for the penetration of robots without the adjustment term in equation (12).
Panel E present results adjusting the penetration of robots using the available data on robot prices (from Robotics Industries of America). Columns 1-3 present long-differences
estimates for changes in log wage bill, 1993-2007. Columns 4-7 present stacked-differences estimates for changes in log wage bill, 1993-2000 and 2000-2007. Columns 8-9 present
long-differences estimates for changes in log value added for 1992-2007 (annualized), and changes in labor share for 1992-2007. The sources of data and their coverage are reported
at the top of the table, and the set of covariates is reported at the bottom of the table. All models control for dummies for manufacturing and light manufacturing, and exposure
to Chinese imports by industry from Acemoglu et al. (2016). In addition, the stacked-differences models control for time period dummies, and in column 5 for industry dummies.
The regressions in columns 1-7 are weighted by baseline industry employment in 1993, and the regressions in columns 8-9 are weighted by baseline value added by industry in 1992.
Standard errors that are robust against heteroskedasticity and serial correlation at the industry level are in parentheses.
Table A5: Industry-level results: different time periods
Estimates for change in log wage bill Estimates for change in log employment
NBER-CES, NBER-CES,
NBER-CES, NBER-CES,
CBP production CBP production
all workers all workers
workers workers
(1) (2) (3) (4) (5) (6)
Panel A. 1993-2010
Adjusted penetration of robots, -1.509 -1.351 -1.618 -1.292 -1.147 -1.295
AP Ri (0.683) (0.665) (0.687) (0.459) (0.419) (0.454)
R-squared 0.93 0.85 0.88 0.94 0.89 0.90
Panel B. 2000- 2010
AP Ri (0.472) (0.441) (0.488) (0.358) (0.291) (0.378)
R-squared 0.82 0.76 0.76 0.84 0.81 0.79
Panel C. 2004-2010
AP Ri (1.770) (1.951) (2.149) (1.365) (1.420) (1.754)
R-squared 0.64 0.45 0.43 0.65 0.46 0.40
Panel D. 2000-2007
AP Ri (0.184) (0.151) (0.140) (0.155) (0.159) (0.157)
R-squared 0.89 0.86 0.89 0.90 0.88 0.90
Panel E. 2004-2007
AP Ri (0.565) (0.746) (0.768) (0.325) (0.507) (0.577)
R-squared 0.79 0.65 0.68 0.64 0.68 0.68
Covariates:
Industry shares X X X X X X
Chinese imports X X X X X X
The table presents estimates of the relationship between adjusted penetration of robots and changes in log wage bill and
log employment across US industries. Columns 1-3 present estimates for changes in log wage bill for 1993-2010 (Panel
A), 2000-2010 (Panel B), 2004-2010 (Panel C), 2000-2007 (Panel D), 2004-2007 (Panel E). Columns 4-6 present estimates
for changes in log employment for 1993-2010 (Panel A), 2000-2010 (Panel B), 2004-2010 (Panel C), 2000-2007 (Panel D),
2004-2007 (Panel E). The sources of data and their coverage are reported at the top of the table, and the set of covariates
is reported at the bottom of the table. All models control for dummies for manufacturing and light manufacturing, and
exposure to Chinese imports by industry from Acemoglu et al. (2016). All regressions are weighted by baseline industry
employment. Standard errors that are robust against heteroskedasticity are in parentheses.
A-26
Table A6: Robots and labor demand: control for other industry trends
Estimates for changes in log wage bill

CBP (all CBP (all
NBER-CES (within manufacturing) NBER-CES (within manufacturing)
industries) industries)
All workers All workers All workers All workers
workers workers
(1) (2) (3) (4) (5) (6)
Panel A. Control for trends in value added
Adjusted penetration -1.387 -1.636 -1.799 -1.446 -1.556 -1.655
of robots, AP Ri (0.443) (0.453) (0.373) (0.259) (0.248) (0.213)
Percent change in value 0.460 0.669 0.671 0.230 0.317 0.317
added, 1992-2007 (0.153) (0.117) (0.119) (0.065) (0.047) (0.057)
R-squared 0.92 0.84 0.89 0.91 0.86 0.90
Panel B. Instrument trends in value added using intermediate import availability
of robots, AP Ri (0.319) (0.408) (0.392) (0.263) (0.282) (0.297)
Percent change in value 0.581 0.583 0.568 0.308 0.319 0.304
added, 1992-2007 (0.093) (0.130) (0.125) (0.042) (0.061) (0.073)
R-squared 0.92 0.83 0.88 0.91 0.86 0.90
First-stage F 11.87 11.71 11.71 8.64 7.44 10.33
Panel C. Control for task offshorability in manufacturing industries
of robots, AP Ri (0.295) (0.273) (0.351) (0.243) (0.257) (0.279)
Share offshorable jobs -2.299 -2.783 -2.613 -1.117 -1.331 -1.259
in 1990 (0.491) (0.467) (0.566) (0.207) (0.194) (0.227)
R-squared 0.93 0.90 0.89 0.91 0.88 0.90
Panel D. Control for industries with significant adoption of robots
of robots, AP Ri (0.415) (0.404) (0.405) (0.261) (0.280) (0.271)
R-squared 0.91 0.71 0.78 0.90 0.80 0.85
Covariates:
Manufacturing dummy X X X X X X
Chinese imports X X X X X X
The table presents estimates of the relationship between adjusted penetration of robots and changes in log wage bill across
US industries. Column 1-3 present long-differences estimates for 1993-2007. Columns 4-6 present stacked differences
estimates for 1993-2000 and 2000-2007. The sources of data are reported at the top of the table. All models include a
dummy for the manufacturing industry and Chinese imports by industry from Acemoglu et al. (2016) as controls. Each
panel lists additional covariates included in the models. Panel A controls for the change in industry value added between
1992 and 2007 (from the BEA input-output tables). Panel B instruments for the change in value added using intermediate
imports in supplier industries (from Feenstra and Hanson, 1999). Panel C controls for the share of offshorable jobs across
industries. Panel D includes a dummy for industries with significant adoption of robots (automotive, plastics and chemicals,
metal products, industrial machinery, food and beverages, basic metals, electronics, miscellaneous manufacturing, and
minerals). All regressions are weighted by baseline industry employment in 1993. Standard errors that are robust against
heteroskedasticity and serial correlation at the industry level are in parentheses.
A-27
Table A7: Industry-level results: labor share, value added and productivity
Long differences, 1992-2007

Excluding
All industries, weighted by
automotive Unweighted
value added in 1992
manufacturing
(1) (2) (3) (4)
Panel A. Change in labor share (in percentage points)
Adjusted penetration of robots, AP Ri -0.864 -0.797 -0.924 -0.656
(0.164) (0.281) (0.331) (0.218)
Observations 19 19 18 19
R-squared 0.36 0.37 0.31 0.35
Panel B. Change in log value added (percent annual change)
Adjusted penetration of robots, AP Ri 0.179 0.128 0.188 0.085
(0.059) (0.061) (0.043) (0.044)
R-squared 0.62 0.72 0.74 0.67
Panel C. Change in log value added per worker (percent annual change)
Adjusted penetration of robots, AP Ri 0.123 0.153 0.220 0.111
(0.055) (0.069) (0.054) (0.055)
R-squared 0.23 0.37 0.42 0.20
Covariates:
Manufacturing dummy X X X X
Light manufacturing dummy X X X
Chinese imports X X X
The table presents estimates of the relationship between adjusted penetration of robots and changes in labor shares and
log value added across US industries. Panel A presents long-differences estimates for changes in labor share of value added,
1992-2007. Panel B presents long-differences estimates for changes in log value added, 1992-2007. Panel C presents long-
differences estimates for changes in log value added per worker, 1992-2007. Column 1 controls for a manufacturing dummy.
Column 2 controls for a light manufacturing dummy and exposure to Chinese imports by industry from Acemoglu et al.
(2016). Column 3 excludes the automotive industry from the sample. The regressions in columns 1-3 are weighted by
baseline industry value added in 1992, and the regression in column 4 is unweighted. Standard errors that are robust
against heteroskedasticity are in parentheses.
A-28
Table A8: Exposure to robots and the location of integrators
Location of robot integrators

Excludes zones
Weighted by population with highest Unweighted
exposure
(1) (2) (3) (4) (5) (6)
Panel A. log one plus number of integrators
Exposure to robots 0.268 0.242 0.193 0.185 0.276 0.094
(0.078) (0.063) (0.032) (0.032) (0.094) (0.019)
Observations 722 722 722 722 712 722
R-squared 0.28 0.74 0.75 0.75 0.74 0.53
Panel B. log one plus number of integrators
Exposure to robots in 0.253 0.224 0.189 0.181 0.239 0.158
automotive industries (0.071) (0.042) (0.027) (0.028) (0.131) (0.022)
Exposure to robots in other 0.325 0.314 0.219 0.214 0.302 0.018
industries (0.171) (0.143) (0.096) (0.089) (0.089) (0.038)
Observations 722 722 722 722 712 722
R-squared 0.28 0.74 0.75 0.75 0.74 0.54
Panel C. log one plus employment in integrators
(0.176) (0.169) (0.101) (0.100) (0.230) (0.069)
Observations 722 722 722 722 712 722
R-squared 0.24 0.68 0.70 0.70 0.71 0.48
Panel D. Dummy for the presence of integrators
(0.033) (0.036) (0.025) (0.026) (0.038) (0.019)
Observations 722 722 722 722 712 722
R-squared 0.12 0.51 0.54 0.54 0.56 0.45
Covariates:
The table presents estimates of the relationship between exposure to robots and the location of robot integrators. The
dependent variable is log of one plus the number of integrators (Panels A and B), log of one plus employment in integrators
(Panel C) and a dummy for the presence of integrators (Panel D). Columns 1-5 present regressions weighted by population
in 1990. Column 5 presents results excluding the top one percent of commuting zones with highest exposure to robots.
Column 6 presents unweighted regressions. The covariates included in each model are reported at the bottom of the table.
Column 1 only includes Census division dummies. Column 2 adds demographic characteristics of commuting zones in 1990
(log population; the share of females; the share of population above 65 years; the shares of population with no college, some
college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics
and Asians). Column 3 adds the shares of employment in manufacturing and light manufacturing, and the female share of
manufacturing employment in 1990. Columns 4-6 add exposure to Chinese imports and the share of employment in routine
jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
A-29
Table A9: Summary statistics: commuting zone data
Means by quartiles of exposure to robots Correlations with exposure to robots

Conditional on
share
All commuting Second Fourth
First quartile Third quartile Raw correlation manufacturing
zones quartile quartile
and Census
division
(1) (2) (3) (4) (5) (6) (7)
Outcomes:
Change employment to population ratio,
0.96 1.97 1.58 0.49 0.59 -0.23 -0.23
1990-2007 (p.p.)
Change log hourly wages, 1990-2007 (log
5.57 9.58 7.24 5.23 3.35 -0.58 -0.41
points, adjusted for composition)
Baseline characteristics:
Employment to population ratio, 1990 0.35 0.32 0.34 0.36 0.37 0.23 0.09
log hourly wages, 1990 2.59 2.53 2.59 2.60 2.62 0.10 0.23
Covariates:
Share Female population in 1990 0.51 0.51 0.51 0.51 0.52 0.17 0.07
Share Hispanic in 1990 0.09 0.17 0.10 0.10 0.05 -0.25 -0.08
Share White in 1990 0.85 0.86 0.83 0.84 0.85 0.02 -0.11
A-30
Share Black in 1990 0.12 0.10 0.13 0.11 0.13 0.06 0.13
Share Asian in 1990 0.03 0.02 0.03 0.04 0.02 -0.17 0.00
Share with no college degree in 1990 0.56 0.54 0.54 0.56 0.58 0.17 -0.07

Share with college degree in 1990 0.13 0.14 0.14 0.14 0.13 -0.18 0.03
Share with masters degree in 1990 0.05 0.05 0.05 0.05 0.05 -0.10 0.07
Share above 65 years of age in 1990 0.12 0.13 0.12 0.12 0.12 -0.04 -0.10
log population in 1990 13.97 13.46 13.97 14.09 14.07 0.09 0.20
Share employment in manufacturing in 1990 0.22 0.13 0.19 0.24 0.26 0.44 .
Female share of manufacturing employment

0.33 0.34 0.35 0.35 0.31 -0.37 -0.39
in 1990 (within manufacturing)
Share light manufacturing in 1990 (within
0.22 0.24 0.25 0.24 0.18 -0.39 -0.38
manufacturing)
Exposure to Chinese imports between 1990
3.35 1.87 2.86 3.97 3.73 0.06 -0.14
and 2007
Share employment in routine jobs in 1990 0.36 0.35 0.36 0.36 0.37 0.20 0.00
Columns 1-5 present sample means for all commuting zones and by quartiles of exposure to robots. Columns 6 and 7 present correlations between the covariate indicated in each
row and exposure to robots across commuting zones. Correlations are weighted by population, and in column 7 are conditional on the share of employment in manufacturing and
Census division dummies. See text for variable definitions and sources.
Table A10: Relationship between exposure to robots and covariates
Dependent variable:
Exposure to robots in
Exposure to robots in
Exposure to robots automotive
other industries
manufacturing
(1) (2) (3) (4) (5) (6)
Share female 0.108 0.107 0.092 0.102 0.064 0.042
(0.053) (0.048) (0.066) (0.053) (0.084) (0.085)
Share Hispanic -0.107 -0.055 -0.074 -0.012 -0.097 -0.101
(0.051) (0.055) (0.065) (0.069) (0.045) (0.056)
Share White 0.001 -0.059 -0.037 -0.108 0.074 0.078
(0.086) (0.102) (0.112) (0.125) (0.071) (0.072)
Share Black 0.088 -0.003 0.096 -0.003 0.011 -0.001
(0.077) (0.083) (0.096) (0.092) (0.093) (0.100)
Share Asian -0.074 0.051 -0.113 0.004 0.055 0.107
(0.098) (0.075) (0.132) (0.101) (0.074) (0.117)
Share with no college 0.024 -0.092 -0.408 -0.510 0.852 0.789
(0.255) (0.285) (0.320) (0.374) (0.212) (0.206)
Share with college degree -0.289 -0.276 -0.546 -0.536 0.415 0.423
(0.254) (0.259) (0.320) (0.327) (0.160) (0.183)
Share with masters degree 0.065 -0.024 -0.081 -0.167 0.304 0.274
(0.107) (0.130) (0.102) (0.141) (0.172) (0.180)
Share above 65 years of age -0.120 -0.111 -0.083 -0.088 -0.110 -0.078
(0.041) (0.047) (0.071) (0.070) (0.087) (0.097)
Log population 0.277 0.303 0.215 0.247 0.205 0.201
(0.127) (0.129) (0.182) (0.195) (0.111) (0.138)
Share employment in manufacturing 1.162 1.051 0.861 0.713 0.944 0.980
(0.309) (0.270) (0.413) (0.374) (0.202) (0.232)
Female share of manufacturing -0.647 -0.556 -0.490 -0.373 -0.504 -0.526
employment (0.248) (0.215) (0.306) (0.264) (0.212) (0.231)
Share employment in light manufacturing -0.323 -0.276 -0.257 -0.213 -0.227 -0.207
(0.066) (0.076) (0.067) (0.066) (0.091) (0.101)
Exposure to Chinese imports -0.115 -0.100 -0.160 -0.137 0.051 0.041
(0.067) (0.061) (0.087) (0.083) (0.069) (0.060)
Share employment in routine jobs 0.083 -0.044 0.183 0.036 -0.169 -0.170
(0.073) (0.052) (0.067) (0.054) (0.082) (0.095)
Observations 722 722 722 722 722 722
R-squared 0.60 0.63 0.44 0.48 0.45 0.46
Additional covariates not reported:
Census divisions X X X
The table presents the relationship between the covariates used in our analysis (measured in 1990) and exposure to robots
(columns 1-2), exposure to robots in the automotive industry (columns 3-4), and exposure to robots in other industries
(columns 5-6) for 1993-2007. The covariates and exposure measures are standardized to ease the comparison of the point
estimates. Columns 2, 4 and 6 also control for a full set of Census division dummies. Standard errors that are robust
against heteroskedasticity and correlation within states are in parentheses.
A-31
Table A11: The effect of robots on employment and wages: excludes the share
of employment in light manufacturing and the female share of manufacturing
employment from covariates.
Long differences, 1990-2007 Stacked differences, 1990-2000 and 2000-2007

Excludes Excludes
Weighted by zones with Weighted by zones with
Unweighted Unweighted
population highest population highest
exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in the employment to population ratio—Excludes the share of employment

in light manufacturing and the female share of employment in manufacturing from covariates.

(0.064) (0.140) (0.121) (0.054) (0.150) (0.096)
Observations 722 712 722 1444 1424 1444
R-squared 0.64 0.63 0.60 0.40 0.39 0.39
Panel B. Change in the employment to population ratio—excludes the female share of

employment in manufacturing from covariates.

(0.058) (0.135) (0.118) (0.052) (0.150) (0.095)
Observations 722 712 722 1444 1424 1444
R-squared 0.66 0.65 0.62 0.41 0.40 0.39
Panel C. Change in log hourly wages—excludes the share of employment

in light manufacturing and the female share of employment in manufacturing from covariates.

(0.124) (0.268) (0.181) (0.178) (0.522) (0.298)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.33 0.32 0.08 0.29 0.27 0.09
Panel D. Change in log hourly wages—excludes the female share of

employment in manufacturing from covariates.

(0.125) (0.270) (0.182) (0.181) (0.543) (0.296)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.33 0.33 0.08 0.29 0.27 0.09
Covariates:
Remaining baseline
X X X X X X
covariates
The table presents estimates of the effects of exposure to robots on employment to population ratios (Panels A and B) and
log hourly wages (Panels C and D). Columns 1-3 present long-differences estimates for the 1990-2007 period. Columns 4-6
present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns 1-2 and 4-5 present regressions
weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent of commuting zones with
highest exposure to robots. Columns 3 and 6 present unweighted regressions. All columns include Census division dummies
(and time period dummies in the stacked-differences specifications), demographic characteristics of commuting zones in 1990
(log population; the share of females; the share of population above 65 years; the shares of population with no college, some
college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and
Asians), the share of employment in manufacturing, exposure to Chinese imports and the share of employment in routine
jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
A-32
Table A12: Coefficients for the main covariates in Tables 2 and 3

Weighted by Weighted by
with highest Unweighted with highest Unweighted
population population
exposure exposure
(1) (2) (3) (4) (5) (6)
(0.059) (0.138) (0.118) (0.132) (0.274) (0.205)
Share manufacturing 1.846 2.603 0.591 -17.649 -18.326 -32.096
employment (3.393) (3.376) (3.657) (6.246) (6.104) (5.398)
Female share of -21.819 -22.380 -14.617 26.298 26.015 55.292
manufacturing employment (7.362) (7.444) (9.362) (13.705) (13.671) (12.221)
Share light manufacturing -7.804 -8.260 -8.297 -13.012 -11.770 -11.985
employment (3.474) (3.298) (2.902) (6.424) (6.510) (4.821)
-0.082 -0.080 -0.049 -0.058 -0.057 -0.070
Exposure to Chinese imports
(0.040) (0.041) (0.032) (0.075) (0.076) (0.045)
Share employment in routine -8.743 -9.677 -17.382 -7.313 -7.321 5.678
jobs (5.206) (5.143) (5.441) (11.399) (11.367) (10.986)
Observations 722 712 722 87100 85776 87100
R-squared 0.67 0.66 0.62 0.33 0.33 0.08
Panel B. Stacked differences, 1990-2000 and 2000-2007
(0.052) (0.150) (0.092) (0.196) (0.586) (0.300)
Share manufacturing 1.490 0.453 -2.062 -5.696 -8.024 -21.129
employment (3.188) (3.065) (1.850) (6.100) (6.004) (3.703)
Female share of -13.630 -10.914 -1.389 12.741 18.417 37.550
manufacturing employment (7.376) (6.525) (4.931) (15.530) (14.654) (8.470)
Share light manufacturing -0.895 -1.044 -3.547 0.651 0.471 -0.730
employment (1.963) (1.952) (2.078) (4.036) (4.503) (3.302)
-0.195 -0.180 -0.128 -0.405 -0.387 -0.249
Exposure to Chinese imports
(0.030) (0.034) (0.029) (0.105) (0.106) (0.082)
Share employment in routine -6.630 -7.278 -7.201 -6.005 -7.312 14.206
jobs (4.192) (4.232) (3.258) (10.473) (10.764) (7.053)
Observations 1444 1424 1444 183607 180819 183607
R-squared 0.41 0.40 0.39 0.21 0.20 0.06
Covariates:
Remaining baseline
X X X X X X
covariates
The table reports the coefficients for the main covariates used in our analysis for models explaining changes in employment
and wages. Panel A presents long-differences estimates for the 1990-2007 period. Panel B presents stacked-differences
estimates for the 1990-2000 and 2000-2007 periods. Columns 1-3 present estimates for changes in the employment to
population ratio. Columns 4-6 present estimates for changes in log hourly wages. The specifications in columns 4-6 are
estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education
and race. Columns 1-2 and 4-5 present regressions weighted by population in 1990. Columns 2 and 5 present results
excluding the top one percent of commuting zones with highest exposure to robots. Columns 3 and 6 present unweighted
regressions. All models include Census division dummies (and time period dummies in the stacked-differences specifications),
demographic characteristics of commuting zones (log population; the share of females; the share of population above 65
years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral
manufacturing, the female share of manufacturing employment, exposure to Chinese imports and the share of employment
in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
A-33
Table A13: The effects of robots on employment and wages: recent periods

Excludes Excludes
Weighted by zones with Weighted by zones with
Unweighted Unweighted
population highest population highest
exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Long-differences estimates, 1990-2014
(0.071) (0.163) (0.114) (0.142) (0.379) (0.282)
Observations 722 712 722 90341 88964 90341
R-squared 0.56 0.56 0.60 0.46 0.44 0.16
Panel B. Long-differences estimates, 1990-2010
(0.068) (0.147) (0.128) (0.137) (0.291) (0.236)
Observations 722 712 722 87417 86089 87417
R-squared 0.64 0.63 0.68 0.37 0.36 0.11
Panel C. Long-differences estimates, 2000-2014
(0.055) (0.158) (0.086) (0.158) (0.441) (0.320)
Observations 722 712 722 106375 104786 106375
R-squared 0.66 0.65 0.46 0.35 0.34 0.09
Panel D. Long-differences estimates, 2000-2007
(0.062) (0.151) (0.123) (0.164) (0.353) (0.331)
Observations 722 712 722 99319 97833 99319
R-squared 0.66 0.63 0.56 0.22 0.22 0.05
Panel E. Long-differences estimates, 2000-2010
(0.057) (0.164) (0.129) (0.151) (0.378) (0.303)
Observations 722 712 722 100547 99063 100547
R-squared 0.69 0.67 0.67 0.27 0.25 0.06
Covariates:
Baseline covariates X X X X X X
The table presents estimates of the effects of exposure to robots on employment and wages for different periods of time.
Panel A presents long-differences estimates for the 1990-2014 period. Panel B presents long-differences estimates for the
1990-2010 period. Panel C presents long-differences estimates for the 2000-2014 period. Panel D presents long-differences
estimates for the 2000-2007 period. Panel E presents long-differences estimates for the 2000-2010 period. Columns 1-3
present estimates for changes in the employment to population ratio. Columns 4-6 present estimates for changes in log
hourly wages. The specifications in columns 4-6 are estimated at the demographic cell × commuting zone level, where
demographic cells are defined by age, gender, education and race. Columns 1-2 and 4-5 present regressions weighted by
population in 1990. Columns 2 and 5 present results excluding the top one percent of commuting zones with highest exposure
to robots. Columns 3 and 6 present unweighted regressions. All models include Census division dummies, demographic
characteristics of commuting zones (log population; the share of females; the share of population above 65 years; the shares
of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the
shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the
female share of manufacturing employment, exposure to Chinese imports and the share of employment in routine jobs.
A-34
Table A14: The effects of robots on manufacturing employment

Excludes zones Control for Excludes zones Control for
with highest Unweighted automotive with highest Unweighted automotive
exposure industry exposure industry
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Change in the employment to population ratio, all workers in manufacturing
(0.042) (0.091) (0.085) (0.078) (0.031) (0.074) (0.082) (0.094)
Observations 722 712 722 722 1444 1424 1444 1444
R-squared 0.77 0.78 0.76 0.78 0.56 0.56 0.60 0.57
Panel B. Change in the employment to population ratio, male workers in manufacturing
(0.031) (0.066) (0.055) (0.057) (0.018) (0.046) (0.053) (0.063)
Observations 722 712 722 722 1444 1424 1444 1444
R-squared 0.69 0.69 0.68 0.69 0.48 0.47 0.52 0.48
Panel C. Change in the employment to population ratio, female workers in manufacturing
(0.015) (0.032) (0.031) (0.030) (0.014) (0.032) (0.030) (0.041)
Observations 722 712 722 722 1444 1424 1444 1444
R-squared 0.86 0.86 0.84 0.86 0.68 0.68 0.68 0.68
Panel D. Change in the employment to population ratio, workers with no college degree in manufacturing
A-35
(0.028) (0.075) (0.075) (0.055) (0.032) (0.066) (0.072) (0.067)
Observations 722 712 722 722 1444 1424 1444 1444
R-squared 0.80 0.80 0.79 0.80 0.59 0.59 0.62 0.59
Panel E. Change in the employment to population ratio, workers with college degree or more in manufacturing
Exposure to robots 0.027 -0.010 -0.050 -0.053 -0.025 -0.017 -0.033 -0.077

(0.023) (0.022) (0.024) (0.042) (0.011) (0.018) (0.019) (0.047)
Observations 722 712 722 722 1444 1424 1444 1444
R-squared 0.45 0.46 0.29 0.48 0.27 0.27 0.15 0.28
Covariates:
Baseline covariates X X X X X X X X
Automotive industry X X
The table presents estimates of the effects of exposure to robots on the manufacturing employment to population ratio. Columns 1-4 present long-differences estimates for the
1990-2007 period. Columns 5-8 present stacked-differences estimates for the 1990-2000 and 2000-2007 period. Panel A is for changes in the manufacturing employment to population
ratio. Panel B is for changes in male manufacturing employment divided by population. Panel C is for changes in female manufacturing employment divided by population. Panel
D is for changes in manufacturing employment of workers with no college degree divided by population. Panel E is for changes in manufacturing employment of workers with a
college degree or more divided by population. Columns 1-2, 4, 5-6, and 8 present regressions weighted by population in 1990. Columns 2 and 6 present results excluding the top
one percent of commuting zones with highest exposure to robots. Columns 3 and 7 present unweighted regressions. All models include Census division dummies (and time period
dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the share of females; the share of population above 65 years;
the shares of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians),
the shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure to Chinese imports and the share of employment in
routine jobs. In addition, Columns 4 and 8 control for exposure to robots in the automotive industry. Standard errors that are robust against heteroskedasticity and correlation
Table A15: The effects of robots on employment: additional outcomes

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in log employment
(0.411) (0.885) (0.588) (0.313) (0.694) (0.388)
Observations 722 712 722 1444 1424 1444
R-squared 0.62 0.61 0.38 0.46 0.45 0.29
Panel B. Change in log manufacturing employment
(0.555) (1.219) (1.019) (0.330) (0.808) (0.881)
Observations 722 712 722 1444 1424 1444
R-squared 0.60 0.61 0.47 0.46 0.45 0.37
Panel C. Change in the employment to population ratio, including self employment
(0.056) (0.134) (0.109) (0.064) (0.162) (0.083)
Observations 722 712 722 1444 1424 1444
R-squared 0.61 0.61 0.54 0.39 0.39 0.32
Panel D. Change in the employment to population ratio, including self employment and public employment
(0.054) (0.137) (0.116) (0.066) (0.172) (0.098)
Observations 722 712 722 1444 1424 1444
R-squared 0.70 0.69 0.67 0.50 0.50 0.44
Panel E. Change in the employment to population ratio from the CBP
(0.152) (0.389) (0.174) (0.144) (0.344) (0.133)
Observations 719 709 719 1438 1418 1438
R-squared 0.49 0.47 0.45 0.48 0.45 0.38
Panel F. Change in non-participation rate
(0.040) (0.121) (0.110) (0.087) (0.189) (0.091)
Observations 722 712 722 1444 1424 1444
R-squared 0.65 0.65 0.54 0.49 0.50 0.32
Panel G. Change in unemployment rate
(0.039) (0.070) (0.049) (0.067) (0.135) (0.078)
Observations 722 712 722 1444 1424 1444
R-squared 0.61 0.59 0.52 0.49 0.41 0.32
Covariates:
The table presents estimates of the effects of exposure to robots on several labor market outcomes. Columns 1-3 present
long-differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000
and 2000-2007 periods. Panel A presents estimates for changes in log (private) employment. Panel B presents estimates
for changes in log manufacturing employment. Panel C presents estimates for changes in employment (including self
employment) to population ratio. Panel D presents estimates for changes in employment (including self employment and
public-sector employment) to population ratio. Panel E presents estimates for changes in the employment to population
ratio computed from the County Business Patterns. Panel F presents estimates for changes in the non-participation rate
(defined as the share of people above 16 years of age who are not in the labor force). Panel G presents estimates for changes
in the unemployment rate (defined as the share of people in the labor force who are not employed). Columns 1-2 and
4-5 present regressions weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent of
commuting zones with highest exposure to robots. Columns 3 and 6 present unweighted regressions. All models include
Census division dummies (and time period dummies in the stacked-differences specifications), demographic characteristics
of commuting zones (log population; the share of females; the share of population above 65 years; the shares of population
with no college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites,
Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the female share of
manufacturing employment, exposure to Chinese imports and the share of employment in routine jobs. Standard errors
that are robust against heteroskedasticity and correlation within states are in parentheses.
A-36
Table A16: The effects of robots on wages: additional outcomes

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in log weekly wages
(0.163) (0.280) (0.274) (0.187) (0.603) (0.330)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.43 0.42 0.12 0.36 0.35 0.11
Panel B. Change in log yearly wages
(0.174) (0.338) (0.309) (0.238) (0.776) (0.404)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.44 0.43 0.13 0.39 0.37 0.13
Panel C. Change in hourly wages (2007 dollars)
(0.029) (0.063) (0.037) (0.034) (0.093) (0.052)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.40 0.40 0.08 0.24 0.23 0.05
Panel D. Change in log wage bill (from the CBP)
(0.649) (1.143) (0.984) (0.521) (1.134) (0.616)
Observations 719 709 719 1438 1418 1438
R-squared 0.67 0.65 0.38 0.62 0.60 0.37
Covariates:
The table presents estimates of the effects of exposure to robots on several labor market outcomes. Columns 1-3 present
long-differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000
and 2000-2007 periods. Panel A presents estimates for changes in log weekly wages. Panel B presents estimates for changes
in log yearly wages. Panel C presents estimates for changes in hourly wages measured in 2007 dollars. The specifications
in panels A to C are estimated at the demographic cell × commuting zone level, where demographic cells are defined by
age, gender, education and race. Panel D presents estimates for changes in log wage bill by commuting zone from the
County Business Patterns. Columns 1-2 and 4-5 present regressions weighted by population in 1990. Columns 2 and 5
present results excluding the top one percent of commuting zones with highest exposure to robots. Columns 3 and 6 present
unweighted regressions. All models include Census division dummies (and time period dummies in the stacked-differences
specifications), demographic characteristics of commuting zones (log population; the share of females; the share of population
above 65 years; the shares of population with no college, some college, college and professional degrees, and masters and
doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing
and light manufacturing, the female share of manufacturing employment, exposure to Chinese imports and the share of
employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in
parentheses.
A-37
Table A17: The effects of robots on government transfers

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in total transfers per capita (2007 dollars)
(13.511) (26.480) (19.823) (9.194) (20.333) (16.546)
Observations 722 712 722 1444 1424 1444
R-squared 0.68 0.69 0.50 0.79 0.78 0.67
Panel B. Change in transfers per capita, SSA retirement benefits (2007 dollars)
(3.701) (7.382) (10.347) (2.297) (7.057) (8.053)
Observations 722 712 722 1444 1424 1444
R-squared 0.61 0.60 0.29 0.51 0.50 0.28
Panel C. Change in transfers per capita, SSA disability benefits (2007 dollars)
(1.763) (3.578) (3.284) (1.313) (3.350) (2.342)
Observations 722 712 722 1444 1424 1444
R-squared 0.70 0.70 0.56 0.66 0.65 0.55
Panel D. Change in transfers per capita, TAA benefits (2007 dollars)
(0.166) (0.158) (0.272) (0.235) (0.355) (0.362)
Observations 722 712 722 1444 1424 1444
R-squared 0.57 0.55 0.50 0.44 0.34 0.35
Panel E.Change in transfers per capita, Unemployment benefits (2007 dollars)
Exposure to robots 0.507 6.469 -1.553 11.570 15.429 11.728
(1.994) (3.545) (1.286) (1.661) (4.316) (2.080)
Observations 722 712 722 1444 1424 1444
R-squared 0.33 0.36 0.29 0.49 0.45 0.22
Panel F.Change in transfers per capita, education and training assistance (2007 dollars)
(2.352) (3.290) (2.602) (1.131) (2.326) (1.621)
Observations 722 712 722 1444 1424 1444
R-squared 0.28 0.31 0.27 0.41 0.40 0.25
Panel G. Change in transfers per capita, Medical benefits (2007 dollars)
(10.362) (23.595) (14.210) (8.919) (19.483) (12.833)
Observations 722 712 722 1444 1424 1444
R-squared 0.65 0.66 0.48 0.72 0.71 0.63
Panel H. Change in transfers per capita, Federal income assistance (2007 dollars)
Exposure to robots -9.088 -7.882 0.085 2.972 5.953 10.645
(2.744) (5.832) (3.925) (1.880) (5.421) (2.592)
Observations 722 712 722 1444 1424 1444
R-squared 0.45 0.44 0.20 0.55 0.55 0.49
Covariates:
The table presents estimates of the effects of exposure to robots on transfers (measured in dollars transferred per capita).
Each panel presents results for transfers from a different program. Columns 1-3 present long-differences estimates for the
1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns 1-2
and 4-5 present regressions weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent
of commuting zones with highest exposure to robots. Columns 3 and 6 present unweighted regressions. All models include
A-38
Table A18: The effects of robots on migration, population, and housing prices
and rents.

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Net migration rate—exclude baseline population from covariates.
Exposure to robots 0.024 0.045 0.049 -0.040 -0.058 0.059
(0.030) (0.067) (0.049) (0.028) (0.075) (0.057)
Observations 722 712 722 1444 1424 1444
R-squared 0.55 0.55 0.30 0.48 0.48 0.32
Panel B. Net migration rate—control for baseline population.
Exposure to robots -0.020 -0.030 0.002 -0.097 -0.158 -0.020
(0.029) (0.066) (0.035) (0.028) (0.069) (0.040)
Log of baseline 0.282 0.287 0.452 0.285 0.293 0.404
population (0.089) (0.094) (0.070) (0.073) (0.078) (0.049)
Observations 722 712 722 1444 1424 1444
R-squared 0.59 0.59 0.41 0.54 0.53 0.43
Panel C. Change in log population—exclude baseline population from covariates.
(0.390) (0.973) (0.630) (0.210) (0.543) (0.332)
Observations 722 712 722 1444 1424 1444
R-squared 0.58 0.58 0.43 0.56 0.55 0.40
Panel D: Change in log population—control for baseline population.
(0.391) (1.003) (0.575) (0.228) (0.534) (0.301)
Log of baseline 2.327 2.413 4.120 1.401 1.452 2.014
population (0.935) (0.996) (0.685) (0.486) (0.521) (0.342)
Observations 722 712 722 1444 1424 1444
R-squared 0.60 0.59 0.52 0.58 0.57 0.47
Panel E: Change in log house prices.
(0.856) (1.894) (0.919) (1.038) (2.893) (1.137)
Observations 601 591 601 1202 1182 1202
R-squared 0.72 0.72 0.55 0.51 0.52 0.31
Panel F: Change in log monthly rents.
(0.219) (0.493) (0.429) (0.507) (0.964) (0.506)
Observations 722 712 722 1444 1424 1444
R-squared 0.42 0.42 0.19 0.57 0.57 0.37
Covariates:
Remaining baseline
X X X X X X
covariates
The table presents estimates of the effects of exposure to robots on migration rates (obtained from the IRS), changes in log
population, and changes in house prices and rents. Columns 1-3 present long-differences estimates for the 1990-2007 period.
Columns 4-6 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns 1-2 and 4-5 present
regressions weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent of commuting
zones with highest exposure to robots. Columns 3 and 6 present unweighted regressions. Panels A and B present estimates
for the net migration rate (from IRS). Panels C and D present estimates for changes in log population. Panel E presents
estimates for changes in log house prices (from an index by the Federal Housing Finance Agency and available for counties
covering 601 commuting zones). Panel F presents estimates for changes in log rents (computed from monthly rents in the
Census and ACS, and including the cost of utilities). Columns 1-2 and 4-5 present regressions weighted by population
in 1990. Columns 2 and 5 present results excluding the top one percent of commuting zones with highest exposure to
robots. Columns 3 and 6 present unweighted regressions. All models include Census division dummies (and time period
dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the
share of females; the share of population above 65 years; the shares of population with no college, some college, college
and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the
shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure
to Chinese imports and the share of employment in routine jobs. Also, the columns in Panels B and D control for baseline
differences in log population. Standard errors that are robust against heteroskedasticity and correlation within states are
in parentheses.
A-39
Table A19: The effects of robots on wage and non-wage income

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in log wage income per capita, BEA
(0.317) (0.724) (0.703) (0.283) (0.769) (0.466)
Observations 722 712 722 1444 1424 1444
R-squared 0.53 0.49 0.46 0.47 0.42 0.32
Panel B. Change in log wage income per capita, IRS
(0.199) (0.442) (0.491) (0.252) (0.668) (0.430)
Observations 722 712 722 1444 1424 1444
R-squared 0.58 0.53 0.62 0.51 0.46 0.39
Panel C. Change in log income per capita, BEA
(0.221) (0.471) (0.479) (0.239) (0.575) (0.388)
Observations 722 712 722 1444 1424 1444
R-squared 0.55 0.50 0.31 0.40 0.33 0.20
Panel D. Change in log non-wage income per capita, BEA
Exposure to robots 0.169 -0.002 0.432 -0.203 -0.189 0.102
(0.384) (1.170) (0.817) (0.542) (1.203) (0.458)
Observations 722 712 722 1444 1424 1444
R-squared 0.47 0.47 0.08 0.22 0.23 0.05
Covariates:
The table presents estimates of the effects of exposure to robots on households’ income. Columns 1-3 present long-differences
estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000 and 2000-2007
period. Panel A presents results for changes in log wage income per capita (from the BEA). Panel B presents results for
changes in log wage income per capita (from the IRS). Panel C presents results for changes in log income per capita (from
the BEA). Panel D presents results for changes in log non-wage income per capita (from the BEA). Columns 1-2 and 4-5
present regressions weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent of
commuting zones with highest exposure to robots. Columns 3 and 6 present unweighted regressions. All models include
A-40
Table A20: Pre-trends 1970-1990: additional outcomes
Long differences, 1970-1990

Excludes zones
exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in manufacturing employment to population ratio
Exposure to robots -0.185 -0.103 0.019 0.055 -0.040 -0.118
(0.088) (0.062) (0.059) (0.050) (0.105) (0.095)
Observations 722 722 722 722 712 722
R-squared 0.44 0.54 0.57 0.63 0.63 0.36
Panel B. Change in the employment to population ratio, including public sector and self employment
(0.075) (0.066) (0.080) (0.073) (0.165) (0.149)
Observations 722 722 722 722 712 722
R-squared 0.12 0.38 0.38 0.41 0.40 0.26
Panel C. Change in non-participation rate
(0.093) (0.086) (0.109) (0.103) (0.260) (0.207)
Observations 722 722 722 722 712 722
R-squared 0.11 0.27 0.28 0.30 0.29 0.24
Panel D. Change in unemployment rate
(0.070) (0.040) (0.041) (0.040) (0.104) (0.067)
Observations 722 722 722 722 712 722
R-squared 0.35 0.55 0.56 0.59 0.58 0.43
Panel E. Change in log weekly wages
Exposure to robots -0.477 -0.310 0.305 0.343 0.640 -0.044
(0.188) (0.210) (0.256) (0.257) (0.532) (0.374)
Observations 59230 59230 59230 59230 58402 59230
R-squared 0.53 0.54 0.55 0.55 0.54 0.28
Covariates:
The table presents estimates of the effects of exposure to robots on past outcomes measured in the 1970-1990 period. For
comparison with our main results, all changes in past outcomes are rescaled to a 14-year equivalent change. Panel A presents
results for changes in manufacturing employment to population ratio 1970-1990. Panel B presents results for changes in
employment (including self employment and public-sector employment) to population ratio 1970-1990. Panel C presents
results for changes in the non-participation rate 1970-1990, defined as the share of people above 16 years of age who are not
in the labor force. Panel D presents results for changes in the unemployment rate 1970-1990, defined as the share of people in
the labor force who are not employed. Panel E present results for changes in log weekly wages 1970-1990. The specifications
for log weekly wages in Panel E are estimated at the demographic cell × commuting zone level, where demographic cells are
defined by age, gender, education and race. Column 1 only includes Census division dummies. Column 2 adds demographic
characteristics of commuting zones in 1970 (log population; the share of females; the share of population above 65 years; the
shares of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and
the shares of Whites, Blacks, Hispanics and Asians). Column 3 adds the shares of employment in manufacturing and light
manufacturing, and the female share of manufacturing employment in 1970. Columns 4-6 add exposure to Chinese imports
A-41
Table A21: Rotemberg weights
Weighted by population Unweighted

Excludes zones Excludes Excludes
Baseline Baseline
Raw data with highest automotive automotive
covariates covariates
exposure manufacturing manufacturing
(1) (2) (3) (4) (5) (6)
Panel A. Long differences, 1990-2007
Metal products .035 -.002 .035 .046 .029 .055
Basic Metals .037 -.018 .008 .03 .017 .035
Plastics and Chemicals .158 .193 .565 .949 .437 .896
Automotive .776 .877 .417 -.001 .542 0
Panel B. Stacked differences, 1990-2000 and 2000-2007

Food and Beverages 00-07 -.006 -.008 -.002 .002 0 .019
Wood and Furniture 00-07 -.005 -.008 -.011 -.004 -.003 .017
Industrial Machinery 90-00 -.002 .005 .01 .01 .007 .008
Mining 90-00 0 .005 .01 .013 .011 .022
Shipbuilding and Aerospace 90-00 .002 .001 .012 .024 .004 .015
Electronics 00-07 .013 .005 .062 .068 .032 .05
Industrial Machinery 00-07 .016 .01 .026 .023 .015 .017
Automotive 90-00 .018 -.035 -.028 0 -.018 -.001
Basic Metals 00-07 .019 .01 .036 .045 .025 .032
Metal products 00-07 .032 .025 .071 .085 .053 .085
Plastics and Chemicals 90-00 .049 .034 .104 .346 .096 .381
Plastics and Chemicals 00-07 .059 .059 .184 .393 .154 .362
Automotive 00-07 .795 .906 .546 -.001 .644 -.001
The table presents Rotemberg weights for the industries (and industries by time period) used in the construction of the
exposure to robots measure, as explained in Goldsmith-Pinkham, Sorkin and Swift (2018). Panel A present these weights for
long-differences specifications, 1990-2007, and panel B report these weights for stacked-differences specifications, 1990-2000
and 2000-2007. In both panels, we report the Rotemberg weights only for industries with a weight above 2% (or 1% in the
stacked-differences specifications) in one of our specifications. Column 1 presents Rotemberg weights for a specification with
no covariates. Columns 2 and 5 present Rotemberg weights for a specification with our baseline covariates (from column 4
in table 2). Column 3 presents Rotemberg weights for a specification that excludes the top one percent of commuting zones
with highest exposure to robots. Columns 4 and 6 present Rotemberg weights for a specification that controls for exposure
to robots in automotive manufacturing. Columns 1-4 are for regression models weighted by population. Columns 5-6 are
for unweighted models.
A-42
Table A22: The effects of robots on employment and wages controlling for trade

(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Controls for exposure to imports from Mexico
(0.059) (0.114) (0.052) (0.093) (0.131) (0.205) (0.181) (0.296)
-0.059 -0.088 0.049 0.108 0.050 0.052 0.071 0.340
Exposure to imports from Mexico
(0.056) (0.074) (0.111) (0.076) (0.153) (0.110) (0.230) (0.180)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.62 0.41 0.40 0.33 0.08 0.29 0.09
Panel B. Controls for exposure to industries with offshorable jobs
(0.065) (0.123) (0.055) (0.096) (0.124) (0.191) (0.175) (0.286)
Exposure to industries with -0.225 -0.300 -0.205 -0.196 -0.316 -0.120 -0.411 -0.183
offshorable jobs (0.072) (0.083) (0.038) (0.047) (0.141) (0.118) (0.096) (0.067)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.68 0.63 0.42 0.41 0.33 0.08 0.29 0.09
A-43
Panel C. Controls for exposure to exports from Germany, Japan, and South Korea
(0.058) (0.119) (0.046) (0.091) (0.135) (0.204) (0.171) (0.280)
Exposure to exports from 0.022 0.191 -0.013 -0.016 -0.172 -0.216 -0.025 -0.075
Germany, Japan, and South Korea (0.106) (0.103) (0.022) (0.012) (0.220) (0.175) (0.049) (0.041)
Observations 722 722 1444 1444 87100 87100 183606 183606

R-squared 0.67 0.62 0.41 0.39 0.33 0.08 0.29 0.09
Covariates:
The table presents estimates of the effects of exposure to robots and changing trade patterns on employment and wages. Columns 1-2 and 5-6 present long-differences estimates
for the 1990-2007 period. Columns 3-4 and 7-8 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns 1-4 present results for employment to
population ratio. Columns 5-8 present results for log hourly wages. The specifications in columns 5-8 for log hourly wages are estimated at the demographic cell × commuting
zone level, where demographic cells are defined by age, gender, education and race. Odd-numbered columns present regressions weighted by population in 1990. Even-numbered
columns present unweighted regressions. In Panel A we control for exposure to imports from Mexico to the US. In Panel B we control for exposure to industries with a high share of
offshorable jobs in 1990 (offshorability is measured as in Autor and Dorn, 2013). In Panel C we control for exposure to exports from Germany, Japan, and South Korea. All models
include Census division dummies (and time period dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the share of
females; the share of population above 65 years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the
shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure to
Chinese imports and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
Table A23: The effects of robots, computers and IT capital on employment and wages

(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Controls for exposure to IT-intensive industries (from ASM)
(0.061) (0.119) (0.050) (0.092) (0.130) (0.204) (0.175) (0.295)
-0.090 -0.026 0.172 0.004 0.183 -0.002 0.536 0.064
Exposure to IT-intensive industries
(0.079) (0.073) (0.128) (0.035) (0.221) (0.128) (0.189) (0.073)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.62 0.41 0.39 0.33 0.08 0.29 0.09
Panel B. Controls for exposure to computer-intensive industries (from CPS)
(0.061) (0.120) (0.051) (0.096) (0.131) (0.200) (0.180) (0.296)
Exposure to computer-intensive 0.047 0.115 0.066 0.098 0.018 0.141 0.092 0.126
industries (0.036) (0.036) (0.043) (0.023) (0.058) (0.055) (0.081) (0.046)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.62 0.41 0.40 0.33 0.08 0.29 0.09
Panel C. Controls for exposure to industry changes in computer use (from CPS)
A-44
(0.067) (0.135) (0.055) (0.103) (0.136) (0.210) (0.180) (0.291)
Exposure to industry changes in -0.226 -0.580 -0.185 -0.372 -0.105 -0.120 -0.413 -0.281
computer use (0.143) (0.154) (0.097) (0.074) (0.253) (0.216) (0.221) (0.143)
Observations 722 722 1444 1444 87100 87100 183606 183606
R-squared 0.67 0.64 0.41 0.41 0.33 0.08 0.29 0.09
Covariates:

The table presents estimates of the effects of exposure to robots and several measures of the deployment of IT technologies and computers on employment and wages. Columns 1-2
and 5-6 present long-differences estimates for the 1990-2007 period. Columns 3-4 and 7-8 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns
1-4 present results for employment to population ratio. Columns 5-8 present results for log hourly wages. The specifications in columns 5-8 for log hourly wages are estimated at
the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Odd-numbered columns present regressions weighted by
population in 1990. Even-numbered columns present unweighted regressions. In Panel A we control for a measure of exposure to IT-intensive industries, constructed by interacting
the baseline share of IT investments in each industry (available for 4-digit SIC87 manufacturing industries from the American Survey of Manufacturing) with its baseline employment
share in the commuting zone. In Panel B we control for a measure of exposure to computer-intensive industries, constructed by interacting the baseline share of employees using
a computer in each industry (obtained from the CPS and available for the 19 IFR industries used in our analysis) with its baseline employment share in the commuting zone. In
Panel C we control for a measure of exposure to changes in computer use by industry, constructed by interacting the change between 1993 and 2003 in the share of employees using
a computer in each industry (obtained from the CPS and available for the 19 IFR industries used in our analysis) with its baseline employment share in the commuting zone. All
models include Census division dummies (and time period dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the
share of females; the share of population above 65 years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral degrees;
and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment,
exposure to Chinese imports and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within states are in parentheses.
Table A24: The effects of robots on employment: alternative constructions of

exposure to robots
Estimates for changes in employment to population ratio
IV estimates,
1990-2007
Weighted by Weighted by Weighted by
population population population
exposure exposure
(1) (2) (3) (4) (5) (6) (7)
(0.059) (0.138) (0.118) (0.052) (0.150) (0.092)
US robot adoption -0.388
(0.091)
Observations 722 712 722 1444 1424 1444 722
Panel B. Exposure computed using employment shares in 1990
(0.107) (0.199) (0.220) (0.138) (0.239) (0.130)
(0.060)
Observations 722 712 722 1444 1424 1444 722
Panel C. Exposure computed using employment shares in 1980 from CBP
(0.108) (0.136) (0.104) (0.113) (0.160) (0.097)
(0.083)
Observations 722 712 722 1444 1424 1444 722
Panel D. Including penetration of robots in Germany
(0.053) (0.137) (0.106) (0.042) (0.141) (0.092)
(0.074)
Observations 722 712 722 1444 1424 1444 722
Panel E. Using penetration of robots among all European countries with data
(0.052) (0.146) (0.108) (0.042) (0.146) (0.100)
(0.067)
Observations 722 712 722 1444 1424 1444 722
Panel F. Using penetration of robots without the adjustment term
(0.048) (0.113) (0.093) (0.043) (0.126) (0.087)
(0.081)
Observations 722 712 722 1444 1424 1444 722
Panel G. Adjusting penetration of robots using robot prices
(0.065) (0.158) (0.129) (0.064) (0.172) (0.094)
(0.085)
Observations 722 712 722 1444 1424 1444 722
Covariates:
Baseline covariates X X X X X X X
The table presents estimates of the effects of exposure to robots on the employment to population ratio. Columns 1-3
present long-differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-
2000 and 2000-2007 periods. Column 7 presents IV estimates for the 1990-2007 period. Panel A presents results for the
baseline construction of the exposure to robots measure. Panel B computes exposure to robots using 1990 employment
shares from Census. Panel C computes exposure to robots using 1980 employment shares from the CBP. Panel D presents
results including the penetration of robots in Germany. Panel E presents results using the data for nine European countries
(Germany, Denmark, Finland, France, Italy, Sweden, Norway, Spain, and the UK) to construct the adjusted penetration
of robots. Panel F presents results for the penetration of robots without the adjustment term in equation (12). Panel G
presents results adjusting the penetration of robots using the available data on robot prices (from Robotics Industries of
America). Columns 1-2, 4-5, and 7 present regressions weighted by population in 1990. Columns 2 and 5 present results
excluding the top one percent of commuting zones with highest exposure to robots. Columns 3 and 6 present unweighted
regressions. All models include Census division dummies (and time period dummies in the stacked-differences specifications),
A-45
Table A25: The effects of robots on wages: alternative constructions of expo-

sure to robots
Estimates for change in log hourly wages
IV estimates,
1990-2007
Weighted by Weighted by Weighted by
population population population
exposure exposure
(1) (2) (3) (4) (5) (6) (7)
(0.132) (0.274) (0.205) (0.182) (0.551) (0.295)
(0.149)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel B. Exposure computed using employment shares in 1990
(0.252) (0.409) (0.313) (0.353) (0.717) (0.395)
(0.105)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel C. Exposure computed using employment shares in 1980 from CBP
(0.214) (0.270) (0.208) (0.317) (0.529) (0.322)
(0.152)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel D. Including penetration of robots in Germany
(0.091) (0.211) (0.155) (0.131) (0.472) (0.238)
(0.115)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel E. Using penetration of robots among all European countries with data
(0.084) (0.193) (0.141) (0.118) (0.450) (0.226)
(0.102)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel F. Using penetration of robots without the adjustment term
(0.083) (0.188) (0.135) (0.139) (0.432) (0.205)
(0.123)
Observations 87100 85776 87100 183606 180818 183606 87100
Panel G. Adjusting penetration of robots using robot prices
(0.125) (0.265) (0.203) (0.196) (0.547) (0.295)
(0.138)
Observations 87100 85776 87100 183606 180818 183606 87100
Covariates:
Baseline covariates X X X X X X X
The table presents estimates of the effects of exposure to robots on the log hourly wages. Columns 1-3 present long-
differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000 and
2000-2007 periods. Column 7 presents IV estimates for the 1990-2007 period. The specifications are estimated at the
demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Panel
A presents results for the baseline construction of the exposure to robots measure. Panel B computes exposure to robots
using 1990 employment shares from Census. Panel C computes exposure to robots using 1980 employment shares from the
CBP. Panel D presents results including the penetration of robots in Germany. Panel E presents results using the data for
nine European countries (Germany, Denmark, Finland, France, Italy, Sweden, Norway, Spain, and the UK) to construct
the adjusted penetration of robots. Panel F presents results for the penetration of robots without the adjustment term in
equation (12). Panel G presents results adjusting the penetration of robots using the available data on robot prices (from
Robotics Industries of America). Columns 1-2, 4-5, and 7 present regressions weighted by population in 1990. Columns 2
and 5 present results excluding the top one percent of commuting zones with highest exposure to robots. Columns 3 and
6 present unweighted regressions. All models include Census division dummies (and time period dummies in the stacked-
differences specifications), demographic characteristics of commuting zones (log population; the share of females; the share
of population above 65 years; the shares of population with no college, some college, college and professional degrees,
and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in
manufacturing and light manufacturing, the female share of manufacturing employment, exposure to Chinese imports and
the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity and correlation within
states are in parentheses.
A-46
Table A26: The effects of robots on employment and wages: the role of outliers
Excludes
Weighted by Excludes observations Robust Median
population Detroit with large regression regression
residuals
(1) (2) (3) (4) (5)
Panel A. Long differences for employment to population ratio, 1990-2007
Exposure to robots -0.448 -0.486 -0.451 -0.493 -0.457
(0.059) (0.103) (0.060) (0.090) (0.115)
Observations 722 721 686 722 722
R-squared 0.67 0.66 0.71 0.65
Panel B. Stacked differences for employment to population ratio, 1990-2000 and 2000-2007
(0.052) (0.113) (0.050) (0.079) (0.106)
Observations 1444 1442 1351 1444 1444
R-squared 0.41 0.40 0.45 0.40
Panel C. Long differences for log hourly wages, 1990-2007
(0.132) (0.190) (0.132) (0.133) (0.129)
Observations 87100 86860 81492 87101 87101
R-squared 0.33 0.32 0.38 0.01
Panel D. Stacked differences for log hourly wages, 1990-2000 and 2000-2007
(0.182) (0.388) (0.180) (0.117) (0.108)
Observations 183606 183124 171897 183607 183607
R-squared 0.29 0.27 0.33 0.03
Covariates:
Baseline covariates X X X X X
estimates for changes in the employment to population ratio, 1990-2007. Panel B presents stacked-differences estimates for
changes in the employment to population ratio, 1990-2000 and 2000-2007. Panel C presents long-differences estimates for
changes in log hourly wages, 1990-2007. Panel D presents stacked-differences estimates for changes in log hourly wages,
1990-2000 and 2000-2007. The specifications in panels C-D are estimated at the demographic cell × commuting zone level,
where demographic cells are defined by age, gender, education and race. Columns 1-3 present regressions weighted by
population in 1990. Column 2 excludes Detroit from the sample. Column 3 excludes observations with a residual above or
below two estimated standard deviations in column 1 and re-estimates the model. Column 4 presents a robust regression
as in Li (1985). Column 5 presents a median regression. All models include Census division dummies (and time period
dummies in the stacked-differences specifications), demographic characteristics of commuting zones (log population; the
share of females; the share of population above 65 years; the shares of population with no college, some college, college
and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the
shares of employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure
to Chinese imports and the share of employment in routine jobs. Standard errors that are robust against heteroskedasticity
and correlation within states are in parentheses.
A-47
Table A27: The effects of robots on employment and wages: additional covari-
ates

exposure exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in the employment to population ratio—controlling for state fixed effects
(0.078) (0.130) (0.098) (0.067) (0.167) (0.094)
Observations 722 712 722 1444 1424 1444
R-squared 0.80 0.79 0.73 0.45 0.44 0.47
Panel B. Change in the employment to population ratio—controlling for baseline employment level
(0.059) (0.106) (0.127) (0.067) (0.147) (0.098)
Observations 722 712 722 1444 1424 1444
R-squared 0.75 0.75 0.68 0.55 0.54 0.45
Panel C. Change in the employment to population ratio—controlling for changes in demographics
(0.058) (0.137) (0.116) (0.049) (0.153) (0.086)
Observations 722 712 722 1444 1424 1444
R-squared 0.75 0.74 0.73 0.56 0.55 0.52
Panel D. Change in log hourly wages—controlling for state fixed effects
(0.154) (0.236) (0.161) (0.216) (0.558) (0.322)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.34 0.34 0.09 0.29 0.28 0.09
Panel E. Change in log hourly wages—controlling for baseline employment level
(0.151) (0.290) (0.233) (0.195) (0.558) (0.316)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.34 0.33 0.08 0.29 0.28 0.09
Panel F. Change in log hourly wages—controlling for changes in demographics
(0.131) (0.267) (0.199) (0.186) (0.531) (0.287)
Observations 87100 85776 87100 183606 180818 183606
R-squared 0.34 0.33 0.08 0.29 0.28 0.09
Covariates:
The table presents estimates of the effects of exposure to robots on employment and wages controlling for additional
covariates. Columns 1-3 present long-differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences
estimates for the 1990-2000 and 2000-2007 periods. Panel A presents estimates for the change in employment to population
ratio controlling for state fixed effects. Panel B presents estimates for the change in employment to population ratio
controlling for baseline employment levels. Panel C presents estimates for the change in employment to population ratio
controlling for changes in demographic characteristics of commuting zones. Panel D presents estimates for the change in log
hourly wages controlling for state fixed effects. Panel E presents estimates for the change in log hourly wages controlling
for baseline employment levels. Panel F presents estimates for the change in log hourly wages controlling for changes in
demographic characteristics of commuting zones. The specifications in panels D-F are estimated at the demographic cell ×
commuting zone level, where demographic cells are defined by age, gender, education and race. Columns 1-2 and 4-5 present
regressions weighted by population in 1990. Columns 2 and 5 present results excluding the top one percent of commuting
zones with highest exposure to robots. Columns 3 and 6 present unweighted regressions. Besides the covariates mentioned
above, all models include Census division dummies (and time period dummies in the stacked-differences specifications),
A-48
Table A28: The effects of robots on employment and wages: covariates selected by LASSO
Long differences, 1990-2007 Stacked differences 1990-2000 and 2000-2007

Covariates included in Lasso procedure: Covariates included in Lasso procedure:
Two-way Two-way
Baseline Extended Extended Baseline Extended Extended
Interactions Interactions
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Change in the employment to population ratio, all workers
(0.059) (0.055) (0.148) (0.064) (0.074) (0.065) (0.196) (0.060)
Observations 722 722 722 722 1444 1444 1444 1444
R-squared 0.62 0.66 0.64 0.64 0.36 0.39 0.41 0.65
Panel B. Change in log hourly wages
(0.114) (0.109) (0.380) (0.095) (0.181) (0.183) (0.780) (0.139)
Observations 87100 87100 87100 87100 183606 183606 183606 183606
A-49
R-squared 0.33 0.33 0.33 0.33 0.28 0.28 0.29 0.30
LASSO + control LASSO+control
Specification details: LASSO LASSO for automotive LASSO LASSO LASSO for automotive LASSO
industry industry
The table presents estimates of the effects of exposure to robots on employment and wages, where the set of covariates is selected using a LASSO procedure (see Belloni et al.

2014). Columns 1-3 present long-differences estimates for the 1990-2007 period. Columns 4-6 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Panel
A presents results for changes in the employment to population ratio. Panel B presents results for changes in log hourly wages. The specifications in this panel are estimated at
the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. The covariates included in each specification are selected
using LASSO, following Belloni et al. (2014). In columns 1 and 5, we start with the baseline set of covariates from Tables 2 and 3 (demographics, industry shares, Chinese import
competition and share routine jobs) and select the subset to be included in the models using LASSO. In columns 2-3 and 6-7, we include all covariates used in the main robustness
checks of the paper (exposure to declining industries, trade, offshoring, and capital deepening), as well as an expanded set of industry employment shares for mining, construction,
agriculture, utilities, and research and development, and select the subset to be included in the models using LASSO. In columns 3 and 7, we separate the role of exposure to
robots in the automotive industry from the role of robots outside this industry. In these models, we include the covariates selected by LASSO as predictors of any of these exposure
measures or the outcome. Finally, in columns 4 and 8, we include all potential two-way interactions between our baseline covariates (including Census division and time dummies)
and the extended set of covariates. We then select the subset of covariates to be included in our model using LASSO. Standard errors that are robust against heteroskedasticity
and correlation within states are in parentheses.
Table A29: The effects of robots on employment and wages by gender
Long differences Stacked differences
Men Women Men Women
Excludes zones Excludes zones Excludes zones Excludes zones
Weighted by Weighted by Weighted by Weighted by
with highest Unweighted with highest Unweighted with highest Unweighted with highest Unweighted
population population population population
exposure exposure exposure exposure
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Panel A. Change in the employment to population ratio
Exposure to robots -0.567 -0.674 -0.652 -0.336 -0.473 -0.385 -0.684 -0.809 -0.972 -0.418 -0.590 -0.515
(0.065) (0.147) (0.148) (0.063) (0.145) (0.112) (0.065) (0.175) (0.134) (0.049) (0.133) (0.079)
Observations 722 712 722 722 712 722 1444 1424 1444 1444 1424 1444
R-squared 0.62 0.59 0.59 0.71 0.71 0.58 0.37 0.35 0.37 0.45 0.45 0.37
Panel B. Change in log of hourly wages
(0.148) (0.291) (0.241) (0.137) (0.306) (0.223) (0.211) (0.601) (0.320) (0.158) (0.512) (0.314)
Observations 43599 42935 43599 43501 42841 43501 92008 90612 92008 91598 90206 91598
R-squared 0.29 0.28 0.07 0.34 0.33 0.08 0.25 0.24 0.08 0.31 0.30 0.11
Panel C. Change in manufacturing employment to population ratio
(0.064) (0.137) (0.112) (0.029) (0.061) (0.062) (0.037) (0.092) (0.105) (0.027) (0.064) (0.060)
Observations 722 712 722 722 712 722 1444 1424 1444 1444 1424 1444
R-squared 0.70 0.70 0.69 0.85 0.85 0.83 0.49 0.48 0.52 0.67 0.67 0.67
A-50
Panel D. Change in non-participation rate
Exposure to robots 0.364 0.276 0.398 0.201 0.205 0.263 0.658 0.951 0.850 0.318 0.578 0.375
(0.063) (0.119) (0.129) (0.036) (0.144) (0.112) (0.096) (0.198) (0.117) (0.082) (0.188) (0.095)
Observations 722 712 722 722 712 722 1444 1424 1444 1444 1424 1444
R-squared 0.51 0.48 0.40 0.71 0.72 0.58 0.55 0.55 0.37 0.43 0.43 0.34
Covariates:
Baseline covariates X X X X X X X X X X X X

The table presents estimates of the effects of exposure to robots on employment, wages and other outcomes by gender. Columns 1-6 present long-differences estimates for the
1990-2007 period. Columns 7-12 present stacked-differences estimates for the 1990-2000 and 2000-2007 periods. Columns 1-3 and 7-9 present results for men. Columns 4-6 and
10-12 present results for women. Panel A presents results for the employment to population ratio. Panel B presents results for log hourly wages. The specifications in this panel
are estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Panel C presents results for manufacturing
employment to population ratios. Panel D presents results for the non-participation rate (defined as the share of the population above 16 years of age who are not in the labor
force). Columns 1-2, 4-5, 7-8 and 10-11 present regressions weighted by population in 1990. Columns 2, 5, 8 and 11 present results excluding the top one percent of commuting
zones with highest exposure to robots. Columns 3, 6, 9 and 12 present unweighted regressions. All models include Census division dummies (and time period dummies in the
stacked-differences specifications), demographic characteristics of commuting zones (log population; the share of females; the share of population above 65 years; the shares of
population with no college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of
employment in manufacturing and light manufacturing, the female share of manufacturing employment, exposure to Chinese imports and the share of employment in routine jobs.
Table A30: Alternative Inference following Borusyak et al. (2018): IV estimates

(1) (2) (3) (4) (5) (6)
(0.116) (0.121) (0.125) (0.269) (0.237) (0.238)
Panel B. Alternative imputation of US data, 1990-2007
(0.120) (0.126) (0.130) (0.279) (0.246) (0.246)
Panel C. Long-differences, 1990-2014
(0.111) (0.055) (0.062) (0.129) (0.113) (0.127)
Panel D. Long-differences, 2000-2007
(0.084) (0.055) (0.063) (0.240) (0.064) (0.069)
Panel E. Long-differences, 2000-2014
(0.147) (0.021) (0.023) (0.078) (0.052) (0.054)
Covariates:
Demographics and
X X X X
industry shares
The table presents IV estimates of the effects of exposure to robots on employment and wages for different time periods.
Panels A and B present results for 1990-2007. Panel C presents results for 1990-2014. Panel D presents results for 2000-2007.
Panel E presents results for 2000-2014. In all models, we instrument the US exposure to robots using exposure to robots
from EU RO5. In Panels A and C-E we rescale the US exposure to robots to match the time period used. In Panel B we
use an alternative imputation strategy for US exposure to robots described in the text. Columns 1-3 present results for
the employment to population ratio. Columns 4-6 present results for log hourly wages. The specifications for log hourly
wages are estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender,
education and race. All IV estimates are from regressions weighted by population in 1990. The covariates included in each
model are reported at the bottom of the table. Columns 1 and 4 only include Census division dummies. Columns 2 and 5
add demographic characteristics of commuting zones (log population; the share of females; the share of population above
65 years; the shares of population with no college, some college, college and professional degrees, and masters and doctoral
manufacturing, and the female share of manufacturing employment. Columns 3 and 6 add exposure to Chinese imports
and the share of employment in routine jobs. Robust standard errors computed following Borusyak et al. (2018).
A-51
Table A31: Alternative Inference following Borusyak et al. (2018): long dif-
ferences
Long differences 1990-2007

Excludes zones
exposure
(1) (2) (3) (4) (5) (6)
Panel A. Change in the employment to population ratio, 1990-2007
(0.053) (0.086) (0.046) (0.043) (0.181) (0.076)
R-squared 0.23 0.12 0.72 0.76 0.28 0.36
Panel B. Change in log hourly wages, 1990-2007
(0.075) (0.092) (0.093) (0.090) (0.180) (0.107)
R-squared 0.70 0.62 0.80 0.80 0.26 0.41
Covariates:
estimates for changes in the employment to population ratio, 1990-2007. Panel B presents long-differences estimates for
changes in log hourly wages, 1990-2007. The specifications in Panel B are estimated at the demographic cell × commuting
zone level, where demographic cells are defined by age, gender, education and race. Columns 1-5 present regressions weighted
by population in 1990. Column 5 presents results excluding the top one percent of commuting zones with highest exposure
to robots. Column 6 presents unweighted regressions. The covariates included in each model are reported at the bottom
of the table. Column 1 only includes Census division dummies. Column 2 adds demographic characteristics of commuting
zones in 1990 (log population; the share of females; the share of population above 65 years; the shares of population with no
college, some college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks,
Hispanics and Asians). Column 3 adds the shares of employment in manufacturing and light manufacturing, and the female
share of manufacturing employment in 1990. Columns 4-6 add exposure to Chinese imports and the share of employment
in routine jobs. Robust standard errors computed following Borusyak et al. (2018).
A-52
Table A32: Alternative Inference following Borusyak et al. (2018): stacked differences
Stacked differences 1990-2000 and 2000-2007

Commuting zone trends
Excludes zones
Weighted by
Weighted by population with highest Unweighted Unweighted
population
exposure
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Change in the employment to population ratio, 1990-2000 and 2000-2007
(0.059) (0.064) (0.069) (0.075) (0.269) (0.107) (0.042) (0.145)
Observations 38 38 38 38 38 38 38 38
R-squared 0.42 0.38 0.40 0.47 0.05 0.48 0.89 0.79
Panel B. Change in log hourly wages, 1990-2000 and 2000-2007
(0.104) (0.124) (0.155) (0.167) (0.574) (0.350) (0.268) (0.689)
Observations 38 38 38 38 38 38 38 38
R-squared 0.65 0.63 0.62 0.66 0.26 0.53 0.79 0.79
A-53
Covariates:
Time period dummies X X X X X X X X
Demographics X X X X X X X
Industry shares X X X X X X

Trade, routine jobs X X X X X
Commuting zone
X X
trends
The table presents estimates of the effects of exposure to robots on employment and wages. Panel A presents stacked-differences estimates for changes in the employment to
population ratio, 1990-2000 and 2000-2007. Panel B presents stacked-differences estimates for changes in log hourly wages, 1990-2000 and 2000-2007. The specifications in Panel
B are estimated at the demographic cell × commuting zone level, where demographic cells are defined by age, gender, education and race. Columns 1-5 and 7 present regressions
weighted by population in 1990. Column 5 presents results excluding the top one percent of commuting zones with highest exposure to robots. Columns 6 and 8 present unweighted
regressions. The covariates included in each model are reported at the bottom of the table. Column 1 only includes Census division dummies and time period dummies. Column 2
adds demographic characteristics of commuting zones (log population; the share of females; the share of population above 65 years; the shares of population with no college, some
college, college and professional degrees, and masters and doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians). Column 3 adds the shares of employment in
manufacturing and light manufacturing, and the female share of manufacturing employment. Columns 4-6 add exposure to Chinese imports and the share of employment in routine
jobs. In addition, columns 7 and 8 include a full set of commuting zone fixed effects. Robust standard errors computed following Borusyak et al. (2018).
Table A33: OLS estimates of the relationship between US exposure to robots

and employment and wages

(1) (2) (3) (4) (5) (6)
Panel A. Long differences between 1990 and 2007
(0.064) (0.039) (0.042) (0.084) (0.085) (0.088)
Observations 722 722 722 87100 87100 87100
R-squared 0.26 0.65 0.65 0.32 0.33 0.33
Panel B. Long differences between 1990 and 2007, alternative imputation of US exposure
(0.066) (0.040) (0.044) (0.087) (0.088) (0.091)
Observations 722 722 722 87100 87100 87100
R-squared 0.26 0.65 0.65 0.32 0.33 0.33
Panel C. Long differences between 1990 and 2014
(0.103) (0.049) (0.052) (0.121) (0.102) (0.106)
Observations 722 722 722 90341 90341 90341
R-squared 0.30 0.55 0.55 0.42 0.46 0.46
Panel D. Long differences between 2000 and 2007
(0.123) (0.074) (0.074) (0.173) (0.137) (0.140)
Observations 722 722 722 99319 99319 99319
R-squared 0.37 0.67 0.67 0.21 0.22 0.22
Panel E. Long differences between 2000 and 2014
(0.163) (0.052) (0.052) (0.136) (0.144) (0.146)
Observations 722 722 722 106375 106375 106375
R-squared 0.37 0.66 0.66 0.33 0.35 0.35
Covariates:
Demographics and
X X X X
industry shares
The table presents OLS estimates corresponding to the IV models in Table 7. Different panels cover different periods, with
the US exposure measure rescaled to match the relevant period. Columns 1-3 present long-differences estimates for changes
in the employment to population ratio. Columns 4-6 present long-differences estimates for changes in log hourly wages.
The specifications in columns 4-6 are estimated at the demographic cell × commuting zone level, where demographic cells
are defined by age, gender, education and race. All regressions are weighted by population in 1990. The covariates included
in each model are reported at the bottom of the table. Columns 1 and 4 only include Census division dummies. Columns
2 and 5 add demographic characteristics of commuting zones (log population; the share of females; the share of population
above 65 years; the shares of population with no college, some college, college and professional degrees, and masters and
doctoral degrees; and the shares of Whites, Blacks, Hispanics and Asians), the shares of employment in manufacturing and
light manufacturing, and the female share of manufacturing employment. Columns 3 and 6 add exposure to Chinese imports
A-54
Table A34: Quantitative results for different parametrizations of our model

γM
Changing value of π0 Changing value of γL
γM γM γM
π0 =0.1 π0 =0.3 π0 =0.5 γL =2 γL =3 γL =4
η = 0.55 η = 0.79 η = 1.21 η = 1.33 η = 0.79 η = 0.52
Baseline ψ, σλ:
ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17
d ln L = -0.77 d ln L = -0.56 d ln L = -0.36 d ln L = -0.65 d ln L = -0.56 d ln L = -0.48
d ln W = -0.59 d ln W = -0.42 d ln W = -0.26 d ln W = -0.49 d ln W = -0.42 d ln W = -0.36
Changing ψ
η = 0.55 η = 0.79 η = 1.21 η = 1.33 η = 0.79 η = 0.52
ψ=0
ε = 0.19 ε = 0.19 ε = 0.19 ε = 0.19 ε = 0.19 ε = 0.19
η = 0.55 η = 0.79 η = 1.21 η = 1.33 η = 0.79 η = 0.52

ψ = 0.1
ε = 0.09 ε = 0.09 ε = 0.09 ε = 0.09 ε = 0.09 ε = 0.09
η = 0.55 η = 0.79 η = 1.21 η = 1.33 η = 0.79 η = 0.52

ψ = 0.15
ε = 0.04 ε = 0.04 ε = 0.04 ε = 0.04 ε = 0.04 ε = 0.04
Changing σ:
η = 0.54 η = 0.75 η = 1.12 η = 1.29 η = 0.75 η = 0.48
σ = 0.5
ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17
η = 0.56 η = 0.83 η = 1.30 η = 1.37 η = 0.83 η = 0.56

σ = 1.5
ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17
Changing λ:
η = 0.58 η = 0.89 η = 1.44 η = 1.59 η = 0.89 η = 0.54
λ = 2.5
ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17
η = 0.54 η = 0.75 η = 1.13 η = 1.23 η = 0.75 η = 0.51

λ = 7.5
ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17 ε = 0.17
The table presents our estimates for η and ε, and the aggregate decline on employment and wages predicted by our model
under different parametrizations. Columns vary the values of π0 and γγM used, as reported at the top of the table. Rows
L
vary the values of ψ, σ and λ used, as indicated in the left column.
A-55

Robots and Jobs (Acemoglu 2019)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Robots and Jobs (Acemoglu 2019)

Uploaded by

Copyright:

Available Formats

Manuscript Click here to

Robots and Jobs: Evidence from US Labor

Daron Acemoglu Pascual Restrepo

June 28, 2019

Journal of Political Economy

2 Robots, Employment and Wages: A Model

2.1 The Effects of Robots in Autarky Equilibrium

in fixed proportions so that

Xci = α−α (1 − α)−(1−α) Aci [ min {xci (s)}]α Kci

We prove in the Appendix that an equilibrium exists and is unique.

Proposition 1 Suppose that πc > 0. Then

where Lci denotes the employment in industry i in commuting zone c.

where ζ disp = (1−α+η)/Λ, ζ prod = (1+η)/Λ, ζc,L

2.2 The Effects of Robots When Commuting Zones Trade

(Ccφ Sc1−φ )1−ψ − 1 B

Proposition 3 Suppose that πc = π0 for all c ∈ C and θi = 0 for all i ∈ I. Then

where the ζ̄’s are functions of the underlying parameters.

2.3 Empirical Specification

dθi γL dMi dYi Mi

d ln Lc =βL · US exposure to robotsc + L W

where AP Ri is the adjusted penetration of robots computed from European countries. We

3.2 Industry Data

3.3 Commuting Zone Data and Exposure to Robots

5 The Effect of Robots across Committing Zones

5.1 Exposure to Robots and Robot-Related Activities

5.2 Reduced-form Results for Employment and Wages

5.3 Other Labor Market Outcomes

5.4 Threats to Validity

5.5 Robots, Capital and Other Technologies

5.6 Other Robustness Checks

5.7 Effects by Industry, Occupation, Gender and Skill

6 IV Estimates and Local and Aggregate Implications

and Worker Skills,” The Quarterly Journal of Economics, 122(4): 1721–1758.

Industrial Robotics: Technology, Programming and Applications, Mcgraw-Hill Inc.

Figure 1: Industrial robots in the United States and Europe.

Figure 2: Adjusted penetration of robots in the US and EU RO5 by industry.

Figure 4: Geographic distribution of exposure to robots 1993-2007

Figure 5: Exposure to robots and the location of robot integrators.

Figure 7: Pre-trends in employment and wages.

Figure 9: Effects of robots on employment and wages by education and gender.

Figure 10: The effects of robots on the wage distribution.

Journal of Political Economy

Table 2: The effects of robots on employment and wages: long differences

Long differences 1990-2007

Stacked differences 1990-2000 and 2000-2007

Journal of Political Economy

Change in the employment to population ratio Change in log hourly wages

Journal of Political Economy

Long differences 1990-2007 Stacked differences 1990-2000 and 2000-2007

Change in the employment to population ratio Change in log hourly wages

Journal of Political Economy

Table 7: The effects of robots on employment and wages: IV estimates

Online Appendix (Not for Publication)

Wc = BCcψ Lεc . (A1)

Market clearing implies

From the production function for robots we also have

RcM = D−1−η (1 + η)−η Mcη . (A4)

Next, define the share of labor in production tasks in industry i as

Cobb-Douglas technology in (2)), and sL

Similarly, the demand for robots can be expressed as

and the demand for capital is

• factor supplies satisfy equations (A3) and (A4);

Proposition A1 An equilibrium of the autarky model exists and is unique.

Proofs of Propositions from Section 2.1

d ln Lc =βL · US exposure to robotsc + L W