Full text of DOI:10.1162/003355300555051

CREATIVE DESTRUCTION AND FIRM
ORGANIZATION CHOICE*
DAVID THESMAR AND MATHIAS THOENIG
Firms’organizational choices are influenced by external conditions such as the
instability of the product market. In order to address this issue in a macroeconomic
perspective, we embed the firm’s choice of organizational structure in a model of
growth through creative destruction, which induces endogenous market volatility.
We find that an increasing supply of skill or globalization may increase the rate of
creative destruction, the skill premium, and the skilled wages, and it may depress
the unskilled wages. We use an original data set to test the empirical relevance of
our theory.
I. INTRODUCTION
In order to cope with increasingly unstable product markets
and to exploit smaller and more rapidly changing customer
niches, many firms in industrialized countries have sought to
change their organization since the mid-1970s. In opposition to
former work organization, the new organizational paradigm is
supposed to achieve greater flexibility, adaptability, and reactivity
through decentralized decision making, product-based hierarchy,
unwritten rules, and multitasking workers.1 These organizational
changes may have had strong consequences at the macroeconomic
level. For example, Piore and Sabel [1984] and Chandler [1990]
discuss in their influential contributions the influence of the
dominant forms of firms’ organization on macroeconomic stability.
A more recent empirical literature claims that reorganization
* We thank Daron Acemoglu, Philippe Aghion, Roland Be´nabou, Eve Caroli,
Daniel Cohen, Bruno Cre´pon, Antoine d’Autume, Mohamad Hammour, Francis
Kramarz, David Margolis, Eric Maurin, J ohn Van Reenen, and Thierry Verdier.
This paper also benefited from insightful remarks from one referee and two editors
of this J ournal. Last, we thank participants of various seminars (Ecole Normale
Supe´rieure, Universite´ Paris I, European Economic Association 1998 Conference,
and the 1999 Workshop on Income Distribution and Growth at the National
Bureau of Economic Research). All remaining errors are ours.
1. A thorough statistical assessment of the extent of these changes is still on
the research agenda. However, a recent empirical literature provides evidence that
these changes cannot be ignored. In 1992 Osterman [1994] conducted a survey of
694 U. S. firms. He found that in 64 percent of all sectors, firms introduced new
working practices concerning at least half of their workforce (for example, 40.5
percent introduced self-directed team working, and 26 percent frequent job
rotations). According to a French survey conducted by N. Greenan for the Ministry
of Industry, 39 percent of French firms (with more than 50 employees) said to have
reorganized between 1988 and 1993 (this rate goes up to 62 percent for firms above
1000 workers). Eighty percent of them aimed at reducing delivery time, while 60
percent wanted to be able to adapt more easily to changes in the production
process.
2000 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
௠
The Quarterly J ournal of Economics, November 2000
1201

1202
QUARTERLY J OURNAL OF ECONOMICS
influences the labor market through a skill upgrading of firms’
occupational structures [Caroli and Van Reenen 1999; Bresnahan,
1998]. Furthermore, Goldin and Katz [1998] show that the
replacement of unskilled labor by skilled workers and machines
began early in the twentieth century. They view this skill upgrad-
ing as the consequence of organizational change, mainly driven by
technological change.
In this paper we develop a theoretical model to assess the
interaction between organizational choice and the macroeconomy.
We argue that firms constantly have to adapt to the pace of
product market change which implies that organization choices
represent a trade-off between efficiency and adaptability. In our
framework this choice feeds back into the product market by
reinforcing its instability. We show that, as a result, the product
market alters labor demand by skill through reorganization.
The core mechanism borrows from Piore and Sabel’s [1984]
insight. We assume that complex organizations can enhance their
productive efficiency through the payment of a product-specific
sunk cost.2 As a result, product market instability, by increasing
the rate of obsolescence of internal procedures, regulations, and
equipment, is an important determinant of organization.3 This
sunk cost modeling strategy is very crude, but it allows us to
assess very clearly the macroeconomic impact and determinants
of organization choice. It is founded on past developments in the
sociology of organizations, where contingency theory [Burns and
Stalker 1961] views firms’ organizational structures as endoge-
nous to the external environment4 and in particular to the product
2. In Piore and Sabel’s words: ‘‘Mass production offered those industries in
which it was developed and applied enormous gains in productivity [. . .] But these
gains had a price. Mass production required large investments in highly special-
ized equipment and narrowly trained workers. In the language of manufacturing,
these resources were ‘‘dedicated:’’ suited to the manufacture of a particular
product—often, in fact to just one make or model. When the market of one
particular product declined, the resources had no place to go.’’
3. Many economists consider available production technology to be the main
determinant of organization choice (see, for example, Caroli and Van Reenen
[1999]).
4. As Daft [1998] puts it: ‘‘When the external environment was stable, the
internal organization was characterized by rules, procedures and a clear hierarchy
of authority. Organizations were formalized. They were also centralized, with most
decisions made at the top. Burns and Stalker called this a mechanistic organiza-
tion system.
In rapidly changing environments, the internal organization was much looser,
free flowing, and adaptative. Rules and regulation often were not written down, or
if written down, were ignored [. . .] The hierarchy of authority was not clear.
Decision making authority was decentralized. Burns and Stalker used the term
organistic to characterize this type of management structure.’’

CREATIVE DESTRUCTION
1203
market instability. According to this literature, environment
matters because there is a fundamental dilemma between organi-
zation and innovation: larger organizations allow substantial
productivity gains (through division of labor, detailed definition of
tasks, specially designed hierarchies, and control) but resist
implementation of innovations for two reasons. First implementa-
tion requires studies and modifications that, by definition, com-
plex organization has not been designed to deal with [March and
Simon 1958]. Second, large organizations favor collusion between
employees, whose rents are endangered when innovations are
introduced [Crozier 1963]. Thus, implementation costs are posi-
tively correlated with the complexity of organizational design
[Perrow 1970].5
We use a Schumpeterian growth model a` la Aghion and
Howitt [1992] where the creative destruction rate provides a
natural interpretation for product market instability: each innova-
tion creates a temporary monopoly rent, while destroying the
incumbent’s market power. Skilled workers participate in produc-
tion and R&D, while unskilled workers can only produce. Environ-
mental instability is captured by a research sector that sells
patents to the producing sector. J ust after purchasing a patent,
new monopolies can choose between an organization that provides
high productivity at the expense of a sunk cost (mechanistic), and
one that provides low productivity without sunk costs (organis-
tic):6 organizational complexity has an innovation-specific compo-
nent. Intuitively, the higher the innovation rate in the economy,
the shorter the project’s life expectancy, and the less the sunk cost
of being well organized proves profitable. This is how the macro-
economy affects organizational choice. On the other hand, changes
in organizational structure, by changing firms’profitability, patent
prices, and research activity, do influence the growth rate of the
economy and the wage distribution, so that organizational choices
feed back into aggregate variables.
This framework provides a macroeconomic theory of interac-
tion between product market instability and organizational choice.
First, it provides a natural macroeconomic foundation for endo-
genizing segregation by skill. Second, our model provides an
5. In the economic literature this innovation versus organization antagonism
has been recently explicitly modeled by Martimort and Verdier [1998] who use
developments in agency theory to analyze the relation between creative destruc-
tion and firm bureaucratization.
6. This terminology is borrowed from Burns and Stalker ’s work.

1204
QUARTERLY J OURNAL OF ECONOMICS
integrated explanation to stylized facts on recent trends of the
U. S. labor market [Gottshalk 1997]: an increase in the relative
supply of skilled workers, an increase in the skill premium, and a
decrease in the unskilled real wage. Third, it predicts that the
degree of specialization in production decreases with market size,
at odds with standard economic literature (Murphy, Shleifer, and
Vishny [1988], for instance). The intuition behind this result is
that increased market size generates more product market insta-
bility, rendering product-specific specialization less affordable to
firms. Finally, we show that the process of globalization, when
arising largely between similar countries, is more likely to have
strong consequences for a firm’s organization and labor demand
by skill because of increased competition.
Our model thus suggests that recent shifts in skilled labor
demand must have been related to rising product market turbu-
lence. The second part of the paper provides original evidence that
this interpretation is empirically relevant. We rely on a French
data set that allows us to disentangle functional and skill
structure at the firm level between 1984 and 1995. At the
aggregate level, we find that half of the increase in skilled labor
demand is accounted for by employment reallocation from (un-
skilled-worker-intensive) production to (skilled-worker-intensive)
development and commercialization. Furthermore, we provide
evidence of an increase in research and creative destruction since
the 1970s, suggesting that product markets have become more
unstable. Then, at the firm level, we show that the degree of
occupational reallocation depends heavily on various measures of
market instability such as the number of researchers and the
share of new products in total sales. The main conclusions of the
model are thus supported by the data, firms have experienced
more product or process turnover, and were thus forced to change
their occupational mix from unskilled-intensive production to
more skill-intensive customization, design of new products, and
trade-related activities.
We see this paper as providing a link between two strands of
the literature. The first, seeking an explanation for the dramatic
rise of U. S. wage inequality for the past two decades, considers
how organizational change feeds back into macroeconomic vari-
ables, in particular the wage distribution and growth (Lindbeck
and Snower [1997], Acemoglu and Newman [1997], and Acemoglu

CREATIVE DESTRUCTION
1205
[1999] among others).7 However, none of these models explicitly
accounts for product market stability. The second strand of
literature emphasizes the relation between uncertainty (techno-
logical or creative destruction) and organization [Piore and Sabel
1984; Aoki 1986; Martimort and Verdier 1998]. In this sense, our
paper is close in spirit to Galor and Moav [2000], who emphasize
that technical change is skill biased in the short run because it
renders low skilled know-how obsolete.
This article is organized as follows. Section II is devoted to
presenting the model. Section III discusses implications for the
wage distribution, economic development, and international trade.
In Section IV we provide macroeconomic and microeconometric
evidence in line with the predictions of our model. Section V
concludes and gives directions for further research.
II. THE MODEL
A. Basic Framework
The framework presented below is in the style of the Schum-
peterian growth model developed by Aghion and Howitt [1992]
and Grossman and Helpman [1991].
There exist two types of labor, skilled and unskilled, whose
exogenous supplies are H and U, respectively. A representative
consumer optimizes her intertemporal utility:
(1)
U ϭ ϱ ln (Cs)eϪ␳(sϪt) ds,
͐
t
t
where Ct is an index of consumption at date t and ␳ is the
subjective rate of time preference. Financial markets are assumed
to be perfect. If we define Et as aggregate spending, straightfor-
˙
ward dynamic utility maximization yields Et/Et ϭ (rt Ϫ ␳). We
then normalize agreggate spending to 1, which ensures that
᭙t, rt ϭ ␳. The consumption index is a function of a continuum of
goods i ʦ [0;1], subject to quality improvements through innova-
tions. It is given by
1
(2)
ln C ϭ
ln (␭s(i) · ct (i)) di,
͐
t
0
where ␭ Ͼ 1, s(i) denotes the number of innovations experienced
7. Another literature attributes temporary increases in wage inequality to
the diffusion of a new technology (see Helpman and Trajtenberg [1994]).

1206
QUARTERLY J OURNAL OF ECONOMICS
by good i since the beginning of time and ct (i) is the quantity of
good i consumed at date t. Under this specification the demand
faced by sector i is given by
(3)
xt(i) ϭ 1/pt(i).
Research laboratories produce innovations according to a
Poisson process ␪, using skilled labor only.8 Since the set of sectors
is of measure 1, ␪ is both the aggregate research effort and the flow
probability that a given sector receives an innovation. In this
activity, research technology is of constant returns; that is, for a
given lab i, ␪ ϭ h RD/b, where h Ri D is the number of employed
i
i
researchers. We assume, for the sake of simplicity, that innova-
tions cannot be perfectly directed toward one particular sector:
hence, while searching, a given lab does not know which sector its
innovation will be designed for. Summing over labs9 yields an
aggregate production of innovations:
b␪ ϭ hRD
.
Once found, patents are sold by laboratories to an infinite number
of potential final good producers; hence a successful lab can
capture the whole value of patent exploitation.10
It must be noted that people working in the R&D sector need
not be thought of as being researchers only. ‘‘R&D’’ refers to all
workers whose job is to create and exploit temporary rents, and as
such includes not only engineers and product designers who
create improved versions of existing products, but also marketing
departments and salesmen who spot new needs, help to design
customized products, and work at recovering market share [Stalk
and Webber 1993; Daft 1998].
In each sector i, risk-neutral producers use both skilled and
unskilled labor according to a constant returns technology y ϭ
a · (h)␣ · (u)1Ϫ␣, where a is the endogenous level of productivity
(see below) and h and u are the number of skilled and unskilled
workers, respectively. Different patent owners in sector i compete
in price to sell their good to the consumer. As is standard in this
8. The only assumption needed in order to establish our results is that R&D
technology is more skill intensive than the productive sector. This assumption is
made just to simplify the calculations.
9. We are assuming independence between the different processes here.
10. As in most Schumpeterian growth models, the producing monopoly
chooses to invest nothing in research because the efficiency effect is strictly
dominated by the replacement effect. This point can be easily shown in our context
of Bertrand competition.

CREATIVE DESTRUCTION
1207
literature, in equilibrium, only one supplier actually produces: the
one with the highest quality to price ratio. Note here that given
the Cobb-Douglas specification, the unit cost function of a firm is
given by
(4)
UC(a,ws,wu ) ϭ uc(ws,wu )/a,
where uc(ws,wu ) ϭ ␣Ϫ␣(1 Ϫ ␣)1Ϫ␣ · w␣s · w1uϪ␣ and (ws,wu ) are the
wages of skilled and unskilled workers, respectively.
Once the firm has bought the patent, an innovation-specific
production process needs to be implemented. J ust after the
purchase of the patent, the firm can choose between two produc-
tion technologies: one with a high productivity a which entails an
innovation-specific sunk cost C,11 or one with a low productivity a,
not incurring any fixed cost.
This assumption can be viewed, as in Murphy, Shleifer, and
Vishny [1989], as the choice between a production technology with
Increasing Returns to Scale (IRS) when the firm pays the sunk
cost, and one with Constant Returns to Scale (CRS). From the
point of view of the organization of production, the CRS firm is less
efficient because it makes little use of labor division, therefore
losing gains from specialization. Also, different production units
may suffer from coordination failures due to a lack of centraliza-
tion. According to Demsetz [1997], externalities between these
units are not internalized enough by the management: bottle-
necks, excess inventories, and inefficient resource (labor, capital)
allocation to the different tasks may occur. Alternatively (or
additionally), the CRS firm can be thought of as a firm investing
little in innovation-specific capital goods. On the other hand, the
IRS firm makes intensive use of innovation-specific machines,
relies on an extensive division of labor, with very precisely defined
tasks and a structured hierarchy; it therefore achieves a higher
degree of efficiency at producing, but incurs coordination costs as
in Becker and Murphy [1992].
As stated above, this paper emphasizes the innovation-
specific nature of coordination costs. The detailed definition of
11. We assume that C is ‘‘wasted money:’’ the lost returns do not show up in
other general equilibrium equations. C can also be viewed as an opportunity cost of
wasting time in adapting a complex organization to the new process. Our results
would be qualitatively unchanged under the assumption that C is a nominal sunk
cost which corresponds to the consumption of final goods (see Appendix 1 for more
details). We also investigated the case where C corresponds to additional labor
consumption: depending on the skill intensity of C with respect to the skill
intensity of the production process, our results would be weakened or enhanced
(we thank Roland Be´nabou for drawing our attention to this point).

1208
QUARTERLY J OURNAL OF ECONOMICS
tasks and design of an appropriate organization needed to orga-
nize the division of labor, the definition of relations between
production units, the scheduling of tasks, the design of an
appropriate hierarchy, and the purchase of ‘‘dedicated’’ equipment
are all, to some extent, up-front innovation-specific costs associ-
ated with a more efficient use of production (mostly unskilled)
workers. The central assumption here is thus that production
(mostly unskilled) labor is complementary to an organizational
structure that entails an innovation-specific sunk cost.
B. Competition at the Microeconomic Level
We first perform a partial equilibrium analysis, in order to
uncover the determinants of firms’ organizations at the sectorial
level. So, in this subsection we take the product market rate of
creative destruction, ␪, as given (␪ is indeed set at the macroeco-
nomic level12). For the sake of simplicity, we omit here the
sectorial subscript i.
First, consider a firm that has just purchased a patent and
has to decide on its production technology. Its main competitor
(the most up-to-date incumbent) has productivity aϪ1. Once the
new entrant has chosen technology a, firms in the sector will
engage in Bertrand competition. In the absence of capacity
constraints, the producer providing the highest quality price ratio
will serve the whole demand, adjusting its price such that its less
efficient competitors make negative profits. Thus, the new en-
trant13 will crowd the incumbent out of the market if
␭a Ͼ aϪ1
.
Hence, the new entrant will always make positive profits,
until another innovation comes out. The entrant’s expected value
of profits is thus given by
(1 Ϫ aϪ1)/␭a
(5)
V(a;aϪ1,␪) ϭ
Ϫ C · މ5aϭa6,
r ϩ ␪
where މ5aϭa is an indicator function.
12. This stems from our assumption that a lab cannot direct its innovations
toward a given sector. In making this assumption, we rule out the strategic
interaction whereby a firm deters further innovations in its own sector, by
influencing the future incumbent’s value and research activity.
13. We make the simplifying assumption that the quality improvement is
large enough to compensate the possibility that the new entrant has chosen a less
productive production technology than the incumbents. In other words, we allow a
workshop to be able to crowd out a Tayloristic factory. This assumption could be
relaxed without altering the main results, although at the cost of some complica-
tions and a lot of expository complication.

CREATIVE DESTRUCTION
1209
For a given ␪, the entrant’s reaction function is easily
computable:
⌬a
a if
a if
␪ Ͼ aϪ1
␪ Ͻ aϪ1
·
·
Ϫ r
Ϫ r
aa␭C
⌬a
a *(aϪ1,␪) ϭ
,
5
aa␭C
with ⌬a ϭ a Ϫ a. For a given ␪, the entrant’s optimal technology
will be more efficient when the incumbent is more efficient; i.e.
there is a strategic complementarity between both competitors. In
economic terms, this means that if the other firm’s productivity is
low, the competitive drawback of not being efficient is less than
the sunk cost of specialization.
Second, if the rate of creative destruction is large enough, the
CRS technology will always be preferred. Indeed, since the
innovation life cycle (ϭ1/␪) is short, the firm is better off adopting
the CRS organization than paying the innovation-specific sunk
cost associated with a more efficient production process: this
effect, at the heart of Piore and Sabel’s [1984] view of flexible
specialization, is the direct consequence of the macro-level exter-
nality R&D exerts on the firm’s choice of organizational structure.
Indeed, an increase in the rate of destruction decreases the
marginal return to being efficient. This externality will play an
important role in the general equilibrium analysis.
In this paper we consider only the symmetric equilibria. We
thus arrive at the following straightforward proposition.
PROPOSITION 1. (Characterization of the possible equilibria).
1. All firms choosing a is an equilibrium if and only if
⌬a
␪ Ͼ
Ϫ r ϭ ␪.
␭Ca
In what follows, we call this the CRS equilibrium.
2. All firms choosing a is an equilibrium if and only if
⌬a
␪ Ͻ
Ϫ r ϭ ␪.
␭Ca
In what follows, we call this the IRS equilibrium.
Hence, the characterization of sectorial equilibrium depends
on the rate of creative destruction. This is a direct consequence of

1210
QUARTERLY J OURNAL OF ECONOMICS
the externality that R&D exerts on a project’s discounted rate of
return. Moreover, the strategic complementarity mentioned above
is strong enough to ensure the existence of multiple equilibria for
intermediate values of ␪, since ␪ Ͻ ␪ for all parameter values.
C. General Equilibrium
In what follows, we restrict ourselves to the study of station-
ary symmetric equilibria; that is, the organizational structure is
the same in all sectors. Since potential innovation buyers compete
a` la Bertrand to buy the patent, its price is exactly the value of the
firm. We get, from equation (5),
1 Ϫ 1/␭
(6)
(7)
V(␪) ϭ
Ϫ C މ5aϭa6
r ϩ ␪
a ϭ a if ␪ Ͻ ␪ and a ϭ a if ␪ Ͼ ␪.
Note that this value function depends on the firm’s organizational
choices by way of the sunk cost C, and not through the firms’
productivity levels a. This is because, in equilibrium, only relative
productivities matter.14 Consequently, for a given ␪, the patent
value is unambiguously higher in the CRS equilibrium than in the
IRS one.
Free entry in R&D equalizes the costs and the benefits of the
activity. Thus,
(8)
bws ϭ V(␪).
Labor market clearing conditions are given by the following
sets of equations:
(9)
H ϭ b␪ ϩ ␣/␭ws,
(10)
U ϭ (1 Ϫ ␣)/␭ · wu.
Hence, the full characterization of the steady state general
equilibrium is given by equations (6)–(8) and (9)–(10).
The right-hand side of equation (9) represents skilled labor
demand expressed in terms of ␪ and ws. By using equations
(6)–(8), we may express this demand in terms of ␪ alone (as shown
in Figure I). When all firms use an IRS production technology, the
supply of skilled workers needed to achieve a given rate of creative
14. This is particulary obvious under Bertrand competition. Nevertheless,
Our results would continue to hold in any general equilibrium model, where
demand depends on relative prices (perfect competition with vintage capital as in
Aghion and Howitt [1994], or monopolistic competition as in Romer [1990]).

CREATIVE DESTRUCTION
1211
FIGURE I
Aggregate Labor Demand in CRS and IRS Equilibria
destruction ␪ is larger, as production is more efficient, and
therefore employs more of them. This is why the labor demand
curve in the IRS regime lies above the CRS one. Thus, strategic
choice of organizational structure renders labor demand nonmono-
tonic, and for intermediate values of the skilled labor supply, the
horizontal supply curve intersects the demand curve twice.
PROPOSITION 2. Description of Equilibria: ᭚(H,H ) such that H Ͻ H
and
1. if H Յ H, then the economy is IRS in all sectors;
2. if H Ն H, then the economy is CRS in all sectors;
3. if H Յ H Յ H, both equilibria coexist.
This result highlights the key role of a second strategic
complementarity that arises in general equilibrium through the
R&D sector. Indeed, let all other firms choose to be CRS. As noted
above, this common strategy raises their value, and hence R&D’s
marginal productivity (recall that the labs’ expected flow of profits
is given by (V/b) · hRD),15 which in turn raises R&D’s output, i.e.,
the growth rate ␪. But from Proposition 1 we know that a higher
rate of creative destruction generates an externality on the
sectors’ choices, rendering more likely the existence of a CRS
equilibrium.
In summary, the externality that firms’ decisions impose on
15. Note that this effect is robust to a Decreasing Returns to Scale specifica-
tion of the R&D technology: ␪ ϭ l␣/b.

1212
QUARTERLY J OURNAL OF ECONOMICS
R&D, and the feedback effect of the level of R&D activity on both
organizations’ comparative advantages, combine to form a strate-
gic complementarity in general equilibrium. This complementar-
ity, along with the sectorial one highlighted in the preceding
subsection are responsible for the coexistence of IRS and CRS
equilibria at the macro level.
III. DISCUSSION
A. Equilibrium Switch
In this section we aim to investigate the pure effect of a
change in organizational structure. In order to do this, we
compare the features of both equilibria for a given labor supply
H ʦ [H; H ]. This allows us to abstract from supply effects. We
present comparative statics results when Reorganization occurs;
i.e., when the economy switches from an IRS to a CRS equilibrium.
Production in the CRS equilibrium is less ‘‘skill consuming’’
(the CRS labor demand curve lies below the IRS one in Figure I): ␪
is unambiguously larger in the CRS equilibrium than in the IRS
equilibrium. As discussed above, for a given ␪, project values are
higher if all firms choose the CRS organization because nobody
pays for the sunk costs, while all firms have the same productivity.
This ensures that the flows of profits are not lower than in the IRS
equilibrium. As a result, R&D’s productivity is higher, as is R&D’s
demand for skilled labor and hence ␪. In other words, firms’
reorganizations, by increasing the price of patents, tend to create
a bias toward R&D activity and innovation.
RESULT 1. The rate of creative destruction is higher in the CRS
equilibrium (␪
Ͼ ␪IRS ), and the stock market value of a
CRS
firm is higher (VCRS Ͼ VIRS ).
The flow of output in equilibrium is given by
(11)
X ϭ a · U1Ϫ␣,(H Ϫ b␪)␣.
As outlined above, skilled workers move from production to
R&D as firms reorganize; at the same time, firms’ reorganization
reduces productivity (a Ͻ a). Hence, the production flow is unam-
biguously lower in the CRS equilibrium for two reasons. First,
productivity goes down. Second, R&D becomes comparatively

CREATIVE DESTRUCTION
1213
more productive, attracting more skilled workers, which leaves
fewer skilled workers in the productive sector.
RESULT 2. The production level is lower in the organistic (CRS)
equilibrium (XCRS Ͻ XIRS ).
One might argue that our results are inconsistent with the
fact that after reorganization, firms often see a massive increase
in their labor productivity (see Askenazy [1998] for the United
States). However, standard measures of productivity are obtained
by dividing value added by the number of workers. An equivalent
measure in our framework would be obtained by multiplying the
number of projects undertaken in a year by their values and
dividing by the workforce. In the CRS equilibrium we thus expect
firms to undertake more projects (since their life cycle is shorter)
that have larger values, and to use fewer employees. Hence, in our
model, we would expect the empirical measure of productivity to
rise when the economy switches from the IRS to the CRS
equilibrium. To sum up, IRS firms have a higher ‘‘static’’productiv-
ity (a Ͻ a), while CRS firms have a higher productivity in a
dynamic sense and they also undertake more projects.
The change of equilibrium affects the real wage distribution
through three channels. First, since the production process is less
efficient, the supply of the final good goes down, which raises its
price. As a result, this ‘‘disorganization’’ effect depresses real
wages for both types of labor. Second, as skilled workers leave
production for research, the marginal gain of hiring production
workers goes down. Third, demand for labor in R&D is higher,
which boosts the skilled wage. This ‘‘innovation push’’ effect has
no direct effect on the unskilled wage. Channels 2 and 3 combine
into what we called above the ‘‘segregation effect;’’ as a result of
widespread reorganization, skilled workers leave the less skill-
intensive activity to join the more skill-intensive one. This result
is consistent with the theory and evidence presented in Kremer
and Maskin [1995], although the source of segregation here is a
change in equilibrium, rather than a change in labor force
composition. However, we show below that a change in labor force
composition can be at the origin of this shift in organizational and
technological paradigm, thereby reconciling our results with
those of Kremer and Maskin.
A switch from an IRS to a CRS equilibrium thus yields an
unambiguous decrease in the unskilled real wage, while the

1214
QUARTERLY J OURNAL OF ECONOMICS
skilled wage may increase if the ‘‘innovation push’’ effect domi-
nates the ‘‘disorganization’’ effect. To summarize, both types of
workers suffer from the lower productivity of the production
sector, but the improved profitability of R&D can outweigh this
lower productivity for skilled workers. The ‘‘ever-changing’’
economy, which emphasizes reactivity and cheap implementation
of short-lived innovations, harmful to the unskilled, who have a
lower productivity, but it can benefit the skilled, who participate
in making the inventions (the ‘‘noise’’).
Next, we measure inequality as the skilled/unskilled wage
ratio. We show in the appendix that the CRS equilibrium is
associated with more inequality and a higher skill premium:
ws
wu
␣
U
(12)
ϭ
·
.
1
2
1 Ϫ ␣ H Ϫ b␪
RESULT 3. Wage inequality, as measured by the skill premium, is
higher in the CRS equilibrium (ws /wu )CRS Ͼ (ws /wu )IRS than
in the IRS equilibrium.
Since innovations in this model appear through quality, and
not productivity, output and real wages are stationary. However,
the instantaneous utility at date ␶ is given by
u (␶) ϭ ␪ · ln ␭ · ␶ ϩ ln X.
Hence, the utility level is increasing in ␶, through successive
improvement of the final goods’ qualities. By considering prefer-
ences instead of output, we again find a positive effect of innova-
tion on both skilled and unskilled workers. Although the real
wages of unskilled workers are decreasing, their instantaneous
utility increases faster after reorganization, at least in the long
run.
B. Skilled Workers and Retooling
According to Goldin and Katz [1998], the fact that a change in
organizational structure and the installation of new machines
require skilled labor, is the source of skill complementarity. In this
section we discuss briefly, in a partial equilibrium context, the
implications of this assessment in our framework.
Consider an extension of our model where the innovation-
specific sunk cost is C units of skilled labor for a mechanistic
organization and C units of skilled labor for an organistic organi-
zation, and C Ͼ C. Aggregate demand for skill is now composed of

CREATIVE DESTRUCTION
1215
demand for producers, researchers, and people who can install
new equipment (retoolers). Let us consider a shift from the
mechanistic equilibrium to the organistic one.
The effects on the first two components of skill demand have
already been discussed above. This is why we focus on the last
component, the demand for retoolers, which is equal to ␪C. In the
organistic equilibrium, demand for retoolers decreases (C goes
down). But, as ␪ increases, firms need to retool more often, which
increases demand for installation. The consequence of an equilib-
rium shift on demand for skilled installers is thus ambiguous, but
if the latter effects dominates, increased product market instabil-
ity is even more favorable to skilled workers because of increased
frequency of retooling.
C. Market Size and Organizational Change
Our model is also able to shed new light on the relation
between market size and the degree of division of labor. To
illustrate this point, we assume in the following sections that the
number of sectors is of measure N. ␪ is now the flow probability
that a given sector receives an innovation (the relevant measure
of creative destruction). ⌰ ϭ N␪ is the aggregate research effort.
All equations are unchanged, except (9).
(13)
H/N ϭ b␪ ϩ ␣b(r ϩ ␪)/(␭ Ϫ 1 Ϫ C␭(r ϩ ␪) · މ5aϭa6 ).
Thus, the choice of the organizational structure depends on
the sectorial ␪, which is in turn determined by the ratio H/N.
In what follows, we call an expansion of market size, an
equiproportional increase in both labor supplies H and L. In our
framework, if N remains fixed, this expansion is favorable to the
CRS equilibrium (Proposition 1), where firms adopt a constant
return to scale production organization, relative to an IRS organi-
zation. Should we be surprised by this result?
According to conventional wisdom since Adam Smith, the
degree of labor division in the economy is limited by the size of the
market. This implies that larger markets will have more competi-
tive IRS firms, because IRS firms more efficiently exploit increas-
ing returns to scale through division of labor. This is not the case
in our model since a larger market here stimulates the R&D
activity and shortens the product life cycle. This renders fixed
costs unsustainable, and therefore CRS firms are more efficient.
This mechanism can be related to the argument suggested by
Becker and Murphy [1992]. They argue that coordination costs

1216
QUARTERLY J OURNAL OF ECONOMICS
within the firm exert a constraint on the degree of labor division
that is more stringent than market size.
This market size effect on growth is standard in models with
imperfect competition. However, J ones [1995] challenges this
mechanism and empirically shows that there is no correlation
between growth and market size. Young [1998] provides a theoreti-
cal framework which allows him to purge Schumpeterian models
of this pure market size effect. He proposes a model of growth in
which both vertical and horizontal differentiation are allowed, but
horizontal differentiation does not benefit from intertemporal
spillovers.
Young’s argument suggests a deeper interpretation of the
effects of market size on the division of labor. Since an equipropor-
tional expansion of labor supply should increase the number N of
commodities available in the economy in the long run, R&D
should remain unchanged, and the choice of organizational struc-
ture should be the same before and after an expansion of size (H/N
is constant in equation (13)). But the story is quite different in the
short run. Young does not consider explicitly the transition
dynamics after a market expansion (in his model, the economy
immediately reaches the steady state equilibrium). But, if we
think of N as a stock variable and ␪ as a forward one, N increases
continuously just after an increase in market size whereas ␪
jumps directly to a higher value and then relaxes. In this case,
instability should temporarily increase and may compel the firm
to switch to a CRS organization. We may thus expect some
hysteresis along the transition path; if firms choose a CRS
organization because of the temporary increase in ␪, they may
remain CRS despite the decline in ␪ that occurs through increas-
ing horizontal differentiation. In sum, if one believes in the
hysteresis effect, market size may have a long-run effect on
organizational choice, and therefore on the degree of labor divi-
sion (see Figure II), even though it has no long-run effect on R&D
activity.
D. Increase in the Skilled Labor Force
Autor, Katz, and Krueger [1998] have documented how the
U. S. college premium sharply increased in the 1980s, while the
college educated share of the workforce also steadily increased
over the period. In addition, Askenazy [1999] shows that among
those U. S. sectors that have reorganized, most did so in the

CREATIVE DESTRUCTION
1217
FIGURE II
Short-Run and Long-Run Effects of a Larger Market Size
mid-1980s. Our model provides a new perspective for understand-
ing these facts.
In order to restrict our attention to pure composition effects in
the labor force, we suppose (in the spirit of Young [1998]) that N ϭ
U ϩ H. An increase in the share of skilled labor may shift the
economy to the CRS equilibrium (equation (13) and Proposition 2).
This has two effects on wage inequality. First, inequality is
reduced through lowering the skilled wage. This is a standard
supply effect. Second, if the skilled labor supply hits the upper
bound H/N, firms face an economywide organizational change. As
a result of this effect, the skill premium rises (see the preceding
section). As an example, Figure III shows the case where both the
supply and organizational change effects are at work, but where
the latter outweighs the former.
Our model is thus able to generate a comovement between the
share of skilled labor and the skill premium. The intuition behind
this outcome lies in the possibility that an increase in skilled labor
supply can induce more than its own demand. This feedback of
supply into demand has been ignored by economic theory until
recently. There is now a small, but growing, body of literature
dealing with these issues. To our knowledge, the first paper
exhibiting these features is Acemoglu [1998a]. He models an
economy where firms post vacancies, choosing in a first step the
quantity of capital associated with each job. Then they search for
workers to fill the vacancies. Labor supply is composed of two
types of workers: productive (skilled) and less productive (un-
skilled). If the supply of skills is scarce, firms will fill their

1218
QUARTERLY J OURNAL OF ECONOMICS
FIGURE III
Increase in Skilled Labor Supply and Equilibrium Shift
vacancies with the first worker they find. In equilibrium, all
posted vacancies will be associated with the same quantity of
capital, since ex ante discrimination is not profitable for firms:
this is his pooling equilibrium. On the other hand, if the supply of
skills is abundant, it will be profitable for firms to create both
highly capital-intensive vacancies, designed for skilled workers
exclusively, and jobs associated with low capital, for unskilled
workers. In this separating equilibrium, skilled wages will be
higher, and unskilled wages lower. In this model, matching
externalities along with hold-up effects, make the expected mar-
ginal return to capital depend on labor force composition.
Our story is quite different. It results from the combination of
two features. First, a CRS economy has a lower demand for
unskilled labor in production, since it does not invest in the sunk
cost. Second, the CRS economy actually has a higher demand for
skilled workers in research. Hence, the economy reacts to a larger
supply of skilled workers by creating even more skilled jobs in
R&D, and by destroying unskilled jobs in production.
E. Economic Integration
The discussion above suggests that our model can provide
some insights into the labor market consequences of international
trade between similar countries. Consider the economic integra-
tion of two countries (1 and 2). Suppose that the skill endowments
of these countries are identical, U1 ϭ U2 and H1 ϭ H2, so that, after
integration, neither country will opt for complete specialization.
The number of goods in each country is Ni ϭ Ui ϩ Hi. We consider

CREATIVE DESTRUCTION
1219
two cases according to the degree of overlap between commodity
markets of both countries (see Grossman and Helpman [1991,
Chapter 9] for a detailed discussion).
At one extreme (case A), both countries produce goods that
are completely different. The number of sectors in the integrated
(relative to autarky) economy increases to (N1 ϩ N2 ) ϭ
(U1 ϩ U2 ϩ H2 ϩ H2); aggregate research effort ⌰ also rises. But
the rate of creative destruction ␪ within each sector does not
change, since the share of skilled labor remains constant (equa-
tion (13)). This case corresponds to a pure composition effect, in
which integration alters neither economic turbulence nor the
choice of organizational structure.
In the opposite case (case B), both economies produce exactly
the same goods before integration. The number of sectors in each
country remains unchanged after integration. However, the rate of
creative destruction ␪ rises, and the integrated economy may shift
from the IRS to the CRS equilibrium. As a result, firms change
organizational structure, and inequality rises. Case B is an
example of a pure size effect.
In this discussion we treated the number of sectors as
exogenous. If we let N adapt to market size as we did above, our
analysis suggests that the size effect should temporarily dominate
in the short run (and possibly also in the long run if we believe in
the hysteresis effect), and economic integration between devel-
oped countries may create markets that are more turbulent, in
which firms are forced to reorganize.
IV. EMPIRICAL EVIDENCE
The supply of educated workers has been increasing sharply
for the past 30 years in most OECD countries [Autor, Katz, and
Krueger 1998]. Our model suggests that this change in the supply
of skilled labor may have affected relative demand for skill from
firms, because product markets may have become too unstable to
support the previously dominant organizational structure. There-
fore, our setup provides a natural framework for interpreting
recent trends in OECD labor markets.
In this section we provide empirical evidence that these
predictions of our model are consistent with recent data. At the
macroeconomic level, we show that skill upgrading of jobs from
1984 to 1995 in the French economy is the result of a reallocation
from functions related to production to functions related to

1220
QUARTERLY J OURNAL OF ECONOMICS
commercialization, customization, and design of new products.16
In addition, there is macroeconomic evidence that creative destruc-
tion on the product market has become more active.
At the firm level, we find a significant correlation between
employment reallocation (away from production-related activi-
ties) and different indicators of creative destruction on the product
market.
A. The Data
The data set we use comes from a French survey of the
structure of employment in firms (Enqueˆte sur la Structure des
Emplois, INSEE). Details concerning the data are provided in
Appendix 2. All sectors are covered except for Household Services
and Nontraded services (Education and Health Care). The data
represent 60,000 firms per year, from 1984 to 1995. It is an
unbalanced panel, since many firms enter and exit the data set
year. At the firm level, the data provide variables measuring total
employment in terms of workers (TOT), sector of activity (four
digits-SECT) and the share of employment in each of the following
five functions: development of new products (DNP: R&D, market-
ing), commercialization of existing products (CO: sales, advertis-
ing, technical sales involving in customization, purchasing), pro-
duction (PR: engineers, technicians, blue-collar workers), logistics
(LO: drivers, maintenance workers, cleaners) and administration
(AD: managers, corporate lawyers, accountants, other employ-
ees). We also know the share of skilled workers employed in each
function.17
Our data make a clear distinction between function and skill
at the firm level. To our knowledge, the literature on occupations
and skills tends to use individual data (U. S. census data have
been used in this context by Katz and Murphy [1992], Goldin and
Katz [1995, 1998], and Murphy and Welch [1993]). On the other
hand, existing firm data (such as those used by Berman, Bound,
and Griliches [1994]) only have the blue-/white-collar classifica-
tion, which mixes skill and occupation. Our data set, providing
16. The theoretical counterpart of this reallocation is the shift of skilled
employment from production to R&D, as would be predicted by a change in
equilibrium.
17. As explained in Appendix 2, the skill measure is derived from the detailed
French socioprofessional classification, which allows us to disentangle skill from
occupation for each job type at the firm level.

CREATIVE DESTRUCTION
1221
FIGURE IV
Share of Skilled J obs in France: 1984–1995
Source Enqueˆte Structure des Emplois (1984–1995).
information on both skill and function at the firm level, allows us
to deal more easily with labor demand issues.18
B. Macroeconomic Evidence
Figure IV plots the share of skilled jobs in aggregate employ-
ment from 1984 to 1995. Over this period, France has seen a skill
upgrading of jobs of the same magnitude as found in other OECD
countries [Machin and Van Reenen 1998]. In our panel the total
increase in the share of skilled workers over the period is 5.8
percent, equivalent to an average increase of 0.5 percent per year.
Table I displays aggregate function shares and skill intensi-
ties across years. We see that production is the largest and most
unskilled labor-intensive function. Development and commercial-
ization are the two smallest functions, but together account for 23
18. We show below that the recent evolution in the skill structure of French
employment exhibits similar trends as other OECD countries. For this reason, if
France behaves like the rest of the world on the skill dimension, then our empirical
analysis should be informative for other developed economies (under the assump-
tion that organizational change is pervasive to some extent).

1222
QUARTERLY J OURNAL OF ECONOMICS
TABLE I
DESCRIPTIVE STATISTICS ON OCCUPATIONAL STRUCTURE
Adminis- Commercial- Produc- Logis- Devel-
Function
tration
ization
tion
tics
opment
Share in total employment (%)
1984
1995
18.5
18.0
11.4
15.4
51.6
44.4
13.7
14.7
4.90
7.50
Share of skilled workers (%)
1984
1995
35.7
43.7
55.8
54.3
14.8
17.2
15.5
15.9
100
100
Source. Employment structure survey [ESE 1984, 1995].
All French Establishments of more than twenty employees except in Household Services, Health Care,
and Education. Employment is measured in terms of distinct individuals. Skill measure is derived from the
detailed French socioprofessional classification and depends on education level and experience.
All numbers are percentage point changes.
percent of aggregate employment in 1995; at the same time they
represent the most skill intensive jobs.
How have the shares of different functions varied with time?
Their evolutions are plotted in Figure V. A large employment
reallocation from production toward development and commercial-
ization (the most skill intensive functions) took place from 1984 to
1995 in the French economy. Indeed, the share of production
workers decreased by 7 percent, while that of development and
commercialization workers increased by 6.5 percent over the
FIGURE V
Occupation Shares in France
Source Enqueˆte Structure des Emplois.
Note. Right axis for Production, Left axis for Administration, Logistics,
Development, and Commercialization. Shares are computed by dividing the
aggregate number of workers in a given function by the sample labor force.

CREATIVE DESTRUCTION
1223
same period. This reallocation is comparable in size to the
aggregate skill upgrading of jobs.
In sum, we see a large employment shift from production (an
unskilled-labor-intensive function) to skilled functions, whose
main purpose is to reap market share through the design of new
products, as well as by customizing and selling them. Consistent
with our model, the French economy seems more and more
involved in the Schumpeterian process of creative destruction and
management of change, and less and less in actual production.
The increase in the share of functions related to commercializa-
tion and development is consistent with the secular trend in U. S.
occupations described in Goldin and Katz [1995]. However, since
they cover the 1900–1940 period, they also witness an increase in
the share of ‘‘operatives,’’ consistent with the diffusion of mass
production during that period.
Now consider the contribution of functional reallocation to
skill upgrading. Functional reallocation seems to be a major
component of the economy’s aggregate skill upgrading. To see this,
we follow the methodology introduced by Berman, Bound, and
Griliches [1994] and decompose aggregate skill upgrading into
two parts. Let H be the total number of skilled jobs and N be total
employment. Let ni be employment and hi be the number of skilled
jobs in sector i. Then
M
H
N
hi ni
ϭ
·
.
͚
ni
N
iϭ1
Taking first differences with respect to time yields
M
M
H
N
hi ni
hi
ni
ni
N
⌬
ϭ
⌬
·
ϩ
· ⌬
,
͚
͚
ni
N
iϭ1
iϭ1
where ⌬ represents the time difference operator, and the upper
bar denotes the intertemporal mean.
In this approach, aggregate skill upgrading can be decom-
posed into 1) a move toward more skill intensive sectors (Between
component) and 2) an upgrading of skills within each sector
(Within component). Using U. S. sectoral data, Berman, Bound,
and Griliches [1994] find that the between component is negligible
when compared with the within term, regardless of the level of
aggregation. Berman, Bound, and Machin [1998] present similar
evidence for other OECD countries. Both studies infer from these
regularities that specialization due to international trade cannot

1224
QUARTERLY J OURNAL OF ECONOMICS
TABLE II
SECTORIAL REALLOCATION AND SKILL UPGRADING
Years
84–90
90–95
84–95
D(Skilled)
Between sector
Within sector
4.02
0.42
3.60
2.06
Ϫ0.03
2.09
6.08
0.09
5.09
Source. Employment structure survey [ESE 1984, 1990, 1995].
All French Establishments of more than twenty employees except in Household Services, Health Care
and Education. Employment is measured in terms of distinct individuals. Skill measure is derived from the
detailed French socioprofessional classification and depends on education level and experience. The table
variables refer to the Berman, Bound, and Griliches [1994] decomposition. The percentage increase in skilled
employment (D(Skilled)) in the sample is equal to the sum of two terms. The first is the sum of net percentage
increases in each sector ’s employment, weighted by this sector ’s share of skilled workers in employment
(Between sector). The second term is the sum of sectorial increases in skilled employment, weighted by each
sector ’s share in total employment (Within sector). To define sectors, we used a two-digit classification (70
sectors).
All numbers are percentage point changes.
be held responsible for the observed widespread skill upgrading of
jobs, and that pervasive Skill Biased Technical Change remains
the best alternative explanation.
Table II gives the within-between sector decomposition of the
aggregate skill upgrading, from 1984 to 1995, and on two five-year
subperiods, using our French enterprise data. We use two-digit
sector classifications.19 The between term is almost negligible,
suggesting that the sectorial reallocation has little explicative
power with regard to the aggregate skill upgrading of French jobs
over the period 1984–1995.20 Our results are thus consistent with
what has been found in the two papers cited above; on the skill
dimension, French data exhibit the same regularities as displayed
by other countries.
The prediction of our model is that functional, instead of
sectorial, reallocation should have taken place, as a result of
increased creative destruction. Table III provides the between-
within function decomposition of the skill upgrading: although we
have only five functions, the functional reallocation explains more
than half of the aggregate skill upgrading in France. Note that the
contributions of production, development, and commercialization
are large: reallocation (the between term) takes place from
19. Berman, Bound, and Griliches’ study involves a classification of 29
sectors. Our two-digit NAF classification provides us with 70 sectors. A similar
decomposition performed on a more disaggregated classification did not change the
result.
20. Also note that between-within firm decomposition also yields a negligible
between firm reallocation.

CREATIVE DESTRUCTION
1225
TABLE III
FUNCTIONAL REALLOCATION AND SKILL UPGRADING
Years
D(skilled)
84–90
4.02
90–95
2.06
84–95
6.08
Between function
Administration
Commercialization
Production
0.10
0.98
Ϫ0.62
0.12
Ϫ0.29
1.25
Ϫ0.56
0.04
Ϫ0.17
2.21
Ϫ1.16
0.16
Logistics
Development
Total
1.17
1.76
1.42
1.88
2.60
3.65
Within function
Administration
Commercialization
Production
Logistics
Development
Total
1.04
0.07
0.97
0.17
.
0.43
Ϫ0.29
0.17
Ϫ0.11
.
1.45
Ϫ0.19
1.12
0.06
.
2.25
0.19
2.43
Source. Employment structure survey [ESE 1984, 1990, 1995].
All French establishments of more than twenty employees except in Household Services, Health Care,
and Education. Employment is measured in terms of distinct individuals. Skill measure is derived from the
detailed French socioprofessional classification and depends on education level and experience. The table
variables refer to a variant of the Berman, Bound, and Griliches [1994] decomposition. The percentage
increase in skilled employment (D(skilled)) in the sample is equal to the sum of two terms. The first is the sum
of net percentage increases in each function employment weighted by this function’s share of skilled workers
in employment (Between function). The second term is the sum of increases in skilled employment within each
function, weighted by this function’s share in total employment (Within function). Recall that functions are
decomposed into five categories, while sectors are decomposed into 70 categories.
All numbers are percentage point changes.
production to development and commercialization. The within
production and within administration terms are also quite large,
suggesting that skill-biased technical change may be taking place
mainly within these two functions.
To sum up, the extent of this functional reallocation from
production to development and commercialization is quantita-
tively important, in that it accounts for almost half of the
economy’s aggregate skill upgrading.21 Murphy and Welch [1993]
find similar results: they estimate, under particular elasticity
assumptions, that a shift in occupations is responsible for one-
third of the increase in demand for skilled labor in the U. S.
economy between 1950 and 1990.
Finally, our theory predicts that economywide reorganization
21. This methodology is purely descriptive; strictly speaking, these results
cannot be interpreted without strong assumptions on the different elasticities of
substitution of the underlying production function (for a more structural analysis,
see Maurin and Thesmar [1999]).

1226
QUARTERLY J OURNAL OF ECONOMICS
FIGURE VI
Patents Applied for in France and Patents Granted in the United States
Source Patent Offices, France and the United States.
should result in a rise in the rate of creative destruction. Accord-
ing to our results mentioned above, the share of total employment
devoted to development of new products (through R&D and
Marketing) rose from 4.9 percent to 7.5 percent (a 50 percent
increase in twelve years). Similar evidence is provided by Machin
and Van Reenen [1998]: they note that, between 1973 and 1989,
the share of value added devoted to R&D in seven OECD countries
(J apan, United States, United Kingdom, France, Germany, Den-
mark, and Sweden) has increased by between 45 and 100 percent.
A consequence of this dramatic rise is a larger patenting rate.
Figure VI shows the number of patents granted by the U. S.
patent office since 1968, along with the number of patents applied
for in France since 1985. In both countries we see a twofold
increase between 1980 and 1998, a sign of a dramatic acceleration
in patenting activity in the past two decades. Although patenting
rate is a crude proxy for innovative activity, this pattern suggests
a substantial increase in the rate of creative destruction.22
Furthermore, Gottschalk and Moffit [1994], using the PSID
22. According to Griliches [1990], the decrease in patents granted during the
1970s is mainly the consequence of a smaller budget for printing the approved
patents at the U. S. patent office. However, as can be seen in Figure VI, the
subsequent rise (from 1980 onward) in patent approval largely surpasses the
number of patents granted in 1970, the peak year before the decline.

CREATIVE DESTRUCTION
1227
data, show an increase in earnings fluctuations and job turnover
from the 1970s to the 1980s. This suggests increasing job creation
and destruction during the period; however, evidence from plant-
level data on this point remains mixed. In our context, we do not
believe that firm-to-firm job reallocation is the relevant counter-
part to our creative destruction rate since product or process
innovations within the firm do not systematically yield worker
replacement. Our point is rather that frequent innovations com-
pel firms to change their functional mix, shifting employment
toward more skill intensive product development and commercial-
ization. Greenwood and J ovanovic [1999] directly measure prod-
uct market instability as the share of market value accounted for
by firms entering and exiting 30 product specific markets. Using
this measure, they find that the rate of turnover has risen from 3
percent in the early 1970s to 6 percent in the late 1980s.
C. Microeconomic Evidence: Creative Destruction
and Organization
Complementarity between production workers and an innova-
tion-specific sunk cost (corresponding to organizational complex-
ity) is the key mechanism in this paper. A direct, testable
implication of this assumption is that firms in sectors experienc-
ing more creative destruction should, in principle, hire fewer
production workers. If our theory is correct, we should therefore
observe a negative correlation between worker allocation to
production and relevant measures of product market instability in
our enterprise data.
As noted above, the most relevant proxies for creative destruc-
tion have to do with innovation. Relations between innovation and
the share of blue-collar workers have been previously investigated
in the skill-biased technical change literature, although the data
available so far lack a proper distinction between function and
skill, and are most often aggregated to the sectorial level. For
example, Machin and Van Reenen [1998] find that the relative
demand for white-collar workers is positively related to existing
capital and R&D expenditures, using two-digit sectorial data in a
panel of OECD countries. Dunne, Haltiwanger, and Troske [1996]
find similar results using plant-level U. S. data between 1972 and
1988. These results can be interpreted as a negative relation
between product turnover (development of new products) and
production employment’s share (blue-collar workers) at the firm
level. However, the predictions of our model concern the total

1228
QUARTERLY J OURNAL OF ECONOMICS
share of production (both skilled and unskilled) workers. Since
our data set allows us to disentangle more accurately function
from skill, this section thus seeks to establish the nature of the
relation between the share of employment devoted to production
and product turnover at the firm level.
In what follows, we use the share of production employees
(both skilled and unskilled) within the firm (PR) as our empirical
measure of what we called ‘‘production workers’’ in our model.23
We use two different measures of creative destruction. The first is
the ratio of products less than three years old in total sales (NPR);
this represents the rate of product turnover. In order to reduce
measurement error, this variable is discretized into four catego-
ries (less than 10 percent, 10–30 percent, 30–70 percent, and more
than 70 percent). The second empirical counterpart for creative
destruction is (more classically) the number of researchers.
To begin, we restrict our attention to firms that were active
between 1986 and 1990. We regressed the average share of
production workers (PR) over the period on our measure of
product turnover. Given that the new product ratio (NPR) is taken
from a separate survey (the Enqueˆte Innovation), the restriction
of a balanced panel of firms present in both surveys reduces the
number of available observations to 11,233. Model 1 regresses PR
on three indicator variables representing the categories of NPR,
while model 2 controls for total employment and sector. Both
models are estimated by OLS with White [1980] correction for
arbitrary heteroskedasticity. The results are given in Table IV; the
reference category for NPR is having less than 10 percent of new
products in total sales. A firm in the 10–30 percent category has on
average a lower (by five points) share of production workers. For
the category more than 70 percent, the difference amounts to 9
percent. All results are highly significant; there is a strong
negative correlation between product turnover and the allocation
of employment to production.
The second regression takes the number of workers devoted
to the function development of new products as a regressor (DNP:
R&D and marketing) instead of the rate of product turnover. In
order to avoid spurious correlation between the dependent vari-
able and the regressor, we replaced PR by the share of production
23. Alternative regressions, using production and administrative workers
(PR ϩ AD) yield qualitatively similar results, and are available from the authors
upon request.

CREATIVE DESTRUCTION
1229
TABLE IV
INNOVATION TURNOVER AND FUNCTION REALLOCATION
Share of production workers
Model 1
Model 2
Share of new products in total sales (86–90)
Between 10 and 30%
Ϫ0.05
(11.44)
Ϫ0.07
(9.54)
Ϫ0.09
(5.31)
—
Ϫ0.05
(10.88)
Ϫ0.06
(8.82)
Ϫ0.08
(4.9)
Ϫ3.58
(1.57)
11233
0.06
Between 30 and 70%
More than 70%
Employment (1e-6)
Observations
R2
11233
0.02
Source. Enqueˆte Innovation and Enqueˆte Structure des Emplois (balanced panel).
The dependent variable is the average of the 1986 and 1990 shares of production workers.
The equations are estimated through OLS with White [1980] correction for arbitrary heteroskedasticity.
Model 2 includes the total employment and fifteen sectorial dummies as additional regressors.
workers in total employment excluding development of new
products. In this regression we use the full, unbalanced panel of
firms, although we require that a firm does not disappear and
later reappear in the data set. This leaves us with 568,146
observations between 1984 and 1995. Model 1 regresses PR on
DNP lagged by one period, while model 2 adds total employment
(TOT) as a second regressor. In both cases, the estimation includes
firm-specific fixed effects. Finally, in order to control for long-term
trends, models 3 and 4 perform the same regression using
five-year differences. As shown in Table V, all models (except
model 3, which is rejected by the data) exhibit a negative relation
between development effort and the share of production workers.
V. CONCLUDING REMARKS
This paper is motivated by the observation that the organiza-
tion structures of firms in all OECD industries experienced
extensive mutations since the 1970s. In a model of Schumpeterian
growth, where production is less skill intensive than research, we
allow patent owners to choose an optimal organizational structure
in response to volatility in the economic environment. The micro-
economic analysis highlights the importance of product market
instability (rate of creative destruction) on a firm’s choice of
organization. A general equilibrium framework reveals that firms’

1230
QUARTERLY J OURNAL OF ECONOMICS
TABLE V
DEVELOPMENT EFFORT AND FUNCTION REALLOCATION
Share of production
Model 1
—
Model 2
Model 3
—
Model 4
Total employment (1e-6)
DNP (1e-6)
6.86
7.05
(8.56)
Ϫ41.1
(4.93)
238932
36.8
(28.34)
Ϫ13.9
(7.47)
568146
106.79
Ϫ5.66
(3.08)
568146
106.6
13.4
(0.42)
238932
0.18
Observation
Fisher statistic
Source. Enqueˆte Structure des Emplois.
Note. DNP corresponds to the number of workers in the function Development of New Products (R&D and
marketing).
Models 1 and 2 use lagged conception employment. They are estimated by OLS with firm-specific effects.
Models 3 and 4 use long differences (five years). They are estimated by OLS.
t-statistics are in parentheses.
organizational decisions could feed back into aggregate instabil-
ity. These two externalities combine to generate a macroeconomic
complementarity, so that mass production itself reinforces the
stability of product markets. Thus, multiple equilibria (mass
production/organistic firms) coexist for certain parameter values.
Our findings mostly deal with the interaction between the
labor market, equilibrium choices of organization, and aggregate
product market instability. Comparative statics show that an
increase in the size of the skilled labor force can shift the economy
into a zone where both equilibria coexist, and possibly to the zone
where only the CRS (organistic) equilibrium exists. This possible
equilibrium shift yields an increase in aggregate instability and
wage inequality. Furthermore, unskilled wages temporarily drop,
while skilled wages jump. Economic integration with similarly
skill endowed countries is shown to have the same effects, at least
in the short run.
Our conclusions are consistent with recent waves of firm
reorganizations, past labor market mutations in several OECD
countries (the United States, the United Kingdom, Australia), and
an increase in the college enrollment rate. The view expressed in
this paper is that there is no obvious reason why technical
progress should be more skilled biased than it was 50 years ago.
However, we argue that a move toward more decentralization and
horizontal communication in the organization of firms intrinsi-
cally favors skilled labor to the disadvantages of the unskilled.
The second part of this paper is devoted to an analysis of
firms’ organizations by way of their functional structures. We