does productivity growth threaten employment ... productivity growth threaten employment? ... 1 does...

Does Productivity Growth Threaten Employment?“Robocalypse Now?”

David Autor 1 Anna Salomons 2

1MIT and NBER

2Utrecht University

NBER Economics of AI Conference14 Sep 2017

Autor & Salomons Does Productivity Growth Threaten Employment? “Robocalypse Now?” 1 / 45

Longstanding concern: Automation threatens employment

Automation and Jobs: 200 Years of Concern

1. Luddites—Skilled weavers in the 19th century

2. U.S. Labor Secretary James Davis in 1927

3. Lyndon Johnson 1964 “Blue-RibbonPresidential Commission on Technology,Automation, and Economic Progress”

4. Wassily Leontief in 1982:Role of workers will diminish — like horses

5. Right now!


Fundamentally, does rising productivity mean fewer jobs?

Citizen, policy-maker, intellectual concern

The more work done by machines, the less work done by people

Steam-powered hammer vs. “steel-driving man”

Professional economic opinion

1 Elastic demand: Advancing sectors may expand (Bessen ’17)

2 Income effects: Rising wealth creates new demands (Clark ’51)

3 Sectoral reallocation: Advancing sectors contract, but labor moves to lagging sectors(Baumol ’67)


Economists appear to be losing confidence in these long-held theories:“Robocalypse Now?”

Labor’s share of national income falling cross-nationallyLabor’s Share of GDP Falling in Four Largest

Economies Since �1980

.55

.6.6

5.7

Labo

r Sha

re

1975 1985 1995 2005 2015

United States

.55

.6.6

5.7

Labo

r Sha

re

1975 1985 1995 2005 2015

Japan

.35

.4.4

5.5

Labo

r Sha

re

1975 1985 1995 2005 2015

China

.55

.6.6

5.7

Labo

r Sha

re1975 1985 1995 2005 2015

Germany

FIGURE II

Declining Labor Share for the Largest Countries

The figure shows the labor share and its linear trend for the four largest economies in the world from 1975.

GLO

BA

LD

EC

LINE

OF

TH

ELA

BO

RSH

AR

E71

Karabarbounis and Neiman, 2014Autor & Salomons Does Productivity Growth Threaten Employment? “Robocalypse Now?” 4 / 45

It’s not just the falling labor share that has scholars worried...

An age of ‘brilliant machines’ (Brynjolfsson-McAfee ’14)

1 Computers managing financial portfolios, beating ’Go’ players

2 Websites and drones eliminating sales workers, warehouse workers

3 Robots leaving the assembly lines, coming for your jobs...


Economists have taken notice...

Emerging understanding makes clear that this can happen

Machines can directly replace specific job tasks, complement workers in other job tasks,possibly spur creation of new labor-using tasks

Autor-Levy-Murnane ’03, Acemoglu-Autor ’11, Acemoglu-Restrepo ’16

Growing literature: Models of labor immiseration

1 Inter-generational market failure: Sachs & Kotlikoff ’12, Berg et al. ’17

2 Task encroachment: No place left to hide (Susskind ’17)

3 New tasks might endogenously be created ‘fast enough’ – or perhaps not (Acemoglu &Restrepo ’16, ’17)


Evidence does not (yet) strongly support immiseration view

Vast literature makes clear that computerization has been skill-biasedAutor-Katz-Kearney ’08; Akerman-Kostol-Mogstad, ’14

But little work on overall employment impact of technological ∆′s

1 Alexopoulos-Cohen ’16: Technological progress strongly employment-creating — but inthe 1910s–1940s

2 Gregory-Salomons-Zierahn ’16: Employment-reducing effects of Routine-ReplacingTechnical Change (RRTC) offset by compensatory demand + local spillover effects

3 Graetz-Michaels ’15: Industrial robots raising wages and value-added, raising demand forskilled workers across Europe (industry-level data)

4 Acemoglu-Restrepo ’17: Industrial robots lowering wages and employment in U.S. locallabor markets


This paper asks: Is recent labor-augmenting technological progresseroding employment?

1 Does productivity growth cause advancing industries to grow orshrink?

2 Do cross-industry spillovers offset or augment direct own-industryeffects—and what’s the net effect?

3 Has the employment-productivity relationship changed in the 2000’s?

4 Is productivity-growth skill-biased—should we worry about jobs orskills?


Is recent labor-augmenting technological progress eroding employment?

Approach

Study the impact of productivity growth on employment across 19 countries, 37 years

Focus on overall productivity growth: (1) output per worker, (2) value-added per worker,(3) total factor productivity

Outcomes

∆ Employment by industry and overall (counts and e/pop)

∆ Final consumption by industry—corroborating productivity effects

∆ Skill inputs within industries

∆ Skill inputs economy-wide—due to induced sectoral shifts


Data sources and the ‘big picture’

Outline

1 Data sources and the ‘big picture’

2 Do ‘advancing’ industries grow or shrink?

3 Reconciling industry and aggregate-level evidence

4 Adding it up

5 Is this time (period) different?

6 Should we worry about jobs or skills?

7 Conclusions and next steps



Big picture: Employment rate usually rises with productivityEmployment growth, productivity growth positively covary, 1970–2007 (cf. Francis-Ramey ’04)

-10

12

34

-2-1

01

2

1970 1990 2010year

1. France

-20

24

6

-4-2

02

4

1970 1990 2010year

2. Germany

-20

24

6

-2-1

01

23

1970 1990 2010year

3. Japan

-50

510

-4-2

02

4

1970 1990 2010year

4. UK

-20

24

-4-2

02

4

1970 1990 2010year

5. USA

-20

24

6

-10

12

3

1970 1990 2010year

6. Mean of all others

Employment to working age pop growth Labor productivity growth

Labo

r pro

duct

ivity

gro

wth

Em

ploy

men

t to

wor

king

age

pop

ulat

ion

grow

th

Figures are for the total economy, excluding agriculture, public administration, private households and extraterritorialorganizations. All growth rates obtained as log changes x 100. Graph 6 reports unweighted mean growth ratesacross the remaining 14 countries. Productivity is gross output per worker.



Not just the ’Big Five’ countries: Employment rates rise with productivityOf course – these macroeconomic correlations are not deeply informative

-50

510

-4-2

02

4

1970 1990 2010year

1. Austria

-20

24

6

-2-1

01

23

1970 1990 2010year

2. Denmark

-4-2

02

46

-4-2

02

4

1970 1990 2010year

3. Italy

-4-2

02

4

-4-2

02

4

1970 1990 2010year

4. Netherlands

-4-2

02

46

-4-2

02

46

1970 1990 2010year

5. Spain

-50

510

15

-10

-50

510

1970 1990 2010year

6. South Korea

Employment to working age pop growth Labor productivity growth

Labo

r pro

duct

ivity

gro

wth

Em

ploy

men

t to

wor

king

age

pop

ulat

ion

grow

th

Figures are for the total economy, excluding agriculture, public administration, private households and extraterritorialorganizations. All growth rates obtained as log changes x 100. Productivity is gross output per worker.



Data sources

Primary: EU KLEMS 1970-2007 (O’Mahony & Timmer ’09)

19 developed countries

AUS, AUT, BEL, DNK, ESP, FIN, FRA, GER, GRC, IRL, ITA, JPN, KOR, LUX, NLD,PRT, SWE, UK, USA

28 industries

All non-farm employment except public administration, private households, andextraterritorial organizations

Employment and labor productivity

Real gross output per worker, real value added per worker, total factor productivity (TFP)by country-industry-year

Additional measures: World Input Output Tables (WIOT)

Measuring consumption responses to productivity gains

Currently adding data through 2014 (EU KLEMS 2016)


Do ‘advancing’ industries grow or shrink?

Outline




4 Adding it up







Testing whether rising productivity raises or lowers employment...

Using KLEMS data for 17 countries, 25 industries, 37 years, fit country- by-industry-by-year stacked first-difference OLS model

∆lnEcit = β0 + β1∆lnLPcit + [αc + δt + γi ] + εcit

∆lnLPcit is growth in labor productivityi indexes industriesc indexes countriest indexes yearsE is employment

Models are weighted by the time-averaged employment shares of industries withincountries




What should happen to industry employment as ∆lnLPcit rises?

1 Lump-of-labor

Could fall one-for-one with labor productivity growth: ∂ ln Ei∂ ln LPi

= −1

2 Demand surge (iPhone, textiles)

Could surge as price/quality improve: ∂ ln Ei∂ ln LPi

> 0

3 Unbalanced growth (Baumol)

Could fall somewhat less than one-for-one: −1 < ∂ ln Ei∂ ln LPi

< 0



Employment robustly falls as labor productivity ↗

dep var: annual log change in employment by country-industry

A. OLS

(1) (2) (3) (4) (5)

∆lnLPcit -0.248** -0.259** -0.275** -0.249** -0.248**(0.024) (0.023) (0.024) (0.024) (0.024)

∆lnpopulationct – – – – 0.895**(0.191)

Country fixed effects NO YES YES YES YESYear fixed effects NO NO YES YES YESIndustry fixed effects NO NO NO YES YES

R2 0.110 0.155 0.201 0.300 0.305N 19,451 19,451 19,451 19,451 19,451

Standard errors in parentheses, ∼ p<0.10, * p<0.05, ** p<0.01



What does happen: Rising labor productivity → Falling industryemployment

-0.4

0-0

.30

-0.2

0-0

.10

0.00

0.10

Est

imat

ed c

oeffi

cien

t

Gross output per worker Value added per worker Gross output based TFP

No FE Country FECountry, year FE Country, year, industry FECountry*year, country*industry,industry*year FE 95% confidence interval

All models are estimated by OLS and control for population growth whenever country-year fixed effects are not included.



Unbalanced growth: Employment in ‘advancing’ sectors shrinks

Cumulative Productivity Growth Cumulative Change in Employment

015

3045

6075

9010

512

0

100

x cu

mul

ativ

e lo

g ch

ange

1970 1980 1990 2000 2010year

Mining, utilities, construction ManufacturingEducation, health Low-tech servicesHigh-tech services

Unweighted average across all 19 countries. Productivity is gross output based.

-20

-15

-10

-50

510

Sha

re c

hang

e (p

erce

ntag

e po

ints

)

1970 1980 1990 2000 2010year

Mining, utilities, construction ManufacturingEducation, health Low-tech servicesHigh-tech services

Shares normalized to 0 in 1970. Unweighted average across all 19 countries.



Rising labor productivity → Falling industry employmentUsing gross-output based labor productivity growth

-0.9

0-0

.45

0.00

Est

imat

ed c

oeffi

cien

t

Min

ing

Util

ities

Con

stru

ctio

n

Food

Text

iles

Woo

dP

aper

Ref

ined

pet

role

umC

hem

ical

sR

ubbe

r, pl

astic

Non

-met

al m

iner

al, n

ecB

asic

met

als

Mac

hine

ry, n

ecE

lect

rical

equ

ipm

ent

Tran

spor

t equ

ipm

ent

Man

uf n

ec, r

ecyc

ling

Edu

catio

nH

ealth

& s

ocia

l wor

k

Mot

or v

ehic

le s

ale

Who

lesa

le tr

ade

Ret

ail;

hh g

oods

repa

irH

otel

s, re

stau

rant

sTr

ansp

ort,

stor

age

Rea

l est

ate

Com

m, s

ocia

l & p

rsnl

svc

s

Pos

t & te

leco

mm

unic

atio

nFi

nanc

ial i

nter

med

iatio

nB

usin

ess

activ

ities

Util's, mining, constructionManufacturing

Educ & healthLow-tech svcs

High-tech svcs95% CI

From a model with a full set of industry interactions in all productivity terms; country, industry, and year fixed effects;and controlling for population growth. Productivity is gross output per worker.



Rising labor productivity → Falling industry employmentUsing value-added based labor productivity growth

-0.90

-0.45

0.00

Estim

ated c

oeffic

ient

Minin

gUt

ilities

Cons

tructi

on

Food

Texti

lesW

ood

Pape

rRe

fined

petro

leum

Chem

icals

Rubb

er, p

lastic

Non-

metal

mine

ral, n

ecBa

sic m

etals

Mach

inery,

nec

Elec

trical

equip

ment

Tran

spor

t equ

ipmen

tMa

nuf n

ec, r

ecyc

ling

Educ

ation

Healt

h & so

cial w

ork

Motor

vehic

le sa

leW

holes

ale tr

ade

Retai

l; hh g

oods

repa

irHo

tels,

resta

uran

tsTr

ansp

ort, s

torag

eRe

al es

tate

Comm

, soc

ial &

prsn

l svc

s

Post

& tel

ecom

munic

ation

Finan

cial in

terme

diatio

nBu

sines

s acti

vities




From a model with a full set of industry interactions in all productivity terms; country, industry, and year fixed effects;and controlling for population growth. Productivity is value added per worker.



Rising labor productivity → Falling industry employmentUsing Total Factor Productivity growth

-0.60

-0.30

0.00

Estim

ated c

oeffic

ient

Minin

gUt

ilities

Cons

tructi

on

Food

Texti

lesW

ood

Pape

rRe

fined

petro

leum

Chem

icals

Rubb

er, p

lastic

Non-

metal

mine

ral, n

ecBa

sic m

etals

Mach

inery,

nec

Elec

trical

equip

ment

Tran

spor

t equ

ipmen

tMa

nuf n

ec, r

ecyc

ling

Educ

ation

Healt

h & so

cial w

ork

Motor

vehic

le sa

leW

holes

ale tr

ade

Retai

l; hh g

oods

repa

irHo

tels,

resta

uran

tsTr

ansp

ort, s

torag

eRe

al es

tate

Comm

, soc

ial &

prsn

l svc

s

Post

& tel

ecom

munic

ation

Finan

cial in

terme

diatio

nBu

sines

s acti

vities




From a model with a full set of industry interactions in all productivity terms; country, industry, and year fixed effects;and controlling for population growth. Productivity is value added based TFP.



Reality check: Is there a consumption response? Check!Consumption of industry output rises with industry productivity, even as employment falls

0.00

0.10

0.20

0.30

0.40

0.50

0.60

Est

imat

ed c

oeffi

cien

t

Gross output per worker Value added per worker

No FECountry FE

Country, year FECountry, year, industry FE95% confidence interval

WIOD, 1995-2009. Models are estimated by OLS; contain country, year, and industry FE; and control for population growth.


Reconciling industry and aggregate-level evidence

Outline




4 Adding it up






Negative employment impact at industry level but seemingly not ataggregate level. Why not?

Reconciling the evidence

Plausible explanation: There are employment spillovers elsewhere in economy

1 Rising final demand — income effects

2 Rising inter-industry demand – input-output linkages

Use industry-level and country-level data to estimate

∆lnEcit = β0 + β1∆lnLPcit +3∑

k=0

β2+k∆lnLPct−k,j 6=i [+αc + δt + γi ] + εcit

LPct−k,j 6=i is aggregate labor productivity excluding own-industry i

LPcit is own-industry labor productivity

c indexes countries

t indexes years



Direct and spillover effects of productivity growthSpillover effects fully offset internal effects: Net impact on emp/pop is weakly positive

-1.2

0-0.8

0 -0.4

00.00

0.40

0.80

1.20

1.60

2.00

Est

imat

ed c

oeffi

cien

t

Direct effect Spillover effect Net effect

Gross output (GO) per workerPeak-to-peak (GO per worker)Trough-to-trough (GO per worker)

Value added per workerGross output based TFP95% confidence interval

All models are estimated by OLS; contain country, year, and industry FE; and control for population growth.



Is all productivity growth equally job-creating?

Industry productivity growth raises aggregate employment onaverage—but does it matter where productivity originates?

We have so far restricted effects of industry productivity to have uniform impacts

But internal and external effects of productivity growth may vary across sectors

Relative weight in the economyProduct market competitionDemand saturationIntegration in international production chains.



Is all productivity growth equally job-creating?

Allow direct effects and spillovers to differ by sector1 Mining, utilities and construction

2 Manufacturing

3 Education and health

4 Low-tech services: Retail, sales, hotels, restaurants, etc.

5 High-tech services: Finance, business services, telecoms

∆lnEict = β0 +5∑

s(i)=1

β1,s(i)∆lnLPict +5∑

s(i)=1

3∑k=0

β2+k,s(i)∆lnLPct−k,s(i),j 6=i

[+αc + δt + γi ] + εict

β1,s(i) are sector-specific effects of own-industry labor productivity

β2+k,s(i) are sector-specific spillovers to other industries



Sizes of direct and spillover effects differ by sectorManufacturing has least negative direct effect; low-tech services has largest positive spillovers

-.5-.4

-.3-.2

-.10

.1.2

.3

Est

imat

ed c

oeffi

cien

t

Direct effect Spillover effect

Mining, utilities & construction ManufacturingEducation & health Low-tech servicesHigh-tech services

Model is estimated by OLS; includes country, industry, and year FE; and controls for population growth.Productivity is gross output per worker.


Adding it up

Outline




4 Adding it up





Adding it up

Translating direct+spillover effects into total emp/pop

What do direct + spillover effects imply for emp/pop in net?

Use estimates to infer how much each sector’s productivity growth has augmented ordecreased total employment-to-population

∆Eict = {Eic,t=base × 1(i ∈ s)× β1,s(i) ×∆lnLPict}

+{Eic,t=base ×5∑

s(i)=1

3∑k=0

β2+k,s(i) ×∆lnLPct−k,s(i),j 6=i}


Adding it up

Implied cumulative net effects of productivity growth on ∆ employmentin % pts, 1970–2007

-20

24

6C

umul

ativ

e pr

edic

ted

∆ em

ploy

men

t (%

)

1970 1980 1990 2000 2010year

Mining, utilities, construction ManufacturingEducation, health Low-tech servicesHigh-tech services Total

Based on model 5 from Table 7; prediction averaged across all 19 countries. Productivity is gross output per worker.


Adding it up

How big are these effects? Not negligible...Actual changes in emp-to-pop vs. contribution of productivity growth: Five largest economies

4850

5254

56

1970 1990 2010year

1. France

5055

6065

1970 1990 2010year

2. Germany

6062

6466

6870

1970 1990 2010year

3. Japan

6065

7075

1970 1990 2010year

4. UK

5560

6570

1970 1990 2010year

5. USA

5055

6065

1970 1990 2010year


Actual Due to productivity growth

Em

ploy

men

t to

wor

king

age

pop

ulat

ion

(per

cent

)

Figures are for the total economy, excluding agriculture, public administration, private households and extraterritorial organizations.Productivity is gross output per worker.


Is this time (period) different?

Outline




4 Adding it up






Productivity ⇒ Job growth: Is this time (period) different?

Productivity and job growth appear to diverge in some countries in2000s (e.g., U.S.)

Consider whether the productivity-employment relationship has changed over time

Why? Changing technologies, growing global production chains, shifting market structure,demand saturation

Add decade-specific effects to baseline equation

∆lnEict = β0 +4∑

d(t)=1

β1,d(t)∆lnLPict +4∑

d(t)=1

3∑k=0

β2+k,d(t)∆lnLPct−k,j 6=i

+αc + δt + γi + εict

where d(t) indicates decades



Internal effect more (–) and spillover less (+) in 2000sBut 2000s do not look very different from the 1980s

-0.5

0-0

.30

-0.1

00.

100.

300.

500.

700.

90

Est

imat

ed c

oeffi

cien

t


1970s1980s

1990s2000s 95% confidence interval

Model is estimated by OLS; contains country, year, and industry FE; and controls for population growth.Productivity is gross output based.



Patterns less encouraging using VA-based productivitySpillover effects declining secularly—net effect becomes weekly negative in the 2000s

-0.5

0-0

.30

-0.1

00.

100.

300.

500.

700.

90

Est

imat

ed c

oeffi

cien

t


1970s1980s

1990s2000s 95% confidence interval

Model is estimated by OLS; contains country, year, and industry FE; and controls for population growth.Productivity is value added based.


Should we worry about jobs or skills?

Outline




4 Adding it up






Even if productivity growth is neutral for employment, may benon-neutral for skill demand

Labor productivity growth may shift skill demands in two ways

1 Skill bias: Firms may differentially eliminate low-, medium-, or high-skill workers

We find that this is not quantitatively important

2 Sector bias: ‘Advancing’ sectors shrink + ‘lagging’ sectors grow

High productivity growth in manufacturing and primary industries may shift the weight ofemployment towards more skill-intensive sectors

This turns out to be quite important



Even if productivity growth is neutral for employment, may benon-neutral for skill demand

Scale predicted employment growth by industry by average shareof low-, middle-, and high- education workers

∆E qic,t=base = {E q

ic,t=base × 1(i ∈ s)× β1,s(i) ×∆lnLPict}

+{E qic,t=base ×

5∑s(i)=1

3∑k=0

β2+k,s(i) ×∆lnLPct−k,s(i),j 6=i}



Productivity growth has been strongly skill-biased 1970-2007 due toinduced sectoral shifts

05

10

Cum

ulat

ive

pred

icte

d ch

ange

(%)

1970 1980 1990 2000 2010year

High-skilled Medium-skilled Low-skilled

Based on model 5 from Table 7; prediction averaged across all 19 countries.Productivity is gross output based.



U.S. stands out for having most ‘polarized’ sectoral shifts: Reallocationtowards high- and low-skill intensive sectors

05

1015

1970 1980 1990 2000 2010year

1. France

02

46

8

1970 1980 1990 2000 2010year

2. Germany

02

46

810

1970 1980 1990 2000 2010year

3. Japan

05

1015

1970 1980 1990 2000 2010year

4. UK

02

46

1970 1980 1990 2000 2010year

5. USA

05

10

1970 1980 1990 2000 2010year


Cum

ulat

ive

pred

icte

d ∆

empl

oym

ent (

%)

Based on model 5 from Table 7. 'Mean of all others' is unweighted average across all remaining 14 countries.Productivity is gross output based.

High-skilled Medium-skilled Low-skilled


Conclusions and next steps

Outline




4 Adding it up






Is productivity growth threatening employment? Not so far...

1 Employment shrinks in advancing sectors—but spillovers offset in lagging sectors

Net effect: Productivity growth modestly contributes to rising employment-to-population—as well as rising consumption

Virtuous relationship may have weakened in the 2000s—“Robocalypse Later?”

2 Distribution of productivity growth across sectors matters

Productivity growth in services produces largest positive spillovers

Good news: Robotics or AI have potential to raise productivity in services

3 Productivity growth good for employment, skill impacts non-neutral

Challenge is not quantity of jobs

Challenge is quality of jobs available to low- and medium-skill workers



Next steps on this project (for Brookings Spring ’2018)

1 Further evidence on “is this time different?”

Extend data to 2014 (new EU Klems)

2 Are we missing intensive margin response?

Investigate hours as well as employment

3 Measuring demand using quantity and price

Study wages by skill group alongside employment

4 Digging deeper on spillovers

1 Final demand spillovers: Assess role of population aging in slackening spillovers

2 Industry linkages: Use World Input Output Tables (WIOT) to identify

5 Does a growing share of final demand accrue to non-OECD producers?

Explore international input-output linkages using WIOT


does productivity growth threaten employment ... productivity growth threaten employment? ... 1 does...

Documents