American Economic Review 2018

The paper should be at least 30 pages, and this is a research paper and the professor told me to use the EBSCOhost website for the citation and sources no other sources. I have also attached the Template that how it should be The topic is Technological changes and their effects on students.

American Economic Review 2018, 108(7): 1737–1772
https://doi.org/10.1257/aer.20161570
1737
* Hershbein: W.E. Upjohn Institute for Employment Research, 300 S Westnedge Avenue, Kalamazoo
49007, and IZA (email: [e-mail]); Kahn: Yale School of Management, 165 Whitney Avenue, PO
Box 208200, New Haven, CT 06511, NBER, and IZA (email: [e-mail]). This paper was accepted to
the AER under the guidance of Hilary Hoynes, Coeditor. We are grateful to Jason Abaluck, Joe Altonji, Dav
Autor, Tim Bartik, David Berger, Nick Bloom, Jeff Clemens, David Deming, David Green, John Horton, Peter
Kuhn, Fabian Lange, Steve Malliaris, Adrien Matray, Alicia Sasser Modestino, Daniel Shoag, Henry Siu, Basit
Zafar, four anonymous referees, and seminar participants at the 2016 AEAs, the Atlanta Fed, Boston
Boston University, Brookings, Georgetown University, Harvard, the LSE, MIT, NBER Summer Institute (Macro
and Labor), Northwestern, Princeton, Rutgers, SOLE/EALE 2015 meetings, Trans Pacific Labor Seminar 2015,
University of British Columbia, University of Chicago, University of Illinois, University of Kansas, University of
Michigan, University of Notre Dame, University of Pennsylvania, University of Pittsburgh, University of South
Florida, and the WEAI 2016 meetings. We are especially indebted to Dan Restuccia, Jake Sherman, and Bled
for providing the Burning Glass data. Jing Cai provided excellent research assistance with CPS data. A previo
version of this paper was titled “Is College the New High School? Evidence from Vacancy Postings.”
† Go to https://doi.org/10.1257/aer.20161570 to visit the article page for additional materials and
disclosure statement(s).
Do Recessions Accelerate Routine-Biased Technological
Change? Evidence from Vacancy Postings†
By Brad Hershbein and Lisa B. Kahn*
We show that skill requirements in job vacancy postings differentially
increased in MSAs that were hit hard by the Great Recession, relative
to lesshard-hit areas. These increases persist through at least the
end of 2015 and are correlated with increases in capital investments,
both at the MSA andfirm levels. We also find that effects are most
pronounced in routine-cognitive occupations, which exhibit relative
wage growth as well. We argue that this evidence is consistent with
the restructuring of production toward routine-biased technologies
and the more-skilled workers that complement them, and that the
Great Recession accelerated this process. (JEL E24, E32, J24, J31,
J63, L23, O33)
The employment shift from occupations in the middle of the skill distribution
toward those in the tails is one of the most important trends in the US labor ma
ket over the last 30 years. Previous research makes the compelling case that a
mary driver of this job polarization isroutine-biased technological change (RBTC),
whereby new machine technologies and overseas labor substitute for middle-s
jobs in the United States and are in turn complementary to high-skill co
jobs.1 Until recently, RBTC had been thought to be a gradual, secular phenomen
However, a long theoretical literature, beginning with Schumpeter’s (1939) “c
ative destruction,” suggests adjustments to technological change may be more
sodic. In boom times, high opportunity costs, or frictions such as adjustment co
may inhibit resources from being reallocated optimally in the face of technolog
1See, for example, the seminal work of Autor, Levy, and Murnane (2003); Goos and Manning (2007); Auto
Katz, and Kearney (2008); and Autor and Dorn (2013).

1738 THE AMERICAN ECONOMIC REVIEW JULY 2018
change. Recessions lower the opportunity cost and can produce large enough s
to overcome these frictions.2
Whether adjustments to new technology are smooth or lumpy is important fo
policy and for our understanding of recoveries. The recoveries from the last thr
US recessions (1991, 2001, 2007–2009) have been jobless: employment was s
to rebound despite recovery in aggregate output. The reasons for jobless recov
are not well understood, but a small theoretical literature points to adjustment
as a potential mechanism, since they can generate reallocation that is concent
downturns (Koenders and Rogerson 2005; Berger 2012; Jaimovich and Siu 2012
Such lumpy adjustment may leave a mass of displaced workers with the
skills for new production. Jaimovich and Siu (2015) provide suggestive evidenc
that countercyclical reallocation, in the form of RBTC, and jobless recove
linked. They show that the vast majority of the declines inmiddle-skill employment
have occurred during recessions and that, over the same time period, recovery
jobless only in these occupations. However, there is still relatively little direct e
dence on how firms restructure employment in the face of technological chang
and, in particular, whether this restructuring is gradual or episodic.3
In this paper we investigate how the demand for skills changed over the Grea
Recession. We use a new dataset collected by Burning Glass Technologie
contains thenear-universe of electronically posted job vacancies in 2007 an
2010–2015. Exploiting spatial variation in economic conditions, we establi
new fact: the skill requirements of job ads increase in metropolitan statistical a
(MSAs) that suffered larger employment shocks in the Great Recession, relative
the same areas before the shock and other areas that experienced smaller sho
Our estimates imply that ads posted in a hard-hit metro area are about 5 perc
points (16 percent) more likely to contain education and experience requireme
and about 2–3 percentage points ( 8–12 percent) more likely to state requirem
for cognitive and computer skills. Moreover, the vast majority of this “upskilling
persists through the end of our sample in 2015. That is, even while most meas
local labor-market strength had converged back to pre-recession levels, differ
in advertised skill demands remain. This is true holding constant a rich set of c
trols for the availability of skilled labor and the composition of ads across firms
occupations. In fact, we find that the very firms that upskilled early in the recov
drive the persistence later in our sample period.
These patterns collectively raise the possibility that a structural shift in
demand for skill occurred disproportionately in harder-hit MSAs. In particular, t
skill requirements we explore (education, experience, cognitive, and computer
known to complement routine-biased technologies (Autor, Levy, and Murnane
Brynjolfsson and McAfee 2011). If a structural shift in line with RBTC is occur-
ring, we would expect changes in these skill requirements to be accompanied b
2Many theoretical papers predict this phenomenon. See, for example, Hall (1991, 2005); Caballer
Hammour (1994, 1996); Mortensen and Pissarides (1994); Gomes, Greenwood, and Rebelo (2001); and Koen
and Rogerson (2005).
3For example, Acemoglu and Restrepo (2017) find that the diffusion of industrial robots across US commu
zones reduced aggregate employment and wages; Harrigan, Reshef, and Toubal (2016) show that job polari
tion was more pronounced in French firms with greater shares oftechnology-related occupations; and Hawkins,
Michaels, and Oh (2015) show that capital investments and employment reductions frequently occur togethe
Korean manufacturing plants, but these papers focus on long-run changes.

1739HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
accelerated adoption of such technologies, as well. Indeed, we find that increas
in skill requirements are correlated with capital investments at both the MSA a
firm levels. First, using the Ci Technology Database from Harte-Hanks, a marke
intelligence firm, we show that harder-hit MSAs exhibited a relative increase in
investments, as measured by the adoption of personal computers, at the same
as they upskilled in job postings. These differences across MSAs emerge only a
the Great Recession and, once again, persist through our sample period. Secon
link firms in our job postings database to those in the Harte-Hanks database, a
as to publicly traded firms in Compustat. We show that the firms increasing the
capital investments, based on PC adoption and physical capital holdings, are al
more likely to upskill. Thus, increased demand for labor skill appears closely lin
to both general and IT capital investment.
If this increased investment is in fact related to routine-biased techno
we would expect to see the strongest changes to labor characteristics for the jo
most susceptible to such technologies, routine ones. We thus additionally
on different types of routine occupations, exploring joint changes in skill requir
ments, employment, and wages. Following Acemoglu and Autor (2011), we dis
guish routine-cognitive occupations (e.g., clerical, administrative, and sales) fr
routine-manual ones (e.g., production and operatives), and we supplement the
ads data with Current Population Survey (CPS) and Occupational Employm
Statistics (OES) data. For routine-manual occupations, we see evidence consis
with firms’ substitution of technology for labor: a sharp increase in layoff risk fo
harder-hit MSAs early on, and persistently depressed employment, with n
ticular impact on skill requirements. This is the traditional view exhibited in the
polarization literature: employment losses concentrated in occupations we exp
to be most readily replaceable by machines. Consistent with Jaimovich a
(2015), we show that these changes also appear to be episodic around the Gre
Recession. However, in contrast to this conventional view of labor substi
routine-cognitive occupations in harder-hit MSAs surprisingly exhibit only mod
est increases in layoff risk and no relative employment losses. Instead, we show
that these occupations experience pronounced upskilling, as well as modest re
wage and employment growth after the recession. That is, rather than disappe
entirely, surviving routine-cognitive occupations appear to have become both
tively higher-skilled and more productive. These occupations thus became epi
cally less routine, and more cognitive, as a result of the Great Recession.
Taken together, our results suggest that firms located in areas more s
affected by the Great Recession were induced to restructure their production to
greater use of technology and higher-skilled workers; that is, the Great Recess
hastened the polarization of the US labor market.
This paper is related to a number of important literatures. First, we pr
evidence that the Great Recession spurred persistent changes in labor in
a manner consistent with technological change. Several classes of model
adjustment costs can rationalize this result. For instance, firms may make prod
tivity-enhancing improvements in a recession because of a decline in the oppo
nity cost of restructuring (Hall 2005), a shift in managerial attention from grow
to efficiency possibly due to an increased risk of closure (Koenders and Rogers
2005; Gibbons and Roberts 2012), or changes in the costs and benefits of mak

1740 THE AMERICAN ECONOMIC REVIEW JULY 2018
layoffs (Berger 2012; Jaimovich and Siu 2012). In addition, recessions may driv
Schumpeterian cleansing, whereby older, less-productive firms die, making wa
newer, more-productive firms. Empirical support for adjustment-cost mod
focused on the impacts of competition or trade shocks on productivity. For exa
Bloom, Draca, and Van Reenen (2016) show that increased Chinese import com
tition in Europe led to technological change within firms.4 Our paper adds to this
literature by highlighting recession-induced changes in the firm-level demand
skill. This may have important consequences for labor market recoveries, since
implies potential for a sudden skill mismatch.
Second, the Burning Glass job postings data provide a unique opportunity to
sure changes in skill requirements both across and within occupations. In contr
the extant literature on job polarization has focused on shifts across occupation
and has therefore been unable to ascertain the importance of the intra-occupa
margin. We show that the bulk of upskilling occurs within occupations, suggest
this margin is quite important. Moreover, our finding that upskilling is concentr
within routine-cognitive occupations and is accompanied by relative wage grow
implies that RBTC occurs both within and across occupations. This result helps
clarify work by Beaudry, Green, and Sand (2014, 2016) and others documentin
“great reversal” in demand for cognitive skill. They show that since 2000, cogn
occupations have seen no gains in employment or wages, and that college grad
have become more likely to work in routine occupations than previously. They
that a decrease in demand for cognitive occupations drove college graduates t
jobs lower in the occupational distribution, squeezing out the high school gradu
who formerly held them. This is something of a puzzle, especially given the com
mon belief that technological change continues and the fact that more-skilled
ers still earn a sizable premium in the labor market (Card, Heining, and Kline 2
Card, Cardoso, and Kline 2016). We hypothesize part of the solution to this puz
that cognitive workers are being drawn into (formerly) routine-task occupation
the skill content of these occupations evolves. These changes make the occupa
more skilled and therefore likely more desirable than before, although probabl
not as desirable as traditional high-skilled jobs.5
Third, we contribute to a growing literature exploiting data on vacancy
ings. Although several studies have used aggregate vacancy data, and even va
microdata, from the Bureau of Labor Statistics’ Job Openings and Labor Market
Turnover (JOLTS) survey (Davis, Faberman, and Haltiwanger 2012, 2013), thes
data contain little information on the characteristics of a given vacancy or the fi
that is posting it. Fewer studies have used vacancy data that contain informatio
on the occupation or specific requirements of the job posted, and these have g
ally used narrow slices of the data (Rothwell 2014), or data that are limited to o
vacancy source (Kuhn and Shen 2013; Marinescu 2017). To our knowledg
4Additionally, Nickell (1996) provides evidence that increased competition is associated with faster total
tor productivity growth; Syverson (2004a, b) shows that productivity is higher in industries and geographies
greater substitutability of products across firms; and Bernard, Redding, and Schott (2011) show firms shift t
higher productivity products upon the liberalization of firm trade. Other examples are cited in each of these
5Our analyses, however, do not explain why employment and wages have not grown in high-skill occupa
Deming (2017) proposes a compelling hypothesis that a rising importance of social skills, especially in conju
with cognitive skills, can help account for this fact.

1741HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
are the first study to use data based on anear-universe of online job postings that
covers every metropolitan area in the United States. Online job vacancies repre
but one slice of the labor market, and, by their nature, will overrepresent grow
firms (Davis, Faberman, and Haltiwanger 2013). Nonetheless, we show that lin
vacancies to data on employment, wages, and capital investments, the last at
firm level, presents consistent evidence on how labor markets changed followin
the Great Recession.
We demonstrate that during the Great Recession firms changed not only who
they would hire in the recovery, but how they would produce. Instead of occurr
gradually, with relatively few workers needing to be reallocated at any given ti
we find support that changes in demand for skill were episodic, resulting in a sw
of displaced workers whose skills were suddenly rendered obsolete as fir
cheted up their requirements. The need to reallocate workers on such a large s
may help drive jobless recoveries. It also has distributional consequences
that low-skill workers are well known to suffer worse employment and wage co
quences in recessions.6 Finally, this type of episodic reallocation likely plays a role
in the well-noted and marked decline in male employment-to-population ratio
the past 25 years, especially since these declines have beenstair-step around reces-
sions (Moffitt 2012).7 The evidence provided in this paper is thus integral for under
standing worker reallocation, and can help inform policymakers about the optim
mix during a downturn of worker retraining and subsidizing job search th
unemployment insurance.
The remainder of this paper proceeds as follows. Section I introduces the dat
while Section II summarizes our methodology. Section III presents new fa
on changes in skill requirements as a function of local labor market con
Section IVinvestigateshow thesechangesare linked to capitalinvestments.
Section V examines cross-occupation heterogeneity in response to local labor
ket shocks on skill requirements, employment, and wages, with a particular foc
routine occupations. Section VI concludes.
I. Data
Our data come from a unique source: microdata from nearly 100 million elec
tronic job postings in the United States that span the Great Recession (
2007 and 2015). These job postings were collected and assembled by Burning
Technologies, an employment analytics and labor market information firm. In t
section, we describe the data and our particular sample construction. We provi
detailed examination of the sample’s characteristics and representativeness in
Appendix A.
6See von Wachter and Handwerker (2008); Hoynes, Miller, and Schaller (2012); and Forsythe (2016).
7Supporting the notion that episodic restructuring drives stair-step declines in male employment, Foote a
Ryan (2014) point out that middle-skill workers, the most vulnerable to RBTC, are most at risk of leaving the
force when unemployed.

1742 THE AMERICAN ECONOMIC REVIEW JULY 2018
A. Burning Glass Overview
Burning Glass Technologies—henceforth, BG or Burning Glass—examines som
40,000 online job boards and company websites to aggregate the job postings,
and deduplicate them into a systematic,machine-readable form, and create labor
market analytic products. Thanks to the breadth of this coverage, BG believes
resulting database captures a near-universe of jobs that were posted online. T
a special agreement, we obtained these posting-level data for the years 2007
from 2010 through 2015, covering every MSA in the United States.8
The two key advantages of our data are its breadth and detail. The broad cov
age of the database presents a substantial strength over datasets based on a s
vacancy source, such as CareerBuilder.com. While the JOLTS asks a nationally
resentative sample of employers about vacancies they wish to fill in the near te
it is typically available only at aggregated levels, and contains relatively little in
mation about the characteristics of vacancies. In contrast, the BG data contain
70 possible standardized fields for each vacancy. We exploit detailed informati
occupation, geography, skill requirements, and firm identifiers. The codified sk
include stated education and experience requirements, as well as thousands of
cific skills standardized from open text in each job posting.9 The data thus allow for
analysis of a key, but largely unexplored, margin of firm demand: skill requirem
within occupation.10 Moreover, they allow for afirm-level analysis, which, as we
show below, is key to disentangling mechanisms for upskilling.
However, the richness of the BG data comes with a few shortcomings. Notab
the database covers only vacancies posted on the internet. First, Davis, Faberm
and Haltiwanger (2013) show that the distribution of vacancies in JOLTS overre
resents growing firms. Although roughly two-thirds of hiring is replacement hir
(Lazear and Spletzer 2012), vacancies in general will be somewhat skewed tow
certain areas of the economy. Second, even though vacancies for availab
have increasingly appeared online instead of in traditional sources, such as new
papers, one may worry that the types of jobs posted online are not representat
of all openings. In online Appendix A, we provide a detailed description o
industry-occupation mix of vacancies in BG relative to other sources (JOLTS, th
Current Population Survey, and Occupational Employment Statistics), an analy
8Our dataset was provided in February 2016. Although BG’s algorithms for removing duplicates and codin
ad characteristics change over time, each iteration is applied to all postings in the data. The database unfort
lacks postings from 2008 and 2009. These years would be useful for completeness and for understanding th
timing over which skill requirements changed; however, since hiring volume fell byone-third in 2008 and did not
begin to recover until 2010 (per JOLTS), and our focus is on longer-term changes in hiring demand, addition
for the recession years are not integral for this paper. We also have data on jobs posted in Micropolitan Stat
Areas, which we do not use for lack of some of the labor market indicators in these areas, and substantial no
the ones that are available. They represent 5.6 percent of all posted ads.
9For example, an ad might ask for a worker who is bilingual or who can organize and manage a team. BG
and codes these and other skills into a taxonomy of thousands of unique, but standardized requirements. Be
with a set of predefined possible skills, BG searches text in an ad for an indication that the skill is required.
example, for team work, they search for the key words “team work” but also look for variations such as “abi
work as a team.”
10Other private-sector firms, such as Wanted Analytics, used by the Conference Board’s Help-Wanted On
Index, also offer disaggregated data, but not skill requirements. State vacancy surveys, conducted by a limit
ber of states, sometimes collect certain skill requirements, but cover only a few geographic areas and are ge
not comparable across states (Carnevale, Jayasundera, and Repnikov 2014; Rothwell 2014).

1743HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
of how it has changed over our sample period, and various validity checks cond
on the data both by us and by other researchers. To briefly summarize, althoug
postings are disproportionately concentrated in occupations and industries tha
ically require greater skill, the distributions are relatively stable across time, an
the aggregate and industry trends in the quantity of vacancies track other sour
reasonably closely.
Another downside of the BG data is that vacancies represent just one
by which a firm may adjust labor inputs: through stated, but not necessarily rea
ized, demand. For a complete picture, one would also like to see hires, separat
wages, and other measures (e.g., incumbent worker training, recruitment inten
(Davis, Faberman, and Haltiwanger 2013)). Thus, we also provide corroborating
evidence on some of these margins using supplemental datasets, as described
We restrict our main BG sample to ads withnon-missing employers that posted
at least 10 ads over the sample period of 2007 and 2010–2015. Employer nam
missing in 40 percent of postings, primarily from those listed on recruiting web
that typically do not reveal the employer.11 Many of our analyses exploitfirm-level
information to distinguish among possible mechanisms for upskilling. We there
choose to focus our entire analysis on the consistent sample of ads with non-m
firms, with a sufficient number of observations per firm to estimate firm-level c
acteristics. However, we have performed analyses not requiring firm-level info
tion on the full dataset and obtain very similar results. Moreover, we have confi
that the probability of satisfying this sample criterion (having a valid firm ident
does not vary over the business cycle (see online Appendix A.8). Thus, our sam
restriction should not confound the estimated relationship between local labor
ket conditions and the skill requirements of postings.
B. Skill Requirements in Burning Glass
In our analyses, we exploit four categories of skill requirements: stated educ
and experience requirements, stated demand for skills that we classify as “cog
tive,” and stated demand for computer skills. We choose these skill requiremen
for two reasons. First, they represent a broad swath of human capital measures
which both employers and economists have interest. Second, they reflect what
economics discipline has learned about technological change over the past 20
(Autor, Levy, and Murnane 2003; Brynjolfsson and McAfee 2011). In particular,
the RBTC literature emphasizes that new information technology or cheap over
labor substitute for routine, algorithmic, middle-skill tasks. These new technolo
are in turn complementary with high-skill cognitive, abstract tasks.12 High-skilled
workers favored by RBTC may be required to work with computers and perform
11When name is available, Burning Glass uses a proprietary algorithm to group name variants into a stan
set: for example, “Bausch and Lomb,” “Bausch Lomb,” and “Bausch & Lomb” would be grouped together. W
perform some additional cleaning on firm name, removing any remaining punctuation, spaces, and a few pro
atic words, such as “Incorporated” or “Inc.” The 10-ad restriction drops about 4 percent of job ads that list a
name. However, employer names with very few ads are likely to be miscoded (for example, capturing a frag
of the city name).
12This literature finds also that RBTC may indirectly affect low-skill, manual tasks (Autor and Dorn 2013)
downside of the BG sample is that low-skill jobs are underrepresented. We thus focus our analysis on the de
which employers shift demand from medium- toward high-skill tasks and workers.

1744 THE AMERICAN ECONOMIC REVIEW JULY 2018
more versatile set of functions. Indeed, thenon-algorithmic tasks that complement
routine-task performing machines or overseas labor involve more complexity,
lem solving, and analytical skills, and the ability to determine which tasks need
be performed at a given moment.
In accord with human capital theory, we believe more-educated workers or t
with greater experience on the job will be better able to perform these function13
In online Appendix A.3, to cross-validate the data, we show that education req
ments strongly correlate with average education levels of employed workers at
MSA and occupation levels.
We categorize cognitive and computer skill requirements based on the open
fields for skills. We designate an ad as requiring computer skills if it contains th
key word “computer” or it is categorized as software by BG.14 We define cognitive
skill requirements based on a set of key words chosen deliberately to match
non-routine analytical job tasks used in Autor, Levy, and Murnane (2003) and s
sequently used by the majority of papers studying RBTC and polarization. We a
ensure that the presence of these key words correlates with external measures
cognitive skill at the occupation level.15
This set of skills (education, experience, cognitive, and computer) aligns wel
with our priors on how jobs change with the availability of computers (Brynjolfs
and McAfee 2011). For example, a sales person who previously devoted most o
or her energy to client relations may now be required to use data analytics to b
ter target packages to clients. This salesperson now needs computer and analy
skills, and some experience in the field may help in mapping data recommenda
to practice. Similarly, thanks to machine vision technology, a quality control op
ator no longer need spend his or her time measuring and identifying the shape
produced goods, but instead can be diverted to other tasks such as troubleshoo
and making judgment calls in design optimization. This set of tasks requires hig
cognitive function and intuition that can be gained by experience.16
Table 1 summarizes data for the primary regression sample.17 In 2007, 34 percent
of the weighted ads list any education requirement (column 1, row 1). Among a
with an education requirement, one-half (17 percent of all ads) specify minimu
education of a bachelor’s degree, another one-quarter ask for a high school dip
13In the raw data, there are two fields each for education and experience requirements: a minimum level
or years of experience) and a preferred level. Postings that do not list an education or experience requireme
these fields set to missing. We use the fields for the minimum levels to generate variables for the presence o
cation or experience requirement as well as the number of years of education or experience required; the m
is much more commonly specified than the preferred, and it is always available when a preferred level is list
14BG includes common software (e.g., Excel, PowerPoint, AutoCAD), as well as less common software and
languages (e.g., Java, SQL, Python).
15Specifically, an ad is categorized as requesting a cognitive skill if any listed skills include at least one of
the following phrases or fragments: “research,” “analy,” “decision,” “solving,” “math,” “statistic,” or “thinki
The fraction of ads at the occupation level that contain each of these skills is strongly correlated with an O*N
measure developed by Deming (2017) meant to categorize cognitive occupations. We obtain this measure f
Deming and Kahn (2018), who categorize a wide range of key words found in the BG job ads into 10 general
including cognitive.
16It has been suggested by Deming (2017) and others that technology complements workers with interpe
skills, since machines are still poor at reading and inferring human emotion. We have also analyzed changes
demand for a composite “social” skill requirement and obtained results very similar to those presented here
cognitive and computer skills.
17In the top two panels, observations are weighted as they are in our regression analyses: we give equal
to ads within an MSA-year, but upweight larger MSAs, based on the size of the labor force in 2006.

1745HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
Table 1—Summary Statistics
Mean (SD)
2007 2010–2015 Change
Panel A. Ad characteristics
Education requirements
Any 0.34 0.57 0.23
(0.06) (0.05)
HS 0.09 0.20 0.10
(0.03) (0.05)
BA 0.17 0.27 0.10
(0.05) (0.08)
>BA 0.03 0.05 0.02
(0.01) (0.01)
Years, conditional on any 14.84 14.67 −0.18
(0.40) (0.44)
Experience requirements
Any 0.32 0.52 0.20
(0.06) (0.07)
0–3 0.13 0.24 0.11
(0.03) (0.03)
3–5 0.14 0.21 0.07
(0.03) (0.04)
>5 0.05 0.08 0.03
(0.02) (0.04)
Years, conditional on any 3.52 3.34 −0.18
(0.47) (0.54)
Skill requirements
Any stated skills 0.73 0.91 0.18
(0.05) (0.04)
Cognitive, conditional on any 0.22 0.34 0.11
(0.05) (0.06)
Computer, conditional on any 0.28 0.39 0.11
(0.06) (0.08)
Panel B. Share of ads in 2010–2015 matching to 2007 and to other datasets
Missing ACS match 0.08
Continuing firm 0.65
In Harte-Hanks, among continuing 0.78
In Compustat, among continuing 0.40
Mean Min Max
Panel C. Cell counts
Number MSAs 381
Posts per MSA-year 21,779 132 1,231,417
Number occupations ( four-digit) 108
Posts per occupation-MSA-year 228 1 194,558
Number firms 170,809
Posts per Firm-MSA-year 13 1 16,413
Notes: Burning Glass data 2007 and 2010–2015. All changes are statistically significant at the
1 percent level. Sample is restricted to ads withnon-missing firms that posted at least ten ads
over our sample period. In the top panel, observations are weighted by the size of the MSA
labor force in 2006. Missing ACS match is the share of weighted observations to MSAs that
cannot be matched to the American Community Survey (weighted by the MSA labor force).
Continuing firms are the fraction of 2010–2015 observations posted by a firm that also posted
in 2007. In Harte Hanks (Compustat) among continuing firms are the share of weighted obser-
vations that post to a firm that can be matched to Harte Hanks (Compustat). All three statistics
are calculated weighting by the firm’s ad share in theMSA-year times the size of the MSA
labor force in 2006.

1746 THE AMERICAN ECONOMIC REVIEW JULY 2018
and the remainder are roughly evenly split between associate degrees (not sho
master’s degrees, and professional degrees or PhDs. Converting the degrees to
modal equivalent years of schooling, the average education requirement, cond
on one being specified, is nearly 15 years.
The second column shows skill requirements averaged over the period 2010
The third column shows the within-MSA change in skill requirements across th
two sample periods. The share of ads specifying an education requirement incr
by 23 percentage points (ppts), on average. This is roughly evenly split across
requiring high school and ads requiring college; because the proportional incre
is slightly larger for high school, the overall (conditional) years of schooling fall
slightly. All differences in means are statistically significant at the 1 percent lev
Experience requirements follow a very similar pattern to education requirem
In 2007, almost one-third of ads specify some amount of experience in the fiel
Among ads with a requirement, the vast majority ask for less than five years, w
much of the remainder asking for between 5 and 10 years. Conditional on post
experience requirement, the average ad asks for 3.5 years. In the later time pe
the propensity to specify an experience requirement increases by 20 ppts. The
increases are again concentrated in the lower categories, so that the average,
tional on specifying any requirement, falls by about one-fifth of a year.
Finally, in 2007, 73 percent of weighted ads specify at least one specific, tex
skill requirement. Among these, 22 percent specify a cognitive skill requiremen
and 28 percent have a computer requirement. In 2010–2015, 91 percent
have at least one text-based skill requirement, and the shares specifying cogn
skills or computer skills increase to roughly one-third and two-fifths, respectiv
In regression analyses, we use the probability of posting a cognitive or comput
skill requirement, conditional on posting a specifictext-based skill, as dependent
variables, rather than the unconditional probabilities, which might instead pick
tendency for ads to become more verbose as posting costs decline.
These increases in stated skill demand could be driven by the nationa
sion that took place between 2007 and the 2010–2015 period. However, they c
also be driven by a variety of other factors, such as changing composition of fir
posting ads online orpreexisting national trends. Because of these issues and the
relatively short panel we have to work with, our regression analyses always con
for year dummies. We therefore fully absorb the overall change in skill requirem
illustrated in Table 1. Instead, we identify differences in the change in skill requ
ments across metro areas as a function of how they weathered the Great Rece
The bottom panel of Table 1 provides an idea of our sample coverage
have a balanced panel of 381 MSAs, which contain an (unweighted) average of
21,779 posts per MSA-year. When we disaggregate to the four-digit occu
level, we have 108 occupations represented, with an average of 228 posts in e
occupation-MSA-year.18 Finally, our data contain roughly 171,000 unique firms,
which translate into an average of 14 posts in each firm-MSA-year.
18Though occupation is available in the BG data at the six-digit Standard Occupation Classification (SOC)
level, we restrict our attention to comparisons across and within four-digit SOC codes, which provid
ads per occupation-MSA-year cell and ensure a balanced panel of occupation-MSAs across years in nearly a
cases. Virtually all ads posted in the 2010–2015 period are in occupation-MSAs that also posted in
within-occupation analyses, we drop the 0.36 percent of ads that cannot be matched.

1747HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
II. Methodology
Our goal is to understand how the Great Recession affected the demand for s
Because we have only a short panel and need to worry about concurren
that may have affected online job ads (e.g., utilization, prices, preexisting nati
trends in upskilling), we exploit cross-sectional geographic variation in the sev
of the Great Recession. Our general approach is to examine temporal changes
skill requirements as a function of an MSA-level employment shock generated
the Great Recession.
Our initial regression specification is shown in equation (1). The term outcomgmt
is any of several different measures associated with changes in labor skill dem
(and eventually changes in other production inputs, as discussed later) in MSAm ,
year t , and sometimes in subgroupg (for example, occupation or firm). The left-hand
side is the difference in the outcome variable between 2007 and year t . The r
sion sample thus includes each post-recession year t ∈ [2010, 2015] . Finally,m
is a measure of the local employment shock generated by the Great Recessiont
are year dummies, controls are additional control variables described in more
below, and εgmt is an error term:
(1) outcom egmt − outcom egm2007 = α0 + [shockm × It ] α1 + It + controls + εgmt .
The variable shoc km is fixed at theMSA-level for our entire sample period; we
describe its construction in detail below. Through an exhaustive set ofshoc km -year
interactions, the regression estimates the impact of the local employment shoc
the change in skill requirements (or other outcomes) for a given MSA (and grou
between 2007 and a subsequent year. The difference specification implicitly co
trols for time-invariant factors at the MSA (or group-MSA) level. We use 2007
the base year in most analyses since this is the only pre-recession year availab
BG. Such a specification allows us to empirically investigate the timing and per
sistence of upskilling in relation to local labor market shocks through the vecto
coefficients, α1 . The inclusion of year fixed effects ( It ) means we identify the key
coefficients purely off of differences across metro areas in the employment sho
rather than relying on the national shock itself.
We cluster standard errors by MSA to address possible serial correlation with
an area. For regressions at the MSA-year level, we weight observations by the
the MSA’s labor force in 2006. This weighting scheme allows us to upweight are
with larger populations, helping with precision, while fixing the weight applied t
each MSA-year. The latter ensures that we identify off of the same MSA weight
mix in each year, regardless of any (endogenous) changes in ads posted. When
further disaggregate to the MSA-year-group level, we weight cells by the produ
the 2006 MSA labor force and the group’s observation share withinMSA-year (the
observation shares sum to unity), so that more aggregate regressions produce
identical to those using more disaggregated data when the underlying specifica
is the same.
The key explanatory variable,shoc km , is theMSA-specific change in projected
annual employment growth between 2006 and 2009, the national peak and tro
years surrounding the Great Recession. We project employment growth in an M

1748 THE AMERICAN ECONOMIC REVIEW JULY 2018
based on its employment shares inthree-digit North American Industry Classification
System (NAICS) industry codes averaged over 2004 and 2005 and national em
ment changes at the three-digit industry level. This type ofshift-share method is
sometimes referred to as a Bartik shock, following the strategy of Bartik (199119
Specifically, we define projected employment growth, Δ Eˆmt in equation (2),
where for K three-digit industries, ϕ is the employment share of industry k
at time τ (in practice, τ is the average of 2004 and 2005), ln Ekt is the log of national
employment in industry k in year t , and ln Ekt−1 is the log of national employment in
the industry one year prior:20
(2) Δ Eˆmt =∑
k=1
K
ϕm, k, τ (ln Ekt − ln Ek, t−1 ), shoc km = Δ Eˆm2009 − Δ Eˆm2006 .
We then define shoc km as the change in projected employment growth from peak
trough (2006 to 2009). The calculated values of shoc km range from about −0.12 to
−0.04 across MSAs, but to make the coefficients easier to interpret, we renorm
this variable so that a one unit change is equal to the difference between the te
and ninetieth percentile MSAs, −0.026 log points; a larger value corresponds to
worse economic shock.
We use this Bartik measure, instead of actual employment growth, for two re
sons. First, actual employment growth at the MSA level is measured with subst
error, while the Bartik measure allows for more precision. Second, actual emplo
ment growth will reflect shocks to labor demand as well as other city-s
shocks, including those to labor supply, which may be problematic.21 We note that
other direct measures of local labor market tightness, such as the local unemp
ment rate, have similar shortcomings in terms of measurement error or revers
sality; for instance, an unemployment rate may be high precisely because a su
demand shift toward more-skilled labor generates structural mismatch. We ex
the robustness of our results to other ways of defining the Bartik shock.
The top left panel of Figure 1 summarizes the relationship between the Barti
employment shock and actual annual (log) employment growth at the MSA lev
(obtained from the BLS State and Metro Area Employment program). We estim
equation (1), which nets out the actual employment growth rate in 2007 throug
the differences specification, for 2000–2015, controlling only for year fixed effe
The coefficients, α1 , thus represent difference-in-differences estimates: the cha
in actual employment growth between a given year t and 2007, for a hard-hi
(ninetieth percentile employment shock) relative to a less hard-hit MSA (tenth
centile employment shock).
19Some other papers utilizing a form of Bartik shock include Blanchard and Katz (1992), Notowidigdo (20
and Yagan (2016).
20We obtain seasonally adjusted national employment for each three-digitindustry-month from the BLS Current
Employment Statistics program, and take an unweighted average over months to obtain Ekt . We construct ϕ using
County Business Patterns data and the algorithm of Isserman and Westervelt (2006) to overcome data supp
the resulting county-level statistics are mapped to MSAs using the definitions provided by the Census Burea
set by the Office of Management and Budget. See http://www.census.gov/population/metro/data/def.html.
21For example, MSAs with secular increases in population due to migration flows may experience employ
changes that are higher than average but still have a weakening labor market. The Bartik shock addresses t

1749HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
We plot the coefficients α1 and 95 percent confidence-interval bars (results fo
Figure 1 are also displayed in columns 1–4 of online Appendix Table C1). For ex
ple, the point estimate of −0.038 in 2009 (top left panel of Figure 1) indicates
one-unit change in the Bartik shock is associated with an additional 0.038 log
drop in employment growth between 2007 and 2009. The difference between 2
and 2009, which corresponds to our Bartik shock definition, that is associated w
one-unit increase in shockm is −0.053 (= −0.038 − 0.015). This is roughly double
the Bartik90–10 gap of −0.026 associated with a one unit change ofshoc km used in
the regression. The actual BLS variables are likely substantially noisier than pro
employment growth and also are influenced by other factors, such as supply sh
The figure also shows that the shock is episodic, such that employment grow
(relative to that in 2007) looks similar across MSAs early in the decade, regardl
of the size of the shock they will eventually face in the Great Recession. Hard-
MSAs peak slightly higher than less hard-hit MSAs in 2005 and 2006, then exp
ence a sharp dip in employment growth from 2007–2010, followed by a recove22
The Bartik shock measure is also highly correlated with movement in the une
ployment rate (obtained from the BLS Local Area Unemployment Statistics pro-
gram). The top-right panel of Figure 1 shows that a hard-hit MSA experiences a
additional 2percentage point increase in the unemployment rate from 2007 to 20
relative to a less hard-hit MSA. Again, areas look very similar in the period befo
the recession, and converge a few years after the recession ends.
22We use the 2006-2009 differential because these are the peak and trough years of our Bartik shock. As
seen in Figure 1, actual employment growth in hard-hit MSAs peaks slightly earlier, in 2005, but remains alm
the same magnitude in 2006.
Figure 1. Labor Market Variables and theMSA-Specific Employment Shock
Notes: We regress the MSA-level change in local labor market variables from 2007 on an exhaustive set of M
employment shock-by-year interactions, controlling for year fixed effects (see equation (1)). Graph plots the
cients on Bartik shock × year, as well as 95 percent CI bars. Unemployment and employment growth rates a
the BLS. Employment-to-population ratios (Epops) are author calculations based on the CPS.
−0.04
−0.02
0
0.02
0.04
2000 2005 2010 2015
Employment growth rate
−0.01
0
0.01
0.02
0.03
2000 2005 2010 2015
Unemployment rate
−0.05
−0.02
0
0.02
0.05
2000 2005 2010 2015
Epop some college or more
−0.05
−0.02
0
0.02
0.05
2000 2005 2010 2015
Epop HS or less
CoefficientCoefficient
Coefficient Coefficient

1750 THE AMERICAN ECONOMIC REVIEW JULY 2018
Our primary regressions of interest involve changes in skill requirements wit
MSA using data that begin in 2007. Although thefirst-difference specification nets
out differences across MSAs in the level of posted skill requirements, we
not control for (or observe) any preexisting trends within MSAs in skill demand.
Identification may be threatened if, for example, preexisting trends in up
were more prevalent in MSAs with industry mixes that would make them more
less susceptible to the demand shock.
The bottom panels of Figure 1 help speak to this concern by examinin
employment-to-population ratios (epop) by education group. We calculate the
variables by MSA using CPS microdata, so they are naturally a bit noisier than t
variables in the top panel (see online Appendix A.7 for details about this sampl
construction). The epops for workers with at least some college (bottom left) a
for workers with a high school diploma or less (bottom right) are fairly similar
across MSAs before the Great Recession. This is shown in the figure by
estimates that are small in magnitude and generally statistically indistinguisha
from zero prior to 2007. As with employment growth, the college epop does pe
slightly higher in 2005 and 2006 for MSAs that will experience a worse shock;
additionally, the epop forless-educated workers fared modestly better in these
MSAs in 2000. However, differences are small and do not appear to be systema
trends.
After 2007, both education groups experience relative drops in epops,
that for college workers is shallow and recovers quickly. The decline in epop fo
less-educated workers is both more severe and more sluggish to recover, and
of roughly 1 ppt still remains in 2015.23 This lack of convergence may suggest
that harder-hit areas had not fully recovered from the Great Recession
We revisit this lack of complete recovery below when we discuss our pr
mechanism.
To alleviate remaining concerns about differential pre-trends, we control, wh
specified, for a wide range of MSA characteristics (including demographics, edu
tional attainment, and economic indicators) obtained from the American Comm
Survey (ACS), averaging years 2005 and 2006.24 These controls help adjust for dif-
ferences across MSAs in their preexisting tendency to upskill, to the extent tha
a tendency is correlated with the skill distribution of the population or the healt
its economy before the Great Recession.
To summarize, we find that our constructed Bartik employment shock is epis
although it is highly correlated with changes in employment growth rates and t
unemployment rate during the Great Recession, the shock is not correlat
23This finding is consistent with Yagan (2016), who uses IRS tax data to show that while unemployment ra
had converged across US commuting zones following the Great Recession, employment probabilities
holding constant a rich set of worker characteristics.
24We chose years just prior to the Great Recession that allow for MSA identification (prior to 2005, the AC
lacks substate identifiers). Specifically, we include the share of the population that is female, black
Asian, married, migrated in the last year, is a high school dropout, has exactly a high school diploma, has so
college, has exactly a bachelor’s degree, is enrolled in school, is less than age 18, is age 19–29, is age 30–39
40–49, and is age 50–64. We also control for theemployment-to-population ratio and the average weekly wage of
full-time workers. We can match all but 8 percent of weighted ads to the ACS (see the middle panel of Table
unmatched ads consisting of small MSAs not identifiable in the ACS. In such cases, we set the ACS controls t
and include an indicator for not matching. In online Appendix B.2, we also include specifications that add con
for changes since 2000 in these variables.

1751HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
these labor market fundamentals before the Great Recession or several years i
the recovery. As we cannot observe skill requirements before 2007, this is reas
ing: the lack of pre-trends in the labor market variables in Figure 1, and for oth
described below, makes it less likely that areas were differentially trending in s
demand before 2007. Below, we explore the relationship between the shock m
and a range of additional labor market variables.
III. Skill Requirements and Local Employment Conditions
A. Main Results
Figure 2 summarizes regression results from equation (1) for our four
dependent variables: the change in the share of ads posting any education req
ment, any experience requirement, any cognitive requirement, and any compu
requirement (results are also displayed in columns 5–8 of online Appendix Tabl
C1). The figures plot the estimated impact of the Bartik shock on the change in
requirements for each year, relative to 2007 (coefficients α1 ), as well as 95 percent
confidence intervals. We use our preferred specification, which includes contro
MSA characteristics and year fixed effects.
Beginning with the top left panel, we find that, relative to 2007 levels, the pr
bility of specifying any education requirement increases by 5.4 ppts in 2010 fo
MSA experiencing a large employment shock (ninetieth percentile) compared t
MSA experiencing a small shock (tenth percentile). This difference-in-differenc
estimate implies an increase of 16 percent of the average requirement in 2007
significant at the 1 percent level. The effect persists at fairly similar magnitude
significance levels for subsequent years, with a small dip in 2012. In 2015, we
mate that the probability of posting an education requirement is still 4.1 ppts la
than it was in 2007 for a hard-hit MSA, compared to a less hard-hit one. That i
76 percent of the initial upskilling effect in 2010 remains five years later. Estim
in each year except 2012 are significant at the 1 percent level.
The remaining panels of Figure 2 exhibit remarkably similar patterns in both
nitudes and statistical significance. The probability of listing an experience req
ment increases by 5.0 ppts (16 percent) between 2007 and 2010, and 85 perce
this increase remains in 2015. The probability of listing a cognitive requi
increases by 2.0 percentage points (12 percent), and this gap widens slightly b
2015. Finally, the probability of listing a computer skill requirement also increa
by roughly 2 ppts and remains elevated through 2015.
These patterns are in stark contrast to the labor market variables in F
For employment growth, the unemployment rate, and the epop for college wor
hard-hit MSAs experience a severe impact of the Great Recession that fully rec
ers within our sample time frame. For illustration, compare Detroit and Pittsbur
The former, ahard-hit MSA, experienced a shock at about the ninetieth percentile
while the latter was at roughly the tenth percentile. Both MSAs had similar skill
requirements in 2007; for example, in both areas about one-third of ads
education requirement. By 2010, skill requirements increased in both MSAs, bu
Detroit’s (actual, not predicted) increase was nearly 10 ppts larger for educatio
and experience requirements and 2–4 ppts larger for cognitive and computer s

1752 THE AMERICAN ECONOMIC REVIEW JULY 2018
requirements. While unemployment rates had converged back topre-recession lev-
els in both MSAs, Detroit’s elevated skill requirements persisted through 2015.
Figure 2 demonstrates that the case of Detroit and Pittsburgh is not isolated
systematic. In terms of their skill requirements, MSAs that looked similar before
Great Recession look quite different from each other in 2015, several years aft
Great Recession ended.
To better understand the mechanisms underlying Figure 2, Table 2 pro
regression results forwithin-occupation changes in skill requirements. In general,
the distribution of postings across high- andlow-skilled jobs may vary for a variety
of reasons. For example, differential job survival, use ofword-of-mouth in recruit-
ing, or time-to-fill across skill groups might generate patterns observed in Figu
especially early in the recovery. However, we find that the primary driver of the
patterns is increased skill requirements within similar types of jobs.
Each column of Table 2 summarizes a separate regression of equation (1), at
occupation-MSA-year level, including MSA characteristics and year fixed effect
The results show significant upskilling effects of a magnitude comparable to th
overall MSA-level effects. For example, within occupation, the propensity to po
an education requirement increases by5.3 ppts in a hard-hit MSA, relative to a less
hard-hit MSA, between 2007 and 2010. Although there is a temporary dip in 20
at least three-quarters of this effect persists from 2013 through 2015. Similar
terns obtain for the remaining skill requirements. Indeed, these within-occupa
increases in skill requirements completely account for the MSA-level ups
effects found in Figure 2; our upskilling results are not driven at all by changes
the occupation mix of postings. (This does not preclude variation in effects acro
0
0.02
0.04
0.06
0.08
2007 2009 2011 2013 2015
Education requirement
0
0.02
0.04
0.06
0.08
2007 2009 2011 2013 2015
Experience requirement
0
0.01
0.02
0.03
0.04
0.05
2007 2009 2011 2013 2015
Cognitive skill requirement
−0.02
0
0.02
0.04
0.06
2007 2009 2011 2013 2015
Computer skill requirement
Coefficient
CoefficientCoefficient
Coefficient
Figure 2. Skill Requirements and the MSA-Specific Employment Shock
Notes: We regress the MSA-level change in BG skill requirements from 2007 on an exhaustive set of MSA em
ment shock-by-year interactions, controlling for year fixed effects and MSA characteristics (see equa
Graph plots the coefficients on Bartik shock × year and 95 percent confidence intervals.

1753HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
occupations, and we examine such heterogeneity, with a particular focus on ro
jobs, in Section V.)
In order to understand within-occupation skill demand changes along the int
sive margin, we also explore the effect of the shock on specific levels of educat
and experience requirements in online Appendix B.1. To summarize, we find eff
throughout the distribution along expected channels: low-skilled jobs become
likely to ask for a high school diploma, higher-skilled jobs become more likely
ask for a college degree, and experience requirement increases are conc
especially within the 1–5 year range.
One hypothesis for these results is that firms may become pickier when labo
and especially skilled labor, becomes more plentiful.25 Then elevated skill require-
ments might reflect opportunistic behavior on the part of firms that cannot ord
attract (or afford)more-skilled workers in a tight market.26 This hypothesis would
be compelling if the market for skilled workers remained more slack toward the
end of our sample period, even while some labor market indicators had recove
However, our results are similar when we include additional controls for local la
25Or, as in Menzio and Shi (2011), firms require a higher-quality match in a recession because of the neg
productivity shock.
26Evidence shows that in downturns workers are more likely to take worse jobs, relative to their skills, bu
is unclear whether this is driven by changes in firm recruitment strategy and/or worker search behavior (De
2002; Kahn 2010; Oreopoulos, von Wachter, and Heisz 2012; Altonji, Kahn, and Speer 2016).
Table 2—Within-Occupation Changes in Skill Requirements
Education Experience Cognitive Computer
(1) (2) (3) (4)
Shock × 2010 0.0526 0.0490 0.0275 0.0203
(0.0135) (0.0134) (0.00726) (0.00859)
Shock × 2011 0.0475 0.0443 0.0281 0.0243
(0.0131) (0.0134) (0.00731) (0.00716)
Shock × 2012 0.0233 0.0253 0.0186 0.0207
(0.0128) (0.0136) (0.00693) (0.00848)
Shock × 2013 0.0400 0.0363 0.0253 0.0252
(0.0120) (0.0122) (0.00642) (0.00664)
Shock × 2014 0.0429 0.0436 0.0265 0.0227
(0.0143) (0.0140) (0.00657) (0.00679)
Shock × 2015 0.0488 0.0468 0.0300 0.0134
(0.0143) (0.0142) (0.00730) (0.00807)
Number Occ-MSA-Year Cells 193,086 193,086 178,176 178,176
R2 0.044 0.069 0.040 0.034
Notes: Regressions are estimated at the MSA-occupation ( four-digit SOC)-year level using BG
data from 2010–2015 (see equation (1)). The dependent variable is the MSA-occupation level
annual change in skill requirements from 2007. All regressions control for year fixed effects and
MSA characteristics from the ACS. Observations are weighted by the size of the MSA labor
force in 2006 multiplied by the occupation’s ad share in theMSA-year. Standard errors are clus-
tered at the MSA level. Shock is the change in projectedyear-over-year employment growth in
the MSA from 2006 to 2009, divided by the 90–10 differential in the variable across all MSAs.
Columns 3 and 4 restrict to the sample of ads that have any specific skill requirements and there-
fore estimate the change in the probability of listing a cognitive or computer skill, conditional
on having any requirement.

1754 THE AMERICAN ECONOMIC REVIEW JULY 2018
market variables by skill level, such aseducation-specific MSA-level unemployment
rates, quit rates, and employment-to-population ratios. These controls accoun
changes in the supply of skilled labor due to, for example, differential quit beha
or changes in educational attainment brought on by the Great Recession or ove
preceding decade (Charles, Hurst, and Notowidigdo 2015). We conclude that o
tunistic upskilling cannot be the primary driver of our results.27
Online Appendix B.2 discusses these and other robustness checks in detail, w
are summarized in online Appendix TablesB1–B4. Our results broadly hold up
to additional controls, different samples, variants of the Bartik shock, and differ
ent weights. For example, our estimates are robust to controls for occupation fi
effects andoccupation-specific time trends, which allow occupations to systemati
cally differ in their change in skill requirements from 2007, as well as in the slo
of the change, across all MSAs. These could be important if some occupations a
both more likely to upskill or accelerate upskilling because of preexisting trend
are disproportionately located in hard-hit MSAs.
We have also explored heterogeneity within and across industry. We sh
online Appendix Table B5 that our results hold up to industry fixed effects and
trends, which is important in light of our identifying variation: industry compos
in an MSA before the Great Recession. Our identification would be threatened b
independent technology shocks concurrent to industries that experienced
employment shocks or by systematic measurement error in industry shares (w
could lead to spurious correlations in the shock across MSAs). The fact that our
results obtain even within industry alleviates these simultaneity and measurem
concerns. We further show in online Appendix Figure B3 that the upskilling effe
tend to be concentrated in industries with locally consumed products, as would
expected given their greater sensitivity to local demand shocks.
Finally, one may be concerned about changes in the use of online job ads ov
our sample period. Rising familiarity with the internet, falling costs of posting jo
increasing labor market tightness in the later period, and other factors may ha
brought more firms online to search for labor. Thewithin-occupation and -industry
results partially address the role that compositional changes in the use of onlin
job ads may have on our results by restricting comparisons to similar types of j
However, they may not adequately control for heterogeneity across firms in, sa
changes to their recruiting strategies or hiring needs. Such variations may be p
ularly pronounced during and after a recession.
In online Appendix B.3 we conduct a formal decomposition exercise to appor
upskilling effects as a function of within- and between-firm responses. Indeed,
do find a large role for substitution between firms that stopped posting after 20
and firms that began posting in 2010. As the latter post for higher skill requirem
on average than the former, this substitution can account for nearly one-half of
27In a pair of related papers and concurrent with our analysis, Sasser Modestino, Shoag, and Ballance (20
b), using a version of our dataset, find evidence of upskilling in harder-hit US counties after the Great Reces
subsequent downskilling as markets improved, and argue that this pattern is driven entirely by firms opport
cally seeking more-skilled workers in a slack labor market. We disagree with this conclusion, which relies he
on the small downward blip in 2012, seen also in our Figure 2, rather than the more careful picture generate
using all available data. In our paper, we also examine heterogeneity within and across firms and occupation
other margins of adjustment, such as capital, employment, and wages. This richer analysis implies more fun
tal changes in production inputs and longer-lasting impacts.

1755HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
full upskilling effect from Figure 2. However, we also find that nearly one-half o
the effects can be attributed to changes in skill requirements within firms, with
minimal role for compositional shifts across firms that post before and a
recession. This suggests that our results are not completely driven by the comp
tional changes mentioned above.
B. Discussion
We thus present strong evidence that employers inharder-hit MSAs were dif-
ferentially induced to increase stated preferences for a range of skills. While m
measures of local labor-market strength had converged back to pre-recession
by 2015, differences in advertised skill demands remained. Furthermore, varia
in the availability of skilled labor and compositional changes in the ads observe
our sample period are unlikely to explain the entire effects that we find.
This set of results raises the possibility that harder-hit MSAs differentially ex
rienced a structural change in demand for skill. In particular, the skill requirem
we investigate are complementary toroutine-biased technologies. Did the Great
Recession push an accelerated adoption of such technologies and accompanied
ing of cognitive workers to complement them? This could explain why skill requ
ments increase even within similar types of jobs. For example, community and
service specialists at a food bank in Washington, DC might be required not only
interact with clients to assist with food security, but may have to understand an
use database software and GIS, as well, to better serve them.28 Simultaneously, in
order to better reach and understand online readers, venerable journalistic org
tions such as the New York Times now hire individuals with science training, no
journalism training, to be chief data officers.29 It is also consistent with our finding
that epops for workers with less education were slow to recover: rapid adoption
new technologies in hard-hit MSAs over this time period may have rendered ce
worker skills obsolete, inducing labor force exit (also see Foote and Ryan 2014
Perhaps the epop of educated workers recovered rapidly precisely becau
increased demand for skill spurred by the Great Recession.
Several theoretical mechanisms may have induced firms to restructure. For e
ple, in the classic Schumpeter (1939) cleansing model, this would occur becaus
low-productivity firms shut down in the recession and resources are reallocate
firms with more-modern production technologies (see also Caballero and Ham
1994, 1996; and Mortensen and Pissarides 1994). Furthermore, this type
sodic restructuring could also occur because firms in harder-hit MSAs experien
a greater negative product-demand shock that: (i) lowers the opportunity cost
adjusting production (Hall 2005); (ii) shifts managerial attention from growth to
efficiency (Koenders and Rogerson 2005), perhaps due to increased pressure f
risk of bankruptcy (as in the canonical principal-agent model (Gibbons and Rob
28See Terrence McCoy, “The Technology That Could Revolutionize the War on Hunger,” The Washi
Post, June 16, 2015 (http://www.washingtonpost.com/local/the-technology-that-could-revolutionize-the-wa
hunger/2015/06/16/056d9d52-1114-11e5-adec-e82f8395c032_story.html).
29See RebeccaGreenfield,“Why the New York Times Hired a Biology ResearcherAs Its
Chief Data Officer,” Fast Company, February 12, 2014 (https://fastcompany.com/3026162/
why-the-new-york-times-hired-a-biology-researcher-as-its-chief-data-sci).

1756 THE AMERICAN ECONOMIC REVIEW JULY 2018
2012)); (iii) alters the costs of making layoffs (Mortensen and Pissarides
Berger 2012);30 and (iv) changes the incentives for a firm to invest in their worker
human capital (Jaimovich and Siu 2012). We do not feel we have the ability to d
entangle these mechanisms or provide strong support for any one model. Inste
point out that these types of workhorse models in macroeconomics can rationa
the results that we see.
If firms are changing how they produce, and not simply whom they hire, chan
to skill demand should persist among the same firms that initially upskilled. No
that the finding of within-firm upskilling, mentioned above, does not nec
ily imply this point, as different firms may have upskilled at different times. The
Burning Glass data provide an unprecedented glimpse at this margin at a detai
level, and we examine this prediction in Figure 3. Here, we divide firms
( posting-weighted) quartiles based on changes in skill requirements between 2
and 2010. We then plot the average skill requirements for each quartile over ti31
Firms began at fairly similar average skill levels in 2007, although this similarit
not imposed by our exercise. By construction there is a sharp contrast across fi
quartiles in 2010, with the darker shaded lines representing firms with larger s
increases. Interestingly, and not by construction, these quartiles remain spread
throughout the remainder of the sample period, and by 2015, firms in the high
quartiles still had substantially higher skill requirements in their new ads than fi
in the lower quartiles.
This within-firm persistence in upskilling holds up in regression analysi
is substantial. Though not shown, we find that, on average, 60–70 perce
firm’s increase in skill requirements between 2007 and 2010 persists through 2
Estimates are even larger when we instrument for the initial increase in upskill
with the Bartik shock. It could have been the case that the majority of firms inc
skill requirements during the recession and reverted back later in the recovery
example, in an attempt to opportunistically recruit while markets were slack), w
higher skill demand in later years unrelated to the recession and driven by diffe
firms. Instead, we find upskilling persists among the same firms both early and
in the recovery.
Furthermore, our finding that a substitution across old and new firms accoun
for some of the upskilling effect could also be consistent with episodic restructu
ing. Substitution from failing ( low-productivity) firms to new ( high-productivity
firms is a hallmark prediction of “cleansing” models of recessions (Schum
1939). In our data, we do not observe firm births and deaths, so the Schumpete
angle is difficult to fully assess. However, we can gain some general intuition b
comparing firms that post in 2007 but not again (possibly representing firm clo
sures) to firms that begin posting in the later period but not in 2007 (possibly r
resenting firm openings). (We readily acknowledge we are abstracting away fro
hiring freezes and migration toward online job postings; we view this exercise a
30Though not formalized, a sufficiently large negative product-demand shock could make layoffs worthw
offsetting any stigma or losses in terms of firm-specific human capital.
31We exploit the subsample of firms in our data that post at least five observations in each of 2007 and 2
comprising 66 percent of weighted observations. Online Appendix A.8 shows that the probability of satisfyin
restriction does not vary with the local labor market shock. Quartiles are defined separately for each skill me
weighting by the firm average number of posts across 2007–2010.

1757HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
illustrative, not definitive.) We find that “opening” firms indeed have higher ski
requirements than “closing” firms, even within occupation, and this is consiste
with the Schumpeterian cleansing view (see online Appendix B.3).32
The episodic restructuring hypothesis has additional predictions that we
take to the data. First, if changes in skill requirements reflect changes in produ
inputs, we should see greater investments in capital, and in particularroutine-labor
replacing technologies, for firms located in harder-hit MSAs. Moreover, th
activities should be linked: the very firms upskilling in their labor demand shou
the ones increasing their investments. Second, routine workers whose skills ca
substituted with these technologies should experience an immediate contractio
labor demand, as well as relative employment declines in the recovery and bey
In contrast, the occupations that are complementary to new technology
become more productive, from the increase in both physical and human capita
thus should exhibit increases in relative wages. We explore these predictions in
next two sections.
IV. Capital
Under episodic restructuring, firms automate routine tasks with technology,
complements skilled labor. If this restructuring is occurring, then firms should a
invest in physical capital around the time that they upskill. Information technol
(IT), in particular, have been linked to RBTC (Michaels, Natraj, and Van Reenen
32Moreira (2017), who shows that firms that begin in a recession are more productive than those that beg
an expansion, also provides support for the Schumpeterian view.
0.3
0.4
0.5
0.6
0.7
0.8
2006 2008 2010 2012 2014 2016
Education requirement
0.2
0.4
0.6
0.8
2006 2008 2010 2012 2014 2016
Experience requirement
0.1
0.2
0.3
0.4
0.5
0.6
2006 2008 2010 2012 2014 2016
Cognitive skill requirement
0.2
0.3
0.4
0.5
0.6
2006 2008 2010 2012 2014 2016
Computer skill requirement
Average requirementAverage requirement
Average requirementAverage requirement
Figure 3. Skill Requirements by Firm, 2007–2010 Change
Notes: Graph plots average BG skill requirement by year and quartile of 2007–2010 firm-level skill change. C
diamonds, triangles, and squares indicate skill change quartile from largest to smallest, respectively.

1758 THE AMERICAN ECONOMIC REVIEW JULY 2018
2014). While investments in capital tend to be procyclical, and production of IT
particular, has exhibited a secular decline (Byrne, Fernald, and Reinsdorf 2016
these trends could mask substantial heterogeneity.
We first investigate whether harder-hit MSAs are more likely to invest
over the Great Recession. To measure IT investment, we use the Ci Tech
Database from Harte-Hanks (now known as Aberdeen), a market intelligence fi
The Harte-Hanks database (hereafter, HH) is created from surveys and intervie
with high-level IT staff at millions of businesses worldwide each year. They col
data primarily to sell to major IT firms like IBM, Dell, and Cisco.33
Following previous work using these data, our primary outcome measure is t
number of personal computers (PCs) at a “site” (akin to business establishmen
have this measure consistently available in even years between 2000 and 2014
normalize by dividing by site employment in thepre-recession period.34 We aggregate
to the MSA-year level by taking an employment weighted average across sites
Figure 4 (and column 9 of online Appendix Table C1) summarizes results from
equation (1), with the MSA-level change in PCs per employee from 2006 as the
come. This graph provides evidence that firms located in harder-hit MSAs are m
likely to intensify IT investment over the same time period. Our estimates impl
sites in a hard-hit MSA add an average of 1.5 PCs (per eachpre-recession employee)
between 2006 and 2012, relative to sites in lesshard-hit MSAs. Though the confi-
dence intervals are wide, this effect is statistically significant at the 5 percent l
in 2008, 2010, and 2012. This differential increase experienced by hard-hit MS
substantial, roughly 60 percent more than the average increase across all MSA
0.93 increase in PCs per employee off a base of 0.75).
By 2014 the point estimate has fallen somewhat and is no longer stat
significant, possibly reflecting the beginning of a more gradual catch-up of tec
nology adoption in lesshard-hit MSAs. However, the point estimate implies that
harder-hit MSAs remain 1 PC per worker ahead of less hard-hit areas, relative
their pre-recession levels.35
Furthermore, the estimated coefficients for 2000, 2002, and 2004 are all clos
zero and statistically insignificant, implying that MSAs that would be severely a
33We thank Nick Bloom for graciously sharing with us extracts of the HH data as used in Bloom, Draca, an
Van Reenen (2016). In that paper, they show that Chinese import penetration increased technological chang
exposed firms in Europe. The data have also been used in several other studies. Bloom, Sadun, and Van Ree
(2012), for example, use HH data to show that US multinationals operating in Europe obtain higher productiv
from IT investments than non-multinationals; Beaudry and Lewis (2014) show that variation in PC adoption
US space can account for variation in declines in the gender pay gap; and Bresnahan, Brynjolfsson, and Hitt
provide evidence that IT use, work organization that shifts more responsibility to workers, and worker skill a
complements in production.
34A measure of PCs per employee is desirable to better understand capital intensity (rather than simply g
in size), but as employment may be varying (endogenously) over this time period, we fix the normalization a
period before the Great Recession: the average of each available year among 2002, 2004, and 2006. This no
ization means that variation in the outcome is strictly due to the numerator (total PCs), and ensures that gre
employment losses in harder-hit MSAs will not mechanically induce a positive association between our PCs
sure and the size of the shock. The fixed normalization requires that our sample be restricted to si
observed both prior and subsequent to the Great Recession, and this covers 65 percent of employment in HH
our sample years. Online Appendix A.8 shows that meeting this restriction is unrelated to our shock measure
35These relatively large magnitudes are in part driven by long right tails in the distribution of PCs per wor
across MSAs and years. To reduce the role of outliers, we have also estimated PC adoption on a sample trim
of the top and bottom 1 percent of observations; we find qualitatively similar patterns of statistically signific
increases in 2008–2012 that gradually decrease to insignificance by 2014, with point estimates that are som
smaller in magnitude.

1759HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
had fairly similar IT investment trends before the Great Recession. If anything,
is a slight relative decline inper-worker PCs in these areas between 2000 and 2006,
mostly in the last two years of this range, but these estimates are somewhat n
Thus, there is no evidence of a capital intensifying trend in harder-hit MSAs be
the recession, as the modest relative movement goes in the opposite direction
with the employment and unemployment rates in Figure 1, it is comforting that
identifying assumptions of the Bartik shock appear to hold. Although we canno
observe skill requirements before 2007, that both employment-to-population r
by education group and IT investments trend similarly across MSAs in the perio
before the Great Recession should reduce concern about preexisting trends.
In Section IV, we showed not only that advertised job skill requirements incre
and persisted inharder-hit MSAs, but that these increases occurred within firms. Si
harder-hit MSAs also intensified their IT investments over the same time perio
next explore whether this investment and upskilling are linked at the firm level
To do so, we link BG job ads at the firm level to two measures of investment
external data sources: PCs per worker from the HH database and capital holdin
from Compustat North America by Standard & Poors (Compustat). Compustat i
most complete database of accounting and balance sheet data among publicly
US firms. Although PCs are a good proxy for overall IT investments, they
miss broader routine-labor replacing investments, such as new machinery, tele
infrastructure, or inventory management systems.36 Thus, a firm’s overall holdings
36The Harte-Hanks database contains other measures of IT investment, including servers (for which we g
erally find results consistent with those from PCs), and specific types of software. Unfortunately, the latter a
consistently available only from 2010 onward.
−1
0
1
2
3
4
Coefficient
2000 2002 2004 2006 2008 2010 2012 2014
Figure 4. PC Adoption and theMSA-Employment Shock
Notes: We regress the MSA-level change in IT investment from 2006 on an exhaustive set of
MSA employment shock-by-year interactions, controlling for year fixed effects and MSA char-
acteristics (see equation (1)). Graph plots the coefficients on Bartik shock × year, as well as
95 percent confidence intervals. MSA-year IT investment is the employment-weighted average
of site-level PCs per pre-recession employment from Harte-Hanks.

1760 THE AMERICAN ECONOMIC REVIEW JULY 2018
of property, plant, and equipment (hereafter, PPENT) from Compustat is a usef
supplement.
We link both datasets to the BG data by firm name. See online Appendix Sec
A.4 and A.5 for details on these mergers. In general, we can match more firms
HH than to Compustat, as the former is meant to cover all businesses while the
ter is restricted to publicly traded companies. Among employers observed in bo
2007 and the later period in BG (which cover 65 percent of postings), we are a
to match about 80 percent of postings to firms in HH and 40 percent of posting
firms in Compustat. Online Appendix A.8 shows that the share of ads matching
these samples does not vary with the local employment shock.
We estimate upskilling regressions at the firm-MSA-year level, defined in eq
tion (3). This equation allows for an additional interaction between the shock-b
variables and the firm-level change in capital investments over the Great Rece
( Capita lf ). Because this is a first-difference specification at the firm-MSA leve
we do not include the main effect ofCapita lf . To reduce measurement error, we
define thesefirm-level changes in capital investment as the difference (PCs) or ra
(PPENT) between the average value in 2010, 2012, and 2014, and the average
in 2002, 2004, and 2006.37 Formally,
(3) outcom efmt − outcomefm07
= α0 + [shockm × It ] α1 + [shoc km × It × Capita lf ] α2 + It + Xm β + εf mt .
Figure 5 plots the estimated coefficients α2 (see also online Appendix Table C2).
To make them easier to interpret, we plot the fitted effect for the 90–10 perce
differential in firm-level capital change. The ninetieth percentile firm in our sam
ple added roughlytwo-thirds of a PC per worker at each of its establishments and
roughly tripled PPENT. In contrast, the tenth percentile firm lost nearlyone-third of
a PC per worker and dropped PPENT holdings by about 20 percent.
We find that firms with larger capital investments differentially and persisten
increase their skill requirements. For example, the top left panel of Figure 5 ( p
shows that between 2007 and 2010, holding the employment shock fixed, a fir
the ninetieth percentile of PC investment increased the likelihood of an educat
requirement in its job postings by 0.7 percentage points more than a firm at th
percentile of PC investment. This differential fluctuates somewhat, but persists
grows to about 1.0 percentage point by 2015. This pattern and approximate re
magnitude hold for experience, cognitive skill, and computer skill require
with statistically significant differentials in mostpost-recession years, usually at the
1 percent level. Overall, we find that inharder-hit MSAs, skill requirements increase
37Note that even though observations are firm-MSA-year cells, the investment change is at the fi
regardless of location. This is a necessary restriction of the Compustat data, which exists only for the firm an
individual establishments; for comparability, we aggregate sites in the HH data to the firm level, weighting b
employment. When we instead measure investment change at the firm-MSA level in the HH data, the result
qualitatively similar for most skill outcomes. However, both for comparability with the Compustat measure a
avoid additional noise from more demanding match criteria, we prefer defining investment change at the fir
As before, for PCs, we normalize this difference by average employment in thepre-period. To limit the influence of
extreme outliers in the PCs measure, we trim the top and bottom 2.5 percent of firms in the full HH database
amounts to roughly 4 percent of weighted observations in our regressions.

1761HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
0
0.005
0.01
0.015
2007 2009 2011 2013 2015
Education requirement
0
0.002
0.004
0.006
0.008
0.01
2007 2009 2011 2013 2015
Experience requirement
0
0.002
0.004
0.006
0.008
2007 2009 2011 2013 2015
Cognitive skill requirement
0
0.002
0.004
0.006
0.008
0.01
2007 2009 2011 2013 2015
Computer skill requirement
0
0.002
0.004
0.006
0.008
0.01
2007 2009 2011 2013 2015
Education requirement
0
0.002
0.004
0.006
0.008
0.01
2007 2009 2011 2013 2015
Experience requirement
0
0.002
0.004
0.006
2007 2009 2011 2013 2015
Cognitive skill requirement
0
0.002
0.004
0.006
0.008
2007 2009 2011 2013 2015
Computer skill requirement
Panel A. PCs (HH)
Panel B. Capital holdings (Compustat)
Figure 5. Differential Upskilling by 90–10 Change in Firm Capital Investments
Notes: We regress the firm-MSA-level change in BG skill requirements from 2007 on an exhaustive set of MS
employment shock-by-year interactions, and triple interactions between the shock, year, and the firm-level c
change. We also control for year fixed effects and MSA characteristics (see equation (3)). Graph plots the co
cients on the triple interactions, fitted to the 90–10 differential in firm capital change, and 95 percent confide
intervals. The capital change variable is the firm level change in average PCs (Harte-Hanks) per pre
employment between 2010–2014 and 2002–2006. Panel B: The capital change variable is the ratio of firm-le
average capital holdings (Compustat) in 2010–2014 to holdings in 2002–2006.

1762 THE AMERICAN ECONOMIC REVIEW JULY 2018
by roughly 30 to 50 percent more inhigh-investment firms than inlow-investment
firms, and these differentials hold over the post-recession period.
We find quantitatively similar effects when using the Compustat PPENT meas
( panel B). For example, inharder-hit MSAs, high-investment firms increase the
likelihood of specifying an education requirement by 0.6 to 0.8 percentage poin
relative to low-investment firms. This reflects roughly 35 percent greater respo
ness in upskilling. Results for the other skills variables are comparable, with mo
impacts significant at the 1 percent level.
Thus, throughout our sample period, firms with larger increases in capital sto
around the time of the Great Recession also had larger increases in their poste
requirements. These patterns are consistent with both human and physical cap
deepening at the firm level.
V. Routine Occupations
Thus far, we have provided evidence that MSAs more severely affected by th
Great Recession experienced persistent increases in the skill demand of job po
as well as greater increases in capital. Moreover, the upskilling and capital inve
ment occurred within the same firms. Both these findings are consistent with e
restructuring. In this section we explore additional predictions of anRBTC-style
restructuring: whether upskilling is more prevalent in more-routine occupation
related trends in employment and wages for these occupations.
Indeed, the literature on job polarization has successfully linked employment
wage shifts across occupations to the tasks performed by workers in the occup
tions. Wages and employment have fallen for occupations in the middle of the
distribution, which, being the most routine, are the kinds of occupations that ca
be replaced by machines or overseas labor.38 Autor (2014) and Jaimovich and Siu
(2015) have noted that employment continued to shift away frommiddle-skill occu-
pations in the Great Recession. Our BG data afford the unique opportunity to m
sure changes in skill requirements within occupations, while the bulk of work o
polarization has measured shifts in employment and wages only across occupa
Therefore, we next ask whether upskilling was relatively concentrated within ro
tine occupations, the very jobs thought to be most affected by technological ch
in recent decades.
To determine an occupation’s routineness, we use Acemoglu and Autor’s (20
routine-cognitive and routine-manual indices, derived from O*NET (see o
Appendix A.6). These indices provide continuous scores based on the intensity
tine tasks performed, are simple to create, distinguish between tasks that use
and physical capacities, and have been used in several other papers (e.g., Aaro
38The original work by Autor, Levy, and Murnane (2003)—henceforth, ALM—uses the US Department of
Labor’s (1977) Dictionary of Occupational Titles (DOT) to categorize tasks (and indirectly occupation
nonroutine manual, routine manual, routine cognitive, and nonroutine cognitive. They chose this categorizat
arguing that new technologies can successfully replace American workers performing routine, algorithmic ta
and are complementary to nonroutine, cognitive/analytical functions. Indeed, this grouping successfully pred
employment changes in the 1990s and has been used in a number of subsequent papers, including Autor, K
Kearney (2008).

1763HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
and Phelan 2017; Keister and Lewandowski 2017).39 Routine-cognitive tasks tend to
be clustered in clerical, administrative, and sales occupations, whileroutine-manual
tasks tend to be found in production and operative occupations. As a whole, em
ment in both types of occupations has been declining for at least the past two d
(Acemoglu and Autor 2011).
We begin by examining changes in skill requirements as a function of these r
tine index scores and the Bartik shock. To simplify our analysis, we focus on th
quartile of routine-cognitive and routine-manual occupations. The general pat
results is similar when we allow for finer distinctions. We estimate regressions
following form, where Routin eo
i is an indicator equal to 1 if occupation,o , is in the
top quartile of categorization,i , wherei ∈ {cognitive, manual} . Similar to equation
(3), we do not include the main effect of Routin eo
i because of thefirst-difference
specification at the occupation-MSA level. The parameter vector α2 in equation (4)
thus captures the additional effect of the shock, each year, for top-quartile rou
occupations relative to the effect in occupations in the bottom three quartiles o
evant routineness index. Henceforth, for exposition, we will refer to these top-
occupations as routine-cognitive or routine-manual, as appropriate. Formally,
(4) outcom eomt − outcomeom07
= α0 + [shoc km × It ] α1 + [shoc km × It × Routin eo
i ] α2 + It + Xm β + εo mt .
Figure 6 (and online Appendix Table C3) plots estimates of α2 for each routineness
index for the four skill requirements. The coefficients forroutine-cognitive occupa-
tions are indicated with blue circles, while coefficients for routine-manual occu
tions (estimated with separate regressions) are shown as maroon squares.
The figure’s primary pattern is a greater degree of upskilling inroutine-cognitive
occupations: these estimates are positive, statistically significant, and persiste
cases (except for experience in 2015). For example, the blue circle in the top le
in 2010 indicates that in hard-hit MSAs, job posts for routine-cognitive occupa
about a 0.5 percentage point larger increase in the probability of having an edu
requirement, relative to other occupations. For 2012 through 2015, routine-co
occupations saw a roughly 1 percentage point larger increase. Comparing thes
ential impacts to the baselinewithin-occupation upskilling effects (Table 2), we find
that routine-cognitive occupations were approximately 25 (education/experien
50 percent (cognitive/computer) more responsive than the average occupation
In contrast, routine-manual occupations do not exhibit a persistent different
in upskilling. In fact, in the case of cognitive and computer skills, these
pations experience relative downskilling compared to occupations that are
39The Acemoglu and Autor (AA) measures use O*Net (the successor of DOT) to essentially update the orig
categorization of ALM. Some papers in the literature (e.g., Autor and Dorn 2013; Autor, Dorn, and Hanson 20
use a simplerroutine-manual-abstract categorization (based on the original ALM categories) that does not allow
a distinction between routine-manual and routine-cognitive occupations; we find this distinction to be impo
Jaimovich and Siu (2015) use broad occupation categories to generate their binary routine classification. For
purposes, the AA measures are preferable because they allow for finer (continuous) distinctions. The Spearm
correlation between our adapted AA measures and the Jaimovich and Siu measures is 0.66 for routine manu
only 0.21 for routine cognitive, indicating that the AA measures likely avoid some miscategorization inheren
the binary definition.

1764 THE AMERICAN ECONOMIC REVIEW JULY 2018
routine-manual.(That is, routine-manualoccupationsupskill less than other
occupations.) For education and experience requirements, routine-manual oc
tions do exhibit temporary differential upskilling, indicated by positive and sign
cant point estimates in 2010 that converge to zero (or negative values) over th
few years. This could reflect opportunistic behavior on the part of firms during
slack market that quickly fades when markets recover.
Upskilling thus appears to be relatively concentrated within routine-cog
jobs. Our hypothesized explanation for this pattern is that the recession accele
technological adoption, but that some types of jobs (routine-cognitive ones) co
be made more complementary to the new technology with additional human ca
while labor for other types ofjobs (routine-manual ones) was more subject to sub-
stitution by the new technology. For example, the use of data analytics may m
salesperson more productive by allowing her to better target customers’ needs
software alone will not close a sale: a salesperson capable of using the softwar
still needed to do the job. On the other hand, machine-vision technology may r
obsolete the manual inspection of parts on an assembly line, essentially replac
that job.40
To investigate this paradigm, we turn to the implications for employment and
wages. If firms do not seek greater skills for routine-manual jobs becaus
jobs can be substituted with technology more readily than work with it,
would expect firms to disproportionately shed these types of jobs through layo
40Indeed, Hawkins, Michaels, and Oh (2015) present recent evidence of this type ofcapital-labor substitution
in the Korean manufacturing sector.
−0.005
0
0.005
0.01
0.015
2007 2009 2011 2013 2015
Education requirement
−0.005
0
0.005
0.01
0.015
2007 2009 2011 2013 2015
Experience requirement
−0.015
−0.01
−0.005
0
0.005
0.01
2007 2009 2011 2013 2015
Cognitive skill requirement
−0.02
−0.01
0
0.01
0.02
2007 2009 2011 2013 2015
Computer skill requirement
CoefficientCoefficient
CoefficientCoefficient
Figure 6. Differential Upskilling for Routine Occupations
Notes: We regress the occupation-MSA-level change in BG skill requirements from 2007 on an exhaustive se
MSA employment shock-by-year interactions, and triple interactions between the shock, year, and whether t
occupation is routine. We also control for year fixed effects and MSA characteristics (see equation (4)). Grap
the coefficients on the triple interactions, and 95 percent confidence intervals. The routineness measures ar
the occupation is in the top quartile of routine-cognitive or routine-manual index scores based on Acemolgu
Autor (2011).

1765HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
with little employment recovery over time. Since firms do seek greater s
routine-cognitive jobs, these workers may complement new technology, and if
their productivity and relative wages should rise. (Predictions for relative wage
routine-manual jobs and relative employment (or job loss) forroutine-cognitive jobs
are less clear cut, and depend on product demand.)
We explore all three margins, involuntary separations, relative employment,
wages, for routine-cognitive and routine-manual occupations. We examine the
of involuntary job loss in the population (not just the unemployed), as measure
the CPS by the propensity to report being a job loser. For employment and wag
we use Occupational Employment Statistics (OES) data. Both datasets allow us
capture trends back to 2000, and the earlier years help to check the validity of
identifying assumption, thatharder-hit MSAs would have been on a similar trend,
in terms of skill demand, if not for the Great Recession. (See online Appendix A
for details on sample construction and an analysis of the overall impact of the B
shock on these outcomes.)
In Figure 7, we focus on the differential impacts of the Bartik shock on layoffs
employment, and wages for routine-manual (red squares) and routine-co
(blue circles) occupations.41
Beginning with the top left panel, we find evidence of a large differential layo
effect for routine-manual occupations. At the peak in 2009, individuals whose
or most recent job was in a routine-manual occupation suffer an additional 1.5
increase in involuntary separations, relative to those not in routine-manual occ
tions, due to a 90–10 percentile MSA shock. For comparison, the same-sized s
increased the probability of having been laid off for those not inroutine-manual
occupations by 0.8 percentage points in 2009 (see column 1 of online Appendi
Table C4 for the full regression output); that is, individuals in routine-manual o
pations experienced nearly triple the chance of involuntary separation as did in
uals in other occupations. Although routine-cognitive occupations also experie
statistically significant differential increase, the magnitude is modest.42 Importantly,
there appears to be littlepre-trend for either routine occupation type, as, except for
a tiny blip in 2003, the layoff rate differential is close to zero in all years from 2
to 2007.
The top right panel (and columns 10 and 11 of online Appendix Table C1) sho
how the share of MSA employment in each type of routine occupation varies w
the Bartik shock. Following the Great Recession, there is a large and persistent
in routine-manual employment and a steady and modest rise inroutine-cognitive
employment.43 At its trough, the employment share ofroutine-manual occupations
fell by about 2 percentage points more in harder-hit MSAs, recovering only one
41OES data are based on athree-year moving average, so annual snapshots are not independent and trends ar
likely smoother than true annual snapshots would be. We have confirmed that this trait does not substantive
our estimates, as our results do not change appreciably if we use data for every third year (which is indepen
instead of annually. Additionally, we get similar results for employment and wages using the CPS or ACS, alb
with fewer included MSAs (CPS) or data years (Census/ACS). We prefer the use of OES for these outcomes fo
fuller coverage across occupations, MSAs, and time.
42It is possible, but not central to our argument, that workers laid off from aroutine-cognitive job have an easier
time finding reemployment and thus do not report their current status as laid off.
43For this result, we estimate versions of equation (1), using as dependent variables the share of employm
the MSA (relative to 2007) that is routine-cognitive or routine-manual.

1766 THE AMERICAN ECONOMIC REVIEW JULY 2018
of this gap by the end of the sample period. In contrast,routine-cognitive occupa-
tions differentially rise as a share of employment inharder-hit MSAs, (or, more
aptly, experience smaller magnitude losses in employment share, relative
hard-hit MSAs) though only modestly. Also, unlike for routine-manual occupat
it is harder to rule out a pre-trend, although it is small and generally statistical
insignificant. This pair of results is generally consistent with Jaimovich an
(2015), who show that employment in routine occupations as a whole fell episo
ically, and more so inharder-hit US states, in each of the past three recessions and
did not recover fully.
The bottom left panel (and column 2 of online Appendix Table C4) shows the
ferential impact on log median hourly wages. Forroutine-cognitive occupations, there
is a slight but persistent rise in wages inharder-hit MSAs beginning after 2010. By
2015, the medianroutine-cognitive worker in ahard-hit MSA has experienced 0.5
percent faster wage growth than a worker in a job that was not routine
Conversely, routine-manual occupations exhibit almost no post-recession
wages evolve similarly in the subsequent period regardless of the MSA shock. I
single major exception to the absence of pre-trending, wages for routine-man
were differentially trending downward before the Great Recession in areas that
experience a more severe shock, even though relative employment trended sim
suggesting other factors (such as declining unionization) were possibly involve
The sharpest predictions of our hypothesis (episodic increases in layoffs and
sistent decreases in employment share for routine-manual workers; and increa
in wages for routine-cognitive workers) are borne out by the data, and for thes
outcomes there is little evidence of differential pre-trending.
−0.005
0
0.005
0.01
0.015
2000 2005 2010 2015
Involuntary separations (CPS)
−0.03
−0.02
−0.01
0
0.01
0.02
2000 2005 2010 2015
Relative employment (OES)
−0.005
0
0.005
0.01
0.015
2000 2005 2010 2015
log median wage (OES)
Coefficient
Coefficient
Coefficient
Figure 7. Differential Employment and Wage Effects for Routine Occupations
Notes: Top left and bottom panels plot coefficients on the triple interactions of shock-year-routine (see equa
and Figure 6). Top right plots coefficients on shock-by-year, where the dependent variable is the MSA chang
employment share of routine occupations (see equation (1)). All regressions control for year fixed effects an
characteristics; we also include 95 percent confidence intervals. The routineness measures are whether the
tion is in the top quartile of routine-cognitive or routine-manual index scores based on Acemolgu and Autor

1767HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
To summarize, in harder-hit MSAs routine-manual occupations experien
sharp increase in layoff risk but there is no evidence of upskilling; rathe
appears to be relative downskilling accompanied by employment losses a
wages. For these occupations, the story is therefore consistent with firms’ subs
tution of technology for labor. This is the traditional view exhibited also
polarization literature: employment losses concentrated in occupations we exp
to be most readily replaceable by machines. Our contribution is to show that th
changes appear episodic around the Great Recession, though we acknowledge
secular relative wage losses preceded the employment shock.
In contrast to this conventional view of labor substitution, routine-cognitive o
pations in hard-hit MSAs surprisingly exhibit only a modest increase in layoffs
no loss in employment share, relative to other occupations. These change
concomitant with a pronounced increase in upskilling. Even as the elevated diff
ential risk of layoff declined, the differential in upskilling persisted, and was me
with modest relative wage and employment growth after the recession. This pa
is quite consistent with an intensive margin restructuring due to technological
tion that effectively shifts out the labor demand curve.
These results are especially enlightening, given recent findings by Beau
Green, and Sand (2014, 2016) that more-educated workers have increased the
presence in lower-skilled jobs since 2000. They term this shift, along with stag
ing employment in cognitive occupations, the “great reversal” in the demand f
cognitive skill (see also Castex and Dechter 2014). They hypothesize that lesse
demand for cognitive occupations induces college graduates to take jobs
in the skill distribution, squeezing out less-educated workers who formerly hel
these jobs. In light of the evidence above, we propose that any declining dema
cognitive occupations was accompanied by an increased demand for cognitive
within routine-task occupations, and this shift accelerated in the Great Recess
Even as employment has shifted away from routine occupations, the task
formed in the routine occupations that remain may be becoming less routine a
more cognitive. Our work thus highlights a complementary hypothesis for
high-skilled workers are increasingly found in lower-skilled occupations: these
ter occupations are becoming more skilled (and more highly paid), and it is pos
ble that less-skilled workers are displaced because they are unable to perform
new duties required.
VI. Conclusion
During the recovery following the Great Recession, anecdotal evidence sugg
that the composition of new hires shifted towardhigher-skilled workers, resulting
in many workers being “overeducated” for their jobs (Burning Glass Technolog
2014). However, it was not clear how broad, deep, or enduring these effects we
or the extent to which they were driven by labor supply or labor demand respo
In particular, firms may have treated the recession as a time of “cleansing,” en
them to restructure their production in a manner consistent with routine-biase
nological change.
In this paper we draw upon detailed job postings data to provide comprehens
broad-based evidence of upskilling (firms demandinghigher-skilled workers) when

1768 THE AMERICAN ECONOMIC REVIEW JULY 2018
the local economy suffers a recession. Using empirical skill measures that refle
what the discipline has learned about technological change andtask-based pro-
duction over the past 20 years (Autor, Levy, and Murnane 2003; Brynjolfsson a
McAfee 2011), we show that job postings in harder-hit MSAs experienced large
increases in their education, experience, cognitive, and computer requirement
lowing the Great Recession. These increases primarily reflected changes in dem
within, and not across, occupations. Furthermore, skill requirements remained
vated through the end of our sample in 2015, even as most measures of labor
ket conditions had converged back to their pre-recession levels. Importa
find that the increases in skill requirements are accompanied by increases in c
investments, both at the MSA andfirm levels. We also show that upskilling is rel-
atively concentrated in routine-cognitive occupations, which exhibit modest w
growth as well. In contrast, routine-manual occupations in harder-hit MSAs ex
a sharp relative decline in employment share following the Great Recession.
We argue that the most likely explanation for this body of results is that the G
Recession did indeed provide firms a catalyst to restructure production accordi
to a paradigm of routine-biased technological change. While firms may respon
changes in labor market conditions through posted skill requirements for a vari
other reasons, these cannot rationalize our full body of findings. For example, fi
may worry that a flood of applicants early in the recovery will create a “bottlen
in screening, and therefore raise requirements to signal that certain (unw
applicants need not apply. Alternatively, firms that typically cannot attrac
afford) more-skilled workers in a tight labor market may opportunistically
them out in a slack one. However, while both of these cyclical behaviors may h
been important early in the recovery, they cannot generate persistent, within-
increases in skill requirements that stand up to controls for the availability of la
by skill group, occur concomitantly with greater investment in multiple measur
of physical capital, and are concentrated in the types of occupations acknowled
to be most susceptible to routine-biased technological change. While it is poss
that labor markets have been slower to recover than indicators such as employ
growth, the unemployment rate, andeducation-specific employment-population
ratios indicate, it is telling that we find little overall convergence in skill require
ments as markets improve, even though the data do indicate such convergenc
some (e.g., routine-manual) occupations.
Simply put, the evidence supports that shifts in skill requirements refle
technologically-driven changes in the means of production, not just changes in
firms seek to hire. Our work is thus consistent with the important, but suggesti
evidence provided by Jaimovich and Siu (2015) that the vast majority of emplo
ment lost in routine occupations was lost during recessions and never recovere
also contributes to the many models in macroeconomics that assume ad
costs and imply that recessions will be times of “cleansing” in terms of product
(Schumpeter 1939; Koenders and Rogerson 2005; Berger 2012). As hypothesiz
by many, these kinds of episodic,productivity-enhancing changes can result in job-
less recovery. Our findings are thus extremely relevant for policymakers, who a
cate billions of taxpayer dollars to subsidize workers’ job searches in a downtur
We also demonstrate how electronic job postings data can provide a u
opportunity to understand real-time changes in skill demand, both across and

1769HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
occupations. This level of detail can provide new insight relative to earlier litera
For example, our result thatroutine-cognitive occupations are apparently becoming
higher skilled and more productive can help to clarify studies by Beaudry, Gree
and Sand (2014, 2016) and others documenting the “great reversal” in deman
cognitive skill. While it may be the case that employment in high-skill occupat
did not grow, on average, over the past decade, our results show that cognitive
ers still retain a substantial advantage over the low-skilled. They are dra
formerly middle-skill jobs, which are becominghigher-skilled. This is indicated by
the persistence in both the relative upskilling and wage growth inroutine-cognitive
occupations located in harder-hit MSAs. Our findings can thus help explain why
skilled workers still earn a premium in the labor market even though the return
cognitive occupations appear to have diminished.
The US economy has seen remarkable changes over the past 30 years, broug
on by the computer revolution and globalization. These changes have led to gr
increases in productivity and wealth, but the benefits have not been shared ac
all workers. Indeed, mounting evidence suggests that a large population of wor
ers, formerly employed in routine-task jobs, have suffered permanent labor ma
health, and social consequences from structural changes in the economy
et al. 2014; Foote and Ryan 2014; Pierce and Schott 2016; Autor, Dorn, and Ha
2017). Our results highlight that a worker’s ability to adjust to these changes m
be especially difficult because the changes are episodic, concentrated in reces
Thus, large numbers of workers can find their skills depreciated at the same tim
This is perhaps evident in the stair-step declines in male labor force participat
that have tended to be concentrated around recessions (Moffitt 2012; Fo
Ryan 2014). If the changes to production instead occurred more gradually, wor
would still need to be retrained, but over a longer time period, and on a much s
scale at any given time. Future policy work should be directed at understanding
to reallocate workers on a large scale following a recession.
REFERENCES
Aaronson, Daniel, and Brian J. Phelan.2017. “Wage Shocks and the Technological Substitution of
Low-Wage Jobs.” Economic Journal. http://dx.doi.org/10.1111/econj.12529.
Acemoglu, Daron, and David Autor.2011. “Skills, Tasks and Technologies: Implications for Employ-
ment and Earnings.” In Handbook of Labor Economics, Vol. 4, edited by David Card and Orley
Ashenfelter, 1043–171. Amsterdam: Elsevier.
Acemoglu, Daron, and Pascual Restrepo.2017. “Robots and Jobs: Evidence from US Labor Markets.”
National Bureau of Economic Research Working Paper 23285.
Altonji, Joseph G., Lisa B. Kahn, and Jamin D. Speer.2016. “Cashier or Consultant? Entry Labor Mar-
ket Conditions, Field of Study, and Career Success.” Journal of Labor Economics 34 (1): S361–
Autor, David H.2014. “Polanyi’s Paradox and the Shape of Employment Growth.” National Bureau o
Economic Research Working Paper 20485.
Autor, David H., and David Dorn.2013. “The Growth of Low-Skill Service Jobs and the Polarization
of the US Labor Market.” American Economic Review 103 (5): 1553–97.
Autor, David H., David Dorn, and Gordon H. Hanson.2015. “Untangling Trade and Technology: Evi-
dence from Local Labour Markets.” Economic Journal 125 (584): 621–46.
Autor, David H., David Dorn, and Gordon H. Hanson.2017. “When Work Disappears: Manufacturing
Decline and the Falling Marriage-Market Value of Men.” National Bureau of Economic Researc
Paper 23173.
Autor, David H., David Dorn, Gordon H. Hanson, and Jae Song.2014. “Trade Adjustment: Work-
er-Level Evidence.” Quarterly Journal of Economics 129 (4): 1799–860.

1770 THE AMERICAN ECONOMIC REVIEW JULY 2018
Autor, David H., Lawrence F. Katz, and Melissa S. Kearney.2008. “Trends in U.S. Wage Inequality:
Revising the Revisionists.” Review of Economics and Statistics 90 (2): 300–23.
Autor, David H., Frank Levy, and Richard J. Murnane.2003. “The Skill Content of Recent Techno-
logical Change: An Empirical Exploration.” Quarterly Journal of Economics 118 (4): 1279–333.
Bartik, Timothy J.1991. Who Benefits from State and Local Economic Development Policies? Kalam
azoo, MI: W. E. Upjohn Institute.
Beaudry, Paul, David A. Green, and Benjamin M. Sand.2014. “The Declining Fortunes of the Young
since 2000.” American Economic Review 104 (5): 381–86.
Beaudry, Paul, David A. Green, and Benjamin M. Sand.2016. “The Great Reversal in the Demand for
Skill and Cognitive Tasks.” Journal of Labor Economics 34 (1): S199–247.
Beaudry, Paul, and Ethan Lewis.2014. “Do Male-Female Wage Differentials Reflect Differences in
the Return to Skill? Cross-City Evidence from 1980-2000.” American Economic Journal: Applied
Economics 6 (2): 178–94.
Berger, David.2012. “Countercyclical Restructuring and Jobless Recoveries.” Unpublished.
Bernard, Andrew B., Stephen J. Redding, and Peter K. Schott.2011. “Multiproduct Firms and Trade
Liberalization.” Quarterly Journal of Economics 126 (3): 1271–318.
Blanchard, Olivier Jean, and Lawrence F. Katz.1992. “Regional Evolutions.” Brookings Papers on
Economic Activity 23 (1): 1–76.
Bloom, Nicholas, Mirko Draca, and John Van Reenen.2016. “Trade Induced Technical Change? The
Impact of Chinese Imports on Innovation, IT and Productivity.” Review of Economic Studies 83
(1): 87–117.
Bloom, Nicholas, Raffaella Sadun, and John Van Reenen.2012. “Americans Do IT Better: US Multina-
tionals and the Productivity Miracle.” American Economic Review 102 (1): 167–201.
Bresnahan, Timothy F., Erik Brynjolfsson, and Lorin M. Hitt.2002. “Information Technology, Work-
place Organization, and the Demand for Skilled Labor: Firm-Level Evidence.” Quarterly Journa
of Economics 117 (1): 339–76.
Brynjolfsson, Erik, and Andrew McAfee.2011. Race against the Machine: How the Digital Revolution
Is Accelerating Innovation, Driving Productivity, and Irreversibly Transforming Employment an
the Economy. Lexington, KY: Digital Frontier Press.
Burning Glass Technologies.2014. Moving the Goalposts: How Demand for a Bachelor’s Degree Is
Reshaping the Workforce. Boston: Burning Glass Technologies.
Byrne, David M., John G. Fernald, and Marshall B. Reinsdorf.2016. “Does the United States Have a
Productivity Slowdown or a Measurement Problem?” Brookings Papers on Economic Activity 4
(1): 109–57.
Caballero, Ricardo J., and Mohamad L. Hammour.1994. “The Cleansing Effect of Recessions.”
American Economic Review 84 (5): 1350–68.
Caballero, Ricardo J., and Mohamad L. Hammour.1996. “On the Timing and Efficiency of Creative
Destruction.” Quarterly Journal of Economics 111 (3): 805–52.
Card, David, Ana Rute Cardoso, and Patrick Kline.2016. “Bargaining, Sorting, and the Gender Wage
Gap: Quantifying the Impact of Firms on the Relative Pay of Women.” Quarterly Journal of Eco
nomics 131 (2): 633–86.
Card, David, Jorg Heining, and Patrick Kline.2013. “Workplace Heterogeneity and the Rise of West
German Wage Inequality.” Quarterly Journal of Economics 128 (3): 967–1015.
Carnevale, Anthony P., Tamara Jayasundera, and Dmitri Repnikov.2014. Understanding Online Job
Ads Data: A Technical Report. Washington, DC: Georgetown University Center on Education a
the Workforce.
Castex, Gonzalo, and Evgenia Kogan Dechter.2014. “The Changing Roles of Education and Ability in
Wage Determination.” Journal of Labor Economics 32 (4): 685–710.
Charles, Kerwin Kofi, Erik Hurst, and Matthew J. Notowidigdo.2015. “Housing Booms and Busts,
Labor Market Opportunities, and College Attendance.” National Bureau of Economic Res
Working Paper 21587.
Davis, Steven J., R. Jason Faberman, and John C. Haltiwanger.2012. “Labor Market Flows in the
Cross Section and Over Time.” Journal of Monetary Economics 59 (1): 1–18.
Davis, Steven J., R. Jason Faberman, and John C. Haltiwanger.2013. “The Establishment-Level
Behavior of Vacancies and Hiring.” Quarterly Journal of Economics 128 (2): 581–622.
Deming, David J.2017. “The Growing Importance of Social Skills in the Labor Market.” Quarterly
Journal of Economics 132 (4): 1593–640.
Deming, David, and Lisa B. Kahn.2018. “Skill Requirements across Firms and Labor Markets: Evi-
dence from Job Postings for Professionals.” Journal of Labor Economics 36 (S1): S337–69.
Devereux, Paul J.2002. “Occupational Upgrading and the Business Cycle.” Labour 16 (3): 423–52.

1771HERSHBEIN AND KAHN: ROUTINE-BIASED TECHNOLOGICAL CHANGEVOL. 108 NO. 7
Foote, Christopher L., and Richard W. Ryan.2014. “Labor-Market Polarization over the Business
Cycle.” In NBER Macroeconomics Annual, Vol. 29, edited by Jonathan A. Parker and Mich
Woodford, 371–413. Chicago: University of Chicago Press.
Forsythe, Eliza.2016. “Why Don’t Firms Hire Young Workers during Recessions?” Unpublished.
Gibbons, Robert, and John Roberts.2012. “Economic Theories of Incentives in Organizations.” In
Handbook of Organizational Economics, edited by Robert Gibbons and John Roberts, 56
Princeton, NJ: Princeton University Press.
Gomes, Joao, Jeremy Greenwood, and Sergio Rebelo.2001. “Equilibrium Unemployment.” Journal of
Monetary Economics 48 (1): 109–52.
Goos, Maarten, and Alan Manning.2007. “Lousy and Lovely Jobs: The Rising Polarization of Work
in Britain.” Review of Economics and Statistics 89 (1): 118–33.
Hall, Robert E.1991. “Labor Demand, Labor Supply, and Employment Volatility.” In NBER Macro-
economics Annual, Vol. 6, edited by Jonathan A. Parker and Michael Woodford, 17–62. Cambri
MA: MIT Press.
Hall, Robert E.2005. “Employment Fluctuations with Equilibrium Wage Stickiness.” American Eco-
nomic Review 95 (1): 50–65.
Harrigan, James, Ariell Reshef, and Farid Toubal.2016. “The March of the Techies: Technology,
Trade, and Job Polarization in France, 1994–2007.” National Bureau of Economic Research Wo
ing Paper 22110.
Hawkins, William B., Ryan Michaels, and Jiyoon Oh.2015. “The Joint Dynamics of Capital and
Employment at the Plant Level.” Unpublished.
Hershbein, Brad, and Lisa B. Kahn.2018. “Do Recessions Accelerate Routine-Biased Technologi-
cal Change? Evidence from Vacancy Postings: Dataset.” American Economic Review. https://d
org/10.1257/aer.20161570.
Hoynes, Hilary, Douglas L. Miller, and Jessamyn Schaller.2012. “Who Suffers during Recessions?”
Journal of Economic Perspectives 26 (3): 27–48.
Isserman, Andrew M., and James Westervelt.2006. “1.5 Million Missing Numbers: Overcoming
Employment Suppression in in County Business Patterns Data.” International Regional Science
Review 29 (3): 311–35.
Jaimovich, Nir, and Henry E. Siu.2012. “The Trend Is the Cycle: Job Polarization and Jobless Recov-
eries.” Unpublished.
Jaimovich, Nir, and Henry E. Siu.2015. “Job Polarization and Jobless Recoveries.” Unpublished.
Kahn, Lisa B.2010. “The Long-Term Labor Market Consequences of Graduating from College in a
Bad Economy.” Labour Economics 17 (2): 303–16.
Keister, Roma, and Piotr Lewandowski.2017. “A Routine Transition in the Digital Era? The Rise of
Routine Work in Central and Eastern Europe.” Transfer: European Review of Labour and Resea
23 (3): 263–79.
Koenders, Kathryn, and Richard Rogerson.2005. “Organizational Dynamics over the Business Cycle:
A View on Jobless Recoveries.” Federal Reserve Bank of St Louis Review 87 (4): 555–79.
Kuhn, Peter, and Kailing Shen.2013. “Gender Discrimination in Job Ads: Evidence from China.”
Quarterly Journal of Economics 128 (1): 287–336.
Lazear, Edward P., and James R. Spletzer.2012. “Hiring, Churn, and the Business Cycle.” American
Economic Review 102 (3): 575–79.
Marinescu, Ioana.2017. “The General Equilibrium Impacts of Unemployment Insurance: Evidence
from a Large Online Job Board.” Journal of Public Economics 150 (C): 14–29.
Menzio, Guido, and Shouyong Shi.2011. “Efficient Search on the Job and the Business Cycle.” Jour-
nal of Political Economy 119 (3): 468–510.
Michaels, Guy, Ashwini Natraj, and John Van Reenen.2014. “Has ICT Polarized Skill Demand? Evi-
dence from Eleven Countries over Twenty-Five Years.” Review of Economics and Statistics 96
60–77.
Moffitt, Robert A.2012. “The Reversal of the Employment-Population Ratio in the 2000s: Facts and
Explanations.” Brookings Papers on Economic Activity 43 (2): 201–50.
Moreira, Sara.2017. “Firm Dynamics, Persistent Effects of Entry Conditions, and Business Cycles.”
U.S. Census Bureau Center for Economic Studies Working Paper 17-29.
Mortensen, Dale T., and Christopher A. Pissarides.1994. “Job Creation and Job Destruction in the
Theory of Unemployment.” Review of Economic Studies 61 (3): 397–415.
Nickell, Stephen J.1996. “Competition and Corporate Performance.” Journal of Political Economy
104 (4): 724–46.
Notowidigdo, Matthew J.2011. “The Incidence of Local Labor Demand Shocks.” National Bureau of
Economic Research Working Paper 17167.

1772 THE AMERICAN ECONOMIC REVIEW JULY 2018
Oreopoulos, Philip, Till von Wachter, and Andrew Heisz.2012. “The Short- and Long-Term Career
Effects of Graduating in a Recession.” American Economic Journal: Applied Economics 4 (1): 1
Pierce, Justin R., and Peter K. Schott.2016. “Trade Liberalization and Mortality: Evidence from U.S.
Counties.” National Bureau of Economic Research Working Paper 22849.
Rothwell, Jonathan.2014. Still Searching: Job Vacancies and STEM Skills. Washington, DC: Brook-
ings Institution.
Sasser Modestino, Alicia, Daniel Shoag, and Joshua Ballance.2016a. “Upskilling: Do Employers
Demand Greater Skill when Skilled Workers Are Plentiful?” Unpublished.
Sasser Modestino, Alicia, Daniel Shoag, and Joshua Ballance.2016b. “Downskilling: Changes in
Employer Skill Requirements over the Business Cycle.” Labour Economics 41 (C): 333–47.
Schumpeter, Joseph A.1939. Business Cycles. New York: McGraw-Hill.
Syverson, Chad.2004a. “Market Structure and Productivity: A Concrete Example.” Journal of Politi-
cal Economy 112 (6): 1181–222.
Syverson, Chad.2004b. “Product Substitutability and Productivity Dispersion.” Review of Economics
and Statistics 86 (2): 534–50.
US Department of Labor.1977. Dictionary of Occupational Titles. 4th ed. Washington, DC: US
Department of Labor.
von Wachter, Till, and Elizabeth Weber Handwerker.2008. “Variation in the Cost of Job Loss by
Worker Skill: Evidence Using Matched Data from California, 1991–2000.” Unpublished.
Yagan, Danny.2016. “The Enduring Employment Impact of Your Great Recession Location.”
Unpublished.

This article has been cited by:
1. Enghin Atalay, Phai Phongthiengtham, Sebastian Sotelo, Daniel Tannenbaum. 2020. The
of Work in the United States. American Economic Journal: Applied Economics 12:2, 1-34.
[View PDF article] [PDF with links]
2. Daniel Dietz, Thomas Zwick. 2020. Training in the Great Recession – Evidence from an In
Perspective. Jahrbücher für Nationalökonomie und Statistik 240:4, 493-523. [Crossref]
3. Matthew Ross. 2020. The Effect of Intensive Margin Changes to Task Content on Employm
Dynamics over the Business Cycle. ILR Review 86, 001979392091074. [Crossref]
4. Amanda Weinstein, Carlianne Patrick. 2020. Recession‐proof skills, cities, and resilience
downturns. Journal of Regional Science 60:2, 348-373. [Crossref]
5. Olga Scrivner, Thuy Nguyen, Kosali Simon, Esmé Middaugh, Bledi Taska, Katy Börner. 20
postings in the substance use disorder treatment related sector during the first five years
expansion. PLOS ONE 15:1, e0228394. [Crossref]
6. Christos Andreas Makridis. 2019. Do Right-to-Work Laws Work? Evidence on Individuals’
Being and Economic Sentiment. The Journal of Law and Economics 62:4, 713-745. [Cross
7. Richard Friberg, Mark Sanctuary. 2019. Exchange rate risk and the skill composition of la
of World Economics 122. . [Crossref]
8. MIAO BEN ZHANG. 2019. Labor‐Technology Substitution: Implications for Asset Pricing. T
Journal of Finance 74:4, 1793-1839. [Crossref]
9. TommasoFerraresi,Andrea Roventini,Willi Semmler.2019. MacroeconomicRegimes,
Technological Shocks and Employment Dynamics. Jahrbücher für Nationalökonomie
239:4, 599-625. [Crossref]
10. Pedro Henrique Melo Albuquerque, Cayan Atreio Portela Bárcena Saavedra, Rafael Lima
Yaohao Peng. 2019. The Robot from Ipanema goes Working: Estimating the Probability of
Automation in Brazil. Latin American Business Review 20:3, 227-248. [Crossref]
11. Wen Zhang. 2019. Deciphering the causes for the post-1990 slow output recoveries. Eco
176, 28-34. [Crossref]
12. Murillo Campello, Janet Gao, Qiping Xu. 2019. Local Taxes and the Demand for Skilled La
Evidence from 27 Million Job Postings. SSRN Electronic Journal . [Crossref]
13. Łukasz Arendt, Wojciech Grabowski. 2019. Technical change and wage premium shifts a
content groups in Poland. Economic Research-Ekonomska Istraživanja 32:1, 3398-3416. [
14. Brad Hershbein. 2018. Discussion for JME special issue: APST paper. Journal of Monetary
97, 68-70. [Crossref]
15. Alexandr Kopytov, Nikolai Roussanov, Mathieu Taschereau-Dumouchel. 2018. Short-run
run gain? Recessions and technological transformation. Journal of Monetary Economics 9
[Crossref]
16. Abi Adams. 2018. Technology and the labour market: the assessment. Oxford Review of
Policy 34:3, 349-361. [Crossref]
17. Dieter Verhaest, Elene Bogaert, Jeroen Dereymaeker, Laura Mestdagh, Stijn Baert.
Employers Prefer Overqualified Graduates? A Field Experiment. Industrial Relations: A Jou
Economy and Society 57:3, 361-388. [Crossref]
18. Christos Makridis, Yuqing Zhou. 2018. Firm Reputation Following Accounting Frauds: Evi
Employee Ratings. SSRN Electronic Journal . [Crossref]

19. Alexander Mallinckrodt. 2018. Automation and Globalisation: The Effects of AI-Based Aut
on the Labour Market Compared to Previous Periods of Globalisation. SSRN Electronic Jou
[Crossref]
20. Aina Gallego, Thomas Kurer, Nikolas Schoell. 2018. Not So Disruptive after All: How Work
Digitalization Affects Political Preferences. SSRN Electronic Journal . [Crossref]
21. Christos Andreas Makridis, Maury Gittleman. 2017. Does Performance Payy Pay? Wage F
Over the Great Recession. SSRN Electronic Journal . [Crossref]