Directed Technical Change

The Review of Economic Studies Ltd.
Directed Technical Change

Author(s): Daron Acemoglu
Source: The Review of Economic Studies, Vol. 69, No. 4 (Oct., 2002), pp. 781-809
Published by: The Review of Economic Studies Ltd.
Stable URL: http://www.jstor.org/stable/1556722
Accessed: 21/05/2009 17:08
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=resl.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.
The Review of Economic Studies Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
The Review of Economic Studies.
http://www.jstor.org
0034-6527/02/00310781$02.00
Review of Economic Studies (2002) 69, 781-809

? 2002 The Review of Economic Studies Limited
Directed
Technical
Change
DARON ACEMOGLU
MIT
First version receivedMay 2001; final version accepted February2002 (Eds.)
For many problems in macroeconomics, development economics, labour economics, and
internationaltrade,whethertechnicalchange is biased towardsparticularfactorsis of centralimportance.
This paper develops a simple frameworkto analyse the forces that shape these biases. There are two
major forces affecting equilibriumbias: the price effect and the market size effect. While the former
encouragesinnovationsdirectedat scarce factors,the latterleads to technicalchange favouringabundant
factors. The elasticity of substitutionbetween differentfactors regulateshow powerfulthese effects are,
determininghow technicalchange andfactorpricesrespondto changesin relativesupplies.If the elasticity
of substitutionis sufficientlylarge, the long runrelativedemandfor a factorcan slope up.
I apply this frameworkto develop possible explanationsto the following questions: why technical
change over the past 60 years was skill biased, and why the skill bias may have acceleratedover the past
25 years?Why new technologiesintroducedduringthe late eighteenthandearly nineteenthcenturieswere
unskill biased? What is the effect of biased technical change on the income gap between rich and poor
countries?Does internationaltradeaffect the skill bias of technicalchange?What are the implicationsof
wage push for technicalchange?Why is technicalchange generallylabouraugmentingratherthancapital
augmenting?
1. INTRODUCTION
There is now a large and influentialliteratureon the determinantsof the aggregate technical
progress (see, among others, Romer (1990), Segerstrom, Anant and Dinopoulos (1990),
Grossmanand Helpman(1991), Aghion and Howitt (1992), Young(1993)). This literaturedoes
not addressquestions related to the direction and bias of technical change. In most situations,
however,technicalchange is not neutral:it benefitssome factorsof productionmore thanothers.
In this paper,I develop a simple frameworkof directedtechnicalchange to studythese biases. In
this framework,profitincentivesdeterminethe amountof researchand developmentdirectedat
differentfactorsand sectors.
To see the potentialimportanceof the biases, considera numberof examples:
(1) Figure 1 plots a measure of the relative supply of skills and a measure of the returnto
skills, the college premium.It shows that over the past 60 years, the U.S. relative supply
of skills has increasedrapidly,but therehas been no tendencyfor the returnsto college to
fall in the face of this large increasein supply-on the contrary,therehas been an increase
in the college premiumover this time period. The standardexplanationfor this pattern
is that new technologies over the post-warperiod have been skill biased. The figure also
shows thatbeginningin the late 1960's, the relativesupplyof skills increasedmore rapidly
than before, and the skill premiumincreased sharplybeginning in the late 1970's. The
standardexplanationfor this patternis an accelerationin the skill bias of technicalchange
(e.g. Autor,Kruegerand Katz (1998)).
(2) In contrast, technical change during the late eighteenth and early nineteenth centuries
appearsto have been unskill biased (skill replacing). The artisanshop was replaced by
the factory and later by interchangeableparts and the assembly line (e.g. James and
Skinner (1985), Goldin and Katz (1998)). Products previously manufacturedby skilled
781
782
REVIEWOF ECONOMICSTUDIES
- 08
o College wage premium

A Rel. supply of college skills
0.6(-
- 0.6
0-5-0.4
0'0.4 -
- -02
0-3 -
-20
39
49
59
69
Year
79
89
96
Relative Supplyof College Skills and College Premium

FIGURE 1
The behaviourof the (log) college premiumand relativesupply of college skills (weeks workedby college equivalents
divided by weeks workedby noncollege equivalents)between 1939 and 1996. Data from MarchCPSs and 1940, 1950
and 1960 censuses
artisansstartedto be producedin factoriesby workerswith relativelyfew skills, and many

previously complex tasks were simplified,reducing the demandfor skilled workers(e.g.
Mokyr(1990, p. 137)).
(3) Over the past 150 years of growth, the prices of the two key factors, capital and labour,
have behavedvery differently.While both in the U.S. and in otherWesterneconomies, the
wage ratehas increasedsteadily,the rentalrateof capitalhas been approximatelyconstant.
This patternindicatesthatmost of the new technologies are labour augmenting.
(4) Beginning in the late 1960's and the early 1970's, both unemploymentand the share of
labourin nationalincome increasedrapidlyin a numberof continentalEuropeancountries.
Duringthe 1980's, unemploymentcontinuedto increase,butthe laboursharestarteda steep
decline, andin manycountries,endedup below its initiallevel. Blanchard(1997) interprets
the firstphase as the responseof these economies to a wage push, and the second phase as
a possible consequenceof capital-biasedtechnicalchange.
These examples document a variety of importantmacroeconomic issues where biased
technical change plays a key role. They also pose a numberof questions: why has technical
change been skill biased over the past 60 years? Why was technical change biased in favourof
unskilledlabourand against skilled artisansduringthe nineteenthcentury?Why has there been
an accelerationin the skill bias of technical change duringthe past 25 years? Why is much of
technologicalprogresslabouraugmentingratherthancapitalaugmenting?Why was thererapid
capital-biasedtechnicalchangein continentalEuropefollowing the wage push by workersduring
the 1970's?
These questionsrequirea frameworkwhere the equilibriumbias of technicalchange can be
studied.The frameworkI presentfor this purposegeneralizesthe existing endogenoustechnical
change models to allow for technical change to be directed towards different factors: firms
ACEMOGLU
DIRECTEDTECHNICALCHANGE
783
can invest resources to develop technologies that complement a particularfactor. The relative
profitabilitiesof the differenttypes of technologies determinethe directionof technicalchange.
I show thatthereare two competingforces determiningthe relativeprofitabilityof different
types of innovation:(i) the price effect, which createsincentivesto develop technologies used in
the productionof more expensive goods (or equivalently,technologies using more expensive
factors); (ii) the market size effect, which encourages the development of technologies that
have a largermarket,more specifically, technologies that use the more abundantfactor.These
two effects are competingbecause, while the price effect implies that there will be more rapid
technological improvementsfavouring scarce factors, the market size effect creates a force
towards innovations complementing the abundantfactor.1 I will show that the elasticity of
substitutionbetween the factors determinesthe relative strengthsof these two effects. When
the elasticity of substitutionis low, scarce factorscommandhigherprices, and the price effect is
relativelymore powerful.
The first majorresult of this frameworkis a "weak induced-biashypothesis":irrespective
of the elasticity of substitutionbetween factors(as long as it is not equal to 1), an increasein the
relativeabundanceof a factorcreatessome amountof technicalchangebiasedtowardsthatfactor.
The second majorresult is a "stronginduced-biashypothesis",and states that if the elasticity of
substitutionis sufficientlylarge (in particular,greaterthan a certainthresholdbetween 1 and 2),
the inducedbias in technologycan overcomethe usual substitutioneffect andincreasethe relative
rewardto the factorthathas become more abundant.Thatis, directedtechnicalchange can make
the long-runrelativedemandcurve slope up. The long runrelativedemandcurvemay be upward
sloping in this set-upbecauseof the underlying"increasingreturnsto scale"in the R&D process:
a new machine,once invented,can be used by many workers.2
Figure 2 illustratesthese results diagrammatically.The relatively steep downward-sloping
lines show the constant-technologyrelativedemandcurves.The economy startsat pointA. In the
absenceof endogenoustechnicalchange,the increasein the supplyshownin the figuremoves the
economy along the constant-technologydemandto point B. The first result of this framework,
the weak induced-biashypothesis, implies that, as long as the elasticity of substitutionbetween
factors is differentfrom 1, the increase in the supply will induce biased technical change and
shift the constant-technologydemand curve out. The economy will thereforesettle to a point
like C. In other words, the (long-run)endogenous-technologydemandcurve will be flatterthan
the constant-technologycurve. The second result, the strong induced-biashypothesis, implies
that the inducedbias in technology can be powerfulenough to createa sufficientlylarge shift in
the constant-technologydemandcurve and take the economy to a point like D. In this case, the
endogenous-technologydemandcurve of the economy is upwardsloping and the relativereward
of the factorthathas become more abundantincreases.
After outlining the general forces shaping the direction of technical change and the main
results, I returnto a number of applications of this framework.I discuss: (1) why technical
change over the past 60 years was skill-biased, and why skill-biased technical change may
have acceleratedover the past 25 years. Also why new technologies introducedduringthe late
eighteenthand early nineteenthcenturieswere labourbiased. (2) Why biased technical change
is likely to increasethe income gap between rich and poor countries.(3) Why internationaltrade
may induce skill-biasedtechnicalchange. (4) Underwhatcircumstanceslabourscarcitywill spur
faster technologicalprogressas suggested by Rothbarth(1946) and Habakkuk(1962). (5) Why
technical change tends to be generally labour augmenting rather than capital augmenting.
1. Anotherimportantdeterminantof the directionof technicalchange is the form of the "innovationpossibilities
frontier"-i.e. how the relative costs of innovation are affected as technologies change. I discuss the impact of the
innovationpossibilities frontieron the directionof technicalchange in Section 4.
2. This is relatedto the nonrivalryin the use of ideas emphasizedby Romer(1990).
REVIEW OF ECONOMIC STUDIES
784
WZ/WL
auippy
auppty sriji
\\ \I
\ \\
Endogenous
!
i,
^ TechnologyDemand
'D\< \ \ Upward-slopingcase
j
N\
\
C
]
Endogenous
TechnologyDemand
ConstantTechnology
B Demand Curves
Z/L
FIGURE 2
Constanttechnology and endogenous technology relative demandcurves. Constanttechnology: A -+ B. Endogenous

technology:A -+ C. Endogenoustechnology with powerfulmarketsize effect: A -+ D
(6) Why a largewage push, as in continentalEuropeduringthe 1970's, may cause capital-biased

technicalchange and affect the factordistributionof income.
This list is by no means exhaustive, and there is much researchto be done to understand
the implications of biased technical change and the determinantsof equilibriumbias of new
technologies. It is partof my aim in this paperto stress the importanceof thinkingaboutbiased
technical change, and to provide a set of tools that are likely to be useful for futureresearchon
these biases.3
Although there is relatively little currentresearch on biased technical change, an earlier
literaturewas devoted to studyingrelatedissues. It was probablyHicks in The Theoryof Wages
(1932) who first discussed the determinantsof equilibriumbias.4 He wrote: "A change in the
relative prices of the factors of productionis itself a spur to invention, and to invention of
a particularkind-directed to economizing the use of a factor which has become relatively
expensive."(pp. 124-125). Hicks' reasoning,thattechnicalchange would attemptto economize
on the more expensive factor, was criticized by Salter (1966) who pointed out that there was
no particularreason for saving on the more expensive factor-firms would welcome all cost
reductions.Moreover,the concept of "moreexpensive factor"did not make much sense, since
all factorswere supposedto be paid theirmarginalproduct.
These questions were revived by the "inducedinnovation"literature.An importantpaper
by Kennedy (1964) introducedthe concept of "innovationpossibilities frontier",capturingthe
trade-offbetween differenttypes of innovations,and arguedthat it is the form of this frontierratherthan the shape of a given neoclassical productionfunction-that determinesthe factor
distributionof income. Kennedy,furthermore,arguedthat induced innovationswould push the
economy to an equilibriumwith a constant relative factor share (see also Samuelson (1965),
Drandakisand Phelps (1965)). Aroundthe same time, Habakkuk(1962) put forththe thesis that
labour scarcity,by increasing wages, induced firms to search for labour saving inventions and
3. The framework here focuses on one type of biased technical change: that which increases the relative
productivityof one factor permanently.Alternatively,as emphasizedby Nelson and Phelps (1966), Schultz (1975) and
especially Galor and Maov (2000), technological progressmay be biased towardsone of the factors, skilled labour,in
the shortrun,because higherability and skilled workersmay be betterat adaptingto a changingenvironment.
4. Marx also touched on these issues. He argued that labour scarcity-the exhaustion of the reserve army of
labour-would induce the capitalistto substitutemachineryfor labourand spurgrowth.See for example the discussion
in Habakkuk(1962, p. 44).
ACEMOGLU
785
spurredtechnological progress (see also Rothbarth(1946)). This literaturewas also criticized

for lack of micro-foundations,however.First, with specificationsas in Kennedy,the production
function at the firm level exhibited increasing returnsto scale because, in addition to factor
quantities, firms could choose "technology quantities".Second, as pointed out by Nordhaus
(1973), it was not clear who undertookthe R&D activities and how they were financed and
priced.These shortcomingsreducedthe interestin this literature,and therewas little researchfor
almost 30 years.
The analysis here, instead, starts from the explicit micro-foundationslaid out by the
endogenous technical change models. In addition to providing an equilibriumframeworkfor
analysing these issues, I demonstratethe presence of the market size effect, which did not
feature in the earlier literature(but see Schmookler (1966)). More explicitly, the frameworkI
present here synthesizes my previous work in Acemoglu (1998, 1999a,b) and Acemoglu and
Zilibotti (2001), as well as work by Kiley (1999) (see also Lloyd-Ellis (1999) and Galor and
Maov (2000), for differentperspectives).The resultsin these papersshow that whethertechnical
change results from quality improvements,expandingvarietyof products,or expandingvariety
of machinetypes is not essential.For this reason,I choose one of the specificationsandhighlight
the modellingchoice thatturnsout to be moreimportant-the form of the innovationpossibilities
frontier.
The rest of the paperis organizedas follows. In the next section, I define some of the terms
thatwill be used throughoutthe paperand clarify the distinctionbetween factoraugmentingand
factor-biasedtechnical change. In this section, I also give a brief overview of the main results.
In Section 3, I introducethe basic frameworkthat determinesthe demand for innovationand
I highlight the price and market size effects on the direction of technical change. Section 4
introducesthe innovationpossibilities frontierand shows how different forms of this frontier
affect the equilibriumbias of technology. Sections 5 and 6 apply the frameworkdeveloped in
Sections 3 and4 to a varietyof situationswherebiased technicalchange appearsto be important.
Section 7 concludes.
2. FACTOR-AUGMENTING,
FACTOR-BIASEDTECHNICALCHANGEAND AN
OVERVIEW
Consideran aggregateproductionfunction,F(L, Z, A), with two inputs,L, labour,and Z, which
could be capital, skilled labour or land. A is a technology index. Without loss of generality
> 0, so a greater level of A correspondsto "bettertechnology" or to
imagine that 3F/la
"technologicalprogress".Technical change is L-augmentingif the productionfunction takes
the more special form F(AL, Z). Z-augmentingtechnicalchange is definedsimilarly.Technical
change is L-biased, on the otherhand,if
F/aL
OaF/aZ
0,
aA
thatis, if technicalchange increasesthe marginalproductof L more thanthatof Z.
To clarifythe differencebetweenthese two concepts,considerthe more specializedconstant
elasticity of substitution(CES) productionfunction
+ (1 - y)(AzZ) a-]-,
Y = [y(ALL)
where AL and Az aretwo separatetechnologyterms,y E (0, 1) is a distributionparameterwhich
determineshow importantthe two factors are, and ar (0, oo) is the elasticity of substitution
between the two factors.When a = oc, the two factorsareperfectsubstitutes,andthe production
functionis linear.When C = 1, the productionfunctionis Cobb-Douglas, andwhen a = 0, there
786
is no substitutionbetweenthe two factors,andthe productionfunctionis Leontieff.Whena > 1,

I referto the factorsas gross substitutes,andwhen a < 1, I referto them as gross complements.5
By construction,AL is L-augmentingand Az is Z-augmenting.I will also sometimes refer to
AL as labourcomplementary,and to Az as Z-complementary.
Whethertechnicalchange is labourbiased or Z-biased, on the other hand, dependson the
elasticity of substitution.To see this, calculatethe relativemarginalproductof the two factors:
1
alI
MPz
MPL
l-y
(A
Y (AL
(1)
The relativemarginalproductof Z is decreasingin the relative abundanceof Z, Z/L. This is

the usual substitutioneffect, leading to a downwardsloping relative demandcurve. The effect
of Az on this relative marginalproductdepends on a, however.If a > 1, an increase in Az
(relativeto AL) increases the relative marginalproductof Z. When cr < 1, an increase in Az
reducesthe relativemarginalproductof Z. Therefore,when the two factorsare gross substitutes,
Z-augmenting(Z-complementary)technicalchange is also Z-biased. In contrast,when the two
factorsaregross complements,then Z-augmentingtechnicalchangeis L-biased.Naturally,when
a = 1, we arein the Cobb-Douglas case, andneithera changein Az nor in AL is biasedtowards
any of the factors.
The intuitionfor why, when a < 1, Z-augmentingtechnicalchange is L-biased is simple:
with gross complementarity,an increase in the productivityof Z increases the demandfor the
other factor,labour,by more than the demandfor Z, effectively creating "excess demand"for
labour.As a result, the marginalproductof labourincreasesby more than the marginalproduct
of Z.
Now to obtainan overviewof the resultsthatwill follow, imaginea set-upwhere AL and Az
are determinedendogenously from the type and quality of machines supplied by "technology
monopolists".One of the major results of the more detailed analysis below will be that the
profitabilityof developingnew Z-complementarymachinesrelativeto the profitabilityof labourcomplementarymachineswill be proportionalto (see equation(17))
(2)
(AZ) a (Z) a
The basic premise of the approach here is that profit incentives determine what types of
innovationswill be developed. So when (2) is high, Az will increaserelativeto AL. Inspection
of (2) shows that when the two factors are gross substitutes(a > 1), an increase in Z/L will
increase the relative profitabilityof inventing Z-complementarytechnologies. To equilibrate
innovationincentives, AZ/AL has to rise, reducing(2) back to its original level. Intuitively,in
this case, of the two forces discussed in the introduction,the marketsize effect is more powerful
thanthe priceeffect, so technicalchangeis directedtowardsthe moreabundantfactor.In contrast,
when the two factors are gross complements(a < 1), an increase in Z/L will lead to a fall in
AZ/AL. However,recall that when a < 1, a lower Az/AL correspondsto Z-biased technical
change. So in this case an increasein Z/L reducesthe relativephysical productivityof factor Z,
but increasesits relativevalue of marginalproduct,because of relativeprice changes. Therefore,
in both cases, an increasein the relativeabundanceof Z causes Z-biased technical change. We
will also see thatif a is sufficientlylarge,this inducedbiasedtechnicalchangecan be so powerful
thatthe increasein the relativeabundanceof a factormay in fact increaseits relativereward-i.e.
the long runrelativedemandcurve for a factormay be upwardsloping.
5. I use this terminologybecause the demandfor Z increases in responseto an increase in the price of L, WL,
holding its price, wz, and the quantityof L constantif and only if a > 1, and vice versa.
ACEMOGLU
787
3. THE DEMAND SIDE

I now develop the basic frameworkfor analysingthe determinantsof the factorbias of technical
change, first focusing on the demand for new technology (innovation).The next section then
introduces"theinnovationpossibilities frontier"and discusses the supply side of innovations.
3.1. The environment
Consideran economy thatadmitsa representativeconsumerwith the usual constantrelativerisk
aversion(CRRA) preferences
- 1
C(t)l
(3)
e-ptdt
where p is the rate of time preferenceand 0 is the coefficient of relative risk aversion (or the
intertemporalelasticity of substitution).I suppressthe time argumentsto simplify the notation,
and I will do so throughoutas long as this causes no confusion. The budget constraintof the
consumeris
C+
+ ___ YYL + (1 Y)YzE
I+ R
E
-1(4)
(4)
where I denotes investment,and R is total R&D expenditure.I also impose the usual no-Ponzi
game condition, requiring the lifetime budget constraint of the representativeconsumer to
be satisfied. The productionfunction in (4) implies that consumption, investment and R&D
expenditurecome out of an outputaggregateproducedfromtwo goods, YLand Yz, with elasticity
of substitutione, and y is a distributionparameterwhich determineshow importantthe two
goods are in aggregateproduction.Of these two goods, YL is (unskilled)labourintensive,while
Yz uses anotherfactor, Z, intensively.In this section and the next, I will not be specific about
what this factoris, but the readermay want to thinkof it as skilled labourfor concreteness.
These two goods have the following productionfunctions6
YL=
XL(j) -dJ )L',
(5)
xz(j)l-fidj)Zf
(6)
and
Yz = 1
where ,f E (0, 1), and L and Z are the total quantitiesof the two factors,assumedto be supplied
inelastically for now. The labour-intensivegood is thereforeproducedfrom labourand a range
intermediatesor machines,the xL's. For simplicity,I will refer to the
of labour-complementary
x's as "machines".The range of machines that can be used with labouris denoted by NL. The
productionfunction for the other good, (6), uses Z-complementarymachines and is explained
similarly.Notice that given NL and Nz, the productionfunctions (5) and (6) exhibit constant
returnsto scale. There will be aggregate increasing returns,however, when NL and Nz are
endogenized.
I assume that machines in both sectors are supplied by "technologymonopolists".In this
section, I take NL and Nz as given, and in the next section, I analyse the innovationdecisions of
these monopolists (the supply of innovations)to determineNL and Nz. Each monopolist sets a
rentalprice XL(I) or xz(j) for the machine it supplies to the market.For simplicity,I assume
6. The firmlevel productionfunctionsare also assumedto exhibit constantreturnsto scale, so thereis no loss of
generalityin focusing on the aggregateproductionfunctions.
788
that all machinesdepreciatefully afteruse, and that the marginalcost of productionis the same
for all machinesand equal to p in termsof the final good.7
The importantpoint to bear in mind is that the set of machines used in the productionof
the two goods are different,allowing technicalchange to be biased. The rangeof machines, NL
and Nz, determineaggregateproductivity,while NZ/NL determinesthe relativeproductivityof
factor Z.
3.2. Equilibrium
An equilibrium(given NL and Nz) is a set of pricesfor machines,XL(j) or xz (j), thatmaximize
the profitsof technology monopolists,machine demandsfrom the two sectors, XL(j) or xz(j),
thatmaximize producers'profits,and factor and productprices, WL, wz, PL, and pz, thatclear
markets.I now characterizethis equilibriumand show that it is unique.8The levels of NL and
Nz will be determinedin the next section once I introducethe innovationpossibilitiesfrontierof
this economy.
The productmarketsfor the two goods are competitive,so marketclearingimplies thattheir
relativeprice, p, has to satisfy
p -
Y_ )
Pz = 1-y
PL
(7)
YL
The greaterthe supply of Yz relativeto YL,the lower is its relativeprice, p. The responseof the
relativeprice to the relativesupply dependson the elasticity of substitution,e.
I choose the price of the final good as the numeraire,so
(8)
]1
[yePL + (1-y) 8p
=1.
Since product markets are competitive, firms in the labour intensive sector solve the
following maximizationproblem
maxL,{xL(j)}PLYL - WLL -
{NL
(9)
XL(j)XL(j)dj,
taking the price of their product,PL, and the rentalprices of the machines, denoted by XL(j),
as well as the range of machines, NL, as given. The maximizationproblem facing firms in
the Z-intensive sector is similar. The first-orderconditions for these problems give machine
demandsas
j (PL
( XLJ)
(and
pZ
l/
Z.
(10)
XL(J)/
\xz(j)
These equationsimply that the desired amountof machine use is increasingin the price of the
product, PL or PZ, and in the firm's employment, L or Z, and is decreasing in the price of
the machine, XL() or Xz(j). Intuitively,a greaterprice for the productincreases the value of
the marginalproductof all factors, including that of machines,encouragingfirms to rent more
xL(j)
xz(j) =(
7. Slow depreciationof machinesdoes not change the BGP equilibrium,and only affects the speed of transitional
dynamics.For example, if machinesdepreciateat some exponentialrate 8, monopolistswill producethe requiredstock
of machines after the discovery of the new variety,and then will replace the machines that depreciate.The rentalprice
will then be a mark-upover the opportunitycost of machinesratherthanover the productioncost. To keep notationto a
minimum,I do not considerthe case of slow depreciation.
8. In this paper,I only characterizeequilibriumallocations.The social planner'ssolutiondiffersfrom equilibrium
allocationsbecauseof monopolydistortions.However,exactly the same equations,in particular,equations(21) and (26),
determinethe bias of technology in the social planner'sallocation.Details availableupon request.
ACEMOGLU
789
machines. A greaterlevel of employment,on the other hand, implies more workersto use the
machines,raisingdemand.Finally,becausethe demandcurvefor machinesis downwardsloping,
a highercost implies lower demand.
Next, the first-orderconditionwith respectto L and Z gives the factorrewardsas
W=
PL(fNL=
wz =
PL
XL()
N
pfp(J
and
dj)L
1
xz(j)1-dj
(11)
My interest is with the determinantsof the direction of technical change. As discussed

above, profit-maximizingfirmswill generatemore innovationsin responseto greaterprofits,so
the first step is to look at the profitsof the technology monopolists.Recall that each monopolist
faces a marginalcost of producingmachines equal to Gr.Therefore,the profitsof a monopolist
supplying labour-intensivemachine j can be writtenas rL() = (XL (j) - )xL (j) Since the
demand curve for machines facing the monopolist, (10), is iso-elastic, the profit-maximizing
price will be a constant mark-upover marginalcost: XL(j) = 1I*. To simplify the algebra,
I normalizethe marginalcost to ~ = 1 - f.9 This implies thatin equilibriumall machineprices
will be given by XL(j) = Xz(j) = 1. Using these prices andthe machinedemandsabove,profits
of technology monopolistsare obtainedas
7rL= lp/
and
7rz = fp/
Z.
(12)
What is relevant for the monopolists is not the instantaneous profits, but the net present
discountedvalue of profits.These net presentdiscountedvalues can be expressedusing standard
dynamicprogrammingequations:
rVL-VL = TL
and
rVz - Vz = 7rz,
(13)
where r is the interestrate, which is potentiallytime varying. The equationsrelate the present
discountedvalue of futureprofits, V, to the flow of profits,7r.The V termtakes care of the fact
thatfutureprofitsmay not equal currentprofits,for example because prices are changing.
To gain intuition,let us startwith a steady state where the V termsare 0 (i.e. profitsand the
interestrate are constantin the future).Then,
L
Z
and
.
Vz=
(14)
r
r
The greater Vz is relative to VL, the greaterare the incentives to develop Z-complementary
machines, Nz, ratherthan labour-complementary
machines, NL. Inspectionof (14) reveals two
forces determiningthe directionof technicalchange:
L=
(1) The price effect: there will be a greaterincentive to invent technologies producingmore
expensive goods, as shown by the fact that Vz and VL are increasingin pz and PL.
(2) The marketsize effect: a largermarketfor the technology leads to more innovation.Since
the marketfor a technology consists of the workers who use it, the market size effect
encouragesinnovationfor the more abundantfactor.This can be seen from the fact that Vz
and VL are increasingin Z and L.
Notice that an increase in the relative factor supply, Z/L, will create both a marketsize
effect and a price effect. The latter simply follows from the fact that an increase in Z/L will
9. This is withoutloss of any generality,since I am not interestedin comparativestatics with respectto f.
790
reduce the relativeprice p

pz/PL. Equilibriumbias in technical change-whether technical
change will favourrelatively scarce or abundantfactors-is determinedby these two opposing
forces.10An additionaldeterminantof equilibriumbias is the form of the innovationpossibilities
frontier,which will be introducedin the next section.
Next, it is useful to investigate the strengthof the price and marketsize effects in more
detail. To do this, let us substitute(10) into the productionfunctions,(5) and (6). This gives
-
YL = 1
and
NLL
YZ= 1
NzZ.
(15)
Substitutingthese into (7) and using some algebra,we obtainthe relativeprice of the two goods
as a functionof relativetechnology,NZ/NL, and the relativefactorsupply,Z/L:
(16)
p (11Y)3e/(NzZ)-/
where e is the elasticity of substitutionbetween the two goods, YL and Yz, while a is the
(derived)elasticity of substitutionbetween the two factors, Z and L, definedas
a =_ ?-(-
1)(l--r).
Note thata > 1 if and only if e > 1-that is, the two factorsaregross substitutesonly if the two
goods are gross substitutes.
Now using (14) and (16) and still assuming steady state, we can write the relative
profitabilityof creatingnew Z-complementarymachinesas
PI
VL
price
effect
(1-y
(Nz=
Y
NL
u(Z
L
(17)
market size
effect
This expressionshows thatthe relativeprofitabilityof the two types of innovationaredetermined

by the price and marketsize effects. An increasein the relativefactorsupply,Z/L, will increase
Vz/ VL as long as the elasticity of substitutionbetween factors, a, is greater than 1 and it
will reduce Vz/ VL if a < 1. Therefore,the elasticity of substitutionbetween the two factors
(and indirectlybetween the two goods) regulateswhetherthe price effect dominatesthe market
size effect so that there are greaterincentives to improve the (physical) productivityof scarce
factors, or whetherthe marketsize effect dominates,creatinggreaterincentives to improvethe
productivityof abundantfactors. When the factors are gross substitutes,the marketsize effect
dominates.And when they are gross complements,the price effect dominates.
Finally, to see anotherimportantrole of the elasticity of substitution,consider the relative
factorrewards,WZ/WL.Using equations(10) and (11), and then substitutingfor (16), we have
wZ
WL
=1/Nz
p/
(Z)
(1Y)(NZ)
NL
NL
(18)
(!8)
First, note that the relative factor reward,WZ/WL, is decreasing in the relative factor supply,
Z/L. This is simply the usual substitutioneffect: the more abundantfactoris substitutedfor the
less abundantone, and has a lower marginalproduct.
10. This discussion emphasizes the role of price and market size effects, while factor prices do not feature
in (14). The early inducedinnovationsliterature,instead,arguedthatinnovationswould be directedat "moreexpensive"
factors(e.g. Hicks (1932), Fellner (1961), Habakkuk(1962)). Using equations(10) and (11), we can re-express(14) as
which show thatan equivalentway of looking at the incentives
and Vz = (1 - P)wZ/rNz,
VL = (1 - 5)LL/rNL
to develop new technologiesis to emphasizefactorcosts, wL and wz (as well as marketsizes, L and Z). As conjectured
by the inducedinnovationsliterature,therewill be more innovationdirectedat factorsthatare more expensive.
ACEMOGLU
791
We also see from equation(18) thatthe same combinationof parameters,(a - 1)/a, which
determineswhether innovation for more abundantfactors is more profitable,also determines
whethera greaterNz/NL increases WZ/WL: when a > 1, it does, but when a < 1, it has the
opposite effect and reduces the rewardto Z relative to labour.This is because Nz/NL is the
relativephysical productivityof the two factors,not theirrelativevalue of marginalproduct.The
latteralso dependson the elasticity of substitutionbetween the two factors(recallequation(1) in
Section 2). Therefore,as in Section 2, Z-biased technical change correspondsto an increase in
(Nz/NL)('-1)/?, not simply to an increase in Nz/NL. Consequently,when a < 1, a decrease
in Nz/NL correspondsto Z-biased technicalchange. This feature,thatthe relationshipbetween
relativephysical productivityand the value of marginalproductis reversedwhen the two goods
(factors)are gross complements,will play an importantrole in the discussion below.
4. THE SUPPLYOF INNOVATIONS:THE INNOVATIONPOSSIBILITIESFRONTIER

The previous section outlinedhow the productionside of the economy determinesthe returnto
differenttypes of innovation-the demandfor innovation.The other side of this equationis the
cost of differentinnovations,or using the term introducedby Kennedy (1964), the "innovation
possibilities frontier".The analysis in this section will reveal that in additionto the elasticity of
substitution,the degree of state dependencein the innovationpossibilities frontierwill have an
importanteffect on the directionof technical change. The degree of state dependencerelates to
how futurerelativecosts of innovationare affectedby the currentcompositionof R&D (current
"state"of R&D). I referto the innovationpossibilities frontieras "statedependent"when current
R&D directedat factor Z reducesthe relativecost of Z-complementaryR&D in the future.
I follow the endogenous growth literaturein limiting attentionto innovationpossibilities
frontiersthat allow steady growthin the long run. Sustainedgrowthrequiresthat the innovation
possibilities frontiertakes one of two forms. The first, which Rivera-Batizand Romer (1991)
referto as the lab equipmentspecification,involves only the final good being used in generating
new innovations. With this specification, steady-state growth is generated with an intuition
similar to the growth model of Rebelo (1991) whereby the key accumulation equation is
linear and does not employ the scarce (non-accumulated)factors, such as labour.The second
possibility is the knowledge-basedR&D specificationof Rivera-Batizand Romer (1991) where
spillovers from past researchto currentproductivityare necessaryto sustaingrowth.It is useful
to distinguish between these two formulationsbecause they naturallycorrespondto different
degrees of state dependencein R&D.
4.1. The directionof technical change with the lab equipmentmodel

Considerthe following productionfunctionfor new machinevarieties
NL = ~rLRL
and
Nz = irzRz,
(19)
where RL is spending on R&D for the labour-intensivegood (in terms of final good), and Rz
is R&D spending for the Z-intensive good. The parameters?rLand r/z allow the costs of these
two types of innovationsto differ. The innovationproductionfunctions in (19) imply that one
unit of final good spent for R&D directedat labour-complementary
machines will generater)L
new varietiesof labour-complementary
machines (and similarly,one unit of final good directed
at Z-complementarymachines will generate r/z new machine types). A firm that discovers a
new machine varietyreceives a perfectly enforced patenton this machine and becomes its sole
supplier-the technology monopolist.Notice that with the specificationin (19), there is no state
792
dependence: (aNz/aRz)/(aNL/aRL)
= rlz/IlL is constant irrespective of the levels of NL
and Nz.
I startwith a balancedgrowthpath (BGP)-or steady-stateequilibrium-where the prices
PL and pz are constant,and NL and Nz grow at the same rate. This implies thatthe V termsin
equation(13) are 0, and Vz/ VLis constant.Moreover,this ratiohas to be equal to the (inverse)
ratio of O's from (19) so that technology monopolists are willing to innovate for both sectors.
This immediatelyimplies the following "technologymarketclearing"condition:11
rLTrL= I1Z7z.
(20)
This condition requiresthat it is equally profitableto invest money to invent labour- and Zcomplementarymachines,so thatalong the BGP NL and Nz can both grow.Definingr/ _ r/z/r/L
to simplify notationand using (12) and (16), this technology marketclearing condition can be
solved for
NL
)Y
(21)
where recall that s is the elasticity of substitution between the goods and r
e - (e-
1)(1 -
is the elasticity of substitutionbetween the factors.Equation(21) shows that with the direction
of technical change endogenized,the relativebias of technology, NZ/NL, is determinedby the
relativefactorsupply and the elasticity of substitutionbetween the two factors.
If a < 1, i.e. if the two factorsare gross substitutes,an increasein Z/L will raise NZ/NL,
hence the physicalproductivityof the abundantfactortends to be higher.Moreover,since a > 1,
a higher level of NZ/NL correspondsto Z-biased technicalchange. Therefore,technology will
be endogenouslybiased in favourof the more abundantfactor.
When a < 1, i.e. when the two factors are gross complements,equation(21) implies that
NZ/NL is lower when Z/L is higher.Nevertheless,because a < 1, this lower relativephysical
productivitytranslatesinto a higher value of marginalproduct.Therefore,even when a < 1,
technology will be endogenously biased in favour of the more abundantfactor, giving us the
"weakinduced-biashypothesis".12
Next considerfactorprices. Using equations(18) and (21), we obtain
z =
WL
-l(1-Y)(Z)-2
y (
L2
(22)
Comparingthis equationto (18), which specified the relativefactorprices as a functionof

relative supplies and technology, we see that the response of relativefactorrewardsto changes
in relativesupplyis always more elastic in (22): cr - 2 > -I/a. This is not surprisinggiven the
LeChatelierprinciple,which states thatdemandcurves become more elastic when otherfactors
adjust.Here, the "otherfactors"correspondto the machinevarietiesNL and Nz.
The more surprisingresult is the "stronginduced-biashypothesis":that if Cais sufficiently
large, in particularif a > 2, the relationshipbetween relativefactor supplies and relativefactor
rewardscan be upwardsloping. Intuitively,with exogenous technology when a factorbecomes
more abundant,its relativerewardfalls: i.e. given NZ/NL, WZ/WLis unambiguouslydecreasing
in Z/L due to the usual substitutioneffect (see equation (18)). Yet, because technology is
endogenously biased towardsmore abundantfactors, the overall effect of factor abundanceon
factorrewardsis ambiguous.
11. This condition assumes no comer solution (where R&D would be directedonly towardsone of the factors),
but does not rule out the economy convergingto a long-runequilibriumwith only one type of innovation.
12. The exception to this result is when cr = 1, i.e. when the productionfunctionis Cobb-Douglas. In this case,
the elasticity of substitutionis equalto 1, andtechnicalchange is neverbiasedtowardsone of the factors(unless different
types of technicalchange affect y differently).
ACEMOGLU
793
Next define the relativefactor sharesas

sz
SL
wzZ
=
-T=r
WL
WLL
I-i(-^V^V)f1
y
(3
(23)
which implies that, with endogenous technology, the share of a factor in GDP will increase in
the relativeabundanceof thatfactoras long as a > 1.
For completeness, it is also useful to determinethe long-rungrowthrate of this economy.
To do this, note that the maximizationof (3) implies gc = 0- (r - p), where gc is the growth
rate of consumptionand recall that r is the interestrate (though there is no capital, agents are
delayingconsumptionby investingin R&D as a functionof the interestrate).In BGP,this growth
rate will also be equal to the growth rate of output, g. So r = Og + p. Next, using (14), the
free-entrycondition for the technology monopolists working to invent labour-complementary
machines is obtainedas rlLVL= 1. In steady state, this condition implies T/LfpL/L/r = 1.
Now using (8), (16), and (21) to substitutefor PL, we obtain13
g = 0- (P[(1 - y)E1zZ)a-1
+ ye(TLL)-1]
_ p).
Finally, it is useful to briefly look at the behaviourof the economy outside the BGP. It
is straightforwardto verify that outside the BGP, there will only be one type of innovation.14
If ?TVZ/VL > 1, there will only be firms creating new Z-complementarymachines, and if
?rVZ/VL < 1, technology monopolists will only undertakeR&D for labour-complementary
machines. Moreover, VZ/VL is decreasingin Nz/NL (recall equation(17)). This implies that
the transitionaldynamics of the system are stable and will returnto the BGP. When NZ/NL
is higher than in (21), there will only be labour-augmentingtechnical change until the system
returnsback to balancedgrowth,and vice versa when NZ/NL is too low.
4.2. The directionof technicalchange with knowledge-basedR&D
With the lab equipment model of the previous subsection, there is no state dependence. I
now discuss an alternativespecification which allows for potential state dependence. In the
lab equipmentspecificationthere are no scarce factors that enter the accumulationequationof
the economy. If there are scarce factors used for R&D, then growth cannot be maintainedby
increasingthe amountof these factorsused for R&D. So for sustainedgrowth,these factorsneed
to become more and more productiveover time. This is the essence of the knowledge-based
R&D specification,whereby spillovers imply that currentresearchers"standon the shoulderof
giants",ensuringthatthe marginalproductivityof researchdoes not decline. Here for simplicity,
I assume that R&D is carriedout by scientists, and thereis a constantsupply of scientists equal
to S.15If therewere only one sector,the knowledge-basedR&D specificationwould requirethat
N/N oc S (proportionalto S). With two sectors, instead,thereis a varietyof specificationswith
differentdegreesof statedependence,becauseproductivityin each sectorcan dependon the state
of knowledge in both sectors. A flexible formulationis
]L =
LN^
SL
and
Nz = tiZNNL
^Z
SZ
(24)
where a < 1 measures the degree of state dependence. When 8 = 0, there is no state
= Trirrespectiveof the levels of NL and Nz-because
dependence-(QNz/dSz)/I(NL/aSL)
13. The no-Ponzi game conditionrequiresthat (1 - 0)g < p.
14. See Proposition 1 in Acemoglu and Zilibotti (2001) for a formal proof that only one type of innovationwill
take place outside the BGP and for a proof of global stabilityin a relatedmodel. The proof here is identical.
15. The resultsgeneralizeto the case where the R&D sectoruses labour(or when Z is takento be skilled labour,a
combinationof skilled and unskilledlabour),but the analysis of the dynamicsbecomes substantiallymore complicated.
794
both NL and Nz create spillovers for currentresearchin both sectors. In this case, the results
are very similar to those in the previous subsection. In contrast, when 8 = 1, there is an
= t7NZ/NL, so an increase in
extreme amountof state dependence:(aNz/aRZ)/(aNL/aSL)
innovationscheaper,but has no effect on the cost
NL today makes futurelabour-complementary
of Z-complementaryinnovations.
The conditionfor technology marketclearingin BGP now changes to
1LNN7IL = ?zNAz,z.
(25)
When 8 = 0, this condition is identical to (20) from the previous subsection. Solving
condition (25) together with equations (12) and (16), we obtain the equilibrium relative
technology as
Y)
-NL=
)i
(26)
L
=
y
Now the relationshipbetween the relative factor supplies and relative physical productivities
depends on 8. This is intuitive:as long as 8 > 0, an increase in Nz reduces the relative cost
of Z-complementaryinnovations,so to restoretechnology marketequilibrium,7rz needs to fall
relativeto TL, and NZ/NL needs to increasemore. Substituting(26) into (18) gives
NL
(1-)E
Wz
WZ=
WL
(I=
_I
(-)
()a
y)
oa-2+8
(
(27)
(27)
L'
Relativefactorsharesare then obtainedas

SZ
-
WZZ
_=
a_-l
1-8
I 1-y
(1- )f
(-<-zs)
Z'
a- 1+5- Ia
I-Sa
(28)
y
rL
It can be verifiedthat when 8 = 0, when there is no state dependencein R&D, these equations
are identicalto theircounterpartsin the previoussubsection.
The growthrate of this economy is determinedby the numberof scientists. In BGP, both
sectors grow at the same rate, so we need NL/NL = NZ/NZ, or irzN-1Sz
=
LN^-1 SL
=
which gives SL = (qz(Nz/NL)^-1S)/(i7Z
and
+ rL(NZ/NL)^1),
g
(?zLqZS)/(ilZ(Nz/
with Nz/NL given by (26).
NL)(18-)/2 + I7L(Nz/NL)(s-1)12)
State dependencein this knowledge-basedR&D specificationimplies that the dynamicsof
the system can be unstable. In particular,there will now only be labour-augmentingtechnical
change if orNSVZ/NA VL < 1, and only Z-augmentingtechnicalchange if tNZ VZ/N VL > 1.
However,irNi VZ/NS VL is not necessarilydecreasingin Nz/NL. Inspectionof (17) shows that
this depends on whethera < 1/8 or not. When a < 1/8, a(N^VZ/N'VL)/a(Nz/NL)
< 0
and transitionaldynamicswill take us back to the BGP.In contrast,when a > 1/8, equilibrium
dynamics are unstableand will take us to a comer where only one type of R&D is undertaken.
Intuitively,a greaterNZ/NL createsthe usual price and marketsize effects, but also affects the
relative costs of futureR&D. If 8 is sufficientlyhigh, this lattereffect becomes more powerful
and creates a destabilizinginfluence.For example, in the extreme state dependencecase where
8 = 1, the system is stable only when a < 1, i.e. when the two factorsare gross complements.
In what follows, I restrictmy attentionto cases where the stabilityconditionis satisfied,so
we have a < 1/8. Next, note from equation(27) thatthe relationshipbetween relativesupplies
and factorprices dependson whether
a > 2 - 8.
(29)
If (29) is satisfied, an increase in the relative abundanceof a factor raises its relative marginal
product.When 8 = 0, this conditionis equivalentto a > 2, which was obtainedin the previous
SL
WLL
ACEMOGLU
795
subsection.It is also clear that when 8 = 1, (29) implies that a > 1/3, so the long-runrelative
demandcurve cannotbe upwardsloping in a stable equilibrium.However,for all values of 3 less
than 1, the conditions a < 1/8 and a > 2 - 3 can be satisfied simultaneously.So the general
conclusions reachedin the previoussubsectioncontinueto apply as long as 8 < 1. In fact, some
degree of state dependencemakes it more likely thatlong-rundemandcurves slope up.
The case of extreme state dependence, 8 = 1, on the other hand, leads to an interesting
special result. In this case, to have innovationsfor both sectors, we need equation (28) to be
satisfied, which implies SZ/SL = wzZ/WLL = r/-1 (and in addition, we need a < 1 for
stability). Technical change is thereforeacting to equate relative factor shares, as conjectured
by Kennedy (1964), who suggested that induced innovationswill push the economy towards
constantfactor shares (see also Samuelson(1965), Drandakisand Phelps (1965)). Here we find
that this result is obtainedwhen the innovationpossibilities frontiertakes the form given in (24)
with an extremeamountof state dependence,3 = 1.
4.3. Discussion
The analysis so far has highlighted two importantdeterminantsof the direction of technical
change. The first is the degree of substitutionbetween the two factors. When the two factors
are more substitutable,the marketsize effect is stronger,and endogenous technical change is
more likely to favour the more abundantfactor. The second determinantof the direction of
technical change is the degree of state dependence in the innovationpossibilities frontier.To
clarify the main issues, I kept the analysis as simple as possible. I now briefly discuss a number
of possible generalizations,and some issues that arise in using this model to think aboutthe real
world.
The first issue to note is that the model here exhibits a "scale effect" in the sense that as
populationincreases,the growthrate of the economy also increases.Jones (1995) convincingly
shows that there is little supportfor such a scale effect, and instead proposes a model where
steady growth in income per capita is driven by populationgrowth. Since the scale effect is
relatedto the marketsize effect I emphasize here, one might wonderwhether,once we remove
the scale effect, the marketsize effect will also disappear.The answeris thatthe scale effect and
the marketsize effect emphasizedhere are distinct, and a very naturalformulationremoves the
scale effect while maintainingthe marketsize effect. Briefly,considerthe following form of the
innovationpossibilities frontier:NL = r?LNISL and NL =
where X E (0, 1]. As long
-rzNZSz,
as X < 1, as in Jones (1995), there is no scale effect, and the long-run rate of growth of the
economy is n/(1 - X), where n is the rate of populationgrowth.However,all the results on the
directionof technical change obtainedabove continue to apply,with the only differencethat in
all the equationsfrom Section 4.2, X replaces B.
Second, the analysishere emphasizesprofit-motivatedR&D. In practice,thereareimportant
advances, perhapsincluding some of the most major scientific discoveries, that are not purely
drivenby profitmotives. Introducingsuch non-profit-motivated
innovationsinto this framework
is straightforward.One possibility would be that non-profitmotives drive what Mokyr (1990)
refers to as "macro-innovations",or what Bresnahan and Trajtenberg(1995) call "generalpurpose technologies", while the profit motives determine how these macro-innovationsare
developed for commercial use. Via this channel, it will be the profit motives that shape the
direction of technical change. Another possibility is to assume that as well as profit-driven
innovations,there is some "technologydrift".For example, the innovationpossibilities frontier
could take the form NL = 4LNL + rtLRL and Nz = 4zNz + rjzRz, where the 4L and ~z
terms capture the technology drift, that is, improvementsin technology that are unrelatedto
R&D directedat differenttypes of innovation.The presence of the ~ terms does not change the
796
marginalreturnto differenttypes of innovation,and for an equilibriumin which there are both

types of R&D, we still need the innovationequilibriumcondition,(20), to hold.
Third,the analysis so far treatedfactor supplies, L and Z, as given. Clearly,these supplies
can, and typically will, respondto relativeprices. I brieflydiscuss the case where Z corresponds
to capital and accumulatesin response to the interestrate in Section 6.1. When Z corresponds
to skilled labour, the relative price of Z, WZ/WL, will determinethe relative supply, Z/L. I
analysedthis case in Acemoglu (1998).16Here, it suffices to note that standardargumentsimply
so when the relative
an upward sloping relative supply of skills, e.g. Z/L = f(wz/wL),
demand for skills is upward sloping, an exogenous increase in Z/L will increase wZ/WL,
encouragingfurtheraccumulationof skills, andyet more skill-biasedtechnicalchange.Note also
thatwhen the demandfor skills is sufficientlyupwardsloping, the relativesupplyandthe relative
demandcurvescould intersectmore thanonce, and multiplelong-runequilibriaare possible (see
Acemoglu (1998)).
Finally,it is useful to brieflydiscuss empiricalevidence on the degree of state dependence.
To the best of my knowledge, there has been no direct investigationof this issue. The data on
patentcitationsanalysedby, amongothers,Jaffe,TrajtenbergandHenderson(1993), Trajtenberg,
Hendersonand Jaffe (1992) and Caballeroand Jaffe (1993) may be useful in this regard.These
papers study the extent of subsequentcitations of patentsand interpreta citation of a previous
patent as evidence that a currentinventionis exploiting informationgeneratedby this previous
invention.This correspondsto the presenceof NL and Nz terms in equation(24) in the model
here. Patent citations data can be used to investigatewhether there is state dependence at the
industrylevel. Unfortunately,it is currentlyimpossible to investigate state dependence at the
factor level, which is more relevantfor the discussion here. This is because, althoughwe have
informationaboutthe industryfor which the patentwas developed,we do not know which factor
the innovationwas directedat. In any case, the results reportedin Table 1 of Trajtenberget al.
(1992) suggest that there is some limited amount of industry state dependence.For example,
patentsare likely to be cited in the same three-digitindustryfrom which they originated;but on
average,they are more often cited in a differentthree-digitindustry(see below).
5. APPLICATIONS(WITHLIMITEDSTATEDEPENDENCE)
In this section, I discuss some of the applicationsof directedtechnicalchange. I will emphasize
both why models with endogenouslybiased technicalchange areuseful in thinkingabouta range
of problemsand how the results depend on the elasticity of substitution.I startwith a range of
applicationswhere a formulationwith limited statedependence,i.e. the case with 8 < 1, appears
to be more appropriate.
5.1. Endogenousskill-biasedtechnicalchange
Figure 1 plots a measureof the relativesupplyof skills (the numberof college equivalentworkers
dividedby noncollege equivalents)and a measureof the returnto skills (the college premium).17
It shows thatover the past 60 years, the U.S. relativesupply of skills has increasedrapidly,and
in the meantime,the college premiumhas also increased.The figure also shows that beginning
in the late 1960's, the relative supply of skills increasedmore rapidlythan before. In response,
16. See Heckman,Lochnerand Taber(1998) for a more detailedanalysis of the responseof the supply of skills to
an increasein the returnsto skills.
17. The samples are constructedas in Katz and Autor(2000). See Katz and Autor(2000) or Acemoglu (2002) for
details.
ACEMOGLU
797
the skill premiumfell duringthe 1970's, and then increased sharplyduringthe 1980's and the
early 1990's.
Existing explanationsfor these facts emphasizeexogenous skill-biasedtechnicalchange.18
Technical change is assumed to naturallyfavour skilled workers, perhapsbecause new tasks
are more complex and generatea greaterdemandfor skills. This skill bias explains the secular
behaviourof the relative supply and returnsto skills. Moreover,the conventionalwisdom is that
there has been an accelerationin the skill bias of technology precisely aroundthe same time as
the relativesupply of skills startedincreasingmuch more rapidly.This accelerationexplains the
increasein the skill premiumand wage inequalityduringthe 1980's.
A model based on directedtechnical change suggests an alternativeexplanation.Suppose
that the second factor in the model above, Z, is skilled labour.Also assume that the innovation
possibilities frontier is given by (24), so that we are in the knowledge-based R&D model.
This is without any loss of generality since for the focus here, the lab equipmentspecification
correspondsto the case with 8 = 0 in termsof (24). Then, equation(26) from Section 4 gives the
equilibriumskill bias and equation(27) gives the long-runskill premium.
The results above, in particularequation(26) and the weak induced-biashypothesis,imply
thatthe increasein the supplyof skills createsa tendencyfor new technologies to be skill biased.
This offers a possible explanationfor the secular skill-biasedtechnical change of the twentieth
centuryirrespectiveof the value of the elasticityof substitution,a (as long as it is not equal to 1),
since the predictionsregardinginduced skill-biasedtechnical change do not dependon whether
the long-runrelativedemandcurve for skill is upwardsloping.
Equally interesting,if a > 2 - 8, then we have the strong induced-biashypothesis:now
the long-runrelationshipbetween the relativesupply of skills and the skill premiumis positive,
and greaterskill abundanceleads to a greaterskill premium.Intuitively,with a largeenough, the
marketsize effect is so powerful that it not only dominatesthe price effect on the directionof
technical change, but also dominatesthe usual substitutioneffect between skilled and unskilled
workersat a given technology (as capturedby equation(18) above). As a result, an increase in
the relativesupplyof skills makes technology sufficientlyskill biased to raise the skill premium.
In this case, the frameworkpredictssufficientlybiased technicalchange over the past 60 years to
actuallyincreasethe skill premium,consistentwith the patterndepictedin Figure 1.
The stronginduced-biashypothesis,i.e. the case with a > 2 - 8, also offers an explanation
for the behaviourof the college premiumduringthe 1970's and the 1980's shown in Figure 1.
Suppose that the economy is hit by a large increase in H/L. If this increase is not anticipated
sufficiently in advance, technology will not have adjusted by the time the supply of skills
increases. The initial response of the skill premiumwill be given by equation(18) which takes
technology as given. Therefore,the skill premiumwill firstdecline, but then as technology starts
adjusting,it will rise rapidly.Figure 3 drawsthis case diagramatically.19
This frameworkalso provides a possible explanationfor why technical change duringthe
late eighteenth and early nineteenthcenturies may have been biased towardsunskilled labour.
The emergence of the most "skill-replacing"technologies of the past 200 years, the factory
18. For example, Autoret al. (1998), Galorand Maov (2000) or Krusell,Ohanian,Rios-Rull and Violante(2000).
19. There was also a very large increase in educationalattainmentin the U.S. duringthe firsthalf of the century.
High school enrolmentand graduationrates doubled in the 1910's (e.g. Goldin and Katz (1995)). The skill premium
fell sharplyin the 1910's. But, despite the even fasterincrease in the supply of high school skills duringthe 1920's, the
skill premiumlevelled off and starteda mild increase.Goldin and Katz (1995) conclude thatthe demandfor high school
graduatesmust have expandedsharplystartingin the 1920's, presumablydue to changes in office technology and higher
demand from new industriessuch as electrical machinery,transportand chemicals (see also Goldin and Katz (1998)).
Therefore,this historicalepisode also suggests that the demandfor skills expandedsharplyin response to the increase
in the supply of skills, giving empirical supportto the weak induced-biashypothesis. However,it does not supportthe
stronginduced-biashypothesis.
798
Relative Wage
Long-runRel Wage
Initial Rel Wage -
\I \^
^ Long-runrelative
demandfor skills
Short-run
Response
\
\
I
l\
Exogenous Shift in
Relative Supply
FIGURE3
Short-runand long-runresponsesof the skill premiumto an increasein the relativesupply of skills when condition(29)
is satisfied
system, coincided with a substantialchange in relative supplies. This time, there was a large
migrationof unskilled workersfrom villages and Irelandto English cities and a large increase
in population(see, for example, Habakkuk(1962), Bairoch (1988) or Williamson(1990)). This
increasecreatedprofitopportunitiesfor firmsin introducingtechnologiesthatcould be used with
unskilled workers.In fact, contemporaryhistoriansconsideredthe incentive to replace skilled
artisansby unskilledlabourersas a majorobjectiveof technologicalimprovementsof the period.
Ure, a historianin the firsthalf of the nineteenthcentury,describesthese incentives as follows:
"It is, in fact, the constantaim and tendency of every improvementin machineryto supersede
human labour altogether,or to diminish its costs, by substitutingthe industryof women and
childrenfor thatof men; of thatof ordinarylabourers,for trainedartisans."(quotedin Habakkuk
(1962, p. 154)). The frameworkdevelopedhere is consistentwith this notion that the incentives
for skill-replacingtechnologies were shaped by the large increase in the supply of unskilled
workers.
Therefore,this model provides an attractiveexplanationfor the skill-replacingtechnical
change of the nineteenthcentury,the secular skill bias of technology throughoutthe twentieth
century,and the recent accelerationin this skill bias. In additionif a > 2 - 8, it explains both
the increase in the skill premiumover the past 60 years and its surge duringthe past 25 years
withoutinvokingany exogenous change in technology.
Whatis the empiricallyplausiblerangefor a - 2 + 87?Firstnote thatFigure 1 suggests the
demandfor skills over the past 50 years has increasedsomewhatfasterthanthe supply of skills.
In the context of the model here, this requiresa - 2 + 8 and (a - 2 + 8)/(l - Sa) to be positive,
but not too large. The elasticity of substitution,a, is generally difficult to estimate, but there
is a relativelywidespreadconsensus that the elasticity between skilled and unskilled workersis
greaterthan 1, most likely, greaterthan 1.4, andperhapsas large as 2 (see, for example,Freeman
(1986)). An interestingstudy by Angrist (1995) uses a "naturalexperiment"arising from the
largeincreasein universityenrolmentsfor Palestinianlabour.His estimatesimply an elasticityof
substitutionbetween workerswith 16 yearsof schooling andthose with less than 12 of schooling
of approximatelya = 2.
What about 87?Unfortunately,the magnitudeof this parameteris much harderto ascertain.
The best we can do is to make a very roughguess based on the evidence presentedby Trajtenberg
ACEMOGLU
799
et al. (1992). Looking at patentcitations,they constructan index of whethera patentfroma given

industryis citing patentsfrom the same industry.Their index, TECHF,takes the value 0 when
the citing and originatingpatentsare in the same three-digitindustry,the value 0-33 when they
are in the same two-digit industry,the value 0.66 when they are in the same one-digit industry,
and the value 1 if they arenot even in the same one-digitindustry.The averageof TECHFin their
sample is approximately0.-31,implying thatthereis some limited degree of state dependenceat
the industrylevel. On the basis of this, a plausible value for 8 would be less than 0-31, perhaps
0-2 (though,of course,it is possible thatstatedependenceat the factorlevel is greaterthanthatat
the industrylevel). Combining8 = 0.2 with a = 1.4, we obtain (a - 2 + 8)/(1 - Sa) = -0-55,
while combining it with a = 2, we obtain (a - 2 + 8)/(1 - 8r) = 0.33. So with relatively
limited state dependence,the plausiblerangeof the long-runelasticity of the skill premiumwith
respectto the relativesupply of skills would lie between -0-55 and 0.33, comparedto the range
of the short-runelasticity of [-0.7, -0-5]. As we consider greatervalues of 8, the range of the
long-runelasticity shifts to the right.20
5.2. Directed technical change and cross-countryincome differences
Many less developedcountries(LDCs) use technologies developedin the U.S. and otherOECD
economies (the North).A numberof economists, including Atkinsonand Stiglitz (1969), David
(1975), Stewart(1977), Basu and Weil (1998), have pointedout that importedtechnologies may
not be "appropriate"to LDCs' needs. Directed technical change increases these concerns: it
implies that technologies will be designed to make optimal use of the conditions and factor
suppliesin the North.Therefore,they will be highly inappropriateto the LDCs' needs (Acemoglu
and Zilibotti, 2001). Because it is still often profitablefor the LDCs to use these technologies
rather than develop their own, the extent of directed technical change will determine how
inappropriatetechnologies used by the LDCs' are to their needs. Via this channel, directed
technical change will influence the income gap between the North and the LDCs. I now use
the above frameworkto discuss this issue.
Suppose thatthe model outlined above applies to a countryI referto as "theNorth"(either
the U.S. or all the OECD countries as a whole). Also to simplify the algebra,I now focus on
the case with no state dependence8 = 0, though all the results generalizeto the case of limited
state dependenceimmediately.Suppose also that there is a set of LDCs in this world economy
thatwill use the technologies developedin the North.TakeZ to be skilled labour,thoughfor the
resultsin this subsection,it could also standfor physical capital.For simplicity,I take all of these
countriesto be identical,with L' unskilledworkersand H' skilled workers.
A key characteristicof the LDCs is that they are less abundantin skilled workersthan the
North,so H'/L' < H/L. I assumethatall LDCs can copy new machinevarietiesinventedin the
North,but firmsface a price of K- /(1-/) ratherthan 1 as in the North.This cost differentialmay
resultfrom the fact thatfirmsin the LDCs do not have access to the same knowledgebase as the
technology monopolistsin the North.Also, thereis no internationaltradebetween the Northand
the LDCs.
An analysissimilarto above immediatelygives the LDC outputlevels (recallequation(15)):
- /1) and
- fi), where p"s denote
=
YH = (p)(1-)/KNHH'/(1
YL (PL)(1- )/t1KNLL'/(1
prices in the LDCs, which differ from those in the Northbecause factorproportionsare different
and there is no internationaltrade. The parameterK featuresin these equations since machine
20. Finally, to see anotherreason why for this applicationthe model with of limited state dependence is more
relevant,recall from equation(28) that if a = 1, in the long run we would have SH/SL = q-1. Therefore,the relative
sharesof skilled and unskilledlabourin GDP whould be constant.This is clearly not consistentwith the patterndepicted
in Figure 1, where both the numberof skilled workersand theirrelativerenumerationhave been increasingover time.
800
costs are differentin the LDCs. It is naturalto think of K as less than 1, so that machineprices
are higher and fewer machines are used in the LDCs than in the North. Notice also that the
technology terms, NL and NH, are the same as in the North, since these technologies are copied
from the North. Using this equationand (4) and (15), the ratioof aggregateincome in the LDCs
to thatin the Northcan be writtenas
K Y((pL)(-)/NLL
Yy(^pL-/Nr
')
+ (1 - y)((p')(1P-)/3NHH')_-. .l-- L) e + (1 - y)(p-)/

NHH)E e
(30)
Some tedious algebra(see Acemoglu (2001)) shows that

aY/Y
aNH/NL
1 -Pf/NH
ocI
A
NLI
-/H"H7F
'(
(31)
Since H'/L' < H/L, this expression implies that when a > 1, i.e. when the two factors are
gross substitutes,an increasein NH/NL raises the income gap between the LDCs and the North
(i.e. reduces Y'/ Y). In contrast,when a < 1, so that the two factorsare gross complements,an
An
increasein NH/NL narrowsthe income gap. The intuitionfor both resultsis straightforward.
increase in NH/NL increases the productivityof skilled workersrelativeto the productivityof
the unskilled.Therefore,everythingelse equal, a society with more skilled workersbenefitsmore
from this type of technical change. However,this change also affects the relative scarcity,and
thereforethe relativeprice, of the two goods. In particular,the skill-intensivegood becomes more
abundantand its relativeprice falls. When a > 1, this effect is weak, and the North still gains
more in proportionalterms,and the ratio of LDC income to income in the Northfalls. However,
when a < 1, the price effect is strong,and as a result, in proportionalterms, the skill-abundant
Northbenefitsless thanthe LDCs.
Next, recall that when a > 1, an increasein NH/NL correspondsto skill-biasedtechnical
change, while with a < 1, it is a decreasein NH/NL thatis skill biased. Therefore,irrespective
of the value of a, skill-biasedtechnicalchange increasesthe income gap between the LDCs and
the North. This extends the results in Acemoglu and Zilibotti (2001) to a slightly more general
model, and also more importantly,to the case where the two factorsare gross complements.
Now to highlight the role of directed technical change, contrast two polar opposite
situations. First, suppose that there are no intellectual propertyrights enforced in the LDCs,
so LDC firmspay no royaltiesto R&D firmsin the North.This implies thatnew technologiesare
developed for the Northernmarket,equation (21) applies and NH/NL oc (H/L) a . Contrast
this situationwith an alternativewhere new technologiescan also be sold in the LDC marketson
exactly the same terms as in the North. In this case the equilibriumtechnologies would satisfy
NH/NL oc (HW/LW)
where HW/LW = (H +KH')/(L
+KL') is the world (effective) ratio
of skilledto unskilledworkers.The parameterK featuresin this equationbecauseit parameterizes

for machinesof LDC workers.
the relativeproductivity/demand
By the assumptionthat H'/L' < H/L, we have HW/LW < H/L. This implies that, as a
directimplicationof the weak induced-biashypothesis,equilibriumtechnologieswill be too skill
biased for the LDCs. In particular,when a > 1, technologiesdevelopedin the Northwill feature
too high a level of NH/NL for the LDCs' needs, and when a < 1, technologies developed in
the Northwill featuretoo low a level of NH/NL for the LDCs (which againcorrespondsto more
skill-biasedtechnology).Hence, irrespectiveof the value of the elasticityof substitution,directed
technical change, by making technologies in the North too skill biased for the LDCs, creates a
force towardsa largerincome gap between the rich and the poor.This resultis intuitive:thereare
ACEMOGLU
801
more skilledworkersin the North,anddirectedtechnicalchangeinducestechnologymonopolists

in the North to develop technologies appropriatefor skilled workers(i.e. when a > 1, higher
NH/NL, and when a < 1, lower NH/NL). These skill-biased technologies are less useful for
LDCs, so LDCs benefitless thanthe North,andthe income gap is largerthanit would have been
in the absence of directedtechnicalchange.21
5.3. The effect of internationaltrade on technical change
The literatureon trade and growth shows how patternsof internationaltrade affect the rate
of technical change (e.g. Grossman and Helpman (1991)). Similarly, when the direction of
technical change is endogenous, changes in the amount of internationaltrade may affect the
type of technologies thatare developed.This may be relevantfor currenteconomic concerns,for
example because, as claimed by Wood (1994), tradeopening may have affectedwage inequality
throughits effect on "defensive"innovations.22
To addressthese issues, consider the same set-up as in the previous subsection with a set
of LDCs that can copy innovations from the North and let us again focus on the case of no
state dependence (i.e. 3 = 0 or the lab equipment specification). Assume first that there is
no internationaltrade, so the equilibriumcharacterizedin the previous subsection applies. In
particular,the relative price of skill-intensive goods in the North is given by equation (16),
and the skill premiumis given by an equation similar to (18). When skill bias of technology
is endogenized, NH/NL is given by (21).
Next, supposethatthereis opening to internationaltrade,with both goods tradedcostlessly.
Assume that the structureof intellectual property rights is unchanged as a result of this
trade opening. Internationaltrade will generate a single world relative price of skill-intensive
goods, pw. To determinethis price, note that the total supply of skill-intensivegoods will be
(H + KH')/(l - fi), and the total supply of labour-intensivegoods will be
(P)(1-)/N
(pW)O'-fi)/# NL (L + KL')/(l - f). Using these expressionsand equation(7), the world relative
price is obtainedas
w
1-y
( [y )
aN
+KH)
(NW(L + KL)
1-y
H
L
IJ-iH
'
(32)
where the last equality defines X = (H/L)/((H + KH')/(L + KcL'))> 1. The fact that X > 1
follows from H'/L' < H/L. I also use the notation NH and NL to emphasize that world
technologies may change from their pre-tradelevels in the North as a result of international
trade.
Since skills are scarcer in the world economy than in the North alone, trade opening
will increase the relative price of skill-intensive goods in the North, i.e. pw > p. This is a
straightforwardapplicationof standardtradetheory.What is differenthere is thatthis change in
productpriceswill also affectthe directionof technicalchange.Recall thatthe two forces shaping
the directionof technical change are the marketsize effect and the price effect. Because trade
does not affect the structureof intellectualpropertyrights, the marketsizes for differenttypes
of technologies do not change. But productprices are affected by trade,so the price effect will
be operational.Since the price effect encouragesinnovationsfor the scarce factor,international
21. Reality presumablylies somewherein between, with imperfectintellectualpropertyrightsenforcementin the
LDCs. It is straightforwardto generalizethe results to that case, and show thattechnologies in that case will also be too
skill biased for the LDCs comparedto the case of full propertyrightsenforcement(see Acemoglu (1999a)).
22. This point is also suggested in Acemoglu (1998) and analysed in detail in Acemoglu (1999a). Xu (2001)
extends the set-upof Acemoglu (1999a), which I use here, to have both goods employ both factors.
802
trade,by makingskills more scarcein the North,will induce moreinnovationsdirectedat skilled

workers.
To see this more formally,once againconsiderthe case of no statedependencein R&D (i.e.
the lab equipmentspecification,or (24) with a = 0). For the technologymarketto clear,we need
condition (20) to be satisfied. Combiningthe relativeprice of skill-intensivegoods, now given
by (32), with technology marketclearingcondition (20), we obtain
Comparingthis equationto (21), we see immediatelythat, because A > 1, tradeincreases the

physical productivityof skilled workers more than that of unskilled workers. As usual, this
increasein NH/NL may not correspondto skill-biasedtechnicalchange if a < 1.
To study the effect of this inducedchange in technologyon factorprices, firstnote that(32)
implies that without a change in technology tradeopening would increase the skill premiumin
the Northto
WH
WL
NH
WL
= (pW)1/
(P )
N(1 = (
\rl
NL
-Yx
Y)
y
(N
A,,
\
H)- (H
H
NL /
(4)
(34)
Comparingthis equationto (18) from before, we see thattradeopening necessarilyincreasesthe

skill premium,which is a simple applicationof standardtradetheory.
With directedtechnicalchange, the skill premiumin the North,instead,changes to
WH
WL
W)
NH
NW
-(H
L!
(35)
This equationdiffers from (34) because the skill bias of technology is now given by (33) rather
thanbeing held constantat the pre-tradelevel.
Comparingthe post-tradeskill premium,(35), to the skill premiumbefore trade opening
(e.g. equation(22)), we see that irrespectiveof whethera > 1 or not, tradeopening increases
the skill premium.However, comparingequations (34) and (35) shows that the trade-induced
technical change will increase the skill premiumby more than predictedby the standardtrade
theory only when a > 1. The intuition for this result is also simple: expression (33) shows
that trade will induce technical change to increase the relative physical productivityof skilled
workers.An increasein relativephysicalproductivitytranslatesinto skill-biasedtechnicalchange
only when a > 1. Therefore, we obtain the result that, as conjecturedby Adrian Wood,
trade opening could induce skill-biased technical change and increase wage inequality more
than predictedby standardtradetheory.Yet, this conclusion is obtainedonly when skilled and
unskilledworkersare gross substitutes(which appearsto be the empiricallyrelevantcase).
5.4. The Habakkukhypothesis
According to the Habakkuk (1962) hypothesis, the rapid technical change or technology
adoption in the U.S. economy during the ninteenthcentury (especially relative to the British
economy) resultedfromthe relativelabourscarcityin the U.S., becauselabourscarcityincreased
wages, and encouraged firms to develop and adopt labour-savingtechnologies. Despite the
prominence of this hypothesis in the economic history literature,there has been no widely
accepted formalization of the argument, and we are consequently unaware of under what
circumstancesHabakkuk'sconclusionwould apply(see, e.g. David (1975)). In addition,standard
economic theory suggests that higher wages may reduce investment, and via this channel,
discourageinnovation.Moreover,in contrastto the basic premise of the Habakkukhypothesis,
ACEMOGLU
803
the endogenous growthliteratureemphasizesthe presence of a scale effect which suggests that

a largerworkforceshould encouragemore innovation.Could the Habakkukhypothesisbe valid
despite these countervailingeffects?
To discuss this case, suppose that Z now stands for land. Moreover,assume that jz = 0,
which implies that there are no land-complementaryinnovations.The only source of technical
change is labour-complementaryinnovations.The rest of the set-up is the same as before, and
to economize on space, I will only study the lab equipmentspecification.The questionis: under
what circumstanceswill greaterlabourscarcity(smaller L/Z) lead to a higher level of NL-i.e.
to more innovations?
From (14), the free-entry condition for technology monopolists, ?ILVL = 1, implies
)LfipL/L/r = 1. Now using (8) and (16), this conditiongives
r7Lp
1NL
a
c1r)2
L---=r,
(36)
where A is a suitably defined constant. Inspection of equation (36) immediately shows that
aNL/ aZ > 0. Therefore,a greaterabundanceof land (for a given level of employment)always
encouragesthe creationof more labour-complementary
technologies. So comparingthe U.S. to
Britain,as Habakkukdid, leads to the conclusion that there should be fastertechnicalchange in
the more land-abundantU.S.
On the other hand, the sign of aNL/aL-i.e. the effect of labour scarcity for a given
supply of land-is ambiguous. If a > 1, so that labour and land are gross substitutes,it can
be verifiedthat aNL/IL > 0, hence in this case, greaterscarcityof labour(for a given supply of
land) discouragesthe developmentof new technologies. In contrast,if a is sufficiently smaller
than 1, i.e. if labourand land are sufficientlycomplementary,we can have aNL/aL < 0. This is
intuitive.A greaterscarcityof labourcreates two forces: the price of the labour-intensivegood
is higher,but also the marketsize for labour-complementary
technology is smaller.Which force
dominatesdependson the strengthof the price effect, which is again a functionof the degree of
substitutability.If a > 1, the price effect is less powerfuland we obtainthe opposite of the result
conjecturedby Habakkuk-it is not labourscarcity,but labourabundancethat spursinnovation
(for a given quantityof land). However, with sufficient complementaritybetween labour and
land, the model gives the result conjecturedby Habakkuk-greater labour scarcity leads to
faster innovation.Therefore, in this frameworkthe Habakkukhypothesis requires labour and
land to be highly complementary.Since the focus of Habakkuk'sanalysis was technical change
in agricultureand low-tech manufacturing,the assumptionof a high degree of complementarity
between labourand otherinputsmay be realistic.
6. TECHNICALCHANGEWITHEXTREMESTATEDEPENDENCE
I have so far discussed a numberof applicationsof directedtechnical change with limited state
dependence.In this section, I discuss two applicationswhere extreme state dependence,a = 1,
may be more appropriate.
6.1. Why is long-run technical change labour augmenting?
Consider anotherregularitydiscussed in the introduction:the share of labour and capital have

been approximatelyconstant in U.S. GDP, while the capital-labourratio has been increasing
steadily. This suggests that technical change has been mostly labour augmenting (unless the
elasticity of substitutionbetween capital and labour happens to be exactly equal to 1). Can a
804
model of directedtechnicalchange be useful in thinkingaboutwhy aggregatetechnicalchange

appearsto be labouraugmenting?
To provide an answer to this question, suppose that Z correspondsto physical capital.23
Assume also that Z = K is growing over time due to capitalaccumulation.Let us startwith the
flexible innovationpossibilities frontiergiven by (24), where 3 parametrizesthe degree of state
dependence.Then equation(28), with K replacing Z, implies that unless a = 1 (the elasticity
of substitutionexactly equal to 1), or 3 = 1 (extreme state dependence),factor shares will not
be constant:with the capital-labourratio growing, the share of capital will be contractingor
expanding,and there exists no BGP. This implies that neither the lab equipmentspecification
nor the knowledge-basedR&D specificationwith limited statedependenceare consistentwith a
constantcapital shareor with purelylabour-augmentingtechnicalchange.
For all technicalchange to be (endogenously)labouraugmenting,we thereforeneed a = 1.
In Acemoglu (1999b), I show that when the innovation possibilities frontier takes the form
in (24) with 3 = 1 and the elasticity of substitutionis less than 1, there exists a unique
technicalchange.This resultcan
equilibriumpathtendingto a BGP with only labour-augmenting
be interpretedas eithera positive or negativeone: on the positive side, it shows thatit is possible
to constructa model where equilibriumlong-runtechnical change is labouraugmenting,even
though capital-augmentingtechnicalchange is also allowed. Moreover,the economy converges
to this equilibrium,andon the transitionpath,therewill typicallybe capital-augmentingtechnical
change. On the negative side, it shows that this result obtains only when there is an extreme
amountof statedependencein R&D, i.e. 3 = 1,24andwhen the elasticityof substitutionbetween
labourand capitalis less than 1.25
6.2. Wagepush and technicalchange
Finally, I use the above frameworkto investigate the technological implications of a "wagepush shock",which Blanchard(1997) and Caballeroand Hammour(1998) arguetook place in
continentalEuropestartingin the late 1960's. Considerthe above frameworkwith Z interpreted
as capital, K, and as in the previous subsection, focus on the case with the elasticity of
substitutionless than 1, i.e. a < 1. In addition,in orderto study the implicationsof wage push,
let me introducea quasi-laboursupply function: L = m(sL), where SL is the relative labour
share, and I assume m' > 0, so that a higher labour share leads to greaterlabour supply.26To
simplify the algebra,I furtherspecialize this supply function,and assume that
L=
(l
= (
(37)
Here r > 0, and tt is a shift parameter.A decrease in It correspondsto a "wage-push"shock,

since it increasesthe labourshareat a given level of employment.A high level of r corresponds
23. One has to be careful with this interpretationsince the model alreadyfeatures"machines"which can also be
thoughtas capital.
24. This leads to the questionof how this result,which requiresextremestatedependence,5 = 1, can be reconciled
with the findingsin the previous subsection, which relied on limited state dependence,i.e. 8 < 1. One possibility is to
develop a hybridframeworkwhere the degree of state dependencevaries across innovationsdirectedat differentfactors
(see Acemoglu (2001)).
25. An elasticity of substitutionless than 1 betweencapitaland labourappearsreasonable.Using time-seriesdata,
Coen (1969), Eisner and Nadiri (1968) and Lucas (1969), for example, all estimateelasticities significantlyless than 1.
AlthoughBerndt(1976) claims thatthe use of higherqualitydataleads to higherestimatesof the elasticityof substitution,
he does not controlfor a time trendin his estimation,biasing his resultstowards1.
26. Note thatthis formulationcapturesboth the laboursupply decisions of workersand bargainingbetween firms
and workers.Forexample, with a bargainingset-up, when unemploymentis high (employmentis low), workerswill have
a weakerbargainingposition, and only receive a smallershareof output,leading to lower sL.
ACEMOGLU
805
to a more inelastic quasi-supplycurve. This formulationis convenient since it yields a simple

expressionfor SL/SK, for given bias of technology,from (18) to be combinedwith (37):
SL
NKK
(l-y
SK
(N
(38)
(38)
In essence (38) shows that, when labour and K are gross complements, i.e. a < 1, as we
have assumed to be the case, an increase in the relative productivity,NK/NL, or relative
abundance,K/L, of the other factor will raise the labour share. So with a < 1, and wagepush shock will increase the labour share and reduce K/L. With a > 1, we would obtain the
opposite result.
What will happenin the long run?There are two marginsof adjustment:the capital-labour
ratio and technology. Even without any adjustmentin technology, the labour share may return
back to its initial level if the capital-labourratio returnsto its initial position. This will be the
case, for example, when the long-runreturnto capital is given by a perfectly elastic supply of
capital (e.g. CRRA preferencesfor the representativeconsumer). However, Blanchard(1997)
shows that changes in capital-labourratio do not accountfor the behaviourof the labourshare.
This raises the possibility thatthe majoradjustmentwas due to changes in technology.To focus
on this, I ignore capital accumulationand normalize K = 1 (see Acemoglu (1999b), for the
analysis of the case where both technology and capitaladjust).
Now combining(37) and (38), we obtain
LSR =
-++l- a
(
N39)a+l-a
Since a < 1, (39) defines a positive relationshipbetween i/ and L. I use the superscript"SR" to
emphasizethat this is the short-runresponse of employmentfor given bias of technology.From
equation(39), a wage-push shock, i.e. a decreasein /z, will reduce employment.Equation(38),
in turn, implies that as long as a < 1, this shock will also increase the labour share in GDP.
Therefore,for a given bias in technology,as arguedby Blanchardand Caballeroand Hammour,
a wage push will reduceemploymentand increasethe labourshare.
The fundamentalimplicationof the frameworkhere is that the wage push will also affect
the equilibriumbias of technical change. As in the previous subsection,consider the case with
extreme state dependence,8 = 1, and combine (37) with equation (28) from Section 4.2 with
Z = K = 1 to obtain the long-runrelative labour share as (sL/SK)LR = 1-1. This equation
states that the technology-marketclearing condition can only be satisfied if the labour share
in GDP returnsback to its initial level. Using this, we also calculate the long-runemployment
level as
LR
Tt Tr,
LLR =
(40)
where the superscript"LR" shows that these expressions refer to the long-run equilibrium.
Next, using equation (26), we obtain the relative technology level consistent with long-run
equilibriumas
NK
aTr-(1-)
T -)
_-=rT
r O/.
/ 1-y
/ \ a
(
(41)
Comparisonof (39) and (40) immediatelyimplies thatthe elasticity of employmentwith respect

to It is greaterin the long run than in the shortrun (a/(ra + 1 - a) < 1/r).27 This again can
be thoughtof as an applicationof the LeChatelierprinciple.Intuitively,with a < 1, despite the
27. When a > 1, equation(40) no longer gives the long-runlabourdemand,since, in this case, the economy is
unstableand will tend to an equilibriumwith only one type of innovation.
806
LabourShare
Quasi-laboursupply
Short-run
response
Initial share
and long-run
share
/
<
Initial
\ sha
Long-rundemand
\
Long-run
empl.
Short-run
\demand
Short- Initial
run
empl.
empl.
Employment
FIGURE 4
Short-runand long-runresponses of the labourshare to a shift in the quasi-laboursupply when capital and labourare
gross complements
increasein the cost of labourresultingfromthe wage push, firmscannotsubstantiallyreducetheir

labourdemandin the shortrunbecause of the low elasticity of substitution.However,the change
in factorprices induces technicalchange that would allow firmsto be less dependenton labour.
Once this technicalchange takesplace, firmsgraduallyreducetheirlabourdemand.28Inspection
of equation (41) also implies that a decline in ,i-i.e. an adverse labour supply shock-will
reduce NK/NL. This correspondsto capital-biasedtechnicalchange since capitaland labourare
gross complements.29Figure 4 draws this case diagrammatically,assuming that technology is
given in the short run when the wage push first occurs, and then traces the adjustmentof the
economy to the shock. It shows diagrammaticallyhow, as in the case of Europeaneconomies,
the wage pushwill firstincreasethe labourshare,andthen graduallyreduceit by creatingcapitalbiased technicalchange. Throughoutthe process, employmentfalls.
7. CONCLUDINGREMARKS
For many problems in macroeconomics, development economics, labour economics, and
internationaltrade, whether technical change is biased towardsparticularfactors is of central
importance.This papersynthesizedsome recentresearchon the determinantsof biased technical
change. The presumptionis that the same economic forces-profit incentives-that affect the
amount of technical change will also shape the direction of technical change, and therefore
determinethe equilibriumbias of technology.I arguedthat this perspectivehelps us understand
a numberof otherwisepuzzling patterns.
I also highlightedthe variousmodelling choices involved in thinkingaboutthe directionof
technical change. I demonstratedthat there are two forces affecting equilibriumbias: the price
28. This idea is relatedto thatproposedby Caballeroand Hammour(1998). They suggest thatduringthe 1980's,
in orderto avoid expropriationby labour,Europeanfirmschose more capital intensiveproductiontechniques,reducing
theirlabourdemand.
29. However,notice thatin this model, the laboursharesimply returnsback to its originallevel. So to explain why
the labour share may have fallen below its original level in many Europeancountries,we may also need some of the
adverselaboursupply shocks to have reversedthemselves (i.e. tz to have fallen back towardsits initial level).
ACEMOGLU
807
effect and the marketsize effect. The elasticity of substitutionbetween differentfactorsregulates

how powerful these effects are, and this has implicationsfor how technical change and factor
prices respondto changes in relativesupplies.
Another important determinantof equilibrium bias of technology is the form of the
innovationpossibilities frontier-how relativecosts of differenttypes of innovationchange with
the currentstate of technology.The choice here is between a specificationthat emphasizes state
dependence,wherebypastinnovationscomplementinga factormakecurrentinnovationsdirected
at that factor cheaper,and a specificationwithout state dependence.It appearsthat in thinking
about skill-biased technical change a specificationwith only limited state dependenceleads to
more plausible results, while in the case of capital- and labour-augmentingtechnical change, a
specificationwith extreme state dependence appearsmore appropriate.Although it is possible
to combine these features to have a unified frameworkthat can be applied both to analysing
the degree of skill bias and capital bias of technology (see Acemoglu (2001)), this framework
is currentlyhighly ad hoc and without micro-foundations.Futurework towardssuch a unified
model would be very useful.
Whether technical change exhibits this type of state dependence, and whether this state
dependenceaffects differentfactors differentially,is ultimatelyan empiricalquestion. Existing
evidence does not enable us to reach firm conclusions. Nevertheless, data on patent citations
seem to providea useful startingplace, and empiricalwork thatcan informmodelling choices in
this field is anotherareafor fruitfulfutureresearch.
Acknowledgements. This paper is a revised version of the Review of Economic Studies lecture delivered at the
2001 Royal Economic Society meeting in Durham.The authorthanks Rub6n Segura-Cayuelafor excellent research
assistanceand OrazioAttanasio,David Autor,OlivierBlanchard,Nobu Kiyotaki,MarkSanders,RubenSegura-Cayuela,
Jaume Ventura,three anonymousreferees, and seminar participantsat MIT, Institutefor InternationalEconomics at
Stockholmand the Royal Economics Society for comments.
REFERENCES
ACEMOGLU, D. (1998), "Why do New Technologies Complement Skills? Directed Technical Change and Wage
Inequality';QuarterlyJournalof Economics,CXIII, 1055-1090.
ACEMOGLU,D. (1999a), "Patternsof Skill Premia"(NBERWorkingPaperNo. 7018).
ACEMOGLU,D. (1999b), "Labor-and Capital-augmentingTechnicalChange"(NBER WorkingPaperNo. 7544).
ACEMOGLU,D. (2001), "DirectedTechnicalChange"(NBER WorkingPaperNo. 8287).
ACEMOGLU,D. (2002), "TechnicalChange, Inequalityand the LaborMarket",Journal of Economic Literature,XL,
7-72.
ACEMOGLU, D. and ZILIBOTTI,F. (2001), "ProductivityDifferences",Quarterly Journal of Economics, CXVI,
563-606.
AGHION,P. andHOWITT,P. (1992), "AModel of GrowthThroughCreativeDestruction",Econometrica,LX, 323-351.
ANGRIST, J. (1995), "The Economic Returnsto Schooling in the West Bank and Gaza Strip",American Economic
Review,85, 1065-1087.
ATKINSON, A. B. and STIGLITZ,J. E. (1969), "A New View of Technological Change",Economic Journal, 79,
573-578.
AUTOR, D., KRUEGER,A. and KATZ, L. (1998), "ComputingInequality: Have ComputersChanged the Labor
Market?"';
QuarterlyJournalof Economics,CXIII, 1169-1214.
BAIROCH,P. (1988) Cities and EconomicDevelopment:Fromthe Dawn of Historyto the Present(Chicago:University
of Chicago Press) translatedby ChristopherBraider.
BASU, S. and WEIL, D. N. (1998), "AppropriateTechnology and Growth",QuarterlyJournal of Economics, CXIII,
1025-1054.
BERNDT, E. (1976), "Reconciling AlternativeEstimates of the Elasticity of Substitution",Review of Economics and
Statistics,58 (1), 59-68.
BLANCHARD,0. (1997), "TheMediumRun",BrookingsPaperson EconomicActivity,89-158.
BRESNAHAN,T. F. and TRAJTENBERG,M. (1995), "GeneralPurposeTechnologies:'Enginesof Growth'?",Journal
of Econometrics,65 (1), 83-108.
CABALLERO, R. J. and HAMMOUR, M. (1998), "Jobless Growth: Appropriability, Factor Substitution
and Unemployment",Carnegie-RochesterConferenceProceedings,Vol. 48, 51-94.
808
CABALLERO,R. J. and JAFFE,A. B. (1993), "How High are the Giants' Shoulders:An EmpiricalAssessment of
Knowledge Spillovers and Creative Destructionin a Model of Economic Growth",in NBER Macroeconomics
Annual 1993 (Cambridge,MA: MIT Press) 15-74.
COEN, R. M. (1969), 'Tax Policy and InvestmentBehavior:A Comment",AmericanEconomicReview,370-379.
DAVID,P. (1975) TechnicalChoice, Innovationand EconomicGrowth:Essays on Americanand British Experiencein
the NineteenthCentury(London:CambridgeUniversityPress).
EconomicJournal,
DRANDAKIS,E. andPHELPS,E. (1965), "AModel of InducedInvention,GrowthandDistribution",
76, 823-840.
EISNER,R. and NADIRI, M. I. (1968), "InvestmentBehaviorand the Neoclassical Theory",Reviewof Economicsand
Statistics,369-382.
FELLNER,W. (1961), "Two Propositionsin the Theory of Induced Innovations",The Economic Journal, 71 (282),
305-308.
FREEMAN,R. (1986), "Demandfor Education",in A. Orley and L. Richard (eds.) Handbookof Labor Economics,
Chapter6, Vol. I (North-Holland)357-386.
GALOR,0. and MAOV,0. (2000), "AbilityBiased TechnologicalTransition,Wage Inequalityand Economic Growth",
QuarterlyJournalof Economics,115, 469-497.
GOLDIN, C. and KATZ,L. (1995), "The Decline of Non-competingGroups:Changes in the Premiumto Education,
1890 to 1940"(NBERWorkingPaperNo. 5202).
GOLDIN, C. and KATZ, L. F. (1998), "The Origins of Technology-skill Complementarity",Quarterly Journal of
Economics,113, 693-732.
GROSSMAN, G. and HELPMAN,E. (1991) Innovationand Growthin the Global Economy(Cambridge,MA: MIT
Press).
HABAKKUK, H. J. (1962) American and British Technologyin the Nineteenth Century:Searchfor Labor Saving
Inventions(CambridgeUniversityPress).
AmericanEconomic
HECKMAN,J., LOCHNER,L. andTABER,C. (1998), "TaxPolicy andHumanCapitalFormation",
ReviewPapersand Proceedings,88, 293-297.
HICKS,J. (1932) The Theoryof Wages(London:Macmillan).
JAFFE,A. B., TRAJTENBERG,M. and HENDERSON,R. (1993), "GeographicLocalizaion of KnowledgeSpillovers
as Evidencedby PatentCitations",QuarterlyJournalof Economics,CVIII, 577-598.
JAMES,J. A. and SKINNER,J. S. (1985), "TheResolutionof the Labor-ScarcityParadox",Journalof EconomicHistory,
45,513-540.
JONES, C. I. (1995), "R&D-basedModels of Economic Growth",Journalof Political Economy,103, 759-784.
KATZ, L. and AUTOR, D. (2000), "Changesin the Wage Structureand EarningsInequality",in 0. Ashenfelterand
D. Card(eds.) TheHandbookof LaborEconomics,Vol. III (Amsterdam:Elsevier).
KENNEDY, C. (1964), "InducedBias in Innovationand the Theory of Distribution",Economic Journal, LXXIV,
541-547.
KILEY,M. (1999), "TheSupply of Skilled Laborand Skill-BiasedTechnologicalProgress",EconomcJournal.
KRUSELL, P., OHANIAN, L., RIOS-RULL, V. and VIOLANTE, G. (2000), "Capital Skill Complementaryand
Inequality",Econometrica,68 (5), 1029-1053.
LLOYD-ELLIS,H. (1999), "EndogenousTechnologicalChangeand Wage Inequality"AmericanEconomicReview,89,
47-77.
in A. Harbergerand M. J. Bailey (eds.) The
LUCAS, R. E. (1969), "Labor-CapitalSubstitutionin U.S. Manufacturing",
Taxationof Incomefrom Capital (Washington,DC: The BrookingsInstitution)223-274.
MOKYR, J. (1990) The Levers of Riches: Technological Creativity and Economic Progress (New York: Oxford
UniversityPress).
NELSON, R. and PHELPS, E. (1966), "Investmentin Humans, Technological Diffusion and Economic Growth",
AmericanEconomicAssociation Papersand Proceedings,Vol. 56, 69-75.
NORDHAUS, W. (1973), "Some Skeptical Thoughts on the Theory of Induced Innovation",Quarterly Journal of
Economics,LXXXVII, 208-219.
REBELO,S. (1991), "Long-runPolicy Analysis and Long-runGrowth",Journalof Political Economy,99, 500-521.
RIVERA-BATIZ,L. A. and ROMER,P. M. (1991), "EconomicIntegrationand EndogenousGrowth",QuarterlyJournal
of Economics,106, 531-555.
ROMER,P. M. (1990), "EndogenousTechnologicalChange",Journalof Political Economy,IIC, S71-S 102.
ROTHBARTH,E. (1946), "Causesof the SuperiorEfficiencyof the USA Industryas Comparedwith BritishIndustry",
EconomicsJournal.
SALTER,W. E. G. (1966) Productivityand TechnicalChange 2nd edition (Cambridge,U.K.: CambridgeUniversity
Press).
SAMUELSON,P. (1965), "A Theory of InducedInnovationsalong Kennedy-WeisackerLines",Reviewof Economics
and Statistics,XLVII, 444-464.
SCHMOOKLER,J. (1966) Inventionand Economic Growth(Cambridge:HarvardUniversityPress).
SCHULTZ,T. (1975), "The Value of the Ability to Deal with Disequalibria",Jouranl of Economic Literature,13,
827--846.
ACEMOGLU
DIRECTED TECHNICAL CHANGE
809
SEGERSTROM,P. S., ANANT, T. and DINOPOULOS,E. (1990), "ASchumpeterianModel of the ProductLife Cycle"'
AmericanEconomicReview,80, 1077-1092.
STEWART,F. (1977) Technologyand Underdevelopment(London,U.K.: The MacmillanPress Ltd.).
TRAJTENBERG,M., HENDERSON, R. and JAFFE,A. B. (1992), "IvoryTower Vs. CorporateLab: An Empirical
Study of Basic Researchand Appropriatability"(NBER
WorkingPaper4146).
WILLIAMSON,J. (1990) Coping with City Growthduring the British IndustrialRevolution(Cambridge:Cambridge
UniversityPress).
WOOD,A. (1994) North-SouthTrade,Employmentand Inequality:ChangingFortunesin a SkillDriven World(Oxford:
ClarendonPress).
XU, B. (2001), "EndogenousTechnologyBias, InternationalTradeand RelativeWages"(Mimeo, Universityof Florida).
YOUNG, A. (1993), "Inventionand BoundedLearningby Doing" Journalof Political Economy,101, 443-472.

Directed Technical Change

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Directed Technical Change

Uploaded by

Copyright:

Available Formats

The Review of Economic Studies Ltd.

Directed Technical Change

Review of Economic Studies (2002) 69, 781-809

o College wage premium

Relative Supplyof College Skills and College Premium

artisansstartedto be producedin factoriesby workerswith relativelyfew skills, and many

REVIEW OF ECONOMIC STUDIES

Constanttechnology and endogenous technology relative demandcurves. Constanttechnology: A -+ B. Endogenous

(6) Why a largewage push, as in continentalEuropeduringthe 1970's, may cause capital-biased

spurredtechnological progress (see also Rothbarth(1946)). This literaturewas also criticized

is no substitutionbetweenthe two factors,andthe productionfunctionis Leontieff.Whena > 1,

The relativemarginalproductof Z is decreasingin the relative abundanceof Z, Z/L. This is

3. THE DEMAND SIDE

+ ___ YYL + (1 Y)YzE

XL(j) -dJ )L',

My interest is with the determinantsof the direction of technical change. As discussed

reduce the relativeprice p

This expressionshows thatthe relativeprofitabilityof the two types of innovationaredetermined

4. THE SUPPLYOF INNOVATIONS:THE INNOVATIONPOSSIBILITIESFRONTIER

4.1. The directionof technical change with the lab equipmentmodel

= rlz/IlL is constant irrespective of the levels of NL

Comparingthis equationto (18), which specified the relativefactorprices as a functionof

Next define the relativefactor sharesas

Relativefactorsharesare then obtainedas

marginalreturnto differenttypes of innovation,and for an equilibriumin which there are both

et al. (1992). Looking at patentcitations,they constructan index of whethera patentfroma given

+ (1 - y)((p')(1P-)/3NHH')_-. .l-- L) e + (1 - y)(p-)/

Some tedious algebra(see Acemoglu (2001)) shows that

where HW/LW = (H +KH')/(L

+KL') is the world (effective) ratio

of skilledto unskilledworkers.The parameterK featuresin this equationbecauseit parameterizes

more skilledworkersin the North,anddirectedtechnicalchangeinducestechnologymonopolists

trade,by makingskills more scarcein the North,will induce moreinnovationsdirectedat skilled

Comparingthis equationto (21), we see immediatelythat, because A > 1, tradeincreases the

Comparingthis equationto (18) from before, we see thattradeopening necessarilyincreasesthe

the endogenous growthliteratureemphasizesthe presence of a scale effect which suggests that

Consider anotherregularitydiscussed in the introduction:the share of labour and capital have

model of directedtechnicalchange be useful in thinkingaboutwhy aggregatetechnicalchange

Here r > 0, and tt is a shift parameter.A decrease in It correspondsto a "wage-push"shock,

to a more inelastic quasi-supplycurve. This formulationis convenient since it yields a simple

Comparisonof (39) and (40) immediatelyimplies thatthe elasticity of employmentwith respect

increasein the cost of labourresultingfromthe wage push, firmscannotsubstantiallyreducetheir

effect and the marketsize effect. The elasticity of substitutionbetween differentfactorsregulates

DIRECTED TECHNICAL CHANGE

You might also like