NBER WORKING PAPER SERIES

IS AUTOMATION LABOR-DISPLACING? PRODUCTIVITY GROWTH, EMPLOYMENT,

AND THE LABOR SHARE

David Autor
Anna Salomons

Working Paper 24871

http://www.nber.org/papers/w24871

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
July 2018

Paper prepared for the Brookings Papers on Economic Activity conference, March 2018. We are
grateful to Daron Acemoglu, Jim Bessen, Uwe Blien, Janice Eberly, Maarten Goos, Pascual
Restrepo, James Stock, Coen Teulings, John Van Reenen, Xianjia Ye, and our discussants, John
Haltiwanger and Richard Rogerson, whose valuable input improved the paper. We thank Pian
Shu for sharing data on approved patent applications and citations by industry, year, and country.
Autor acknowledges funding from IBM Global Universities, Schmidt Sciences, and the Smith-
Richardson Foundation. Salomons acknowledges funding from the Netherlands Organisation for
Scientific Research. The views expressed herein are those of the authors and do not necessarily
reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2018 by David Autor and Anna Salomons. All rights reserved. Short sections of text, not to
exceed two paragraphs, may be quoted without explicit permission provided that full credit,
including © notice, is given to the source.
Is Automation Labor-Displacing? Productivity Growth, Employment, and the Labor Share
David Autor and Anna Salomons
NBER Working Paper No. 24871
July 2018
JEL No. D33,J23,O33,O57

ABSTRACT

Many technological innovations replace workers with machines, but this capital-labor substitution
need not reduce aggregate labor demand because it simultaneously induces four countervailing
responses: own-industry output effects; cross-industry input–output effects; between-industry
shifts; and final demand effects. We quantify these channels using four decades of harmonized
cross-country and industry data, where we measure automation as industry-level movements in
total factor productivity (TFP) that are common across countries. We find that automation
displaces employment and reduces labor's share of value-added in the industries in which it
originates (a direct effect). In the case of employment, these own-industry losses are reversed by
indirect gains in customer industries and induced increases in aggregate demand. By contrast,
own-industry labor share losses are not recouped elsewhere. Our framework can account for a
substantial fraction of the reallocation of employment across industries and the aggregate fall in
the labor share over the last three decades. It does not, however, explain why the labor share fell
more rapidly during the 2000s

David Autor
Department of Economics, E52-438
MIT
77 Massachusetts Avenue
Cambridge, MA 02139
and NBER
[email protected]

Anna Salomons
Utrecht School of Economics
Utrecht University
Kriekenpitplein 21-22
3584 EC Utrecht
The Netherlands
[email protected]
Introduction
It is a widely held view that recent and incipient breakthroughs in artificial intelligence and

dexterous, adaptive robotics are profoundly shifting the terms of human vs. machine

comparative advantage. In light of these advances, numerous scholars and popular writers

anticipate the wholesale elimination of a vast set of currently labor-intensive and cognitively

demanding tasks, leaving an ever-diminishing set of activities in which labor adds significant

value (Brynjolfsson and McAfee, 2014; Ford, 2015; Frey and Osborne, 2017). The displacement of

labor from production could take (at least) two forms: employment displacement, meaning the
elimination of aggregate employment; or labor-share displacement, meaning the erosion of

labor’s share of value-added in the economy.

Whether technological progress ultimately proves employment- or labor-share-displacing

depends proximately on two factors: how technological innovations shape employment and
labor’s share of value-added directly in the industries where they occur; and how these direct

effects are augmented or offset by employment and labor-share changes elsewhere in the

economy that are indirectly spurred by these same technological forces. The first of these

phenomena—the direct effect of technological progress on employment and labor-share in the

specific settings in which it occurs—is often readily observable, and we suspect that observation

of these direct labor-displacing effects shapes theoretical and empirical study of the aggregate

impact of technological progress. The indirect effects of technological progress on these same

outcomes, however, are likely more challenging to observe and quantify, and hence may receive

short shrift in economic analysis and in the wider public debate.2

To see the challenge this creates, consider the two panels of Figure 1, which reports bivariate

scatters of the relationship between industry-level Total Factor Productivity (TFP) growth over

the 1970 - 2007 period and contemporaneous industry-level log employment growth (Figure 1A)

and industry-level changes in log labor share (Figure 1B), defined as the log ratio of the wagebill

2Caselli and Manning (forthcoming) observe that many recent analyses of the potential impact of new technology on
workers implicitly rely on models that omit general equilibrium effects.

1
to value-added.3 Both figures reveal a well-determined downward slope: industries

experiencing faster measured TFP growth on average exhibit steep relative declines in

employment and labor share over this period. It would be tempting to infer from these figures

that technological advances (captured by TFP growth) erode aggregate employment and labor’s

share of national income.

But theory makes clear that there is no direct mapping between the evolution of

productivity and labor demand at the industry level and the evolution of labor demand in the
aggregate (Foster et al., 2017). A long-standing literature, starting with Baumol (1967), has

considered reallocation mechanisms for employment, showing that labor moves from

technologically advancing to technologically lagging sectors if the outputs of these sectors are

not close substitutes. Further, Ngai and Pissarides (2007) and Acemoglu and Guerrieri (2008)
show that such ongoing unbalanced productivity growth across sectors can nevertheless yield a

balanced growth path for labor and capital shares. Indeed, one of the central stylized facts of

modern macroeconomics, immortalized by Kaldor (1961), is that during a century of

unprecedented technological advancement in transportation, production, and communication,

labor’s share of national income remained roughly constant (Jones and Romer, 2010). This

empirical regularity, which Keynes (1939) deemed “a bit of a miracle,” has provided

economists—though not the lay public—with grounds for optimism that, despite seemingly

limitless possibilities for labor-saving technological progress, automation need not displace

labor as a factor of production.

Table 1 confirms the broad relevance of these theoretical observations. Aggregate

employment grew dramatically in all countries in this time interval even as relative employment

fell in the industries experiencing the fastest productivity growth. Yet, conversely, labor’s share
of value-added was steady or rising in the 1970s, declined modestly in the 1980s and 1990s, and

then fell steeply in the 2000s in many countries. These facts thus highlight the pitfalls of

3Our data sources and methods are documented in detail in Section 1. The figures above average across the 19
developed countries in our sample encompassing 28 market industries. Each industry is weighted by its own-country
average share of employment (Figure 1A) or value-added (Figure 1B) over the full time interval. Patterns are similar
when instead using decadal changes in employment or labor’s share and previous-decade TFP growth starting in the
1980s.

2
extrapolating from direct, first-order technological relationships (here, observed at the industry

level) to labor market outcomes in the aggregate, because the latter incorporate both direct and

indirect consequences of technological progress (as well as many non-technological factors).

This paper applies harmonized cross-country and cross-industry data to explore the

relationship between technological change and labor market outcomes over four decades. A

first contribution of the paper is to attempt to overcome the tension, endemic to this area of

work, of using microeconomic variation to afford identification while attempting to speak to

macroeconomic outcomes. This tension arises here because we study the relationship between

productivity growth, innovation, and labor displacement at the country-industry level. As

Figure 1 underscores, naively extrapolating from industry-level to aggregate-level relationships

is potentially fallacious. The alternative, directly estimating effects at the macro-level, often
suffers from under-identification and low statistical power, and furthermore is silent on the

microeconomic channels through which aggregate effects come about.

To overcome these pitfalls, we empirically model three micro-macro linkages that, in

combination with the industry-level estimates, allow us to make broader inferences about
aggregate labor-displacement effects.4 The first link uses harmonized data from the World

Input-Output Database (Timmer et al., 2015) enumerating cross-industry input-output linkages

to trace the effects of productivity growth in each industry to outcomes occurring in customer

industries and in supplier industries—that is, industries for which, respectively, the originating

industry is upstream or downstream in the production chain.5 The second link connects

aggregate economic growth and sectoral labor demands. Recognizing that productivity growth

in each industry augments aggregate income and hence indirectly raises final demand, we

estimate the elasticity of sectoral demand emanating from aggregate income growth and then
apply our TFP estimates to infer the indirect contribution of each industry’s productivity

growth to final demand. Third, our analytic framework recognizes that uneven productivity

4Our approach here builds on our earlier work (Autor and Salomons 2017), in which we incorporate only one of
these linkages.
5Our analysis follows many recent works exploiting these linkages to study the propagation of trade and technology
shocks (e.g. Acemoglu et al. 2016; Pierce and Schott, 2016; Acemoglu, Akcigit and Kerr, 2016).

3
growth across industries yields shifts in industry shares of value-added, which in turn

potentially alter labor’s share of aggregate value-added.6

Our net estimates of the impact of productivity growth and innovation on aggregate

outcomes of interest therefore sum over (1) direct industry-level effects; (2) indirect customer

and supplier effects in linked sectors; (3) final demand effects accruing through the effect of

productivity growth on aggregate value-added; and (4) composition effects accruing through

productivity-induced changes in industry shares of value-added. We believe that this simple

accounting framework can be usefully applied to other data sets and sources of variation.

Distinct from earlier work that focuses on specific measures of technological adoption or

susceptibility (e.g., robotics, routine task replacement), a second contribution of the analysis is

to employ total factor productivity growth (TFP), an omnibus measure of technological progress
(Solow, 1956). Using TFP as our baseline measure potentially overcomes the challenge for

consistent measurement posed by the vast heterogeneity of innovation across sectors and

periods. TFP is also applicable to our analysis for a second reason: because all margins of

technological progress ultimately induce a rise in TFP—either by increasing the efficiency of

capital or labor in production or by reallocating tasks from labor to capital or vice versa—our

empirical approach is not predicated on a specific mechanism through which technological

progress affects outcomes of interest. But the flipside of this agnosticism is that merely

observing a change in TFP in any industry or time period does not tell us which channel

(augmentation, reallocation) is operative. Using information on output, employment, earnings,

and labor’s share of value added, however, we can infer these channels. Specifically, we study

how changes in industry-level TFP affect output (value-added) quantities and prices,

employment, earnings, and labor’s share of value-added economy-wide, to draw inferences on

both industry-level and aggregate labor-augmenting and labor-displacing effects of

technological progress.

It is well understood that estimates of TFP may also be confounded with business cycle

effects, industry trends, and cross-industry differences in cyclical sensitivity (Basu and Fernald,

6This mechanism is akin to skill-biased structural change in the Buera et al. (2015) framework, though here we focus
on labor share rather than skill composition.

4
2001). We confront these issues directly. We purge the simultaneity between an industry’s

estimated TFP growth and changes in other industry-level measures that serve as inputs into

the TFP calculation (e.g., output, wagebill, employment) by replacing own-country-industry

TFP with the mean TFP of the corresponding industry observed in other countries in the same

year.7 We purge the potential cyclicality of TFP by including a set of distributed lags as well as

country-specific business-cycle indicators, which absorb business cycle variation in productivity

measures. We address the opaqueness of TFP as a measure of technological progress by

complementing it with an alternative, directly observable measure of industry-level

technological advancement: patent awards by industry and country (Autor et al. 2017a). Patent

awards—and even more so, patent citations—prove to be strong predictors of industry-level

TFP growth. Using patent awards in place of TFP growth, we obtain strongly comparable
estimates of the relationships between technological progress, employment, wagebill, and value

added, which we view as useful corroborative evidence.

TFP’s virtue as an omnibus technology measure is also its shortcoming as a specific

technology measure. Because TFP incorporates productivity growth arising from all sources,
our analysis cannot directly answer the question of whether recent or specific technologies—

such as industrial robotics or artificial intelligence—are more or less labor-complementing or

labor-displacing than earlier generations of technology. By the same token, our analysis cannot

distinguish between the impacts of automation- versus non-automation-based sources of TFP

growth, which may in turn have distinct (or even countervailing) effects on employment or on

labor’s share of value added. We refer to readers to recent studies focusing on specific

technological advances for this evidence (e.g., Graetz and Michaels, forthcoming; Acemoglu

and Restrepo, 2017; Dauth et al., 2017; Chiacchio et al. 2018).

Our work builds on an active recent literature that questions the optimistic implications of

the longstanding Kaldor facts by offering models where aggregate labor displacement is a

potential consequence of advancing technology. Acemoglu and Restrepo (2018 and

forthcoming) consider models in which two countervailing economic forces determine the

7This strategy leverages the fact that changes in other-country, same-industry TFP are highly predictive of the
evolution of own-country-industry TFP but are not intrinsically correlated with its evolution.

5
evolution of labor’s share of income: the march of technological progress, which gradually

replaces ‘old’ labor-using tasks, reducing labor’s share of output and possibly diminishing real

wages; and endogenous technological progress that generates novel labor-demanding tasks,

potentially reinstating labor’s share. The interplay of these forces need not necessarily yield a

balanced growth path: that is, labor’s share may decline. Susskind (2017) develops a model in

which labor is ultimately immiserated by the asymptotic encroachment of automation into the

full spectrum of work tasks—contrary to Acemoglu and Restrepo (2018b), labor immiseration is
guaranteed because falling labor scarcity does not spur the endogenous creation of new labor-

using tasks or labor-complementing technologies.8

A central empirical regularity that underscores the relevance of this recent work is that

labor’s share of national income has indeed fallen in many nations in recent decades, a trend
that may have become more pronounced in the 2000s (e.g., Elsby, Hobijn and Sahin, 2013;

Karabarbounis and Neiman, 2013; Piketty 2014; Barkai, 2017; Autor et al. 2017b; Dao et al. 2017;

Gutiérrez and Philippon, 2017). Reviewing an array of within- and cross-country evidence,

Karabarbounis and Neiman (2014) argue that labor’s falling share of value-added is caused by a
steep drop in the quality-adjusted equipment prices of Information and Communication

Technologies (ICT) relative to labor. Though Karabarbounis and Neiman’s work is controversial

in that it implies an aggregate capital-labor substitution in excess of unity—which is a non-

standard assumption in this literature—their work has lent empirical weight to the hypothesis

that computerization may erode labor demand. Related work by Eden and Gaggl (2018)

calibrates an aggregate production function and similarly attributes part of the decline in the

U.S. labor share to a rise in the share of income paid to ICT capital.

8 The conceptual frameworks of both papers build on Zeira (1998), Autor, Levy, and Murnane (2003) and Acemoglu
and Autor (2011), who offer models in which advancing automation reduces labor’s share by substituting machines
(or computers) for workers in a subset of activities (which Autor, Levy, and Murnane designate as ‘tasks’). Other
labor-displacement mechanisms are found in Sachs and Kotlikoff (2012) and Berg, Buffie and Zanna (2018), who
develop overlapping-generation models in which rapid labor-saving technological advances generate short-run gains
for skilled workers and capital owners, but in the longer run, immiserate those who are not able to invest in physical
or human capital. Stansbury and Summers (2017) present time-series evidence that productivity growth and wage
growth are positively correlated.

6
 A growing micro-econometric literature presents a mixed set of findings on whether such

erosion has occurred recently or in the past. Focusing on the first half of the twentieth century,

Alexopoulos and Cohen (2016) find that positive technology shocks raised productivity and

lowered unemployment in the United States between 1909 and 1949. Using contemporary

European data, Gregory, Salomons, and Zierahn (2016) test whether Routine-Replacing

Technical Change has reduced employment overall across Europe and find that though this

type of change has reduced middle-skill employment, this reduction has been more than offset
by compensatory product demand and local demand spillovers. In work closely related to ours,

Dao et al. (2017) analyze sources of the trend decline in labor share in a panel of 49 emerging

and industrialized countries. Using cross-country and cross-sector variation in the prevalence of

occupations potentially susceptible to automation (as per Autor and Dorn, 2013), Dao et al. find
that countries and sectors initially more specialized in routine-intensive activities have seen a

larger decline in labor share, consistent with the possibility of labor displacement.9

Concentrating on industrial robotics, arguably the leading edge of workplace automation,

Graetz and Michaels (forthcoming) conclude that industry-level adoption of industrial robots
has raised labor productivity, increased value-added, augmented workers’ wages, had no

measurable effect on overall labor hours, and modestly shifted employment in favor of high-

skill workers within EU countries. Conversely, using the same underlying industry-level

robotics data but applying a cross-city design within the U.S., Acemoglu and Restrepo (2017)

present evidence that U.S. local labor markets that were relatively exposed to industrial robotics

experienced differential falls in employment and wage levels between 1990 and 2007.10

Our analysis proceeds as follows. Section 1 summarizes the data and measurement

framework and presents the simple shift-share decomposition that undergirds our accounting
framework. Section 2 presents our estimates for the direct effects of productivity growth

9Using an analogous approach, Michaels, Natraj, and Van Reenen (2014) find that ICT adoption is predictive of
within-sector occupational polarization in a country-industry panel sourced from EUKLEMS covering 11 countries
observed over 25 years.
10Dauth et al. (2017) and Chiacchio et al. (2018) apply the Acemoglu-Restrepo approach to German and E.U.-wide
data respectively. Dauth et al. find that robot adoption leads to worker reallocation but has no net impact on
employment or wages. Chiacchio et al. affirm the Acemoglu-Restrepo results for employment to population though
not for wages.

7
(measured initially by TFP, in section 2.1; and by patents in section 2.2) on labor input, value-

added, and labor’s share of value-added, across a range of model specifications. Section 3 then

presents our main results accounting for both direct (`own-industry’) effects, and for indirect

effects operating through input-output linkages and final demand. Section 4 quantifies the

aggregate implications of these direct and indirect effect estimates for employment, hours

worked, and labor’s share of value added to assess whether technological progress has in net

been either augmenting or displacing of aggregate employment or of the labor share. We also
consider in this section whether our accounting approach can explain cross-industry patterns of

employment change and aggregate, time-series changes in the evolution of labor-share between

and within industries.

To briefly summarize our results, automation (as embodied in TFP growth) has been
employment-augmenting yet labor-share-displacing over the last four decades. As implied by the

scatter plot in Figure 1A, industries with persistent gains in relative productivity secularly

contract as a share of aggregate employment, meaning that the direct effect of rising

productivity has been to reduce labor input in the sectors where it originates. But this direct
effect is more than fully offset by two indirect effects: first, rising TFP within supplier industries

catalyzes strong, offsetting employment gains among their downstream customer industries;

and second, TFP growth in each sector contributes to aggregate growth in real value-added and

hence rising final demand, which in turn spurs further employment growth across all sectors.

Conversely, we find that productivity growth is directly labor-share displacing in the

industries where it originates, and moreover, this direct effect is not offset by indirect effects

spurred by input-output linkages, compositional shifts, or final demand increases. Thus, we

conclude that productivity growth has contributed to an erosion of labor’s share of value-
added. Notably, this labor-share eroding effect was not present in the first decade of our

sample, the 1970s, but then became strongly evident thereafter. Our analysis therefore broadly

supports the hypothesis that the decline in the labor share since the 1980s is consistent with a

shift towards more labor-displacing technology commencing in the 1980s. But the acceleration
in the labor share decline observed during the 2000s is left unaccounted for by this mechanism.

8
 In the Conclusion, we briefly consider the interpretation of our findings, focusing in

particular on the relationship between the industry-level and aggregate outcomes observed in

our data, and the underlying unobserved firm-level dynamics that may contribute to these

outcomes.

1. Data and measurement

Our analysis draws on EU KLEMS, an industry level panel dataset covering OECD

countries since 1970 (see O’Mahony and Timmer, 2009, http://www.euklems.net/). We use the

2008 release of EU KLEMS, supplemented with data from the 2007 and 2011 releases to

maximize data coverage. Our primary analytic sample covers the period of 1970 – 2007. We
limit our analysis to 19 countries: the developed countries of the European Union, i.e. excluding

Eastern Europe; and Australia, Canada, Japan, South Korea, and the United States. These

countries and their years of data coverage are listed in Appendix Table A1. The EU KLEMS

database contains detailed data for 32 industries in both the market and non-market economies,

summarized in Appendix Table A2. We focus on non-farm employment, and we omit the

poorly measured Private household sector, and Public administration and defense, and

Extraterritorial organizations, which are almost entirely non-market sectors.11 The end year of

our analysis is dictated by major revisions to the industry definitions in EU KLEMS that were
implemented from the 2013 release onwards. These definitional changes inhibit us from

extending our consistent 1970 – 2007 analysis through to the present, though we analyze 2000 –

2015 separately using the 2017 release of EU KLEMS for a smaller subset of countries for which

these data are available.12

Table 1 summarizes trends in aggregate hours of labor input and labor’s share of value-

added by decade for the 19 countries in our analysis. As with all analyses in the paper, these

statistics are calculated using the 28 market industries that comprise our analytic sample and

11Although EU KLEMS classifies healthcare and education as non-market sectors, they are a substantial and growing
part of GDP across the developed world and, in many countries (e.g., the United States), also encompass a large
private sector component. We therefore choose to retain these sectors in our analysis.
12Austria, Belgium, Denmark, Finland, France, Germany, Italy, the Netherlands, Spain, Sweden, the United Kingdom,
and the United States.

9
are annualized to account for the fact that years of data coverage differ by country. With very

few exceptions, aggregate labor hours rise in all countries and time periods. The growth rate of

labor hours is most rapid in the 1980s, slower in the 1990s, and slower still in the 2000s. Distinct

from aggregate labor hours, trends in labor’s share of value-added differ by country and time

period. On average, the aggregate labor share rises in the 1970s and then falls during the

subsequent three decades, with by far the sharpest annual rate of decline in the 2000s.

Table 2 reports analogous statistics for trends in hours of labor input and labor’s share of
value-added among the 28 industries in our sample. There is a substantial diversity of

experiences among industries. Employment fell steeply in mining and quarrying, textiles and

related products, and refining, while growing rapidly in many business and personal services.

Labor’s share of value-added declined in the majority of sectors, with the steepest fall in heavy
industry. TFP growth, meanwhile, was most rapid in manufacturing and was negative in

several service industries.

Table 3 summarizes trends in employment, hours, wages, value-added, labor share, and

TFP by industry over the four decades of our sample. We quantify these trends overall, by
broad sector, and by decade by estimating regression models for the change in country-

industry-year outcomes (multiplied by 100). In this table, and throughout the paper, regression

models are weighted by time-averaged shares of the relevant weighting variable—employment,

hours, or value-added—within countries multiplied by time-varying country shares of the

weighting variable. As such, we weight by country size in our main estimates, and show in the

appendix that our main results are not sensitive to this choice.

Panel A of Table 3 reports estimates for all industries and time periods. Panel B reports

these relationships separately by decade, and panel C reports them separately for five broad
sectors encompassing the 28 industries in our analysis. As detailed in the rubric in Appendix

Table A2, these sectors are: mining, utilities, and construction; manufacturing; education and

health; low-tech services (including personal services, retail, wholesale and real estate); and

high-tech services (including post and telecommunications, finance, and other business
services). The reported regression coefficients, which correspond to within-industry changes,

reflect a number of key trends in the data. Employment growth, measured in workers or hours,

10
is positive in all decades but slows substantially across consecutive decades. Employment

growth is negative in manufacturing; modestly positive in mining, utilities, and construction;

and strongly positive in services, with the most rapid growth evident in high-tech services,

followed by education and health, and finally low-tech services. Like employment, growth of

real hourly wages is positive in all periods but secularly slowing.

Consistent with results reported in much recent work (e.g. Elsby, Hobijn, and Sahin 2013;

Karabarbounis and Neiman 2014; Autor et al. 2017b), trends in the labor share of value added
vary across decades. Labor’s share of value-added trends modestly upward in the 1970s, then

falls in each decade of the 1980s, 1990s, and 2000s. This trend is most pronounced in

manufacturing and in mining, utilities, and construction. It is modest in high-tech services, and

in the education and health sector, and it is absent in the low-tech services sector.
The descriptive statistics given in Table 3 focus on within-industry changes in the labor

share and its components. But of course, changes in the aggregate labor share may stem from

both (1) within-industry shifts in labor’s share of value-added; and (2) changes in the share of

value-added accounted for by industries that differ in their labor shares. Our analysis will
assess the contribution of technological change to both margins. To quantify the importance of

within- versus between-industry shifts, we implement a simple shift-share decomposition as

follows. Let 𝐿"#$ = ∑( 𝜔(#$ 𝑙(#$ equal the aggregate log labor share in country 𝑐 in year 𝑡, defined as

the weighted sum of log labor shares 𝑙(#$ in each industry 𝑖, where weights 𝜔(#$ correspond to

industry 𝑖′𝑠 share in value-added in its respective country and year.13 Let ∆𝐿"#0 equal the change

in aggregate log labor share in country 𝑐 over time interval 𝜏, equal to 1970-80, 1980-90, 1990-00,
̅ = 4𝑙(#,$6 − 𝑙(#,$8 9/2 and 𝜔
or 2000-07, where Δ is the first difference operator. Finally, let 𝑙(#0 <(#0 =

4𝜔(#,$6 − 𝜔(#,$8 9/2. We can then decompose the observed labor share change in each decade as:

∆𝐿"#0 = = 𝜔 ̅ Δ𝜔(#0 ,
<(#0 Δ𝑙(#0 + = 𝑙(#0
( (
(1)

where the first term to the right of equal sign is the contribution of within-industry changes in
labor share to the aggregate change and the second term is the contribution to the aggregate

change due to shifts in value-added shares across industries.

13 Per our convention, this calculation includes only the 28 market industries featured in our analysis.

11
 The results of this decomposition, reported in Table 4, indicate that the majority, but not the

entirety, of the change in aggregate labor share in each decade is accounted for by within-

industry shifts, consistent with evidence for the U.S. (Eden and Gaggl 2018). Focusing first on

the country size-weighted calculations, we find that more than all of the rise in labor share in

the 1970s is due to within-industry changes, whereas between 51 percent and 72 percent of the

fall in the labor share in the subsequent three decades is accounted for by within-industry

declines. If we instead weight each country equally in the shift-share decomposition (shown in
the right-hand panel of Table 4), we reach similar conclusions about the importance of within-

industry labor share movements. Further, if we decompose the change in the mean level of labor

share rather than the mean log level (Appendix Table A3), we find a similar time pattern as for

the log labor share and a similarly outsized role played by within-industry changes.
These decomposition results suggest that the within-industry determinants of changes in

the aggregate labor share are of greater analytic interest compared with the between-industry

drivers, though we explore both margins below. The 2000s stand out, however, for having a

roughly even distribution of aggregate labor share changes into within- and between-industry
components. Consistent with the observations of Rognlie (2015) and Gutiérrez (2017), this

pattern reflects the outsized growth of the Real Estate industry´s value added in numerous

countries—particularly during the 2000s—and this industry has an extremely low share of labor

in value-added (see Appendix Table A4). If we eliminate Real Estate from the analysis,

however, we find that the fall in the aggregate labor share in the 2000s is reduced by less than

one quarter (from −0.86 to −0.64 per year); the within-industry component of the labor share

decline explains no less than 90 percent of the total in each decade; and the annual rate of

decline in the labor-share during the 2000s is still more than twice as rapid as in the 1990s.14
Thus, the rising share of real-estate in value-added is not the primary driver of the falling labor

share.

Figure 2 adds country-level detail to these calculations by plotting the evolution of the

aggregate labor share of value added for all countries in our sample. Each panel contains two

14 Supplemental tables are available upon request from the authors.

12
series: in the first series industry shares are permitted to vary by year; and the second series

holds these shares constant at their within-country, over-time averages. The fact that these

series closely correspond for almost all countries reinforces the inferences from the

decomposition that most of the aggregate changes in the labor share observed in the data stem

from within-industry movements in this share.

2. Main estimates
Before making estimates, we tackle two remaining issues: simultaneity and timing. The

simultaneity issue arises because labor’s share of value-added features in the construction of

TFP, inducing a mechanical correlation between TFP growth and shifts in the labor share.15 To
overcome this pitfall, we construct industry-level TFP growth for each industry-country pair as

the leave-out mean of industry-level TFP growth in all other countries in the sample. This

approach eliminates the mechanical correlation between TFP and labor share and arguably

exploits movements in the technology frontier that are common among industrialized

economies. Confirming the utility of this strategy, we show in Appendix Table A5 that other-

country-same-industry TFP is a strong predictor of own-country-industry TFP: in a set of

regressions of own-country-industry TFP on other-country-industry TFP that includes a large

number of country, year, sector, and business cycle main effects, we obtain a prediction
coefficient that ranges from 0.32 to 0.57, with a t-value above five in all specifications. Based on

this reasoning and evidence, we employ the leave-out TFP measure in place of own-industry

TFP in all the analyses given below.

The second issue—timing—arises because contemporaneous productivity innovations are

unlikely to induce their steady state effects immediately, meaning that a lag structure is needed

for estimating the relationship between TFP and outcomes of interest (Ramey 2016). To explore

a suitable structure, we estimate simple local projection models in the spirit of Jordà (2005),

which involve regressing a series of first-differences of increasing length of the outcome

15In EU KLEMS, TFP growth is calculated as the log change in industry value-added minus the log change in labor
and capital inputs, weighted by the average start and end period of their respective factor shares (Timmer et al. 2007).
In a regression of the change in labor share on TFP growth, the change in labor share used in the TFP calculation
enters the right-hand side of the equation, leading to a mechanical relationship.

13
variable of interest on the explanatory variable of interest (here, TFP growth) and a set of

controls. We estimate

ln 𝑌(#,$KL − ln 𝑌(#,$M6
L

= 𝛽8 + 𝛽6 Δ ln 𝑇𝐹𝑃(,#R#((),$M6 + = 𝛽UV ∆ ln 𝑇𝐹𝑃(,#R#((),V (2)

VW8

+ 𝛽X Δ ln 𝑇𝐹𝑃(,#R#((),$MU + 𝛽Y Δ ln 𝑌(#,$MU + 𝛼#$ + 𝛾\ + 𝜀(#$ ,

where ln 𝑌(#,$KL denotes the log outcome of interest in industry 𝑖, country 𝑐, and year 𝑡, and 𝐾

denotes the time horizon for the local projection. The dependent variables therefore reflect the

log change in outcome 𝑌 from base year 𝑡 − 1 up to year 𝑡 + 𝐾. The impulse variable is the log
change in other-country-industry TFP between years 𝑡 − 2 and 𝑡 − 1, Δ ln 𝑇𝐹𝑃(,#R#((),$M6 . These

effects are estimated while controlling for lagged values of both TFP growth (Δ ln 𝑇𝐹𝑃(,#R#((),$MU )

and of outcome variable growth (Δ ln 𝑌(#,$MU ) – that is, conditional on the lagged history of both

TFP and outcome growth at the path start time. This allows for feedback dynamics within the

system and controls for them through the inclusion of the lagged variables. Each model further
controls for a set of country--year fixed effects (𝛼#$ ), as well as fixed effects for five broad sectors

(𝛾\ , as outlined in Appendix Table A2). Following the approach of Teulings and Zubanov

(2014), we also control for subsequent TFP innovations occurring between 𝑡 = 0 and 𝑡 = 𝐾,

which reduce the influence of serial correlation in TFP innovations on estimates of 𝛽6 . Finally,
standard errors are clustered by country-industry.

Figure 3 reports local projection estimates and confidence intervals for the relationship

between a TFP innovation shock, measured as an increase in TFP of 1 standard deviation,

occurring between periods 𝑡 = −1 and 𝑡 = 0 and ensuing industry-level changes ΔV ln 𝑌(# ≡

ln 𝑌(#,$KV − ln 𝑌(#,$M6 for 𝐾 ∈ {0, … ,5}.16 For all the outcome variables considered (employment,

hours, wagebill, value-added, and labor share), the local projection estimates indicate that TFP

growth predicts small or negligible contemporaneous changes in the outcome of interest that

cumulate in ensuing years. In all cases, however, these effects plateau after three years,

implying that no more than four lags of the independent variable are needed to capture the

16 The standard deviation of TFP growth is 2.6 log points (as reported in Appendix Table A6).

14
impulse response of a contemporaneous shock. For completeness, we include five lags in our

main specifications, though we shorten the lag structure when analyzing sub-intervals of the

data.

2.1. Within-industry direct effects: Own-industry TFP and own-industry outcomes

Our initial estimates, reported in Table 5, consider the within-industry ‘direct’ effects of TFP

growth on own-industry outcomes. We fit OLS first-difference models of the form

∆ ln 𝑌(#$ = 𝛽8 + = 𝛽6V ∆ ln 𝑇𝐹𝑃(,#R#((),$MV + 𝛼# + 𝛿$ + 𝛼# × 𝑡 + 𝛼# × (𝑡 = 𝑝𝑒𝑎𝑘)

VW8 (3)
+ 𝛼# × (𝑡 = 𝑡𝑟𝑜𝑢𝑔ℎ ) + 𝜀(#$ ,
where ∆ ln 𝑌(#$ is an outcome of interest and as above, 𝑖 indexes industries, 𝑐 indexes countries,

and 𝑡 indexes years; and the log change in TFP (contemporaneous plus five distributed lags) is

the explanatory variable of interest. Because equation (3) is a first-difference specification

estimated at the industry-country-time level, it implicitly eliminates industry-country effects.

We additionally include country and year indicator variables, which correspond to linear

country and time trends in the first-difference model; country-time interaction terms, which

allow country trends to accelerate or decelerate over the sample interval; and country-specific

cyclical peak and trough indicators interacted with country indicators to account for country-

specific business cycle effects. All models are weighted by industries´ time-averaged shares of

the relevant weighting variable—employment, hours, or value-added—within countries

multiplied by time-varying country shares of the weighting variable, and standard errors are
clustered at the level of country-industry pairs.

The first panel of Table 5 presents estimates for industry-level employment, measured as the

(log) number of workers (encompassing both employees and the self-employed). We estimate

that industries experiencing relative gains in productivity exhibit relative declines in

employment. The point estimate of −2.07 in column 1, corresponding to the sum of the six 𝛽6V

coefficients, implies that an increase of 1 standard deviation in own-industry TFP (2.58 log

points) predicts a fall in own-industry employment of approximately 2 log points. This estimate

15
implies that the estimated elasticity of employment to TFP growth is below unity (0.80 =

2.07/2.58)—that is, there is a partial industry-level demand offset (cf. Bessen 2017).

Columns 2 and 3 stress-test this estimate by adding five major sector-group fixed effects and

by replacing the country-trend and country-specific business-cycle controls with an exhaustive

set of country-year indicator variables. The inclusion of sector-group trends reduces the point

estimate from −2.07 to −1.13 and increases precision. This pattern suggests that TFP

innovations may spill-over across industries within a sector. We subsequently model these
spillovers in the next section, when we add input-output linkages to the regression model;

meanwhile, we add sector-group dummies (reflecting sector-group trends in the log level

models) to all subsequent models, so our primary identification comes from within-sector,

between-industry comparisons. Conditional on the inclusion of these sector-group trends, the

addition of a full set of country-year dummies in column 3 has almost no impact on the

magnitude or precision of the point estimates. This insensitivity is worth bearing in mind

because we do not include exhaustive country-year dummies in our main models; these

dummies would interfere with the identification of input-output linkages, which have much
lower country-year variability than own-industry TFP.

Panel B of Table 5, which reports analogous estimates for log hours of labor input, finds an

almost identical slope as for employment, indicating that most of the employment adjustment

to productivity changes occurs on the extensive margin. Panel C explores the relationship

between TFP and nominal industry wagebill changes. These point estimates are also similar to

those for hours and employment, suggesting that industry (relative) nominal wages are not

much affected by TFP changes; rather, the industry-level relationship between TFP and

wagebill changes stems from employment shifts.

We turn to output measures in Panels D and E of Table 5. Rising industry TFP predicts

significant relative declines in industry-level nominal value-added (Panel D) and significant

relative rises in real industry value-added (Panel E), implying (logically) that rising industry

productivity lowers industry prices.

Comparing the estimates in Panels D and E reveals that a rise in industry TFP predicts a

smaller (less negative) change in nominal value-added than in the wagebill. This suggests that

16
rising TFP predicts a (relative) fall in labor’s share of industry value-added.17 Panel F of the

table confirms this implication: a rise in TFP of 1 standard deviation predicts a fall in an

industry´s labor share of value-added of approximately 0.55 percentage points over a five-year

horizon.

We have implemented a large number of tests of the robustness of these estimates, which

are reported in Table 6. These include: weighting all countries equally rather than by their

value-added shares (panel A); eliminating the contemporaneous TFP term from the distributed
lag model (panel B); eliminating the self-employed from our employment, wagebill, and labor-

share models (panel C); imputing zeros to the TFP measures in cases where the reported values

are negative (panel D)18; estimating eqn. (3) using five-year long first-differences in place of

annual first-differences (panel E) 19; and using data from the 2000 – 2015 period from the 2017
release of the EU KLEMS data (Van Ark and Jäger, 2017), thus adding eight additional outcome

years at the cost of dropping prior decades and several countries (panel F).20 Results are

remarkably stable across these many sets of estimates, though precision is much lower for

models fit using the short 2000-2015 panel.

The robust negative industry-level relationships between TFP and both employment and

labor’s share of value-added seen in Tables 5 and 6 are central inputs into our subsequent

analysis. We stress that these findings do not by themselves imply that productivity growth

depresses either employment or the labor share in the aggregate. Indeed, these direct within-

industry relationships do not at present incorporate any of the potentially countervailing effects

17Because the wagebill regression is weighted by hours shares and the value-added regression by value-added
shares, the precise impact of TFP growth on the labor share cannot be directly inferred from a comparison of these
two columns.
18Thirty-six percent of all country-industry-year TFP growth observations are negative. This is most frequently the
case for Renting of machinery and equipment and other business activities (code 71t74), Other community, social and
personal service activities (code O), Hotels and restaurants (code H), and Real estate activities (code 70). But it occurs
in all industries to some extent. The likely cause is that annual frequency TFP calculations incorporate a fair amount
of measurement error, leading to short-run intervals where nominal value-added rises less rapidly than the share-
weighted growth of labor and capital inputs.
19These estimates are obtained from full-length five-year intervals (1970-1975, 1975-1980, …, 2000-2005) only; and the
reported coefficients reflect the effect of TFP growth occurring over the previous five-year interval.
20More recent EUKLEMS releases cover a smaller set of countries and rely on back-casting data prior to 1995. We use
a balanced panel of 12 countries (Austria, Belgium, Denmark, Finland, France, Germany, Italy, the Netherlands,
Spain, Sweden, U.K., and U.S.) over 2000-2015.

17
operating through other channels, including input-output linkages, compositional shifts, and

final demand effects. Before incorporating these links in the next section, we perform a validity

test on our main technology measure.

2.2. Applying direct measures of technological progress

Our omnibus measure of productivity-augmenting technological change, TFP, has the

advantage that it is not bound to a specific set of technologies or their associated measurement

challenges. But TFP’s strength is also its weakness. Because it is an accounting residual, one can

only speculate on the underlying sources of technological progress that contribute to rising TFP.

To partially address this concern, we test whether our key results above hold when we focus on

a specific margin of technological advancement: industry-level patenting flows (as in

Acemoglu, Akcigit, and Kerr 2017).

Using data from Autor et al. (2017a), who match patent grants to their respective corporate
owners and then to industry codes based on corporate owners’ industry affiliations, we

construct counts of patent grants and patent citations by year for patents granted to both US

and non-US inventors using U.S. Patent and Trademark Office (USPTO) data by U.S. Standard

Industrial Classification (SIC) industry, cross-walked to the EUKLEMS industry level.

Aggregate summary statistics for standardized patent counts and patent citations are reported

in the lower panel of Table A6, while Appendix Table A7 reports the mean log number of

patent grants and patent citations by industry and by inventor nationality (U.S. versus non-

U.S.), and Appendix Table A8 summarizes industry-level trends by decade and sector. These

tables highlight the substantial heterogeneity in patent flows across industries and over time,

with the highest levels of patenting occurring in Chemicals and Electrical equipment, and the

lowest occurring in Education. Patent grants rise across decades while citations fall in the most

recent decade, reflecting the substantial lag between patent grants and patent citations.

Although citations are likely a better measure of innovation than the raw count of patent grants

(Trajtenberg 1990), citations may understate innovation in the final years of the sample because

they arrive with a lag. In what follows, we report results using both measures of patenting

activity.

18
 Given that patenting activity is an input into the industry-level innovation and automation

process, it should predict TFP growth. To verify this supposition, we estimate industry-level

descriptive regressions of the form:

X

∆ ln 𝑇𝐹𝑃(#$ = 𝛽8 + = 𝛽6V ln 𝑃𝐴𝑇(,#R#((),$MV + 𝛼# + 𝛿$ + 𝛼# × (𝑡 = 𝑝𝑒𝑎𝑘)

VW8 (4)
+ 𝛼# × (𝑡 = 𝑡𝑟𝑜𝑢𝑔ℎ ) + 𝜀(#$ ,

where ∆ ln 𝑇𝐹𝑃(#$ is the measured change in industry-level TFP, and ln 𝑃𝐴𝑇(,#R#((),$ is log count

of industry-level patents, which are normalized to have a standard deviation of 1. Paralleling

the specifications given above, we include both contemporaneous patenting activity and a set of
annual distributed lags. Analogous to our strategy of using other-country (‘leave out’) TFP

growth by industry, we use patenting activity by non-U.S. inventors as predictors of U.S. TFP

growth and, similarly, use patenting activity by U.S. inventors as predictors of non-U.S. TFP

growth.
The estimates of eqn. (4), reported for patent counts in the upper panel of Table 7 and for

patent citations in the lower panel, confirm that patent flows are a strong predictor of industry

TFP growth. A one-standard deviation higher rate of industry patents or patent citations

predicts approximately 0.6 log points faster industry TFP growth (𝑡 = 2.9). This relationship is
robust: adding year effects (column 2), country-specific business-cycle effects (column 3), and

country- year effects (column 4) to these first-difference models has almost no impact on the

magnitude or precision of the predictive relationship.

Table 8 explores the relationship between patenting activity and the evolution of industry-
level labor input, value-added, and factor payments. Following the template of the tables

presented above, we report regressions of industry-level first-differences in outcome variables

on log industry patent counts or patent citations—contemporaneous and five annual

distributed lags—and the full set of controls used in Table 7.21 Comparable to the pattern of
results for TFP, we find that industry-level patent citation flows predict a fall in own-industry

21Since the majority of variation in patenting reflects stable cross-industry differences rather than over-time, within-
industry fluctuations, we exclude sector-specific indicators from these models (which would otherwise absorb most
identifying variation). Due to this limited variation, we confine our patent analysis to direct (own-industry) effects.

19
employment and hours, a decline in nominal value-added, a rise in real value-added, and, most

importantly, a fall in own-industry labor share.22 These findings hold for both measures of

patenting activity— patent counts and patent citations. Though precision is far lower for the

patent- than TFP-based estimates—likely because we effectively have patenting data for only

two countries, U.S. and ‘non-U.S.’—we view these findings as supportive of our main results.

3. Linking micro to macro

As underscored by Figure 1A, it would be erroneous to conclude that because relative

employment declines in industries experiencing rising productivity, aggregate employment falls

as productivity rises. To move from this cautionary observation to a rigorous quantification of

how industry-level productivity growth affects employment and labor share in the aggregate,

we next add three micro-macro linkages to our estimation and accounting framework:

customer-supplier linkages; final demand effects; and composition effects.

3.1. Accounting for customer-supplier linkages

The effect of productivity growth occurring in an industry is unlikely to be confined to the

sector in which it originates. Industries facing lower input prices or higher quality inputs from

their suppliers may increase purchases; similarly, industries whose customers are experiencing

rising productivity may face rising or falling output demands. We account for these input-

output linkages by adding two terms to eqn. (3):

e e e

∆ ln 𝑌(#$ = 𝛽8 + = 𝛽6V ∆ ln 𝑇𝐹𝑃(,#R#((),$MV + stR(,#,$MV

= 𝛽UV ∆ ln 𝑇𝐹𝑃 uvw stR(,#,$MV
+ = 𝛽XV ∆ ln 𝑇𝐹𝑃 xvuy

VW8 VW8 VW8 (5)

+ 𝛼# + 𝛿$ + 𝛾\ + 𝛼# × 𝑡 + 𝛼# × (𝑡 = 𝑝𝑒𝑎𝑘) + 𝛼# × (𝑡 = 𝑡𝑟𝑜𝑢𝑔ℎ ) + 𝜀(#$ .

stR(,#,$
These additional terms, 𝑇𝐹𝑃 uvw stR(,#,$
and 𝑇𝐹𝑃 xvuy
, measure the weighted sum of TFP growth in all

other domestic industries 𝑗 ≠ 𝑖, which are, respectively, the suppliers and customers of industry

𝑖:23

22Due to the differences in underlying units, the magnitude of coefficients cannot be directly compared between the
TFP and patents models.
23We eliminate the on-diagonal (own-industry) term from the input-output measures since these are captured by the
direct TFP terms (𝛽6V ).

20
 ~

stR(,#,$
∆ ln 𝑇𝐹𝑃 | |
= = 𝑤𝑒𝑖𝑔ℎ𝑡tR(,# |
× ∆ ln 𝑇𝐹𝑃tR(,#,$ , ∀ 𝐿 ∈ 𝑆𝑈𝑃, 𝐶𝑈𝑆𝑇, (6)
tW6

The supplier and customer weights used for this calculation are obtained from input-output
coefficients from the World Input-Output Database (WIOD) and are averaged over 1995-2007.

The supplier weights are equal to each domestic supplier industry 𝑗′s value-added as a share of

the value-added of industry 𝑖, capturing the importance of supplier industries 𝑗 in the

production of industry 𝑖’s output. Analogously, the customer weights are the shares of value-

added of each industry 𝑖 that are used in domestic industry 𝑗’s final products, capturing the

importance of industries j as end-consumers of industry i’s output. 24 These weights account not

only for shocks to an industry’s immediate domestic suppliers or buyers but also for the full set

of input-output relationships among all connected domestic industries (i.e., the Leontief
inverse). We renormalize both the customer and supplier TFP terms to have a standard

deviation of 1, with summary statistics reported in Appendix Table A6. As with our main

(direct) measure of TFP, these supplier and customer TFP linkage terms are calculated using

industry-level leave-out means of TFP growth in all other countries in the sample.

The estimates of equation (5), reported in the upper panel of Table 9, indicate that

productivity growth emanating from supplier industries predicts steep increases in the

employment and hours of labor input of customer industries (though not in their nominal

wagebill, value-added, or labor share). Specifically, the point estimate of 0.97 on the supplier-
industry TFP term in column 1 indicates that a one standard deviation rise in an industry’s

supplier productivity predicts an employment gain of 97 log points. This effect is almost

identical in magnitude but opposite in sign to the estimated direct effect of TFP growth of −0.95

on own-industry employment. Thus, this input-output linkage reveals a first channel by which

direct effects of productivity growth on own-industry outcomes may be offset by effects

accruing outside the originating sector.

24While every industry is potentially both a customer and supplier to every other industry, the terms customer and
supplier refer to the direction of flow of inputs and outputs: suppliers produce outputs that are purchased by
(downstream) customers; and customers purchase inputs produced by (upstream) suppliers.

21
 Conversely, productivity growth emanating from customer industries (row 3 of the upper

panel of Table 9) generally has negligible and always insignificant estimated effects on

employment, hours, wagebill, value-added, and labor share in supplier industries. This result is

consistent with the simple Cobb-Douglas input-output framework developed in Acemoglu,

Akcigit, and Kerr (2016), where productivity innovations in a given industry lead to output gain

in its customer industries—those benefiting from its price declines—but have no net effect on its

supplier sectors, where price and quantity effects are offsetting.

A third important pattern revealed by Table 9 is that our earlier estimates of the relationship

between TFP growth and own-industry outcomes are essentially unaffected by the inclusion of

the customer and supplier terms (compare the point estimates in Tables 5 and 9). Thus, our

initial findings for the relationship between TFP growth and own-industry employment and
labor share are unaltered.

3.2. Accounting for final demand effects

The lower panel of Table 9 adds a third channel of response: final demand effects accruing

through the contribution of productivity growth to aggregate value-added. To capture these

final demand effects, we estimate the relationship between country-specific aggregate economic

growth (contemporaneous and five distributed lags) and industry-specific inputs using the

following specification:
e

∆ ln 𝑌(#$ = 𝜆8 + = 𝜆6V ∆ ln 𝑉𝐴tR(,#,$MV + 𝛼\ + 𝜀(#$ , (7)

VW8

The explanatory variable of interest in this equation, ∆ ln 𝑉𝐴tR(,#,$ , is the growth of own-country

real or nominal value-added, where the subscript 𝑗 ≠ 𝑖 highlights that we exclude own-industry

output from the explanatory measure for each industry to eliminate any mechanical correlation

between aggregate growth and industry outcomes. These stacked first-difference regression

models drop the country, year, trend, and business-cycle indicators used in equation (5), so that
identification largely arises from country and year time-series. Since these are first-difference

models, however, they implicitly eliminate industry-country effects.

The estimates of equation (7), reported in the lower panel of Table 9, document a second

countervailing effect of industry-specific productivity innovations on aggregate outcomes: each

22
log point gain in country-level real value-added predicts an approximately 0.6 log point rise in

same-country, other-industry employment and hours. Similarly, each log point gain in country-

level nominal value-added predicts essentially a one-for-one rise in same-country, other-

industry wagebill and nominal value-added, as well as a very modest but statistically

significant rise in same-country, other-industry labor-share (the estimated elasticity is 0.071).

Because TFP growth emanating from any one sector raises the real aggregate value-added in the

country where it occurs, these estimates imply that each industry’s productivity growth
contributes to aggregate labor demand across all other sectors.25

3.3. Accounting for compositional (between-sector) effects

The estimates given in Table 9 reveal one further mechanism by which sectoral productivity

gains affect the aggregate labor share: by shifting relative sector sizes. Column 4 of panel A

finds that a rise in own-industry TFP growth predicts a fall in industry-level nominal value-
added with an elasticity of −0.58. This finding implies that sectors with rising productivity will

tend to shrink as a share of nominal value-added. Figure 4 confirms this intuition by depicting a

scatter plot of the bivariate relationship between industry-level TFP growth and the change in

industries’ log shares of own-country nominal value-added (averaged over years and across

countries). On average, industries that experience one log point faster TFP growth than the

economy-wide average lose about 0.6 log point as a share of nominal economy-wide value-

added.

Applying this observation to the Oaxaca decomposition equation above (eqn. 1), it is

immediately clear that uneven productivity growth across industries will shift the aggregate

labor share through changes in relative sector sizes. If rapid productivity growth occurs in

industries with relatively low labor shares (e.g., manufacturing industries), this will indirectly

25We report a pure stacked country-level time series version of these estimates in Appendix Table A9 in which we
eliminate industry level variation entirely and use instead only country--year observations. These point estimates are
similar to those in panel B of Table 9, which we prefer because they eliminate the mechanical relationship between
own-industry and country-level aggregate outcomes.

23
raise the aggregate labor share; conversely, relatively rapid productivity growth in labor-

intensive sectors (e.g., education and health) will have the opposite effect.26

4. Quantitative implications
With these estimates in hand, we now quantify the implied contribution of TFP growth to

employment and labor share evolutions in the aggregate accruing through the four channels

outlined above: own-industry, supplier and customer, final demand, and composition. We start

with employment and hours, then turn to the labor share.

4.1. Aggregate employment and hours effects

The effect of TFP growth on employment and hours combines the first three of these effects:

the own-industry (or `direct’) effect, the supplier and customer effects, and the final demand

effect. The first (own-industry) effect is equal to the sum of the 𝛽6V coefficients in eqn. (5)

multiplied by their corresponding ∆ ln 𝑇𝐹𝑃(,#R#((),$ terms, and aggregated by weighting these

industry-level predictions by the time-averaged share of each industry in total employment or

hours:
e Š
𝜕 ln 𝑌#$
Δ ln 𝑌#$†‡ˆ ≡ = = 𝛽6V = 𝜔(# × ∆ ln 𝑇𝐹𝑃(,#R#((),$ (8)
𝜕 ln 𝑇𝐹𝑃(,#R#((),$
VW8 (W6

Here, ln 𝑌#$ is log employment or hours in county 𝑐 in year 𝑡; ∑eVW8 𝛽6V is the sum of coefficients

in eqn. (5); 𝜔(# is the time-averaged employment or hours share of industry 𝑖 in its respective
country; and ∆ ln 𝑇𝐹𝑃(,#R#((),$ is own-industry TFP growth.

The supplier and customer effects are, analogously, equal to the sum of the 𝛽UV and 𝛽XV
stR(,#,$
coefficients multiplied by their corresponding 𝑇𝐹𝑃 uvw stR(,#,$
and 𝑇𝐹𝑃 xvuy
terms, and then

aggregated to the national level by weighting each by its time-averaged industry employment

or hours shares (𝜔(# ):

e Š
𝜕 ln 𝑌#$
Δ ln 𝑌#$uvw ≡ stR(,#,$
= = 𝛽UV = 𝜔(# × ∆ ln 𝑇𝐹𝑃 uvw
and (9)
stR(,#,$
𝜕 ln 𝑇𝐹𝑃 uvw
VW8 (W6

26The upstream and downstream linkages estimated in eqn. (5) can also contribute to the between-industry
component of the labor share change through their effects on industry nominal output shares, though we estimate
these effects to be comparatively small and statistically insignificant.

24
 e Š
𝜕 ln 𝑌#$
Δ ln 𝑌#$xvuy ≡ stR(,#,$
= = 𝛽XV = 𝜔(# × ∆ ln 𝑇𝐹𝑃 xvuy
.
stR(,#,$
𝜕 ln 𝑇𝐹𝑃 xvuy
VW8 (W6

The third component that we calculate is the final demand effect of TFP growth in each industry
on employment or hours economy-wide, Δ𝑌#$•Ž . For any one industry, this contribution is equal

to the product of four terms: (1) the effect of TFP growth in 𝑖 on 𝑖′𝑠 real value-added

(∑eVW8 𝛽6,••
V
); (2) the effect of growth in 𝑖′𝑠 real value-added on total value-added (𝜙(# ); (3) the

effect of growth in real-value added on employment or hours in each industry 𝑗 ≠ 𝑖 (∑eVW8 𝜆6V );

and (4) the size of industry 𝑗 relative to overall employment or hours in the economy (𝜔(# ).27 To

obtain the aggregate effect (summing across industries), we calculate:

e Š
𝜕 ln 𝑌#$ 𝜕 ln 𝑉𝐴#$ 𝜕 ln 𝑉𝐴#$ 𝜕 ln 𝑉𝐴(#$
Δln𝑌#$•Ž ≡ = = 𝜆6V = 𝜔(# ’ × “ (10)
𝜕 ln 𝑉𝐴#$ 𝜕 ln 𝑇𝐹𝑃(,#R#((),$ 𝜕 ln 𝑉𝐴(#$ 𝜕 ln 𝑇𝐹𝑃(,#R#((),$
VW8 (W6
e e Š

= ”= 𝜆6V × V
= 𝛽6,•• • = 𝜔(# × 𝜙(# .
VW8 VW8 (W6

In this expression, ln 𝑌#$ is log employment or hours in county 𝑐 in year 𝑡 as before; ∑eVW8 𝜆6V is

the estimated effect of aggregate real value added on outcome 𝑌 from eqn. (7) reported in

column 5 of the lower panel of Table 9; ∑eVW8 𝛽6,••

V
is the estimated direct effect of ∆ ln 𝑇𝐹𝑃 in

eqn. (5) on own-industry real value-added (reported in column 5 of Table 9’s upper panel); 𝜔(#
is the time-averaged employment or hours share of industry 𝑖 in its respective country; and 𝜙(#

is the time-averaged value-added share of industry 𝑖 in country 𝑐.28

Figures 5A and 5B display the results of these calculations for overall employment and

hours of labor input, respectively. The first bar in Figure 5A corresponds to the direct effect of

TFP growth on own-industry employment. Its height of −0.22 implies that on average,

productivity growth reduced own-industry employment by approximately 8.2 percent over the

full 37-year period (0.22 × 37). The second bar (“supplier effect”), with height 0.35, indicates

27In calculating Δ𝑌#$•Ž , we also include the customer and supplier TFP effects estimated in eqn. 5, though we suppress
those terms above to conserve notation.
28This last term 𝜙𝑖𝑐 is derived by differentiating the sum of industry log value-added at the country level with
respect to the log value-added of industry 𝑖 in country 𝑐, which is simply equal to 𝑖′𝑠 share in country 𝑐′𝑠 value-
added. Note that the sum of industry shares is less than one since we exclude non-market industries from the
analysis, though they are (logically) included in aggregate national value-added.

25
that the countervailing effect of rising supplier productivity on employment in customer

industries more than offset this direct effect. The third bar (“customer effect”) with height

0.05 indicates that rising productivity in customer industries exerted a very modest positive

employment effect in supplier industries. The fourth bar with height 0.30 reflects the substantial

contribution of rising productivity to overall employment operating through final demand. The

fifth bar (“net effect”) sums these four components to estimate a net positive effect of

productivity gains on aggregate employment, totally approximately eighteen log points

(0.48 × 37 = 17.8) over the outcome period.

Figure 5B reports the analogous exercise for hours of labor input rather than employment.

We find comparable effects on this outcome: although rising productivity reduces relative

employment in the sectors in which it occurs, it augments employment in (downstream)

customer sectors (as captured by the supplier effect) and boosts aggregate demand through its

contribution to overall value-added. As with employment, the net effect on hours is strongly

positive.

To provide insight into how rising TFP spurs (relative) employment declines in directly-
affected industries while simultaneously generating rising employment in the aggregate,

Appendix Tables A11A and A11B report the contributions to employment growth by industry

operating through each channel estimated above: direct effects, input-output linkages, and final

demand effects. These contributions, underlying the aggregate employment growth predictions

in Figures 5A and 5B, can be analyzed from two complementary perspectives. The first,

reported in Appendix Tables A11A , calculates the contribution of TFP growth originating in

each industry to the predicted aggregate change in employment. The second, reported in

Appendix Table A11B, enumerates the predicted effect of TFP growth originating throughout
the economy on predicted employment growth in each destination industry, scaled by that

industry’s weight in aggregate employment.29

For the direct effect, the contributions to employment in the originating and destination

industry are the same by definition since these direct effects operate only within industries. As

We do not separately report contributions for hours worked because they are nearly identical to those for
29

employment.

26
shown in Tables A11A and A11B, the negative direct effects that we estimate for employment

originate in industries that have experienced strong TFP growth (such as Electrical and optical

equipment, and Transport and storage), or industries that make up a large share of total value-

added (such as Retail), or both.

Conversely, TFP growth originating in supplier and customer industries leads to

employment and hours growth elsewhere in the economy through input-output linkages. The

supplier-customer contribution of any given industry to aggregate employment depends on

three terms: the industry’s rate of TFP growth; the weight that industry has as a supplier or

customer of other industries; and, in turn, the weight that those customer and supplier

industries have in aggregate employment. Industries such as Post and telecommunications,

Wholesale trade, Financial intermediation, and Transport and storage, produce important
positive employment spillovers to other industries, in part because they are suppliers to a

variety of service industries, which are themselves a large share of total employment. These

industries highlight the potential of productivity growth in service industries to induce sizable

positive employment spillovers. On the other hand, Other business activities, an important
supplier industry, exhibits declining productivity and thus contributes a meaningful negative

employment spillover. Finally, manufacturing industries such as Chemicals, Basic and

fabricated metals, and Electrical and optical equipment, make a large indirect contribution to

employment in customer industries due to their rapid productivity growth.30

Finally, each industry’s TFP growth potentially contributes to employment via its effect on

final demand. This effect depends on two terms: the originating industry’s rate of TFP growth

and its share in national value-added. Hence, productivity growth in industries that make up a

large share of value-added has a larger effect on overall income. Electrical and optical
equipment, Post and telecommunications, Financial intermediation, Transport equipment,

30The indirect employment contribution made by productivity gains in customer industries is much smaller than the
corresponding effect of productivity gains in supplier industries, and it is primarily driven by TFP growth in
Electrical and optical equipment, Transport equipment, and Machinery.

27
Chemicals, and Wholesale trade are the largest contributors by TFP source to final demand,

reflecting their rapid productivity growth and substantial weight in aggregate value-added.31

How successful is our approach in capturing the evolution of employment observed in the

data? Figures 6A and 6B answer this question by comparing the industry-level employment

predictions of our statistical model to observed employment changes, averaged across country-

years. In each figure, employment growth predictions, obtained by summing across all channels

in the model, are reported on the horizontal axis, while observed employment growth is
reported on the vertical axis. Figure 6A plots the predicted versus observed log employment

change by industry, while Figure 6B plots the predicted versus observed contribution that each

industry makes to aggregate employment growth.32 These figures make evident that our model

can account for a substantial part of the variation in employment growth by industry (Figure
6A), and the extent to which these industry effects contribute to aggregate job growth (Figure

6B). Each of the three channels featured in the model contributes to its predictive power. A

regression of the observed contribution of each industry to aggregate employment growth on

its predicted value based only on the direct (own-industry) effect yields an R2 value of 0.34.
Adding customer and supplier effects to this prediction raises this R2 to 0.45. Incorporating the

final demand effect raises it further to 0.61. Given that the model exclusively uses variation in

TFP across industries to form predictions, we consider this as strong confirmation of the utility

of our accounting framework.

4.2. Aggregate labor-share effects

We now perform the analogous exercise for the implied effect of rising TFP on labor’s share

of value-added. In this calculation, the own-industry, inter-industry, and final demand effects

31Observe that the contribution of final demand growth to employment and hours worked in destination industries
reported in Appendix Table A11B is directly proportional to the size of the industry in total employment.
32The predicted versus observed employment contribution (Figure 6B) depends on the proportional growth in each
industry multiplied by its weight in overall employment, whereas the predicted versus observed employment change
(Figure 6A) depends on only the first of these terms.

28
are obtained analogously to those for employment and hours.33 However, the labor share

calculation includes a fourth channel: TFP-induced compositional shifts in value-added shares

across industries. This between-industry composition effect is calculated as

Š

Δ ln 𝑌#$x†–w ≡ =—Δω ̅ š
™ 𝑖𝑐 × 𝑙(# (11)
(
Š
𝜔(# exp4∑eVW8 𝛽6,••
V
× ∆ ln 𝑇𝐹𝑃(,#R#((),$ 9
= = ’› ̅ “
− 𝜔(# Ÿ × 𝑙(#
Š e V
∑( —𝜔(# exp4∑VW8 𝛽6,•• × ∆ ln 𝑇𝐹𝑃(,#R#((),$ 9š
(

Here, Δω ̅ is
™ (# is the predicted change in the value-added share of industry 𝑖 in country 𝑐, and 𝑙(#
V
the time-averaged log labor share in industry 𝑖 in country 𝑐. The terms 𝜔(# and 𝛽6,•• are defined

as in equation (10), again adjusted for the labor-share model: the time-averaged weights 𝜔(# are

shares of nominal value-added rather than shares of employment or hours worked, and the
V
coefficients 𝛽6,•• reflect nominal rather than real value-added coefficients (shown in column 4 of

Table 9). Concretely, this prediction reflects the sum of induced shifts in each industry’s share of

own-country nominal value-added (Δ𝜔(# , the expression in braces) multiplied by that industry’s

labor share.34

We report quantitative implications for labor’s share of value-added in Figure 7. The first
bar reflects the labor-share effect associated with own-industry productivity growth. Its height

of −0.14 suggests that on average, own-industry productivity growth reduced the labor share

by some 5.2 percent over the 37-year period (0.14 × 37). Unlike for employment and hours

worked, however, there are no positive countervailing effects from inter-industry linkages or

final demand; rather, these additional channels also serve to decrease the aggregate labor share,

albeit by small amounts compared to the direct effect (−0.01, −0.06 and −0.02 log points

annually for respectively the supplier, customer, and final demand effects). Finally, industry

33The weights (𝜔(# ) used in eqns. (8), (9), and (10) are now time-averaged industry value-added shares rather than
employment or hours shares; and the final demand effect is calculated using aggregate increases in nominal rather
than real value-added. Hence, the coefficients ∑eVW8 𝜆6V and ∑eVW8 𝛽6,••
V
in eqn. (10) are taken from column 4 (rather than
column 5) of, respectively, the lower and upper panels of Table 9.
34As with prior calculations, we incorporate customer and supplier TFP effects into this calculation but suppress
them from the equation for simplicity. We have also estimated models that allow the aggregate income elasticities
estimated in the lower panel of Table 9 to vary by broad sector (thereby potentially admitting non-homotheticities).
This has negligible effects on the predicted composition changes, and we therefore do not report these specifications.

29
composition shifts resulting from a reallocation of value-added across industries also predict a

small net labor share decline: this effect amounts to approximately 1.7 percent over the entire

period (0.046 × 37).

Taken together, all four channels operating on the labor share—direct, supplier/customer,

final demand, and composition—predict a decline of −0.27 log points annually, or around 10

percent over the entire period (0.27 × 37). Most of this effect stems from the direct labor-

displacing effect operating within industries, combined with an absence of countervailing

effects operating within-industries. Compositional shifts modestly reinforce this trend. Given an

initial average labor share of around 67 percent in our 19 countries (Table 1), this corresponds to

a non-negligible predicted decline of 6 percentage points over 1970-2007, of which the large

majority (0.225 log points annually—that is, 8.3 percent, or around 5.5 percentage points, over
the entire period) is predicted to occur within industries.

Table 10 reports the separate industry-level contributions made to these overall

predictions.35 The first column shows each industry´s contribution to the total predicted within-

industry effects (that is, the predicted effects for own-industry TFP growth; inter-industry
linkages; and final demand taken together, which are largely driven by the own-industry

effect). The second column analogously shows the contribution of each industry to the

predicted between-industry effect shown in Figure 7. Table 10 highlights that most industries

experience a negative within-industry labor-share effect. Predictably, some of the largest

contributions are made by industries that have witnessed strong productivity growth, such as

Electrical and optical equipment, Chemicals, Basic and fabricated metals, and Post and

telecommunications. However, industries with more modest productivity growth but

comprising relatively large shares of value added— such as Wholesale trade, and Transport and
storage— also contribute substantially to the aggregate within-industry effect. Real estate, and

Other business activities are the only industries that contribute a small countervailing effect:

here, positive within-industry labor-share changes are predicted since these sectors have

experienced negative TFP growth on average. Finally, several (public) services such as

35 Unlike for employment and hours, most effects for the labor share are driven by the direct effect. As a result, there
is no need to separately consider the industry contributions by source of TFP growth.

30
Education, Health and social work, and Other personal services contribute almost nothing to

the predicted aggregate labor-share decline since they have experienced virtually no measured

productivity growth.

Table 10 also shows that the industry-specific contributions to the composition (i.e., between-

industry) effect are quite heterogeneous. In general, the predicted shift away from capital-

intensive Mining, Utilities, and Manufacturing industries tends to increase labor’s share: in

isolation, these industries contribute a predicted increase in the labor share of around 1.6
percent cumulated over the period (0.036 × 37). This is reinforced by contributions from

(mostly high-tech) services such as Post and telecommunications, Financial intermediation, and

Transport and storage. However, Real estate single-handedly contributes a large negative

compositional effect of, on average, 0.086 log points annually, or over 3 percent across the entire
period. This prediction is consistent with the aggregate labor decomposition reported in Table 4

and stems from three distinctive features of the Real estate industry: a very low labor share

relative to the economy-wide average; a rising share of value-added; and zero or negative TFP

growth.

4.3. Why has the fall in labor share accelerated?

Our results imply that technological progress, broadly construed, has been employment-

augmenting but labor-displacing—that is, generating net employment gains while serving to

reallocate value-added away from labor and towards other factors. But this observation raises a

puzzle. If automation has been consistently labor-displacing, why has the evolution of labor’s

share differed so sharply across decades—rising during the 1970s, declining in the 1980s and

1990s, and then falling more steeply in the 2000s? We briefly take up this question here.

Table 11 reports our baseline model’s predictions separately by decade. Panel A reports the

observed annual log labor share change in each decade, both within and between industries,

while panel B reports the changes predicted by our baseline model. This table highlights the fact

that, although our baseline approach explains a substantial part of the aggregate labor share fall

observed since the 1980s, it fails to match two key features of the decade-specific patterns: the

positive sign of the within-industry effect operating in the 1970s, and the observed acceleration

31
of the within-industry log labor share decline in the 2000s. The proximate reason for both

mismatches is clear: the bulk of the model’s explanatory power for the labor share derives from

the so-called ‘direct effect’—the differential decline of the labor share in industries with faster

TFP growth; thus, for the baseline approach to explain the time pattern of rising and then falling

labor share across decades, it would need to be the case that TFP growth was negative in the

1970s, became positive in the 1980s and 1990s, and then accelerated in the 2000s. This does not

match the time pattern of TFP growth, however (see Table 3). The model does slightly better at
capturing the time pattern of between-industry effects—predicting larger compositional shifts

in the 1970s and 2000s, which is approximately consistent with the data—but our explanatory

power is limited here as well.36

Our empirical framework admits several mechanisms through which the effect of
technological progress on the labor share may differ over time. One mechanism is that an

acceleration of TFP growth will lead to a more rapid fall in the labor share. But as noted above,

this explanation is a non-starter because TFP growth decelerated in the 2000s even as the fall in

the labor share accelerated. Second, the locus of productivity growth may be differently
distributed among industries in different eras. To the extent that industries experiencing rapid

TFP gains are more (or less) labor-intensive or make up a larger (or smaller) share of the total

economy, the aggregate labor share will decline more (or less) strongly through, respectively,

composition effects and within-industry effects. But Table 11 suggests that these explanations

have a limited bite. Allowing the sources of TFP growth to differ across decades, as we do in the

table, does not explain the sharp decadal differences in the between- and within-industry

contributions to the fall in the labor share.

A third possibility is that, all else being equal, a given amount of overall productivity
growth might have different effects in different eras if the source of that productivity growth is

changing—for example, if productivity growth increasingly stems from technologies that are

relatively less labor-augmenting and relatively more labor-displacing. Figure 8 suggest that this

36The substantial between-industry component of the falling labor share in the 2000s is, as above, due to the rapid
growth of the Real estate industry in value-added, a phenomenon that is unlikely to be attributable to technological
progress.

32
explanation has some promise. Akin to Figure 1 above, Figure 8 presents bivariate scatters of

the relationship between industry-level TFP growth and changes in, respectively, industry-level

log employment (Figure 8A) and industry-level log labor-share (Figure 8B). Distinct from earlier

figures, the two panels of Figure 8 depict separate slopes by decade. The upper panel shows a

consistently stable, downward-sloping relationship between industry-level TFP growth and

relative declines in employment, with a somewhat steepening slope after the 1970s. By contrast,

Figure 8B shows a much more noticeable shift in the relationship between productivity and
labor’s share over time. During the 1970s, there is no appreciable link between industries’

productivity growth and their labor share changes. A clear negative relationship emerges in the

1980s, however, and remains in place during the 1990s and 2000s. This pattern suggests that a

shift towards more labor-displacing productivity growth is a possible explanation for the fall in
the labor share commencing in the 1980s.

To explore this possibility more rigorously, we estimate a set of distributed lag models

where the own-industry impact of TFP growth is allowed to vary by decade. Across a range of

specifications, we find that the 1970s stand out as a period when own-industry TFP growth had
a less negative effect on labor’s share. We do not find much evidence of statistically significant

heterogeneity in coefficients for the decades thereafter, consistent with the broad patterns

shown in Figure 8B. Table 12 provides estimates of the direct effect of TFP growth on our range

of outcomes, estimated separately for the 1970s and the three subsequent decades. As shown in

Panel F, there is a statistically insignificant positive relationship between own-industry TFP

growth and own-industry labor share changes during the 1970s, which turns statistically

significant and negative for the three more recent decades. Appendix Table A12 provides

additional detail by estimating these models separately by decade, applying a five-year lagged
long-difference specification.37

To assess the quantitative importance of these decadal differences, Table 13 reports a set of

decade-specific predictions based on Table 12. These predictions are constructed by allowing

the 𝛽6V coefficients in eqn. (8) and the 𝛽6,••

V
coefficients in eqn. (11) to be different in the 1970s

37We are severely limited in our ability to estimate distributed lag models for the decade of the 1970s since no
country enters the EU KLEMS data prior to 1970, and several enter later (see Appendix Table A1).

33
compared to the other three decades, thereby allowing both the effect of TFP growth on the

within-industry and between-industry components of the aggregate labor-share to change over

time.38 A drawback of performing predictions with these estimates is that, relative to our main

estimates, the estimated TFP slopes are shallower across all periods, likely because

identification of the distributed lag terms is weak in short panels. Nevertheless, the predicted

within-industry pattern now qualitatively matches the turnaround after the 1970s: productivity

growth is predicted to modestly increase labor’s share during the 1970s and to decrease it
thereafter. The model is also somewhat successful at predicting the increase in the between-

industry component of the falling labor share in the 2000s. The model is not successful, however,

in explaining the acceleration of the within-industry fall in the labor share in the 2000s.

Summarizing, our analysis broadly supports the hypothesis that the decline in the labor
share since the 1980s is consistent with a shift towards more labor-displacing technology

commencing in the 1980s. But the acceleration in the labor share decline observed during the

2000s is left unaccounted for by this mechanism. We hypothesize that a closer study of specific

technologies may yield additional insights into these periods. At the same time, we do not
assume that technological factors are the sole contributor to the changing secular pattern of the

labor share decline or its recent deceleration. Instead, what our findings make clear is that

technological progress has been broadly employment-augmenting and labor-displacing for at

least three decades. The consistency of the evidence, rather than its over-time acceleration or

deceleration, is what gives us confidence in the utility of our approach for tracing through the

within-industry, between-industry, and aggregate consequences of productivity growth

originating in all industries.

5. Concluding remarks
Theory makes clear that there is no direct mapping between the evolution of productivity

and labor demand at the industry level and the evolution of labor demand in the aggregate.

Theory gives less guidance about how to draw this indirect mapping. We present an empirical

38We restrict our attention here to the direct effect since we find this to be the main driver of aggregate labor share
changes, irrespective of the time period under consideration.

34
approach for mapping the industry-level effects of technological progress to aggregate

employment and labor-share outcomes, taking into account both the direct effects of

productivity growth in advancing industries and the indirect effects from inter-industry

demand linkages, between-industry compositional change, and increases in final demand. Our

findings indicate that these indirect effects are sizable and are countervailing for employment.

We find that technological progress is broadly employment-augmenting in the aggregate. But

this is not so for labor’s share of value-added, where direct labor-displacing effects dominate.
Our simple framework can account for a substantial fraction of both the reallocation of

employment across industries and the aggregate fall in the labor share over the last three

decades. It does not, however, explain why the share of labor in value-added fell more rapidly

during the 2000s than in prior decades. Nor can it distinguish between the contributions of
automation- versus non-automation-based sources of productivity growth, which may

plausibly exert distinct effects on either employment or on labor’s share of value added.

Although our empirical exploration of labor displacement has linked effects at the industry

level to aggregate outcomes, this high-level representation is consistent with a variety of within-
and between- firm adjustments. At one extreme, every firm in an industry undergoing

technological progress might substitute capital for labor in a subset of tasks. Alternatively,

absent any within-firm change in task allocation, a technological advance might spur an

increase in industry market share among relatively capital-intensive firms and a concomitant

decline among relatively labor-intensive firms.39 Under either scenario, labor’s share in industry

value-added would fall. Our analysis cannot speak to these within-firm versus between-firm

dynamics. Nevertheless, we believe that the scope of the evidence presented here complements

more granular, but narrower firm-level and establishment-level studies.

39See Decker et al. (2017), Autor et al. (2017b), and Foster et al. (2017 and 2018) for further explorations of the linkage
between firm-level dynamics and aggregate productivity.

35
6. References
Acemoglu, Daron, David Autor, David Dorn, Gordon Hanson, and Brendan Price. 2016.
“Import Competition and the Great U.S. Employment Sag of the 2000s.” Journal of Labor
Economics, 34, S1 (Part 2), S141–S198.

Acemoglu, Daron, Ufuk Akcigit, and William Kerr. 2017. “Networks and the Macroeconomy:
An Empirical Exploration.” In Martin Eichenbaum and Jonathan A. Parker (eds.), NBER
Macroeconomics Annual, 31, Chicago: University of Chicago Press.

Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implications for
Employment and Earnings.” in O. Ashenfelter and D. Card, eds., The Handbook of Labor
Economics Volume IV, Amsterdam: Elsevier.

Acemoglu, Daron, and Veronica Guerrieri. 2008. “Capital Deepening and Nonbalanced
Economic Growth.” Journal of Political Economy 116 (3): 467–98.

Acemoglu, Daron and Pascual Restrepo. 2017. “Robots and Jobs: Evidence from U.S. Labor
Markets.” NBER Working Paper No. 23285, March.

Acemoglu, Daron and Pascual Restrepo. 2018. “Artificial Intelligence, Automation and Work.”
NBER Working Paper No. 24196, January.

Acemoglu, Daron and Pascual Restrepo. Forthcoming. “The Race Between Machine and Man:
Implications of Technology for Growth, Factor Shares and Employment.” American Economic
Review.

Alexopoulos, Michelle, and Jon Cohen. 2016. “The Medium Is the Measure: Technical Change
and Employment, 1909–1949.” Review of Economics and Statistics 98(4): 792–810.

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): 1553–1597.

Autor, David, David Dorn, Gordon H. Hanson, Gary Pisano, and Pian Shu. 2017a. “Foreign
Competition and Domestic Innovation: Evidence from U.S. Patents.” MIT Working Paper,
December.

Autor, David, David Dorn, Lawrence F. Katz, Christina Patterson, and John Van Reenen. 2017b.
“The Fall of the Labor Share and the Rise of Superstar Firms.” NBER Working Paper No.
23396, May.

Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of Recent
Technological Change: An Empirical Exploration.” Quarterly Journal of Economics 118(4):
1279–1333.

Autor, David H. and Anna M. Salomons. 2017. “Robocalypse Now—Does Productivity Growth
Threaten Employment?” European Central Bank Sintra Forum conference paper, June 2017.

36
Barkai, Simcha. 2017. “Declining Labor and Capital Shares.” Working Paper, University of
Chicago.

Basu, Susanto, and John Fernald. 2001. “Why is Productivity Procyclical? Why Do We Care?” In
Charles R. Hulten, Edwin R. Dean and Michael J. Harper, eds., New Developments in
Productivity Analysis, Chicago: Univ of Chicago Press, 225–302.

Baumol, William J. 1967. “Macroeconomics of Unbalanced Growth: The Anatomy of Urban

Crisis.” American Economic Review 57(3): 415–26.

Berg, Andrew, Edward F. Buffie, and Luis-Felipe Zanna. 2018. “Robots, Growth, and Inequality:
Should We Fear the Robot Revolution? (The Correct Answer is Yes).” Working Paper no.
18/116. Washington: International Monetary Fund.

Bessen, James. 2017. “Automation and Jobs: When Technology Boosts Employment.” Boston
University School of Law: Law & Economics Paper No. 17-09, April.

Brynjolfsson, Erik, and Andrew McAfee. 2014. The Second Machine Age: Work, Progress, and
Prosperity in a Time of Brilliant Technologies. New York and London: W.W. Norton &
Company.

Francisco J. Buera, Joseph P. Kaboski, Richard Rogerson. 2015. “Skill Biased Structural Change.”
NBER Working Paper No. 21165, May.

Caselli, Francesco, and Manning. Forthcoming. “Robot Arithmetic: New Technology and
Wages.” American Economic Review: Insights.

Chiacchio, Francesco, Georgios Petropoulos, and David Pichler. 2018. “The Impact of Industrial
Robots on EU Employment and Wages: A Local Labour Market Approach.” Bruegel
Working Paper, April.

Dao, Mai, Mitali Das, Zsoka Koczan, and Weicheng Lian. 2017. “Why is Labor Receiving a
Smaller Share of Global Income? Theory and Empirical Evidence,” IMF Working Paper.

Dauth, Wolfgang, Sebastian Findeleisen, Jens Südekrum, and Nicole Wößner. 2017. “German
Robots – The Impact of Industrial Robots on Workers,” University of Mannheim Working
Paper, December.

Decker, Ryan A., John Haltiwanger, Ron S. Jarmin, and Javier Miranda. 2017. “Declining
Dynamism, Allocative Efficiency, and the Productivity Slowdown.” American Economic
Review: Papers & Proceedings, 107(5), 322–326.

Eden, Maya, and Paul Gaggl. 2018. “On the Welfare Implications of Automation.” Review of
Economic Dynamics, 29, 15–43.

Elsby, Mike, Bart Hobijn, and Aysegul Sahin. 2013. “The Decline of the U.S. Labor Share.”
Brookings Papers on Economic Activity, 1-42.

37
Ford, Martin. 2015. Rise of the Robots: Technology and the Threat of a Jobless Future. New York: Basic
Books.

Foster, Lucia S., Cheryl Grim, John Haltiwanger, and Zoltan Wolf. 2017. “Macro and Micro
Dynamics of Productivity: From Devilish Details to Insights.” NBER Working Paper No.
23666, August.

Foster, Lucia S., Cheryl Grim, John Haltiwanger, and Zoltan Wolf. 2018. “Innovation,
Productivity Dispersion, and Productivity Growth.” NBER Working Paper No. 24420,
March.

Frey, Carl Benedikt, and Michael A. Osborne. 2017. “The Future of Employment: How
Susceptible are Jobs to Computerisation?” Technological Forecasting and Social Change 114:
254–280.

Graetz, Georg, and Guy Michaels. 2018. “Robots at Work.” Review of Economics and Statistics,
forthcoming.

Gregory, Terry, Anna M. Salomons, and Ulrich Zierahn. 2016. “Racing with or Against the
Machine? Evidence from Europe.” Utrecht School of Economics, Tjalling C. Koopmans
Research Institute, Discussion Paper Series 16-05, July.

Gutiérrez, Germán. 2017. “Investigating Global Labor and Profit Shares,” NYU Working Paper,
October.

Gutiérrez, Germán and Thomas Philippon. 2017. “Declining Competition and Investment in the
U.S,” NBER Working Paper No. 23583, July.

Jones, Charles I, and Paul Romer. 2010. “The New Kaldor Facts: Ideas, Institutions, Population,
and Human Capital.” American Economic Journal: Macroeconomics 2: 224–45.

Jordà, Òscar. 2005. “Estimation and Inference of Impulse Responses by Local Projections.”
American Economic Review, 95(1), 161-182.

Kaldor, Nicholas. 1961. “Capital Accumulation and Economic Growth.” in The Theory of Capital,
ed. F. A. Lutz and D. C. Hague, 177–222. New York: St. Martins Press.

Karabarbounis, Loukas and Brent Neiman. 2014. “The Global Decline of the Labor Share.”
Quarterly Journal of Economics, 129(1), 61-103.

Keynes, John Maynard, 1939. “Relative Movements of Real Wages and Output.” Economic
Journal 49, 34-51

Michaels, Guy, Ashwini Natraj, and John Van Reenen. 2014. “Has ICT Polarized Skill Demand?
Evidence from Eleven Countries over Twenty-Five Years.” Review of Economics and Statistics
96 (1): 60–77.

38
Ngai, L. Rachel, and Christopher A. Pissarides. 2007. “Structural Change in a Multisector Model
of Growth.” American Economic Review 97(1): 429-443.

O’Mahony, Mary and Marcel Timmer. 2009. “Output, Input and Productivity Measures at the
Industry Level: the EU KLEMS Database,” Economic Journal 119(538), F374-503.

Pierce, Justin R., and Peter K. Schott. 2016. “The Surprisingly Swift Decline of US
Manufacturing Employment.” American Economic Review 106(7), 1632–1662.

Piketty, Thomas. 2014. Capital in the Twenty-First Century. Harvard University Press.

Ramey, Valerie A. 2016. “Macroeconomic Shocks and Their Propagation.” Handbook of

Macroeconomics. 1st ed. Vol. 2. Elsevier B.V., 71-162.

Rognlie, Matthew. 2015. “Deciphering the Fall and Rise in the Net Capital Share: Accumulation
or Scarcity?” Brookings Papers on Economic Activity, 1-69.

Sachs, Jeffrey D., and Laurence J. Kotlikoff. 2012. “Smart Machines and Long-Term Misery.”
NBER Working Paper 18629, December.

Solow, Robert M. 1956. “A Contribution to the Theory of Economic Growth.” Quarterly Journal
of Economics 70 (1): 65–94.

Stansbury, Anna and Lawrence H. Summers. 2017. “Productivity and Pay: Is the Link Broken?”
NBER Working Paper No. 24165, December.

Susskind, Daniel. 2017. “A Model of Technological Unemployment.” Oxford University

Working Paper, April.

Teulings, Coen N. and Nikolay Zubanov. 2014. “Is Economic Recovery a Myth? Robust
Estimation of Impulse Responses.” Journal of Applied Econometrics 29(3), 497–514.

Timmer, Marcel P., Erik Dietzenbacher, Bart Los, Robert Stehrer, and Gaaitzen J. de Vries. 2015.
“An Illustrated User Guide to the World Input–Output Database: The Case of Global
Automotive Production,” Review of International Economics 23: 575–605.

Timmer, Marcel, Ton van Moergastel, Edwin Stuivenwold, Gerard Ypma, Mary O’Mahony and
Mari Kangasniemi. 2007. “EU KLEMS Growth and Productivity Accounts Version 1.0: Part I
Methodology.” University of Groningen Mimeo, March.

Trajtenberg, Manuel. 1990. “A Penny for Your Quotes: Patent Citations and the Value of
Innovations.” The RAND Journal of Economics 21(1): 172–187.

Van Ark, Bart, and Kirsten Jäger (2017). “Recent Trends in Europe's Output and Productivity
Growth Performance at the Sector Level, 2002-2015,“ International Productivity Monitor 33,
Fall 2017.

39
Zeira, Joseph. 1998. “Workers, Machines, and Economic Growth,” Quarterly Journal of Economics,
113(4): 1091–1117.

40
7. Figures

Figure 1: Industry-Level Annual Average TFP Growth 1990 – 2007 vs. Industry-Level Annual Changes
in (A) Log Employment; and (B) Labor’s Share of Value-Added: Scatter Plots

1A. Log TFP Growth versus Log Changes in Industry Employment

1B. Log TFP Growth versus Log Changes in Industry Labor Share

41
 Figure 2: Trends in Labor Share by Country, 1970 – 2007:
Observed and Reweighted to Hold Constant Industry Value-Added Shares

42
Figure 3: Local Projection Estimates of the Relationship Between Productivity Growth and Outcome
Variables, 1970 – 2007:

43
Figure 4: Log TFP Growth Changes in Industries’ Shares of Country-Level Nominal Value-Added, 1970 – 2007

44
Figure 5: Predicted Effects of TFP Growth on Aggregate Employment and Hours of Labor Input, 1970 –
2007

5A: Predicted Effects of TFP Growth on Aggregate Employment

5B: Predicted Effects of TFP Growth on Aggregate Hours of Labor Input

45
Figure 6: Predicted vs. Observed Log Employment Changes for (A) Industry-Level Changes; and (B)
Industry-Level Contributions to Aggregate Changes, 1970 – 2007

Figure 6A: Predicted vs. Observed Changes in Log Employment by Industry

Figure 6B: Predicted vs. Observed Industry Contributions to Aggregate Log Employment Change

46
Figure 7: Predicted Effects of TFP Growth on Aggregate Labor Share, 1970 – 2007

47
Figure 8: Scatter Plots of Industry-Level TFP Growth 1990 – 2007 vs. Industry-Level Growth in
(A) Employment and (B) Log Labor-Share by Decade: 1970s, 1980s, 1990s, and 2000s

Figure 8A: Changes in TFP vs Log Employment by Industry

Figure 8B: Changes in TFP vs Log Labor Share by Industry

48
8. Tables

Table
1970s 1.1970s
Trends in Hours Worked
1970s 1970s and Labor-Share1970s
by Country
1970s and Decade 1970s
Average across years: 100 × Annualized log hours worked change in: 100 × Annualized log laborshare change in:
Country Log hrs Laborshare VA share 1970s 1980s 1990s 2000s 1970s 1980s 1990s 2000s
Australia 9.41 64.8% 2.0% 1.77 2.48 2.32 3.03 -0.22 -1.07 0.01 -0.27
Austria 8.61 67.2% 0.9% 0.52 0.48 1.42 1.95 -0.72 -1.12 -0.79 -1.16
Belgium 8.50 64.1% 1.1% -0.96 0.23 1.82 1.79 0.92 -1.27 0.22 2.52
Canada 9.82 59.4% 2.8% 2.59 2.38 1.82 2.55 -0.40 -0.02 -1.13 0.42
Denmark 8.20 67.6% 0.7% -0.07 0.18 1.18 2.02 0.14 -0.37 -0.96 0.59
Finland 8.10 68.3% 0.6% 0.26 1.34 -0.30 2.17 -0.10 0.35 -2.53 0.17
France 10.36 67.9% 6.3% 0.04 0.37 1.07 1.80 -0.37 -1.07 -0.81 -0.44
Germany 10.82 66.6% 9.3% -0.60 0.29 1.13 0.80 0.42 -1.18 0.15 -1.52
Ireland 7.77 55.9% 0.7% 3.71 5.32 4.37 0.17 -2.15 0.78
Italy 10.39 68.2% 5.2% 1.20 1.21 0.84 1.94 0.54 -0.52 -1.82 -0.53
Japan 11.57 56.6% 19.6% 1.17 0.80 -0.27 0.97 2.38 -0.43 -0.76 -0.71
Luxembourg 5.82 55.4% 0.1% 3.52 4.86 3.99 1.21 -0.84 -0.08
Netherlands 9.06 68.3% 1.7% -0.59 1.26 3.26 1.77 -1.73 -0.47 0.09 -0.85
Portugal 8.87 59.4% 0.4% 1.43 -1.23 0.76 0.88 3.01 2.26 -0.56 -0.55
South Korea 10.33 69.5% 1.7% 6.46 3.43 1.82 1.56 -0.07 0.44 -1.22 0.93
Spain 9.85 62.8% 2.7% 0.81 1.28 2.72 4.06 0.10 -0.11 0.31 -0.94
Sweden 8.72 67.9% 1.5% 1.50 0.29 1.30 -0.61 -0.91 0.40
United Kingdom 10.65 70.5% 5.9% 0.11 1.46 0.92 2.38 -0.34 0.36 -0.91 0.32
United States 12.08 63.7% 36.6% 2.39 2.70 2.50 0.70 0.12 -0.38 0.36 -1.46
Weighted average 1.424 1.699 1.553 1.350 0.513 -0.459 -0.263 -0.861
Notes: See Appendix Table 1 for data availability by country. Changes are annualized long differences by decade. Weighted averages are constructed using time-
averaged hours worked weights for hours and time-averaged value added shares for the labor share.

49
 Table 2. Trends in Hours Worked, Labor Share, and TFP by Industry
Time-averaged 100 × annual log change in:
ISIC code Description VA share Hrs worked Laborshare TFP
C Mining and quarrying 1.52% -2.45 -1.22 0.29
15t16 Food, beverages, and tobacco 2.63% -0.52 -0.08 0.72
17t19 Textiles, textile , leather, and footwear 1.15% -3.96 0.18 2.07
20 Wood and wood products 0.51% -1.34 -0.32 2.12
21t22 Pulp, paper, paper, printing, and publishing 2.17% -0.25 -0.19 1.10
23 Coke, refined petroleum and nuclear fuel 0.59% -1.54 -1.60 -0.49
24 Chemicals and chemical products 2.23% -0.78 -0.44 3.19
25 Rubber and plastics 0.95% 0.67 0.21 2.56
26 Other non-metallic mineral 0.86% -1.33 -0.18 1.68
27t28 Basic metals and fabricated metal 2.91% -0.87 -0.22 1.72
29 Machinery, not elsewhere classified 2.25% -0.60 0.03 1.86
30t33 Electrical and optical equipment 3.33% -0.28 -0.10 4.49
34t35 Transport equipment 2.42% -0.12 -0.27 2.42
36t37 Manufacturing not elsewhere classified; recycling 0.81% -0.58 -0.03 1.09
E Electricity, gas, and water supply 2.50% -0.28 -0.65 1.29
F Construction 6.75% 0.94 0.04 0.20
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 1.39% 0.95 -0.05 0.11
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles 6.39% 0.67 -0.28 1.07
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods 5.18% 0.73 -0.16 1.18
H Hotels and restaurants 2.77% 1.80 -0.09 -0.88
60t63 Transport and storage 4.49% 0.91 -0.17 1.24
64 Post and telecommunications 2.48% 0.52 -1.18 3.04
J Financial intermediation 6.35% 1.70 -0.46 0.95
70 Real estate activities 11.32% 3.08 0.70 -0.66
71t74 Renting of machinery & equipment and other business activities 9.79% 4.63 0.68 -1.65
M Education 5.71% 1.67 -0.01 -0.14
N Health and social work 6.61% 2.89 0.05 -0.22
O Other community, social and personal service activities 3.97% 2.16 0.11 -1.02
Notes: Weighted by country size (hours worked weights for hours worked; value added weights for laborshare and TFP).

50
 Table 3. Within-Industry Trends in Key Variables Used in the Analysis

Employ- Hours Nominal Real hrly Nominal TFP

ment worked hrly wage wage value added Labor share measure
100 x mean annual log change 1.337** 1.001** 6.472** 1.700** 6.917** -0.051 0.619**
(0.166) (0.171) (0.171) (0.116) (0.209) (0.104) (0.150)

100 x mean annual log change for:

1970s 2.035** 1.572** 11.556** 2.472** 11.952** 0.503* 0.440*
(0.204) (0.210) (0.401) (0.261) (0.290) (0.217) (0.201)
1980s 1.661** 1.365** 6.550** 1.728** 7.372** -0.265* 0.994**
(0.198) (0.214) (0.238) (0.170) (0.351) (0.123) (0.168)
1990s 0.996** 0.656** 3.815** 1.319** 4.035** -0.141 0.603**
(0.190) (0.220) (0.178) (0.157) (0.262) (0.171) (0.154)
2000s 0.382* 0.174 3.043** 1.127** 3.269** -0.395** 0.360**
(0.190) (0.210) (0.226) (0.176) (0.268) (0.152) (0.133)

100 x mean annual log change for:

Mining, utilities & construction 0.625 0.521 6.441** 1.641** 6.142** -0.391~ 0.405**
(0.393) (0.443) (0.809) (0.573) (0.527) (0.223) (0.114)
Manufacturing -0.810** -0.984** 7.068** 2.246** 5.344** -0.157* 2.185**
(0.170) (0.182) (0.266) (0.160) (0.212) (0.080) (0.180)
Education & health 2.566** 2.351** 6.422** 1.664** 8.342** -0.058 -0.190**
(0.203) (0.196) (0.412) (0.245) (0.402) (0.089) (0.023)
Low-tech services 1.676** 1.227** 6.158** 1.403** 6.815** 0.162 0.150
(0.148) (0.168) (0.251) (0.168) (0.280) (0.237) (0.193)
High-tech services 3.286** 3.091** 6.324** 1.608** 8.817** -0.095 -0.022
(0.379) (0.379) (0.380) (0.274) (0.402) (0.222) (0.452)

N 18,062 18,062 18,062 18,062 18,062 18,062 18,062

Model weighted by: Employment Hours Hours Hours VA VA VA
Notes: Excludes agriculture, public administration, private households, and extra-territorial organizations. All models weighted by time-
averaged industry shares of the weighting variable within countries, multiplied by time-varying country shares in total annual value of the
weighting variable. The number of observations is equal to the number of country-industry cells multiplied by the number of years.
Standard errors are clustered by country-industry and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

51
 Table 4. Shift-Share Analysis of the Log Changes in Labor Share by Decade

100 x Annualized decadal log change in laborshare by country

Weighted by country size Unweighted
Between Within Between Within
Decade Mean industry industry Mean industry industry
1970s 0.513 -0.187 0.700 0.230 -0.146 0.376
-36% 136% -63% 163%
1980s -0.459 -0.183 -0.276 -0.201 -0.121 -0.080
40% 60% 60% 40%
1990s -0.263 -0.075 -0.188 -0.750 -0.304 -0.446
28% 72% 41% 59%
2000s -0.861 -0.425 -0.436 -0.126 -0.018 -0.104
49% 51% 14% 83%

52
 Table 5. Estimates of the Relationship Between Productivity Growth
and Industry-Level Outcomes, 1970 – 2007: Direct Effects
dependent variable: annual change in log outcome by country-industry

A. Employment B. Hours C. Wagebill

(1) (2) (3) (4) (5) (6) (7) (8) (9)
Σ Δ ln Own-Industry TFP (i, c, t-k) -2.073** -1.132** -1.117** -1.989** -1.048** -1.028** -1.848** -1.078** -1.029**
(0.172) (0.144) (0.147) (0.187) (0.160) (0.162) (0.272) (0.220) (0.225)
R2 0.223 0.271 0.359 0.203 0.239 0.359 0.414 0.426 0.530
Model weighted by: Employment Hours Hours
D. Nominal VA E. Real VA F. Laborshare
(10) (11) (12) (13) (14) (15) (16) (17) (18)
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.332** -0.629** -0.609** 0.641 1.214** 1.238** -0.504** -0.571** -0.541**
(0.221) (0.180) (0.191) (0.494) (0.401) (0.405) (0.128) (0.148) (0.152)
R2 0.299 0.313 0.368 0.105 0.137 0.183 0.063 0.064 0.147
Model weighted by: Nominal VA Nominal VA Nominal VA
Fixed effects for all models:
Country YES YES YES YES YES YES YES YES YES
Year YES YES YES YES YES YES YES YES YES
Sectorgroup NO YES YES NO YES YES NO YES YES
Country × Timetrend YES YES NO YES YES NO YES YES NO
Country × Business cycle YES YES NO YES YES NO YES YES NO
Country × Year NO NO YES NO NO YES NO NO YES
N 15,520 15,520 15,520 15,520 15,520 15,520 15,520 15,520 15,520
Notes: TFP is other-country-same-industry TFP, rescaled to have a unit standard deviation. Estimates are the sum of coefficients
for the contemporaneous effect and 5 annual distributed lags (k=0, ..., 5). The number of observations is equal to the number of
country-industry cells multiplied by the number of years. Standard errors are clustered by country-industry and reported in
parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

53
 Table 6. Robustness Tests for Direct Productivity Effect Estimates in Table 5
dependent variable: annual change in log outcome by country-industry

Employment Hours Wagebill Nominal VA Real VA Laborshare

(1) (2) (3) (4) (5) (6)
A. All countries given equal weight
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.038** -0.996** -0.888** -0.603** 1.040** -0.426**
(0.123) (0.125) (0.146) (0.147) (0.182) (0.108)
R2 0.331 0.335 0.565 0.395 0.218 0.104
B. Excluding contemporaneous effects (k=1,…,5)
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.038** -0.985** -1.039** -0.719** 0.947* -0.423**
(0.142) (0.153) (0.198) (0.157) (0.367) (0.145)
R2 0.358 0.358 0.530 0.367 0.174 0.146
C. Excluding the self-employed
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.156** -1.056** -0.996** -0.609** 1.238** -0.528**
(0.156) (0.163) (0.218) (0.191) (0.405) (0.142)
R2 0.384 0.386 0.580 0.368 0.183 0.147
D. Setting negative TFP growth to zero
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.109** -0.962** -0.880** -0.490* 1.880** -0.690**
(0.223) (0.234) (0.309) (0.228) (0.575) (0.186)
R2 0.350 0.352 0.528 0.367 0.186 0.145
E. Five-year annualized long differences, lagged effects
Δ ln Own-Industry TFP (i, c, T-1) -0.683** -0.636** -0.713** -0.472** 0.631** -0.348**
(0.090) (0.097) (0.119) (0.115) (0.231) (0.104)
R2 0.505 0.490 0.787 0.687 0.263 0.119
F. EUKLEMS (2017) database, 2000-2015
Σ Δ ln Own-Industry TFP (i, c, t-k) -1.194** -0.943** -0.904* 0.070 0.896 -0.633~
(0.304) (0.310) (0.359) (0.286) (0.562) (0.368)
R2 0.365 0.492 0.331 0.272 0.304 0.094
All models weighted by: Employment Hours Hours VA VA VA

Notes: TFP is other-country-same-industry TFP, rescaled to have a unit standard deviation. Except for panels B and E,
estimates are the sum of coefficients for the contemporaneous effect and 5 annual distributed lags (k=0, ..., 5). The
number of observations is 15,520 for panels A, B, C, and D; 2,820 for panel E; and 3,148 for panel F. All panels contain
country, year, and country-by-year fixed effects (where for panel E, years are defined as five-year panels); panels A, B,
C, D, and E additionally contain sectorgroup fixed effects. Standard errors are clustered by country-industry and
reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

54
Table 7. Predictive Relationships between Industry Patenting Activity and TFP Growth, 1970 – 2007
dependent variable: 100 x annual change in log TFP by country-industry

(1) (2) (3) (4)

Σ ln patents (i, c, t-k) 0.574** 0.602** 0.602** 0.603**
(0.197) (0.202) (0.202) (0.204)
R2 0.061 0.137 0.138 0.142
N 16,518 16,518 16,518 16,518

Σ ln patent citations (i, c, t-k) 0.608** 0.647** 0.648** 0.649**

(0.208) (0.229) (0.230) (0.233)
R2 0.054 0.139 0.140 0.143
N 16,479 16,479 16,479 16,479
Fixed effects for all models:
Country YES YES YES YES
Year NO YES YES YES
Country × Timetrend NO NO YES NO
Country × Business cycle NO NO YES NO
Country × Year NO NO NO YES
Notes: Log patents and log patent citations rescaled to have a unit standard deviation.
Estimates are the sum of coefficients for the contemporaneous effect and 3 annual
distributed lags (k=0, ..., 3). The number of observations is equal to the number of
country-industry cells multiplied by the number of years. Standard errors are clustered by
country-industry and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

55
 Table 8. The Relationship Between Patenting
and Industry-Level Outcomes, 1970 – 2007
dependent variable: 100 x annual change in log outcome by country-industry

A. Employment B. Hours C. Wagebill

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Σ ln patents (i, c, t-k) -0.328~ -0.261 -0.267 -0.303 -0.243 -0.243 -0.420~ 0.039 0.029
(0.187) (0.197) (0.201) (0.192) (0.205) (0.209) (0.219) (0.222) (0.226)
R2 0.036 0.130 0.223 0.032 0.137 0.261 0.117 0.387 0.492

Σ ln patent citations (i, c, t-k) -0.327 -0.239 -0.246 -0.338 -0.206 -0.211 -0.769** 0.097 0.087
(0.208) (0.230) (0.235) (0.213) (0.239) (0.244) (0.272) (0.263) (0.269)
R2 0.089 0.280 0.336 0.023 0.104 0.150 0.006 0.062 0.147

Models weighted by: Employment Hours Hours

D. Nominal VA E. Real VA F. Laborshare

(10) (11) (12) (13) (14) (15) (16) (17) (18)

Σ ln patents (i, c, t-k) -0.437~ -0.133 -0.135 0.607* 0.678** 0.672* -0.263* -0.213~ -0.222
(0.230) (0.209) (0.210) (0.250) (0.256) (0.261) (0.133) (0.129) (0.135)
R2 0.039 0.130 0.222 0.033 0.137 0.261 0.160 0.388 0.492

Σ ln patent citations (i, c, t-k) -0.729** -0.099 -0.121 0.553* 0.738** 0.731* -0.329* -0.242~ -0.235~
(0.259) (0.215) (0.217) (0.272) (0.285) (0.291) (0.145) (0.141) (0.142)
R2 0.116 0.284 0.341 0.021 0.106 0.152 0.008 0.066 0.153

Models weighted by: Nominal VA Nominal VA Nominal VA

Fixed effects for all models:
Country YES YES YES YES YES YES YES YES YES
Year NO YES YES NO YES YES NO YES YES
Country × Timetrend NO YES NO NO YES NO NO YES NO
Country × Business cycle NO YES NO NO YES NO NO YES NO
Country × Year NO NO YES NO NO YES NO NO YES
Notes: Log patents and log patent citations rescaled to have a unit standard deviation. Estimates are the sum of coefficients for the
contemporaneous effect and 5 annual distributed lags (k=0, ..., 5). The number of observations is 15,456 for patent models and 15,417 for
patent citation models. Standard errors are clustered by country-industry and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

56
 Table 9. Estimates of the Relationship Between Productivity Growth and Industry-Level Outcomes,
1970 – 2007: Direct Effects, Supplier and Customer Effects, and Aggregate Effects
dependent variable: annual change in log outcome by country-industry

A. Industry effects
Employment Hours Wagebill Nominal VA Real VA Laborshare
(1) (2) (3) (4) (5) (6)
Σ Δ ln Own-Industry TFP (i, c, t-k) -0.951** -0.869** -1.052** -0.579** 1.243** -0.584**
(0.144) (0.160) (0.233) (0.201) (0.398) (0.171)
Σ Δ ln Supplier-Industry TFP (j≠i, c, t-k) 0.971** 1.028** 0.196 0.376 0.269 -0.029
(0.223) (0.237) (0.313) (0.291) (0.426) (0.269)
Σ Δ ln Customer-Industry TFP (j≠i, c, t-k) 0.097 0.159 -0.121 -0.410~ 0.253 -0.110
(0.128) (0.152) (0.202) (0.243) (0.221) (0.178)
Fixed effects:
Country YES YES YES YES YES YES
Year YES YES YES YES YES YES
Sectorgroup YES YES YES YES YES YES
Country × Timetrend YES YES YES YES YES YES
Country × Business cycle YES YES YES YES YES YES
R2 0.280 0.252 0.428 0.317 0.142 0.069
N 15,520 15,520 15,520 15,520 15,520 15,520
Model weighted by: Employment Hours Hours VA VA VA
B. Aggregate elasticities
Employment Hours Wagebill Nominal VA Real VA Laborshare
(1) (2) (3) (4) (5) (6)
Σ Δ ln aggregate real VA (j≠i, c, t-k) 0.633** 0.558** 0.907**
- - -
(0.073) (0.083) (0.084)
Σ Δ ln aggregate nominal VA (j≠i, c, t-k) 1.083** 1.030** 0.071**
- - -
(0.026) (0.024) (0.025)
Fixed effects:
Sectorgroup YES YES YES YES YES YES
R2 0.227 0.194 0.414 0.300 0.110 0.006
N 15,520 15,520 15,520 15,520 15,520 15,520
Model weighted by: Employment Hours Hours VA VA VA
Notes: All TFP terms refer to other-country TFP, and are rescaled to have a unit standard deviation. Estimates are the sum of
coefficients for the contemporaneous effect and 5 annual distributed lags (k=0, ..., 5). The number of observations is equal to the
number of country-industry cells multiplied by the number of years. Standard errors are clustered by country-industry and
reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

57
Table 10. Industry-Level Contributions to Predicted Within- and Between-Industry Components of the
Change in Aggregate Labor Share, 1970 – 2007
Within- Between-
ISIC code Description industry industry
C Mining and quarrying -0.003 0.001
15t16 Food, beverages, and tobacco -0.006 -0.005
17t19 Textiles, textile , leather, and footwear -0.009 0.001
20 Wood and wood products -0.004 0.001
21t22 Pulp, paper, paper, printing, and publishing -0.009 0.001
23 Coke, refined petroleum and nuclear fuel 0.000 0.000
24 Chemicals and chemical products -0.019 0.010
25 Rubber and plastics -0.008 0.002
26 Other non-metallic mineral -0.005 0.001
27t28 Basic metals and fabricated metal -0.021 0.008
29 Machinery, not elsewhere classified -0.013 0.000
30t33 Electrical and optical equipment -0.038 0.009
34t35 Transport equipment -0.016 0.000
36t37 Manufacturing not elsewhere classified; recycling -0.003 0.000
E Electricity, gas, and water supply -0.010 0.009
F Construction -0.006 -0.008
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel -0.002 0.000
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles -0.023 0.008
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods -0.018 0.002
H Hotels and restaurants 0.003 -0.003
60t63 Transport and storage -0.018 0.005
64 Post and telecommunications -0.018 0.012
J Financial intermediation -0.017 0.009
70 Real estate activities 0.013 -0.086
71t74 Renting of machinery & equipment and other business activities 0.017 -0.008
M Education 0.001 -0.002
N Health and social work 0.001 -0.005
O Other community, social and personal service activities 0.006 -0.004
Total -0.225 -0.046
Notes: Predictions based on Table 9.

58
Table 11. The Contribution of TFP Growth to the Within and Between-Industry Components of the
Change in Aggregate Labor Share by Decade, 1970 – 2007
Annual change in laborshare in log points
A. Actual B. Predicted
Between Within Between Within
Decade Total industry industry Total industry industry
1970s 0.513 -0.187 0.700 -0.294 -0.124 -0.169
1980s -0.459 -0.183 -0.276 -0.365 -0.005 -0.360
1990s -0.263 -0.075 -0.188 -0.202 0.005 -0.207
2000s -0.861 -0.425 -0.436 -0.231 -0.091 -0.140
Notes: Predictions based on Table 9.

59
 Table 12. The Relationship Between Productivity Growth and Industry-Level Outcomes:
Allowing for Decade-Specific Direct Effects
dependent variable: annual change in log outcome by country-industry

A. Employment B. Hours C. Wagebill

Σ Δ ln Own-Industry TFP (i,c,t-k) impact in period: (1) (2) (3) (4) (7) (8)
1970s -0.834** 0.042 -0.774** 0.071 -0.464 0.244
(0.194) (0.148) (0.205) (0.154) (0.283) (0.259)
1980s-2000s -2.135** -1.182** -2.062** -1.109** -1.855** -1.014**
(0.170) (0.148) (0.183) (0.166) (0.261) (0.208)
Models weighted by: Employment Hours Hours
D. Nominal VA E. Real VA F. Laborshare
Σ Δ ln Own-Industry TFP (i,c,t-k) impact in period: (7) (8) (9) (10) (11) (12)
1970s -0.469~ -0.089 0.627~ 1.126** 0.146 0.289
(0.244) (0.329) (0.364) (0.363) (0.234) (0.335)
1980s-2000s -1.452** -0.685** 0.640 1.224** -0.386** -0.423**
(0.221) (0.191) (0.487) (0.415) (0.102) (0.144)
Models weighted by: Nominal VA Nominal VA Nominal VA
Fixed effects for all models:
Country YES YES YES YES YES YES
Year YES YES YES YES YES YES
Sectorgroup NO YES NO YES NO YES
Country × Timetrend YES YES YES YES YES YES
Country × Business cycle YES YES YES YES YES YES
Country × Year NO NO NO NO NO NO
Notes: TFP is other-country-same-industry TFP, rescaled to have a unit standard deviation (across the entire period).
Estimates are the sum of coefficients for the contemporaneous effect and 2 annual distributed lags (k=0, ..., 2) for the
1970s; and the contemporaneous effect and 5 annual distributed lags (k=0, ..., 5) for the 1980s-2000s. Models estimated
separately by sub-period, the number of observations is 3,520 for the 1970s; and 13,341 for the 1980s-2000s. Standard
errors are clustered by country-industry and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

60
Table 13. The Contribution of TFP Growth to the Within and Between-Industry Components of the
Change in Aggregate Labor Share by Decade, 1970 – 2007
Annual change in laborshare in log points
A. Actual B. Predicted
Between Within Between Within
Decade Total industry industry Total industry industry
1970s 0.513 -0.187 0.700 0.030 -0.020 0.050
1980s -0.459 -0.183 -0.276 -0.201 -0.022 -0.179
1990s -0.263 -0.075 -0.188 -0.125 -0.016 -0.109
2000s -0.861 -0.425 -0.436 -0.150 -0.085 -0.065
Notes: Predictions based on Table 12.

61
9. Appendix Tables

Appendix Table A1. EUKLEMS data coverage by country

ISO code Country Years
AUS Australia 1970-2007
AUT Austria 1970-2007
BEL Belgium 1970-2006
CAN Canada 1970-2004
DNK Denmark 1970-2007
ESP Spain 1970-2007
FIN Finland 1970-2007
FRA France 1970-2007
GER Germany 1970-2007
IRL Ireland 1988-2007
ITA Italy 1970-2007
JPN Japan 1973-2006
KOR South Korea 1970-2007
LUX Luxembourg 1986-2007
NLD Netherlands 1970-2007
PRT Portugal 1970-2006
SWE Sweden 1981-2007
UK United Kingdom 1970-2007
USA United States 1970-2005
Notes: Data coverage for EUKLEMS database, 2008
release supplemented with information from 2009
and 2007 releases. Greece excluded for lack of TFP
data.

62
 Table A2. EUKLEMS industry list

ISIC code Description Sector grouping

C Mining and quarrying Mining, utilities, and construction
15t16 Food, beverages, and tobacco Manufacturing
17t19 Textiles, textile , leather, and footwear Manufacturing
20 Wood and wood products Manufacturing
21t22 Pulp, paper, paper, printing, and publishing Manufacturing
23 Coke, refined petroleum and nuclear fuel Manufacturing
24 Chemicals and chemical products Manufacturing
25 Rubber and plastics Manufacturing
26 Other non-metallic mineral Manufacturing
27t28 Basic metals and fabricated metal Manufacturing
29 Machinery, not elsewhere classified Manufacturing
30t33 Electrical and optical equipment Manufacturing
34t35 Transport equipment Manufacturing
36t37 Manufacturing not elsewhere classified; recycling Manufacturing
E Electricity, gas, and water supply Mining, utilities, and construction
F Construction Mining, utilities, and construction
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel Low-tech services
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles Low-tech services
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods Low-tech services
H Hotels and restaurants Low-tech services
60t63 Transport and storage Low-tech services
64 Post and telecommunications High-tech services
J Financial intermediation High-tech services
70 Real estate activities Low-tech services
71t74 Renting of machinery & equipment and other business activities High-tech services
M Education Health and education
N Health and social work Health and education
O Other community, social and personal service activities Low-tech services
Notes: ISIC revision 3 codes. We exclude agriculture (industry AtB), public administration (industry L), private households (P) and
extra-territorial organizations (Q) from our analyses.

63
 Appendix Table A3. Shift-Share Analysis of Labor Share Changes in Levels by Decade

Annualized decadal percentage point change in laborshare by country

Weighted by country size Unweighted
Between Within Between Within
Decade Mean industry industry Mean industry industry
1970s 0.268 -0.084 0.352 0.123 -0.050 0.172
-31% 131% -40% 140%
1980s -0.256 -0.024 -0.232 -0.164 0.003 -0.167
9% 91% -2% 102%
1990s -0.126 -0.010 -0.116 -0.268 -0.088 -0.180
2% 98% 33% 67%
2000s -0.405 -0.130 -0.274 -0.116 0.019 -0.136
33% 67% -16% 117%

64
 Appendix Table A4. Contribution of Each Industry to Between- and Within-Industry Components of
Change in Aggregate Mean Log Labor Share by Decade
Contribution to between effect Contribution to within effect
ISIC code Description 1970s 1980s 1990s 2000s 1970s 1980s 1990s 2000s
C Mining and quarrying -0.18 0.23 0.04 -0.10 -0.11 0.03 -0.03 -0.05
15t16 Food, beverages, and tobacco 0.02 0.02 0.02 0.03 0.05 -0.04 -0.01 0.02
17t19 Textiles, textile , leather, and footwear 0.03 0.01 0.01 0.02 0.02 -0.01 0.00 0.00
20 Wood and wood products 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.01
21t22 Pulp, paper, paper, printing, and publishing 0.00 0.00 0.01 0.04 0.02 -0.02 -0.01 -0.01
23 Coke, refined petroleum and nuclear fuel 0.02 0.03 -0.01 -0.04 -0.01 0.00 -0.01 -0.03
24 Chemicals and chemical products 0.02 0.00 0.01 0.02 0.03 -0.03 -0.01 -0.02
25 Rubber and plastics 0.00 0.00 0.00 0.01 0.01 0.00 0.00 0.00
26 Other non-metallic mineral 0.01 0.01 0.01 0.01 0.02 -0.01 0.00 0.00
27t28 Basic metals and fabricated metal 0.07 0.04 0.02 0.01 0.04 -0.01 0.00 -0.04
29 Machinery, not elsewhere classified 0.00 0.02 0.01 0.01 0.02 -0.01 0.02 -0.02
30t33 Electrical and optical equipment 0.01 -0.01 0.00 0.05 0.04 -0.03 -0.01 0.00
34t35 Transport equipment -0.01 0.01 0.01 0.00 0.01 -0.01 -0.01 0.00
36t37 Manufacturing not elsewhere classified; recycling 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00
E Electricity, gas, and water supply -0.06 0.01 0.03 0.01 0.02 -0.04 -0.02 -0.02
F Construction 0.02 0.01 0.03 -0.02 0.08 -0.04 0.05 -0.08
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 0.00 0.00 0.00 0.00 0.01 0.00 -0.02 -0.01
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles -0.01 0.01 -0.01 0.00 0.03 0.01 -0.04 -0.09
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods 0.01 0.00 0.00 0.01 0.05 -0.04 -0.03 -0.01
H Hotels and restaurants 0.00 0.00 0.00 0.00 0.03 -0.01 -0.02 -0.03
60t63 Transport and storage 0.02 0.02 0.01 0.01 0.06 -0.02 -0.02 -0.03
64 Post and telecommunications -0.01 0.00 -0.02 0.01 0.00 -0.04 -0.01 -0.06
J Financial intermediation -0.02 -0.05 -0.05 0.00 0.06 -0.02 -0.09 -0.05
70 Real estate activities -0.04 -0.40 -0.10 -0.45 0.17 0.05 0.04 0.08
71t74 Renting of machinery & equipment and other business activities -0.04 -0.11 -0.09 -0.02 0.05 0.03 0.06 -0.01
M Education 0.00 0.00 0.00 0.00 -0.01 -0.02 -0.01 0.06
N Health and social work -0.02 -0.01 -0.02 -0.04 0.02 0.01 -0.02 -0.03
O Other community, social and personal service activities -0.02 -0.02 -0.01 0.00 0.02 -0.02 0.02 0.01
Total -0.187 -0.183 -0.075 -0.425 0.699 -0.276 -0.188 -0.436

65
 Appendix Table A5. Predictive Industry-Level Relationship Between Other-Country TFP Growth and
Own-Country TFP Growth
dependent variable: annual change in own-country log TFP by industry

(1) (2) (3) (4) (5) (6) (7) (8)

Δ ln TFP (ict) in other countries 0.567** 0.567** 0.549** 0.554** 0.572** 0.517** 0.322** 0.496**
(0.079) (0.079) (0.082) (0.083) (0.084) (0.065) (0.064) (0.064)
Fixed effects:
Country NO YES YES YES YES YES YES YES
Year NO NO YES YES YES YES YES YES
Country × Timetrend NO NO NO YES NO NO NO YES
Country × Business cycle NO NO NO YES NO NO NO YES
Country × Year NO NO NO NO YES YES YES NO
Sectorgroup NO NO NO NO NO YES NO YES
Industry NO NO NO NO NO NO YES NO
R2 0.060 0.064 0.073 0.085 0.117 0.127 0.156 0.095
N 15,007 15,007 15,007 15,007 15,007 15,007 15,007 15,007
Notes: All models weighted by industry value added shares within countries, multiplied by time-varying country shares in total value added.
The number of observations is equal to the number of country-industry cells multiplied by the number of years. Standard errors are
clustered by country-industry and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

66
 Appendix Table A6. Summary Statistics for Standardized TFP Measures and Patents
A. Total Factor Productivity
Weighted by country size Not weighted by country size
Mean Sd Mean Sd
Δ ln Own-Industry TFP (i, c, t) 0.622 2.585 0.636 2.484
Δ ln Supplier-Industry TFP (j≠i, c, t) 0.196 0.359 0.167 0.315
Δ ln Customer-Industry TFP (j≠i, c, t) 0.108 0.318 0.098 0.252

B. Patents and patent citations

Weighted by country size Not weighted by country size
Mean Sd Mean Sd
Δ ln Own-Industry patents (i, c, t) 4.334 2.591 4.641 2.475
Δ ln Own-Industry patent citations (i, c, t) 3.224 2.809 3.692 2.698
Notes: All variables weighted by industry shares in value added.

67
 Appendix Table A7. Trends in Patent Grants and Patent Citations by Industry for US and non-US
Inventors, 1970 – 2007
Mean log nr of patents Mean log nr of patent citations
by US by non-US by US by non-US
ISIC code Description inventors
6.23 inventors
4.50 inventors
5.10 inventors
3.14
C Mining and quarrying 6.23 4.50 5.10 3.14
15t16 Food , beverages, and tobacco 5.27 3.47 3.87 1.62
17t19 Textiles, textile , leather, and footwear 5.18 4.07 4.22 2.78
20 Wood and wood products 4.02 3.14 2.95 1.84
21t22 Pulp, paper, paper, printing, and publishing 6.56 4.50 5.73 3.26
23 Coke, refined petroleum and nuclear fuel 7.38 6.02 6.05 4.32
24 Chemicals and chemical products 8.31 7.23 7.22 5.88
25 Rubber and plastics 5.98 4.49 4.91 3.03
26 Other non-metallic mineral 5.83 3.42 4.79 1.96
27t28 Basic metals and fabricated metal 6.54 5.09 5.29 3.53
29 Machinery, not elsewhere classified 7.43 6.32 6.35 5.00
30t33 Electrical and optical equipment 8.54 8.15 7.65 6.88
34t35 Transport equipment 7.36 6.70 6.25 5.37
36t37 Manufacturing not elsewhere classified; recycling 5.75 4.11 4.83 2.87
E Electricity, gas, and water supply 3.01 3.15 1.75 1.49
F Construction 4.57 3.05 3.80 1.75
50 Sale, maintenance and repair of motor vehicles; retail sale of fuel 3.09 2.20 1.87 0.53
51 Wholesale trade and commission trade, except of motor vehicles 3.40 2.76 2.09 1.01
52 Retail trade, except of motor vehicles; repair of household goods 5.04 3.52 4.19 2.18
H Hotels and restaurants 3.55 2.23 2.53 0.86
60t63 Transport and storage 4.10 2.77 3.05 1.44
64 Post and telecommunications 6.72 4.83 5.82 3.67
J Financial intermediation 4.83 3.86 4.10 2.29
70 Real estate activities 1.68 1.58 0.48 0.17
71t74 Renting of machinery & equipment and other business activities 7.40 6.59 6.42 5.37
M Education -0.95 -2.04 -1.45 -3.12
N Health and social work 2.79 1.59 2.03 0.79
O Other community, social and personal service activities 5.30 3.91 4.09 2.15
Notes: Average across 1970-2007.

68
 Appendix Table A8. Trends in Patent Grants and Patent Citations by Industry for US and non-US
Inventors: Overall and by Decade and Sector
dependent variable: 100 x log outcome by country-industry-year

Nr of patents Nr of patent citations

by US by non-US by US by non-US
inventors inventors inventors inventors
(1) (2) (3) (4)
Mean 4.578** 3.513** 3.583** 2.191**
(0.097) (0.093) (0.108) (0.101)

Mean for:
1970s 3.761** 2.605** 2.531** 1.221**
(0.194) (0.159) (0.213) (0.159)
1980s 4.489** 3.363** 4.016** 2.469**
(0.177) (0.172) (0.174) (0.178)
1990s 4.973** 4.013** 4.951** 3.320**
(0.183) (0.179) (0.172) (0.188)
2000s 5.385** 4.308** 2.604** 1.464**
(0.196) (0.220) (0.262) (0.276)

Mean for:
Mining & utilities & construction 4.471** 3.284** 3.543** 1.913**
(0.207) (0.135) (0.262) (0.216)
Manufacturing 6.675** 5.360** 5.583** 3.926**
(0.078) (0.102) (0.098) (0.120)
Education & health 1.091** -0.186 0.448 -1.143**
(0.282) (0.290) (0.322) (0.344)
Low-tech services 3.471** 2.581** 2.342** 1.098**
(0.117) (0.084) (0.153) (0.111)
High-tech services 6.318** 5.276** 5.447** 3.924**
(0.167) (0.200) (0.193) (0.235)

N 1,059 1,042 1,057 1,026

Notes: 1970-2007. All models weighted by time-averaged industry value added shares, averaged across
countries. The number of observations is equal to the number of industry cells multiplied by the
number of years. Robust standard errors are reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

69
 Appendix Table A9. Aggregate Elasticity Estimates: Country-Level Growth and Industry-Level
Employment, Hours, Wagebill, and Output
dependent variable: annual change in log outcome by country

Employment Hours Wagebill Nominal VA Real VA Laborshare

(1) (2) (3) (4) (5) (6)
Σ Δ ln aggregate real VA (c, t-k) 0.414** 0.301** 0.436**
- - -
(0.041) (0.051) (0.134)
Σ Δ ln aggregate nominal VA (c, t-k) 0.929** 0.858** 0.054**
- - -
(0.025) (0.027) (0.021)
R2 0.301 0.173 0.685 0.656 0.056 0.051
N 552 552 552 552 552 552
Model weighted by: Employment Hours Hours VA VA VA
Notes: Estimates are the sum of coefficients for 5 annual distributed lags (k=1, ..., 5). The number of observations is equal to
the number of countries multiplied by the number of years. Standard errors are reported in parentheses, ~ p<0.10, * p<0.05,
** p<0.01.

70
Appendix Table A10. Estimates of the Relationship Between Productivity Growth and Industry-Level
Outcomes, 1970 – 2007 Excluding Contemporaneous TFP Measure from Distributed Lag Model: Direct
Effects, Supplier and Customer Effects, and Aggregate Effects
dependent variable: annual change in log outcome by country-industry

A. Industry effects
Employment Hours Wagebill Nominal VA Real VA Laborshare
(1) (2) (3) (4) (5) (6)
Σ Δ ln Own-Industry TFP (i, c, t-k) -0.895** -0.871** -1.061** -0.688** 0.941** -0.452**
(0.137) (0.151) (0.204) (0.172) (0.357) (0.156)
Σ Δ ln Supplier-Industry TFP (j≠i, c, t-k) 0.797** 0.678** 0.021 0.130 -0.091 0.048
(0.214) (0.229) (0.317) (0.322) (0.421) (0.271)
Σ Δ ln Customer-Industry TFP (j≠i, c, t-k) 0.105 0.165 -0.098 -0.241 0.310 -0.208
(0.125) (0.151) (0.189) (0.201) (0.225) (0.188)
Fixed effects:
Country YES YES YES YES YES YES
Year YES YES YES YES YES YES
Sectorgroup YES YES YES YES YES YES
Country × Timetrend YES YES YES YES YES YES
Country × Business cycle YES YES YES YES YES YES
R2 0.275 0.244 0.427 0.314 0.135 0.067
N 15,520 15,520 15,520 15,520 15,520 15,520
Model weighted by: Employment Hours Hours VA VA VA
B. Aggregate elasticities
Employment Hours Wagebill Nominal VA Real VA Laborshare
(1) (2) (3) (4) (5) (6)
Σ Δ ln aggregate real VA (j≠i, c, t-k) 0.450** 0.344** 0.434**
- - -
(0.066) (0.077) (0.095)
Σ Δ ln aggregate nominal VA (j≠i, c, t-k) 0.984** 0.917** 0.082**
- - -
(0.024) (0.024) (0.024)
Fixed effects:
Sectorgroup YES YES YES YES YES YES
R2 0.183 0.132 0.371 0.252 0.046 0.005
N 15,520 15,520 15,520 15,520 15,520 15,520
Model weighted by: Employment Hours Hours VA VA VA
Notes: All TFP terms refer to other-country TFP, and are rescaled to have a unit standard deviation. Estimates are the sum of
coefficients for 5 annual distributed lags (k=1, ..., 5). The number of observations is equal to the number of country-industry cells
multiplied by the number of years. Standard errors are clustered by country-industry and reported in parentheses, ~ p<0.10, *
p<0.05, ** p<0.01.

71
 Appendix Table A11A. Industry Contributions to Predicted Employment Effects
by Source of TFP Growth, 1970 – 2007
ISIC code Description Direct Supplier Customer Final dem. Net
C Mining and quarrying 0.000 0.003 0.000 0.002 0.005
15t16 Food, beverages, and tobacco -0.007 0.013 0.004 0.013 0.022
17t19 Textiles, textile , leather, and footwear -0.025 0.008 0.003 0.015 0.002
20 Wood and wood products -0.006 0.013 0.000 0.006 0.013
21t22 Pulp, paper, paper, printing, and publishing -0.009 0.031 0.001 0.014 0.037
23 Coke, refined petroleum and nuclear fuel 0.000 -0.002 0.000 -0.002 -0.004
24 Chemicals and chemical products -0.014 0.060 0.004 0.035 0.085
25 Rubber and plastics -0.010 0.032 0.001 0.013 0.037
26 Other non-metallic mineral -0.006 0.020 0.000 0.008 0.022
27t28 Basic metals and fabricated metal -0.021 0.056 0.001 0.027 0.063
29 Machinery, not elsewhere classified -0.017 0.016 0.006 0.024 0.029
30t33 Electrical and optical equipment -0.051 0.063 0.012 0.070 0.094
34t35 Transport equipment -0.019 0.015 0.012 0.037 0.044
36t37 Manufacturing not elsewhere classified; recycling -0.006 0.005 0.002 0.006 0.007
E Electricity, gas, and water supply -0.004 0.043 0.001 0.017 0.057
F Construction -0.005 0.003 0.002 0.007 0.007
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel -0.001 0.003 0.000 0.001 0.003
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles -0.024 0.077 0.003 0.035 0.091
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods -0.049 0.047 0.004 0.032 0.034
H Hotels and restaurants 0.019 -0.015 -0.004 -0.014 -0.013
60t63 Transport and storage -0.021 0.069 0.002 0.030 0.080
64 Post and telecommunications -0.017 0.121 0.003 0.042 0.149
J Financial intermediation -0.014 0.089 0.002 0.034 0.111
70 Real estate activities 0.003 -0.065 -0.003 -0.033 -0.099
71t74 Renting of machinery & equipment and other business activities 0.049 -0.313 -0.003 -0.088 -0.355
M Education 0.004 -0.001 0.000 -0.003 0.000
N Health and social work 0.006 -0.001 -0.001 -0.006 -0.001
O Other community, social and personal service activities 0.023 -0.035 -0.003 -0.020 -0.035
Total -0.222 0.353 0.051 0.301 0.482
Notes: Predictions based on Table 9.

72
 Appendix Table A11B. Industry Contributions to Predicted Employment Effects
by Destination of Employment Growth, 1970 – 2007
ISIC code Description Direct Supplier Customer Final dem. Net
C Mining and quarrying 0.000 0.002 0.001 0.002 0.004
15t16 Food, beverages, and tobacco -0.007 0.014 0.000 0.008 0.014
17t19 Textiles, textile , leather, and footwear -0.025 0.022 0.001 0.010 0.008
20 Wood and wood products -0.006 0.005 0.001 0.002 0.002
21t22 Pulp, paper, paper, printing, and publishing -0.009 0.009 0.002 0.006 0.007
23 Coke, refined petroleum and nuclear fuel 0.000 0.001 0.000 0.000 0.001
24 Chemicals and chemical products -0.014 0.004 0.001 0.004 -0.005
25 Rubber and plastics -0.010 0.010 0.002 0.003 0.005
26 Other non-metallic mineral -0.006 0.006 0.001 0.003 0.004
27t28 Basic metals and fabricated metal -0.021 0.016 0.005 0.010 0.010
29 Machinery, not elsewhere classified -0.017 0.023 0.001 0.007 0.015
30t33 Electrical and optical equipment -0.051 0.014 0.001 0.009 -0.026
34t35 Transport equipment -0.019 0.021 0.000 0.006 0.008
36t37 Manufacturing not elsewhere classified; recycling -0.006 0.013 0.000 0.004 0.011
E Electricity, gas, and water supply -0.004 0.002 0.001 0.002 0.001
F Construction -0.005 0.064 0.001 0.026 0.086
50 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel -0.001 0.010 0.001 0.006 0.015
51 Wholesale trade and commission trade, except of motor vehicles and motorcycles -0.024 0.009 0.006 0.018 0.010
52 Retail trade, except of motor vehicles and motorcycles; repair of household goods -0.049 0.010 0.006 0.034 0.000
H Hotels and restaurants 0.019 0.029 0.002 0.019 0.069
60t63 Transport and storage -0.021 0.007 0.004 0.014 0.004
64 Post and telecommunications -0.017 -0.001 0.001 0.005 -0.012
J Financial intermediation -0.014 -0.009 0.002 0.012 -0.010
70 Real estate activities 0.003 0.002 0.000 0.004 0.009
71t74 Renting of machinery & equipment and other business activities 0.049 0.034 0.009 0.024 0.116
M Education 0.004 0.005 0.001 0.021 0.031
N Health and social work 0.006 0.020 0.000 0.023 0.050
O Other community, social and personal service activities 0.023 0.012 0.002 0.018 0.056
Total -0.222 0.353 0.051 0.301 0.482
Notes: Predictions based on Table 9.

73
 Appendix Table A12. Direct Effects by Decade
dependent variable: annual change in log outcome by country-industry

Employment Hours Wagebill Nominal VA Real VA Laborshare

(1) (2) (3) (4) (5) (6)
Δ ln Own-Industry TFP (ci) in previous 5-year period
1970s -0.256 -0.206 -0.446~ -0.457 0.465 -0.215
(0.171) (0.187) (0.233) (0.300) (0.400) (0.299)
N 419 419 419 419 419 419
1980s -0.533** -0.421* -0.619** -0.266 0.435 -0.368*
(0.187) (0.210) (0.232) (0.270) (0.340) (0.161)
N 894 894 894 894 894 894
1990s -1.088** -1.061** -0.645** -0.413* 0.835* -0.455**
(0.149) (0.161) (0.189) (0.187) (0.348) (0.156)
N 1,005 1,005 1,005 1,005 1,005 1,005
2000s -0.605** -0.662** -0.741** -0.472* 1.046** -0.397*
(0.162) (0.176) (0.210) (0.219) (0.222) (0.167)
N 502 502 502 502 502 502
Fixed effects for all models
Country YES YES YES YES YES YES
Sectorgroup YES YES YES YES YES YES
Models weighted by: Employment Hours Hours VA VA VA
Notes: TFP is other-country-same-industry TFP, and are rescaled to have a unit standard deviation (across the
entire period). Models estimated separately by decade. The number of observations is equal to the number of
country-industry cells multiplied by the number of years. Standard errors are clustered by country-industry
and reported in parentheses, ~ p<0.10, * p<0.05, ** p<0.01.

74