The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Preface: Technology, Growth,
and the Labor Market
DONNA K. GINTHER AND MADELINE ZAVODNY
Ginther is an associate professor of economics at the University of Kansas and a former research
economist and associate policy adviser at the Atlanta Fed. Zavodny is an associate professor
of economics at Occidental College and is currently on leave from her position as a research economist
and associate policy adviser at the Atlanta Fed. They extend many thanks to presenters and discussants
(whose names appear in this preface) as well as conference participants at the January 2002 “Technology,
Growth, and the Labor Market” conference. The views expressed in the conference proceeding papers
presented in this issue of the Economic Review are the authors’ and should not be attributed to any
institutions in the Federal Reserve System or any other institutions with which the authors are affiliated.1
n 1998, Federal Reserve Board Chairman Alan
Greenspan posed the provocative question, Is
there a “new economy”? He described the new
economy’s characteristics as including technological innovations that raise productivity and
that have, accordingly, removed pricing power
from the world’s producers on a more lasting basis
(Greenspan 1998). Although the 2001 recession
quelled the discussion about whether the United
States, and perhaps even the world, had entered a
new era characterized by sustained high levels of economic growth, researchers continue to investigate the
effects of technological change on the economy.
This issue of the Economic Review contains
four papers that examine the underpinnings of the
new economy—technology and its effects on
macroeconomic growth and the labor market.
These papers were among those presented at the
“Technology, Growth, and the Labor Market” conference sponsored by the research department of
the Federal Reserve Bank of Atlanta and the
Andrew Young School of Policy Studies at Georgia
State University in January this year. This introduction summarizes all the speeches, papers, and discussant comments presented at the conference.2
Researchers were quick to examine the new
economy, but many of their early conclusions
remain open to debate. Macroeconomists, including
Martin N. Baily (2001), Stephen D. Oliner and
Daniel E. Sichel (2000), and Kevin Stiroh (2001),
I
argued that the technological change embodied in
increased computer investment contributed substantially to the surge in productivity growth experienced in the United States between 1995 and
2000. Although productivity traditionally declines
during recessions, labor productivity remained high
during the recession that officially began in March
2001, perhaps because the large investments in
equipment and software made during the late 1990s
continued to boost output for several years after the
purchases were made. However, the advent of the
2001 recession and research by skeptics, such as
Robert J. Gordon (2000), indicate that the effect of
technology on current and future productivity
growth remains an open question.
Labor economists have also investigated technology’s impact on the wage structure. The generally
accepted hypothesis among labor economists is that
skill-biased technological change has increased the
relative demand for skilled workers, causing the
observed increase in earnings inequality in the 1980s
(Council of Economic Advisers 1997; Katz and Autor
1999). Articles by, among others, Alan B. Krueger
(1993), Eli Berman, John Bound, and Zvi Griliches
(1994), and Ann P. Bartel and Nachum Sicherman
(1998) noted an association between computerization and higher wages for skilled workers. However,
the skill-biased technological change hypothesis has
been difficult to prove because of the paucity of data
on workers’ use of technology in the workplace.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
v
Other research has attempted to forge a link
between human resource practices, computerization, productivity, and the returns to skill. Casey
Ichniowski and Kathryn Shaw (1995, 1999, and
2000) argued that innovative human resource practices raise worker productivity in a variety of contexts. Peter Cappelli and William Carter (2000)
evaluated the relative contributions of computerization and high-performance workplace practices,
concluding that higher wages are associated with
both technology (as represented by computers)
and high-performance workplace practices. This
research suggests that, in addition to technology,
human resource practices may be contributing to
higher productivity growth.
Productivity and the Macroeconomy
he conference included two plenary talks by economists with firsthand experience in determining
how productivity, inequality, and other such factors
should be taken into account when setting monetary
policy. The speeches by Edward M. Gramlich and
Alice M. Rivlin framed the questions addressed by
conference participants. Gramlich discussed why
understanding the role of technology in the economy
is important to economists and monetary policymakers. He raised many issues, including what stage of an
“information transformation” the U.S. economy is in,
why productivity defied past patterns by holding up
during the 2001 recession, the relative merits of public versus private investment, and why the United
States experienced a much larger productivity spurt
during the late 1990s than Western European nations
that had access to the same technologies.
Rivlin discussed the relevance of the new economy
paradigm and whether the economic recovery in the
United States will continue to feature high productivity growth and low inflation and unemployment.
She indicated that the Internet, combined with a number of advances in business practices, has led to an
increase in economic potential. One of the key implications of being in a new economy is that inflation has
become less of a concern for monetary policymakers
because employers are able to raise wages without
passing higher labor costs along via price increases.
Instead, excessive investments and overvalued equity
markets are central concerns going forward. Unfortunately, she noted, monetary policymakers have less
influence over such factors than over inflation.
T
Productivity Growth and Technology:
What the Future Holds
T
vi
he conference included two papers, printed in
this Review, that discussed the sources of the
surge in labor productivity growth during the latter
half of the 1990s and presented forecasts of labor
productivity growth rates during the next few
years. The two papers are similar in their methodologies and findings and also dovetail with recent
research by Baily (2001).
Dale W. Jorgenson, Mun S. Ho, and Kevin J.
Stiroh reviewed recent studies on the sustainable
rate of labor productivity growth and quantified
the source of growth, focusing on information
technology (IT). Using an augmented growth
accounting framework, they concluded that the
resurgence of labor productivity growth during the
late 1990s remains intact despite the 2001 recession. They projected that trend labor productivity
growth during the next decade will be about 2.2 percent per year, with a range of 1.3 percent to 2.9 percent, and output growth will be about 3.3 percent
per year, with a range of 2.3 percent to 4.0 percent. Jorgenson, Ho, and Stiroh found that IT, particularly semiconductors, played a large role in
growth during the second half of the 1990s, a
trend that is expected to continue but is nonetheless uncertain.
Stephen D. Oliner and Daniel E. Sichel used a
similar growth accounting framework to explore
the role of IT in labor productivity growth. They
also analyzed the steady-state properties of a multisector growth model in order to estimate the longrun rate of labor productivity growth and to calculate
to what extent technical progress drives productivity improvements. Oliner and Sichel concluded that
the likely annual rate of labor productivity growth
is about 2 to 2.75 percent, depending on the pace
of technological advances in the semiconductor
industry. This conclusion implies that the rates of
labor productivity growth achieved in the United
States during the second half of the 1990s are
sustainable.
The discussion of these two papers by John
Fernald noted that the estimates of labor productivity growth might be on the conservative side
because the papers do not account for adjustment
costs. The high levels of investment in IT during the
second half of the 1990s presumably led to sizable
adjustment costs, which lowered both output
growth and productivity growth. Fernald also pointed
out that much about the role of IT in future growth
is unknown at this point, raising questions such as,
Will the rate of technical change slow? How elastic
is the demand for IT? Will the relative price of IT
goods continue to fall? What will happen in the
non-IT sector, which accounts for 94 percent of
the economy?
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Skill-Biased Technological Change
and Wage Inequality
he conference also included two papers about
the role of technological advances in changes in
inequality in the labor market. The authors examined whether inequality should be viewed as a
causal result of skill-biased technological change or
whether there is a missing link—or perhaps no
link—between changes in technology and changes
in wage inequality.
David H. Autor, Frank Levy, and Richard J.
Murnane began by examining the contributions of
changes in labor supply and labor demand to wage
inequality during the 1940s through the 1990s. The
authors discussed why computers increase the
demand for more educated workers, arguing that
computers have transformed the importance of
manual versus cognitive tasks and routine versus
nonroutine tasks. The data they used indicate that
demand shifts are an important contributor to
recent trends in inequality although supply shifts
also exerted considerable influence during the
entire period. The authors then explored several
pieces of indirect evidence that computerization is
responsible for the higher growth in relative demand
for skilled workers during recent decades, including
the timing of increases in computerization compared
with the timing of the rise in wage inequality and
trends in educational upgrading within industries.
Several puzzles emerge from Autor, Levy, and
Murnane’s paper, such as whether relative demand
for skilled workers began accelerating during the
1970s or during the 1980s and why the growth in
relative demand for skilled workers decelerated
during the 1990s. As Donna Ginther’s discussion
noted, the major contribution of Autor, Levy, and
Murnane’s work is that it provides a mechanism by
which computers and information technology could
lead to skill-biased technological change.
David Card and John E. DiNardo, in a paper printed
in this Review, examined whether the increase in
wage inequality during the 1980s was caused by skillbiased technological change. They focused on the
merits and limitations of the skill-biased technological
change hypothesis, namely, that an increase in
T
demand for skilled workers has led to an increase in
wage dispersion between skilled and unskilled workers. Card and DiNardo noted that the supply of skilled
workers has increased, so there must have been a
more-than-offsetting change in demand to account
for the observed rise in wage inequality during the
1980s. They investigated whether different aspects of
the wage structure are consistent with the possibility
that technical change underlies the changes in
demand that must have occurred.
Card and DiNardo pointed out many inconsistencies that make it difficult to reconcile all of the
observed trends with the skill-biased technological
change hypothesis. As the discussion by Ginther
noted, Card and DiNardo provide a good start at
critically examining the skill-biased technological
change hypothesis; however, she argued that it is an
oversimplification to suggest that skill-biased technological change is a “unicausal” explanation for the
many changes in the wage structure since 1980.
Technology and Productivity in the Firm
ohn Haltiwanger presented a paper that complements those by Autor, Levy, and Murnane and
Card and DiNardo. The latter two papers used data
on wage inequality from the perspective of workers
while Haltiwanger’s paper used data from wages on
the establishment side.
Haltiwanger discussed the correlation between
technology investments and wage dispersion and productivity dispersion, both of which increased since the
1980s. He found that both phenomena occurred at the
between-plant, within-industry level, suggesting that
the changes in economic forces were not industrywide
but occurred at a more micro level. Another implication of Haltiwanger’s findings is that workers have
become more segregated by skill level, a proposition
directly tested in a related paper by Lengermann
(2001). Haltiwanger concluded that changes in plants’
investments in computers and other forms of capital
account for a substantial proportion of the increases in
wage and productivity dispersion.
The discussion by Robert A. Eisenbeis cautioned
that some of the paper’s findings are sensitive to the
time periods analyzed and that the empirical model
J
1. The authors also thank Lynn Foley, Peter Hamilton, Vanessa Jordan, C. Anitha Manohar, Elizabeth McQuerry, Pierce Nelson,
Nancy Pevey, Avani Raval, Vivian Wilkins, and especially Jess Palazzolo for their invaluable assistance with organizing the conference. Finally, they thank Jack Guynn, president of the Federal Reserve Bank of Atlanta; Roy Bahl, dean of the Andrew
Young School of Policy Studies at Georgia State University; and James Alm, chair of the economics department at Georgia
State University, for making the funds available to host the conference.
2. Presentations included seven papers by distinguished economists, discussant comments on those papers, and speeches by
Edward M. Gramlich, a member of the Board of Governors of the Federal Reserve System, and Alice M. Rivlin, a senior fellow at the Brookings Institution and former vice chair of the Federal Reserve Board. The entire conference proceedings,
including the papers published in this Economic Review, will be published later this year by Kluwer Academic Publishers.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
vii
does not explain much of the wage and productivity
dispersion between plants. Eisenbeis also noted
that the role of macroeconomic cyclical forces versus secular changes is unclear.
Edward N. Wolff, in the final paper from the conference printed in this Review, used industry-level
data to examine the relationships between productivity and the computerization, educational attainment, and skill levels of workers at the industry
level. Perhaps surprisingly, he found no evidence
that education is linked to productivity growth.
However, cognitive skills—as measured by job-skill
requirements from the Dictionary of Occupational
Titles—are related to productivity growth, albeit
modestly. Wolff’s results also indicate that computers and related IT investments are not significantly
associated with productivity growth at the industry
level. Paula Stephan’s discussion of Wolff’s paper
questioned whether the failure to find a relationship
between computerization and productivity would
be robust to including data from the 1990s and
whether other measures might better capture skill,
particularly with regard to IT skills.
Kathryn Shaw investigated the roles of investment
in IT and changes in human resource management
practices in corporate performance. Using traditional
case study techniques, Shaw documented the relationship between changes in human resource practices and productivity gains in the steel industry. The
paper argues that IT lowers the costs of providing
information to workers as well as greater problemsolving capacities on the part of skilled workers. In
her discussion, Stephan noted that Shaw’s paper
makes an important contribution by linking the literature on IT and performance with the literature on
workplace practices and performance.
The papers presented at the conference add to
our understanding of the role of technological change
in the economy, both in recent years and in the
decades ahead.
REFERENCES
Baily, Martin N. 2001. Macroeconomic implications of the
new economy. In Economic policy for the information
economy. Kansas City: Federal Reserve Bank of Kansas City.
———. 1999. The effects of human resource systems on
productivity: An international comparison of U.S. and
Japanese plants. Management Science 45:704–22.
Bartel, Ann P., and Nachum Sicherman. 1998. Technological change and the skill acquisition of young workers.
Journal of Labor Economics 16, no. 4:718–55.
———. 2000. TQM practices and innovative HRM practices: New evidence on adoption and effectiveness. In
The quality movement in America: Lessons from
theory and research, edited by Robert Cole and Richard
Scott. New York: Russell Sage.
Berman, Eli, John Bound, and Zvi Griliches. 1994. Changes
in the demand for skilled labor within U.S. manufacturing:
Evidence from the Annual Survey of Manufactures. Quarterly Journal of Economics 109, no. 2:367–97.
Cappelli, Peter, and William Carter. 2000. Computers,
work organization, and wage outcomes. National Bureau
of Economic Research Working Paper 7987.
Council of Economic Advisers. 1997. Economic report
of the president. Washington, D.C.: Government Printing Office.
Gordon, Robert J. 2000. Does the “new economy” measure up to the great inventions of the past? Journal of
Economic Perspectives 14 (Fall): 49–74.
Greenspan, Alan. 1998. Question: Is there a new economy?
Presented at the University of California-Berkeley, September 4, 1998. <www.federalreserve.gov/boarddocs/
speeches/1998/19980904.htm> (June 5, 2002).
Ichniowski, Casey, and Shaw, Kathryn. 1995. Old dogs and
new tricks: Determinants of the adoption of productivityenhancing work practices. Brookings Papers: Microeconomics, 1–65.
viii
Lengermann, Paul A. 2001. Is it who you are, where you
work, or with whom you work? Reassessing the relationship between skill segregation and wage inequality. University of Maryland. Photocopy.
Katz, Lawrence F., and David H. Autor. 1999. Changes in
the wage structure and earnings inequality. In Handbook
of labor economics, vol. 3, edited by O. Ashenfelter and
D. Card. Amsterdam, New York City, and Oxford: Elsevier
Science.
Krueger, Alan B. 1993. How computers have changes the
wage structure: Evidence from micro data, 1984–1989.
Quarterly Journal of Economics 108, no. 1:33–60.
Oliner, Stephen D., and Daniel E. Sichel. 2000. The
resurgence of growth in the late 1990s: Is information
technology the story? Journal of Economic Perspectives 14 (Fall): 3–32.
Stiroh, Kevin. 2001. Information technology and the U.S.
productivity revival: What do the industry data say? Federal Reserve Bank of New York. Photocopy, December
(American Economic Review, forthcoming).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Projecting Productivity
Growth: Lessons from the
U.S. Growth Resurgence
DALE W. JORGENSON, MUN S. HO, AND KEVIN J. STIROH
Jorgenson is the Frederic Eaton Abbe Professor of Economics at Harvard
University. Ho is a visiting scholar at Resources for the Future. Stiroh is a
research officer at the New York Fed. They thank John Fernald for helpful comments and Jon Samuels for excellent research assistance. The Bureau of Labor
Statistics and the Bureau of Economic Analysis provided data and advice.
he unusual combination of more rapid
output growth and lower inflation from
1995 to 2000 has touched off a strenuous
debate among economists about whether
improvements in U.S. economic performance can be sustained. This debate has
intensified with the recession that began in March
2001, and the economic impacts of the events of
September 11 are still imperfectly understood. Both
factors add to the considerable uncertainties about
future growth that currently face decision makers
in both the public and private sectors.
The range of informed opinion can be illustrated
by the projections of labor productivity growth
reported at the August 2001 Symposium on
Economic Policy for the Information Economy,
organized by the Federal Reserve Bank of Kansas
City. J. Bradford Delong, professor of economics
at the University of California at Berkeley, and
Lawrence H. Summers, president of Harvard University and former Secretary of the Treasury, offered
the most optimistic perspective with a projection of
labor productivity growth of 3 percent per year.1 A
more pessimistic tone was set by Martin N. Baily
(2001), former chairman of the Council of Economic Advisers, who speculated that labor productivity would average near the low end of the 2 to
2.5 percent per year range.
This uncertainty is only magnified by the observation that recent productivity estimates remain
T
surprisingly strong for an economy in recession. The
Bureau of Labor Statistics (BLS) (2002) estimates
that business sector productivity grew 1.9 percent
per year during 2001 while business sector output
grew only 0.9 percent per year as the U.S. economy
slowed during the 2001 recession. Growth of both
labor productivity and output, however, appears
considerably below trend rates, partially reflecting
the collapse of investment spending that began
toward the end of 2000, continued through 2001,
and seems likely to be maintained well into 2002.
This paper reviews the most recent evidence and
quantifies the proximate sources of growth using
an augmented growth accounting framework that
allows us to focus on information technology (IT).
Despite the downward revision to gross domestic
product (GDP) and investment in some IT assets
in the annual GDP revisions by the Bureau of
Economic Analysis (BEA) in July 2001, we conclude that the U.S. productivity revival remains
largely intact and that IT has played a central role.
For example, the capital deepening contribution
from computer hardware, software, and telecommunications equipment to labor productivity
growth for the 1995–2000 period exceeded the contribution from all other capital assets. We also find
increases in total factor productivity (TFP) in both
the IT-producing sectors and elsewhere in the economy although the non-IT component is smaller than
in earlier estimates.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
1
The paper then turns to the future of U.S. productivity growth, concluding that the projections of
Jorgenson and Stiroh (2000), prepared more than
eighteen months ago, are largely on target. Our new
base-case projection of trend labor productivity
growth for the next decade is 2.21 percent per year,
only slightly below the average of the 1995–2000
period of 2.36 percent per year. The projection of
output growth for the next decade, however, is only
3.31 percent per year, compared with the 1995–2000
average of 4.6 percent, as a result of slower projected
growth in hours worked.
Projecting growth for periods as long as a decade
is fraught with uncertainty. Our pessimistic projec-
Our new base-case projection of trend labor
productivity growth for the next decade is
2.21 percent per year, only slightly below
the average of the 1995–2000 period of
2.36 percent per year.
tion of labor productivity growth is only 1.33 percent
per year, while our optimistic projection is 2.92 percent. For output growth, the range is from 2.43 percent in the pessimistic case to 4.02 percent in the
optimistic. These ranges result from fundamental
uncertainties about future technological changes in
the production of information technology equipment
and related investment patterns, which Jorgenson
(2001) traced to changes in the product cycle of
semiconductors, the most important IT component.
The starting point for projecting U.S. output
growth is the projection of future growth of the labor
force. The 2.24 percent per year growth of hours
worked from 1995 to 2000 is not likely to be sustainable because labor force growth for the next decade
will average only 1.1 percent. An abrupt slowdown in
growth of hours worked would have reduced output
growth by 1.14 percent, even if labor productivity
growth had continued unabated. We estimate that
labor productivity growth from 1995 to 2000 also
exceeded its sustainable rate, however, leading to an
additional decline of 0.15 percent in the trend rate of
output growth so that the base-case scenario projects
output growth of 3.31 percent for the next decade.
The next section reviews the historical record,
extending the estimates of Jorgenson and Stiroh
(2000) to incorporate data for 1999 and 2000 and
revised estimates of economic growth for earlier
2
years. We employ the same methodology and summarize it briefly. Then we present projections of the
trend growth of output and labor productivity for
the next decade and compare these with projections based on alternative methodologies.
Reviewing the Historical Record
he methodology for analyzing growth sources
is based on the production possibility frontier
introduced by Jorgenson (1996, 27–28). This
framework captures substitution between investment and consumption goods on the output side
and between capital and labor inputs on the input
side. Jorgenson and Stiroh (2000) and Jorgenson
(2001) have recently used the production possibility
frontier to measure the contributions of information
technology to U.S. economic growth and the growth
of labor productivity.
The production possibility frontier. In the
production possibility frontier, output (Y ) consist of
consumption goods (C) and investment goods (I )
while inputs consist of capital services (K) and labor
input (L). Output can be further decomposed into
IT investment goods—computer hardware (Ic), computer software (Is ), communications equipment (Im ),
and all other non-IT output (Yn ). Capital services
can be similarly decomposed into the capital service
flows from hardware (Kc ), software (Ks ), communications equipment (Km ), and all other capital services (Kn ).2 The input function (X) is augmented
by total factor productivity (A). The production
possibility frontier can be represented as
T
(1) Y(Yn , Ic , Is , Im ) = A × X(Kn , Kc , Ks , Km , L).
Under the standard assumptions of competitive
product and factor markets and constant returns to
scale, equation 1 can be transformed into an equation
that accounts for the sources of economic growth:
(2) wYn ∆ ln Yn + wIc ∆ ln Ic + wIs ∆ ln I s + wIm ∆ ln I m =
vK n ∆ ln K n + vK c ∆ ln K c + vK s ∆ ln K s
+ vK m ∆ ln K m + vL ∆ ln L + ∆ ln A,
– denotes the average output
where ∆ x = xt – xt–1, w
–
shares, v denotes the average input shares of the
– +w
– +w
– = v–
– +w
subscripted variables, and w
Yn
Ic
Is
Im
Kn
–
–
–
–
+ v Kc + v Ks + v Km + v L = 1. The shares are averaged over
periods t and t – 1. We refer to the share-weighted
growth rates in equation 2 as the contributions of
the inputs and outputs.
Average labor productivity (ALP) is defined as
the ratio of output to hours worked, so that ALP =
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
y = Y/H, where the lower-case variable (y) denotes
output (Y) per hour (H). Equation 2 can be rewritten
in per hour terms as
This process enables us to decompose aggregate
TFP growth as
(4) ∆ ln A = uIT ∆ ln AIT + ∆ ln An ,
(3) ∆ ln y = vK n ∆ ln kn + vK IT ∆ ln kIT
+ vL (∆ ln L − ∆ ln H ) + ∆ ln A,
where v–KIT = v–Kc + v–Ks + v–Km and ∆ln kIT is the growth of
all IT capital services per hour.
Equation 3 decomposes ALP growth into three
sources. The first is capital deepening, defined as
the contribution of capital services per hour,
which is decomposed into non-IT and IT components. The interpretation of capital deepening is
that additional capital makes workers more productive in proportion to the capital share. The second
factor is labor quality improvement, defined as
the contribution of labor input per hour worked.
This factor reflects changes in the composition of
the workforce and raises labor productivity in proportion to the labor share. The third source is total
factor productivity growth, which raises ALP growth
point for point.
In a fully developed sectoral production model,
like that of Jorgenson, Ho, and Stiroh (2002), TFP
growth reflects the productivity contributions of
individual sectors. It is difficult, however, to create
the detailed industry data needed to measure
industry-level productivity in a timely and accurate
manner. The Council of Economic Advisers (CEA)
(2001), Jorgenson and Stiroh (2000), and Oliner and
Sichel (2000, 2002) have employed the price dual of
industry-level productivity to generate estimates of
TFP growth in the production of IT assets.
Intuitively, the idea underlying the dual approach
is that declines in relative prices for IT investment
goods reflect fundamental technological change
and productivity growth in the IT-producing
industries. We weight these relative price declines
by the shares in output of each of the IT investment goods in order to estimate the contribution
of IT production to economywide TFP growth.
– represents IT’s average share of output,
where u
IT
∆ln AIT is IT-related productivity growth, and
– ∆ln A is the contribution to aggregate TFP from
u
IT
IT
IT production. ∆ln An reflects the contribution to
aggregate TFP growth from the rest of the economy,
which includes TFP gains in other industries as well
as reallocation effects as inputs and outputs are
shifted among sectors.
We estimate the contribution to aggregate TFP
growth from IT production, u–IT ∆ AIT, by estimating
output shares and productivity growth rates for
computer hardware, software, and communications
equipment. Productivity growth for each investment
good is measured as the negative of the rate of price
decline relative to the price change of capital and
labor inputs. The output shares are the final expenditures on these investment goods, divided by total
output.3 This estimate likely understates IT output
because it ignores the production of intermediate
goods, but this omission is relatively small. Finally,
the non-IT contribution to aggregate TFP growth,
∆ An, is estimated as a residual from equation 4.
Data. This section briefly summarizes the data
required to implement equations 1–4; more detailed
descriptions are available in Ho and Jorgenson (1999)
and the appendices of Jorgenson and Stiroh (2000).
The output measure is somewhat broader than the
one used in the official labor productivity statistics,
published by the BLS (2001a, 2001b) and employed
by Gordon (2000) and Oliner and Sichel (2000, 2002).
Our definition of the private U.S. economy includes
the nonprofit sector and imputed capital service flows
from residential housing and consumer durables.
The imputations raise the measure of private output
by $778 billion in current dollars, or 9 percent of nominal private GDP, in 2000.
The output estimates reflect the revisions to the
U.S. National Income and Product Accounts (NIPA)
1. DeLong and Summers (2001, 21) do not actually provide a point estimate but state that “it is certainly possible—if not probable—that when U.S. growth resumes, trend productivity will grow as fast or faster than it did in the late 1990s.” The 3 percent estimate is attributed to Summers in a review of the symposium in The Economist, September 8, 2001.
2. Note that the output and capital service flow concepts include the service flows from residential structures and consumer
durables. See Jorgenson and Stiroh (2000) for details.
3. Output shares include expenditures on consumption, investment, government, and net exports for each IT asset. Note that
the use of the price dual to measure technological change assumes competitive markets in IT production. As pointed out by
Aizcorbe (2002) and Hobijn (2001), the market for many IT components, notably semiconductors and software, is not perfectly competitive, and part of the drop in prices may reflect oligopolistic behavior rather than technological progress.
Aizcorbe, however, concludes that declining markups account for only about one-tenth of the measured declines in the price
of microprocessors in the 1990s, so the use of prices to measure technological progress seems a reasonable approximation.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
3
released in July 2001. These revisions included a
downward adjustment to software investment as
well as a new quality-adjusted price index for local
area networks. Both of these revisions are incorporated into the estimates of IT investment.
The capital service estimates are based on the
Tangible Wealth Survey, published by the BEA and
described in Herman (2001). This survey includes
data on business investment and consumer durable
purchases for the U.S. economy through 2000. We
construct capital stocks from the investment data
by the perpetual inventory method and assume that
the effective capital stock for each asset is the average of the current and lagged estimates. The data
Changes in the underlying trend growth rate
of productivity are likely to be permanent, but
cyclical factors such as strong output growth
or extraordinarily rapid investment are more
likely to be temporary.
on tangible assets from the BEA are augmented
with inventory data to form the measure of the
reproducible capital stock. The total capital stock
also includes land and inventories.
Finally, we estimate capital service flows by multiplying rental prices and effective capital stocks, as
originally proposed by Jorgenson and Griliches
(1996). The estimates incorporate asset-specific
differences in taxes, asset prices, service lives, and
depreciation rates. This method is essential for
understanding the productive impact of IT investment because IT assets differ dramatically from
other assets in rates of decline of asset prices and
depreciation rates.
The difference between the growth in aggregate
capital service flows and effective capital stocks is
referred to as the growth in capital quality. That is,
(5) ∆ln KQ = ∆ln K – ∆ln Z,
where KQ is capital quality, K is capital service flow,
and Z is the effective capital stock. The aggregate
capital stock, Z, is a quantity index over seventy different effective capital stocks plus land and inventories using investment goods prices as weights.
The aggregate flow of capital services, K, is a quantity index of the same stocks using rental (or service) prices as weights. The difference in growth
4
rates is the growth rate of capital quality, KQ. As
firms substitute among assets by investing relatively
more in assets with relatively high marginal products,
capital quality increases.
Labor input is a quantity index of hours worked
that takes into account the heterogeneity of the work
force among sex, employment class, age, and education levels. The weights used to construct the index
are the compensation of the various types of workers. In the same way as for capital, we define growth
in labor quality as the difference between the growth
rate of aggregate labor input and hours worked:
(6) ∆ln LQ = ∆ln L – ∆ln H,
where LQ is labor quality, L is the labor input index,
and H is hours worked. As firms substitute among
hours worked by hiring relatively more highly skilled
and highly compensated workers, labor quality rises.
The labor data incorporate the Censuses of
Population for 1970, 1980, and 1990, the annual
Current Population Surveys (CPS), and the NIPA.
This study takes total hours worked for private
domestic employees directly from the NIPA
(Table 6.9c), self-employed hours worked for the
nonfarm business sector from the BLS, and selfemployed hours worked in the farm sector from
the Department of Agriculture.
Results. Table 1 reports the estimates of the
components of equation 2, the sources of economic
growth. For the period as a whole, output grew approximately 3.6 percent per year. Capital input made the
largest contribution to growth of 1.8 percentage
points, followed by approximately 1.2 percentage
points from labor input. Less than 20 percent of
output growth, 0.7 percentage point, directly reflects
TFP. These results are consistent with the other
recent growth accounting decompositions like CEA
(2001), Jorgenson and Stiroh (2000), and Oliner
and Sichel (2000, 2002).
The data also show the substantial acceleration in
output growth after 1995. Output growth increased
from 3 percent per year for the 1973–95 period to
4.6 percent for the 1995–2000 period, reflecting large
increases in IT and non-IT investment goods. On
the input side, more rapid capital accumulation contributed 0.84 percentage point to the post-1995 acceleration while faster growth of labor input contributed
0.30 percentage point and accelerated TFP growth
the remaining 0.47 percentage point. The contribution of capital input from IT increased from 0.36 percentage point per year for the 1973–95 period to 0.85
for the 1995–2000 period, exceeding the increased
contributions of all other forms of capital.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 1
Sources of Growth in Private Domestic Output, 1959–2000
1959–73
1973–95
1995–2000
1995–2000
less
1973–95
3.61
4.24
2.99
4.60
1.61
3.30
0.16
0.09
0.07
4.10
0.07
0.03
0.05
2.68
0.17
0.09
0.06
3.79
0.37
0.26
0.17
1.12
0.20
0.18
0.11
1.80
1.44
0.19
1.99
1.81
0.09
1.54
1.18
0.20
2.38
1.52
0.47
0.84
0.34
0.28
Communications (Km )
Contribution of labor (L)
Aggregate total factor productivity (TFP)
0.09
0.08
1.16
0.66
0.03
0.06
1.12
1.13
0.09
0.07
1.12
0.33
0.25
0.13
1.42
0.80
0.16
0.06
0.30
0.47
Contribution
Contribution
Contribution
Contribution
0.47
1.33
0.28
0.88
0.34
1.65
0.39
0.73
0.41
1.14
0.23
0.89
1.09
1.28
0.17
1.26
0.69
0.15
–0.06
0.37
1959–2000
Growth in private domestic output (Y)
Contribution of selected output components
Other output (Yn )
Computer investment (Ic )
Software investment (Is )
Communications investment (Im )
Contribution of capital and CD services (K)
Other (Kn )
Computers (Kc )
Software (Ks )
of
of
of
of
capital and CD quality
capital and CD stock
labor quality
labor hours
Note: A contribution of an output or input is defined as the share-weighted, real growth rate. “CD” stands for consumer durables.
Source: Authors’ calculations based on BEA, BLS, Census Bureau, and other data
The last four rows in Table 1 present an alternative decomposition of the contribution of capital and labor inputs using equations 5 and 6. Here,
the contribution of capital and labor reflects the
contributions from capital quality and capital stock
as well as labor quality and hours worked, respectively, as
(7) ∆ ln Y = vK ∆ ln Z + vK ∆ ln KQ + vL ∆ ln H
+ vL ∆ ln LQ + ∆ ln A.
Table 1 shows that the revival of output growth
after 1995 can be attributed to two forces. First, a
massive substitution toward IT assets in response to
accelerating IT price declines is reflected in the rising contribution of capital quality while the growth of
capital stock lagged considerably behind the growth
of output. Second, the growth of hours worked
surged as the growth of labor quality declined. A fall
in the unemployment rate and an increase in labor
force participation drew more workers with relatively
low marginal products into the workforce. We employ
equation 7 in projecting sustainable growth of output
and labor productivity in the next section.
Table 2 presents estimates of the sources of ALP
growth, as in equations 3 and 4. For the period as a
whole, growth in ALP accounted for nearly 60 percent of output growth, due to annual capital deepening of 1.13 percentage points, improvement of
labor quality of 0.28 percentage point, and TFP
growth of 0.66 percentage point. Growth in hours
worked of 1.54 percentage points per year accounted
for the remaining 40 percent of output growth.
Looking more closely at the post-1995 period, one
sees that labor productivity increased by 0.92 percentage points per year from 1.44 for the 1973–95
period to 2.36 for the 1995–2000 period, and hours
worked increased by 0.68 percentage points from an
annual rate of 1.55 for the 1973–95 period to 2.24 for
the 1995–2000 period. The labor productivity growth
revival reflects more rapid capital deepening of 0.52
percentage point and accelerated TFP growth of
0.47 percentage point per year; the contribution of
labor quality declined. Nearly all of the increase in
capital deepening was from IT assets with only a
small increase from other assets. Finally, we estimate that improved productivity in the production
of IT-related assets contributed 0.27 percentage
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
5
TABLE 2
Sources of Growth in Average Labor Productivity, 1959–2000
Output growth (Y)
Hours growth (H)
Average labor productivity growth (ALP)
Capital deepening
IT capital deepening
Other capital deepening
Labor quality
TFP growth
IT-related contribution
Other contribution
1959–2000
1959–73
3.61
1.54
2.07
1.13
0.32
0.82
0.28
0.66
0.23
0.43
4.24
1.27
2.97
1.44
0.16
1.28
0.39
1.13
0.10
1.03
1973–95
2.99
1.55
1.44
0.88
0.32
0.56
0.23
0.33
0.24
0.08
1995–2000
1995–2000
less
1973–95
4.60
2.24
2.36
1.40
0.76
0.64
0.17
0.80
0.51
0.29
1.61
0.68
0.92
0.52
0.44
0.08
–0.06
0.47
0.27
0.20
Note: A contribution of an output or input is defined as the share-weighted, real growth rate.
Source: Authors’ calculations based on BEA, BLS, Census Bureau, and other data
point to aggregate TFP growth while improved productivity growth in the rest of the economy contributed the remaining 0.2 percentage point. These
results suggest that IT had a substantial role in the
revival of labor productivity growth through both
capital deepening and TFP channels.
Our estimate of the magnitude of the productivity revival is somewhat lower than that reported in
earlier studies by BLS (2001a), Jorgenson and Stiroh
(2000), and Oliner and Sichel (2000). These studies
were based on data reported prior to the July 2001
revision of the NIPA, which substantially lowered
GDP growth in 1999 and 2000. Our estimates of the
productivity revival are also lower than the estimates
in BLS (2001b), however, which does include the
July 2001 revisions in GDP.
BLS (2001b) reports business sector ALP growth
of 2.68 percentage points for 1995–2000 and 1.45
for 1973–95, an increase of 1.23 percentage points,
compared to our estimated acceleration of 0.92 percentage point. This divergence results from a combination of a slower acceleration of our broader
concept of output and our estimates of more rapid
growth in hours worked. BLS (2001b), for example,
reports that hours grew 1.95 percent per year for
the 1995–2000 period in the business sector while
our estimate is 2.24.
Our estimate of private domestic employee
hours is taken directly from the NIPA and includes
workers in the nonprofit sector, and the BLS estimate does not. In addition, BLS (2001b) has revised
the growth in business sector hours in 2000 downward by 0.4 percentage point on the basis of new
data from the 2000 Hours at Work Survey. Our
estimate of labor quality change is also slightly
6
different from BLS (2001a) because of the different methods of estimating the wage-demographic
relationships and our use of only the March CPS
data as opposed to the monthly CPS data used by
BLS. These differences ultimately appear in our
estimated contribution to TFP from non-IT sources
because this cannot be observed directly without
detailed industry data, and we therefore estimate
it as a residual.
Projecting Productivity Growth
hile there is little disagreement about the
resurgence of ALP growth after 1995, there
has been considerable debate about whether this
is permanent or temporary. Changes in the underlying trend growth rate of productivity are likely
to be permanent, but cyclical factors such as strong
output growth or extraordinarily rapid investment
are more likely to be temporary. This distinction is
crucial to understanding the sources of the recent
productivity revival and projecting future productivity growth.
This section presents projections of trend rates
of growth for output and labor productivity over the
next decade, abstracting from business cycle fluctuations. The key assumptions are that output and
the reproducible capital stock will grow at the same
rate and that labor hours will grow at the same rate
as the labor force.4 These features are characteristic of the U.S. and most industrialized economies
over periods of time longer than a typical business
cycle. For example, U.S. output growth averaged
3.6 percent per year for the 1959–2000 period while
our measure of the reproducible capital stock grew
3.9 percent per year.5
W
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
We begin by decomposing the aggregate capital
stock into the reproducible component, ZR, and
business sector land, LAND, which we assume to be
fixed. This decomposition implies that
(8) ∆ ln Z = µ R ∆ ln ZR + (1 − µ R )∆ ln LAND
= µ R ∆ ln ZR ,
where u–R is the value share of reproducible capital
stock in total capital stock.
We then employ our projection assumptions to
construct estimates of trend output and productivity
growth, which are conditional on the projected growth
of the remaining sources of economic growth. More
formally, if ∆lnY = ∆lnZR, then combining equations
3, 4, 7, and 8 implies that trend labor productivity
and output growth are given by
– )∆ln H + v– ∆ln LQ
(9) ∆lny = [v–K ∆ln KQ – v–K(1 – µ
R
L
– )
–
+ uIT ∆ln A IT + ln An ] / (1 – v–K µ
R
∆lnY = ∆lny + ∆ln H.
Equation 9 is a long-run relationship that averages
over cyclical and stochastic elements and removes
the transitional dynamics relating to capital accumulation. The second part of a definition of trend
growth is that the unemployment rate remains constant and hours growth matches labor force growth.
Growth in hours worked was exceptionally rapid in
the 1995–2000 period as the unemployment rate
fell from 5.6 percent in 1995 to 4 in 2000, so output
growth was considerably above its trend rate.6 To
estimate hours growth over the next decade, we
employ detailed demographic projections based on
Census Bureau data.
To complete intermediate-term growth projections based on equation 9 requires estimates of capital and labor shares, IT output shares, reproducible
capital stock shares, capital quality growth, labor
quality growth, and TFP growth. Labor quality growth
and the various shares are relatively easy to project,
but extrapolations of the other variables involve much
greater uncertainty. Accordingly, we present three
sets of projections—a base-case scenario, a pessimistic
scenario, and an optimistic scenario.
We hold labor quality growth, hours growth, the
capital share, the reproducible capital stock share,
and the IT output share constant across the three
scenarios and refer to these as the “common assumptions.” We vary IT-related TFP growth, the contribution to TFP growth from non-IT sources, and
capital quality growth across these scenarios and
label them “alternative assumptions.” Generally
speaking for these variables, the base-case scenario
incorporates data from the 1990–2000 business
cycle, the optimistic scenario assumes the patterns
of the 1995–2000 period will persist, and the pessimistic case assumes that the economy reverts to
1973–95 averages.
Common assumptions. Hours growth (∆ln H)
and labor quality growth (∆ln LQ) are relatively
An important difficulty in projecting capital
quality growth from recent data is that investment patterns in the 1990s may partially reflect
an unsustainable investment boom in response
to temporary factors like Y2K investment and
the Nasdaq stock market bubble.
easy to project. The Congressional Budget Office
(CBO) (2001a), for example, projects growth in
the economywide labor force of 1.1 percent per
year based on Social Security Administration projections of population growth. Potential hours
growth is projected at 1.2 percent per year for the
nonfarm business sector for 2001–11 based on
CBO projections of hours worked for different
demographic categories of workers. The CBO estimate of potential hours growth is a slight increase
from earlier projections due to incorporation of
recent data from the 2000 census and changes in
the tax laws that will modestly increase the supply of labor. The CBO (2001a) does not employ
the concept of labor quality.
We construct our own projections of demographic
trends. Ho and Jorgenson (1999) have shown that
the dominant trends in labor quality growth are due
to rapid improvements in educational attainment in
the 1960s and 1970s and the rise in female participation rates in the 1970s. Although the average
educational level continued to rise as younger and
4. The assumption that output and the capital stock grow at the same rate is similar to a balanced growth path in a standard
growth model, but our actual data with many heterogeneous types of capital and labor inputs make this interpretation only
an approximation.
5. Reproducible assets include equipment, structures, consumer durable assets, and inventories but exclude land.
6. These unemployment rates are annual averages for the civilian labor force sixteen years and older from the BLS.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
7
better-educated workers entered the labor force
and older workers retired, the improvement in educational attainment of new entrants into the labor
force largely ceased in the 1990s.
Growth in the population is projected from the
Bureau of the Census demographic model, which
breaks the population down by individual year of
age, race, and sex.7 For each group, the population
in period t is equal to the population in period t – 1,
less deaths plus net immigration. Death rates are
group-specific and are projected by assuming a
steady rate of improvement in health. The population of newborns in each period reflects the number
of females in each age group and the age- and race-
Our optimistic scenario puts labor productivity
growth just below 3 percent per year and
reflects the assumption of continuing rapid
technological progress.
specific fertility rates. These fertility rates are projected to fall steadily.
We observe labor force participation rates in the
last year of our sample period and then project the
work force by assuming constant participation rates
for each sex-age group. The educational attainment
of workers aged α in period t is projected by assuming that it is equal to the attainment of the workers
of age α – 1 in period t – 1 for all those who are over
thirty-five years old in the last year of the sample.
For those who are younger than thirty-five, we
assume that the educational attainment of workers
aged α in forecast period t is equal to the attainment of workers aged α in the base year.
The index of labor quality is constructed from
hours worked and compensation rates. We project
hours worked by multiplying the projected population in each sex-age-education group by the annual
hours worked per person in the last year of the sample. The relative compensation rates for each group
are assumed to be equal to the observed compensation in the sample period. With these projected
hours and compensation we forecast the quality
index over the next twenty years.
Our estimates suggest that hours growth (∆lnH)
will be about 1.1 percent per year over the next ten
years, which is quite close to the CBO (2001a) estimates, and 0.8 percent per year over a twenty-year
8
period. We estimate that growth in labor quality
(∆lnLQ) will be 0.27 percent per year over the next
decade and 0.17 percent per year over the next two
decades. This estimate is considerably lower than the
0.49 percent growth rate for the 1959–2000 period,
which was driven by rising average educational
attainment and stabilizing female participation.
The capital share (v–K) has not shown any obvious trend over the last forty years. We assume it
holds constant at 42.8 percent, the average for
1959–2000. Similarly, the fixed reproducible capital
– ) has shown little trend, and we assume it
share (µ
R
remains constant at 80.4 percent, the average for
1959–2000.
We assume the IT output share (u–IT) stays at 5.1
percent, the average for the 1995–2000 period. This
estimate is likely conservative because IT has steadily
increased in importance in the U.S. economy, rising
from 2.1 percent of output in 1970 to 2.7 percent in
1980, 3.9 percent in 1990, and 5.7 percent in 2000.
On the other hand, there has been speculation that
IT expenditures in the late 1990s were not sustainable because of Y2K investment, the Nasdaq bubble,
and abnormally rapid price declines.8
Alternative assumptions. IT-related productivity growth (∆ln AIT) has been extremely rapid
in recent years with a substantial acceleration
after 1995. For the 1990–95 period productivity
growth for production of the three IT assets averaged 7.4 percent per year while the 1995–2000 average growth rate was 10.3 percent. These growth
rates are high but quite consistent with industrylevel productivity estimates for high-tech sectors.
For example, BLS (2001a) reports productivity
growth of 6.9 percent per year for the 1995–99
period in industrial and commercial machinery,
which includes production of computer hardware,
and 8.1 percent in electronic and other electric
equipment, which includes semiconductors and
telecommunications equipment.
Jorgenson (2001) argues that the large increase
in IT productivity growth was triggered by a much
sharper acceleration in the decline of semiconductor prices that can be traced to a shift in the product cycle for semiconductors in 1995 from three
years to two years, a consequence of intensifying
competition in the semiconductor market. It would
be premature to extrapolate the recent acceleration
in productivity growth into the indefinite future,
however, because this depends on the persistence
of a two-year product cycle for semiconductors.
To better gauge the future prospects of technological progress in the semiconductor industry, we
turn to the International Technology Roadmap
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
for Semiconductors (2000). This projection, performed annually by a consortium of industry associations, forecasts a two-year product cycle through
2003 and a three-year product cycle thereafter. The
Roadmap is a reasonable basis for projecting the
IT-related productivity growth of the U.S. economy.
Moreover, continuation of a two-year cycle provides
an upper bound for growth projections while reversion to a three-year cycle gives a lower bound.
Our base-case scenario follows the Roadmap
and averages the two-year and three-year cycle projections with IT-related growth of 8.8 percent per
year, which equals the average for the 1990–2000
period. The optimistic projections assume that the
two-year product cycle for semiconductors remains
in place over the intermediate future so that productivity growth in the production of IT assets averages 10.3 percent per year, as it did for 1995–2000.
The pessimistic projection assumes the semiconductor product cycle reverts to the three-year cycle
in place during the 1973–95 period when IT-related
productivity growth was 7.4 percent per year. In all
cases, the contribution of IT to aggregate TFP
growth reflects the 1995–2000 average share of
about 5.1 percent.
The TFP contribution from non-IT sources (∆ An)
is more difficult to project because the post-1995
acceleration is outside of standard growth models.
Therefore, we present a range of alternative estimates that are consistent with the historical record.
The base case uses the average contribution from
the full business cycle of the 1990s and assumes a
contribution of 0.2 percentage point for the intermediate future. This base case assumes that the
myriad of factors that drove TFP growth in the
1990s—such as technological progress, innovation,
resource reallocations, and increased competitive
pressures—will continue into the future. The optimistic case assumes that the contribution for
1995–2000 of 0.29 percentage point per year will
continue for the intermediate future while our pessimistic case assumes that the U.S. economy will
revert to the slow-growth 1973–95 period, when
this contribution averaged only 0.08 percent per year.
The final step in our projections is to estimate
the growth in capital quality (∆ln KQ). The workhorse aggregate growth model with one capital
good has capital stock and output growing at the
same rate in a balanced growth equilibrium, and
even complex models typically have only two capital
goods. The U.S. data, however, distinguish between
several dozen types of capital, and the historical
record shows that substitution between these types
of capital is an important source of output and productivity growth. For 1959–2000, for example, capital quality growth contributed 0.47 percentage
point to output growth as firms substituted toward
short-lived assets with higher marginal products.
This contribution corresponds to a growth in capital
quality of about 1 percent per year.
An important difficulty in projecting capital quality
growth from recent data, however, is that investment
patterns in the 1990s may partially reflect an unsustainable investment boom in response to temporary
factors like Y2K investment and the Nasdaq stock
Our primary conclusion is that a consensus
has emerged about trend rates of growth for
output and labor productivity.
market bubble, which skewed investment toward IT
assets. Capital quality for the 1995–2000 period grew
at 2.5 percent per year as firms invested heavily in IT,
for example, but there has been a sizable slowdown
in IT investment in the second half of 2000 and in
2001. Therefore, we are cautious about relying too
heavily on the recent investment experience.
The base case again uses the average rate for
1990–2000, which was 1.75 percentage points for
capital quality; this rate effectively averages the high
substitution rates in the late 1990s with the more
moderate rates of the early 1990s and uses evidence
from the complete business cycle of the 1990s. The
optimistic projection ignores the belief that capital
substitution was unsustainably high in the late 1990s
and assumes that capital quality growth will continue
at the 2.45 percent annual rate of the 1995–2000
period. Our pessimistic scenario assumes that the
growth of capital quality reverts to the 0.84 percent
annual growth rate seen for the 1973–95 period.
Output and productivity projections. Table 3
assembles the components of the projections and
presents the three scenarios. The top section shows
the projected growth of output, labor productivity,
and the effective capital stock. The middle section
7. See Bureau of the Census (2000) for details of the population model.
8. See McCarthy (2001) for determinants of investment in the late 1990s.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
9
TABLE 3
Output and Labor Productivity Projections
Scenarios
1995–2000
Pessimistic
Base-Case
Optimistic
Output growth
ALP growth
Effective capital stock
4.60
2.36
2.94
2.43
1.33
1.96
Projections
3.31
2.21
2.66
4.02
2.92
3.23
Hours growth
Labor quality growth
Capital share
IT output share
Reproducible capital stock share
2.240
0.299
0.438
0.051
0.798
1.100
0.265
0.428
0.051
0.804
Common assumptions
1.100
0.265
0.428
0.051
0.804
1.100
0.265
0.428
0.051
0.804
7.39
0.37
0.08
0.84
Alternative assumptions
8.78
0.44
0.20
1.75
TFP growth in IT
Implied IT-related TFP contribution
Other TFP contribution
Capital quality growth
10.33
0.52
0.29
2.45
10.28
0.52
0.29
2.45
Notes: In all projections, hours growth and labor quality growth are from internal projections, capital share and reproducible capital stock
shares are 1959–2000 averages, and IT output shares are for 1995–2000. The pessimistic case uses 1973–95 average growth of capital quality, IT-related TFP growth, and non-IT TFP contribution. The base case uses 1990–2000 averages, and the optimistic case uses 1995–2000
averages.
reports the five factors that are held constant across
scenarios—hours growth, labor quality growth, the
capital share, the IT output share, and the reproducible capital stock share. The bottom section
includes the three components that vary across
scenarios—TFP growth in IT, the TFP contribution
from other sources, and capital quality growth.
Table 3 also compares the projections with actual
data for the same series for 1995–2000.
The base-case scenario puts trend labor productivity growth at 2.21 percent per year and trend
output growth at 3.31 percent per year. Projected
productivity growth falls just short of our estimates
for the 1995–2000 period, but output growth is considerably slower due to the large slowdown in projected hours growth; hours grew 2.24 percent per
year for the 1995–2000 period compared to our projection of only 1.1 percent per year for the next
decade. Capital stock growth is projected to fall in
the base case to 2.66 percent per year from 2.94
percent for the 1995–2000 period.
Our base-case scenario incorporates the underlying pace of technological progress in semiconductors
embedded in the Roadmap forecast and puts the
contribution of IT-related TFP below that of the
1995–2000 period as the semiconductor industry
eventually returns to a three-year product cycle. The
slower growth is partially balanced by larger IT out10
put shares. Other TFP growth also makes a smaller
contribution. Finally, the slower pace of capital input
growth is offset by slower hours growth so that strong
capital deepening brings the projected growth rate
near the observed rates of growth for 1995–2000.
Our optimistic scenario puts labor productivity
growth just below 3 percent per year and reflects
the assumption of continuing rapid technological
progress. In particular, the two-year product cycle
in semiconductors is assumed to persist for the
intermediate future, driving rapid TFP in production of IT assets as well as continued substitution
toward IT assets and rapid growth in capital quality.
In addition, other TFP growth continues the relatively rapid contribution seen after 1995.
Finally, the pessimistic projection of 1.33 percent annual growth in labor productivity assumes
that many trends revert to the sluggish growth rates
of the 1973–95 period and that the three-year product cycle for semiconductors begins immediately.
The larger share of IT, however, means that even
with the return to the three-year technology cycle
and slower TFP growth, labor productivity growth
will equal the rates seen in the 1970s and 1980s.
Alternative Methodologies and Estimates
T
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
his section briefly reviews alternative approaches
to estimating productivity growth trends from
the historical record and projecting productivity
growth going forward. We begin with the econometric methods for separating trend and cyclical
components of productivity growth employed by
Gordon (2000), French (2001), and Roberts (2001).
A second approach is to control for factors that
are most likely to be cyclical, such as factor utilization, in the augmented growth accounting
framework of Basu, Fernald, and Shapiro (2001).
In a third approach, the CBO (2001a, 2001b) calibrates a growth model to the historical record and
uses the model to project growth of output and
productivity. Finally, Oliner and Sichel (2002) present a projection methodology based on a growth
accounting framework; this paper appears in this
issue of the Economic Review and is not discussed
in detail here.
Econometric estimates. We begin with the
studies that employ econometric methods for
decomposing a single time series between cyclical
and trend components. Gordon (2000) estimates
that of the 2.75 percent annual labor productivity
growth rate during the 1995–99 period, 0.5 percent
can be attributed to cyclical effects and 2.25 percent to trend. The post-1995 trend growth rate is
0.83 percent higher than the growth rate in the
1972–95 period. Capital and labor input growth and
price measurement changes account for 0.52 percent, and TFP growth in the computer sector
accounts for 0.29 percent, leaving a mere 0.02 percent to be explained by acceleration in TFP growth
in the other sectors of the private economy. In this
view the productivity revival is concentrated in the
computer-producing sector.
Other studies have employed state-space models
to distinguish between trend and cycles for output.
Roberts (2001) uses time-varying parameter methods to model the growth of labor and total factor
productivity. He represents trend productivity as a
random walk with drift and allows the drift term to
be a time-varying parameter. These estimates suggest
that trend labor productivity growth has increased
from 1.6 percent per year during the 1973–94 period
to 2.7 percent by 2000 while trend TFP growth rose
from 0.5 percent during the 1985–95 period to 1.1 percent during the 1998–2000 period. This estimate of
trend labor productivity falls between our base-case
and optimistic projections.
French (2001) uses a Cobb-Douglas production
function to model trends and cycles in total factor
productivity growth. He considers filtering methods
and concludes that they are all unsatisfactory
because of the assumption that innovations are normally distributed.9 He applies a discrete innovations
model with two high-low TFP growth regimes and
finds that the trend TFP growth after 1995 increases
from 1.01 percent to 1.11 percent.
Finally, Hansen (2001) provides a good primer on
recent advances in the alternatives to random walk
models—testing for infrequent structural breaks in
parameters. Applying these methods to the U.S.
manufacturing sector, he finds strong evidence of a
break in labor productivity in the mid-1990s, the
break date depending on the sector being analyzed.
We do not compare his specific estimates because
they are only for manufacturing.
Our second conclusion is that trend growth
rates are subject to considerable uncertainty.
For the U.S. economy this can be identified with
the future product cycle for semiconductors
and the impact on other high-tech gear.
Augmented growth accounting. Basu, Fernald,
and Shapiro (2001) present an alternative approach
to estimating trend growth in total factor productivity by separately accounting for factor utilization
and factor accumulation. They extend the growth
accounting framework to incorporate adjustment
costs, scale economies, imperfect competition, and
changes in utilization. Industry-level data for the
1990s suggest that the post-1995 rise in productivity
appears to be largely a change in trend rather than
a cyclical phenomenon since there was little change
in utilization in the late 1990s. While Basu, Fernald,
and Shapiro are clear that they do not make predictions about the sustainability of these changes, their
results suggest that any slowdown in investment
growth is likely to be associated with a temporary
increase in output growth as resources are reallocated away from adjustment and toward production.
Calibration and projection. The CBO (2001a)
presents medium-term projections for economic
growth and productivity for the 2003–11 period for
both the overall economy and the nonfarm business
sector. The CBO’s most fully developed model is for
the nonfarm business sector. Medium-term projections are based on historical trends in the labor
force, savings and investment, and TFP growth.
9. Both Roberts (2001) and French (2001) employ the Stock and Watson (1998) method of dealing with the zero bias.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
11
These projections allow for possible business cycle
fluctuations, but the CBO does not explicitly forecast fluctuations beyond two years (CBO 2001a, 38).
For the nonfarm part of the economy, the CBO
(2001a) projects potential output growth of 3.7 percent per year and potential labor productivity of
2.5 percent per year. For the economy as a whole,
the CBO projects potential labor productivity growth
of 2.1 percent per year, which is quite close to our
estimates.
For the nonfarm business economy, the CBO
(2001a) utilizes a Cobb-Douglas production function
without labor quality improvement. The CBO’s relatively high projection of labor productivity growth
for the nonfarm business sector reflects projections
of capital input growth of 4.8 percent per year and
TFP growth of 1.4 percent per year.10 The CBO’s
relatively rapid rate of capital input growth going
forward is somewhat slower than their estimate of
5.2 percent for the 1996–2000 period but considerably faster than their estimate of 3.9 percent annual
growth for the 1990–2000 period. These estimates
reflect the model of savings and investment used by
the CBO as well as the expectation of continued
substitution toward short-lived IT assets. Potential
TFP growth of 1.4 percent per year reflects an estimated trend growth of 1.1 percent per year augmented by the specific effects of computer quality
improvement and changes in price measurement.
Conclusion
ur primary conclusion is that a consensus has
emerged about trend rates of growth for output and labor productivity. Our central estimates of
2.21 percent for labor productivity and 3.31 percent
for output are very similar to those of Gordon
(2000) and the CBO (2001a) and only slightly more
optimistic than Baily’s (2001).11 Our methodology
assumes that trend growth rates in output and
O
reproducible capital are the same and that hours
growth is constrained by the growth of the labor
force to form a balanced growth path. While productivity is projected to fall slightly from the pace
seen in late 1990s, we conclude that the U.S. productivity revival is likely to remain intact for the
intermediate future.
Our second conclusion is that trend growth rates
are subject to considerable uncertainty. For the U.S.
economy this can be identified with the future
product cycle for semiconductors and the impact
on other high-tech gear. The switch from a threeyear to a two-year product cycle in 1995 produced
a dramatic increase in the rate of decline of IT
prices. This is reflected in the investment boom of
the 1995–2000 period and the massive substitution
of IT capital for other types of capital that took
place in response to price changes. The issue that
must be confronted by policymakers is whether this
two-year product cycle can continue and whether
firms will continue to respond to the dramatic
improvements in the performance/price ratio of IT
investment goods.
As a final point, we have not tried to quantify
another important source of uncertainty, namely,
the economic impacts of the events of September 11.
These impacts are already apparent in the slowdown of economic activity in areas related to travel
and increased security as well as higher government expenditures for the war in Afghanistan and
enhanced homeland security. The cyclical effects
will likely produce only a temporary reduction in
productivity as civilian plants operate at lower utilization rates. Even a long-term reallocation of
resources from civilian to public goods or to security operations, however, should produce only a
one-time reduction in productivity levels rather
than a change in the trend rate of growth of output and productivity.
10. See CBO (2001b) for details. Note also that the CBO assumes a capital share of 0.3, which is substantially smaller than our
estimate of 0.43.
11. Note that our output concept is slightly different so the estimates are not directly comparable. Nonetheless, the broad predictions are similar.
12
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
REFERENCES
Aizcorbe, Ana. 2002. Why are semiconductor prices
falling so fast? Industry estimates and implications for
productivity measurement. Federal Reserve Board of
Governors, February. Photocopy.
Herman, Shelby W. 2001. Fixed assets and consumer
durables for 1925–2000. Survey of Current Business
(September): 27–38.
Baily, Martin Neal. 2001. Macroeconomic implications of
the new economy. In Economic policy for the information economy, the proceedings of a symposium presented
by the Federal Reserve Bank of Kansas City, Jackson
Hole, Wyo., August 30.
Basu, Susanto, John G. Fernald, and Matthew G. Shapiro.
2001. Productivity growth in the 1990s: Technology, utilization, or adjustment? Carnegie-Rochester Conference
Series on Public Policy 55 (December): 117–65.
Bureau of the Census. 2000. Methodology and assumptions for the population projections of the United States:
1999 to 2100. <www.census.gov/population/www/
projections/natproj.html> (November 2001).
Bureau of Labor Statistics (BLS). 2001a. Multifactor productivity trends, 1999. USDL 01-125, May 3.
———. 2001b. Productivity and costs, third quarter
2001. USDL 01-452, December 6.
Congressional Budget Office (CBO). 2001a. The budget
and economic outlook: An update. Washington, D.C.:
GPO.
International technology roadmap for semiconductors,
2000 update. 2000. Austin, Tex.: Sematech Corporation.
<http://public.itrs.net> (October 2001).
Jorgenson, Dale W. 1996. The embodiment hypothesis.
In Postwar U.S. economic growth. Cambridge, Mass.:
MIT Press.
———. 2001. Information technology and the U.S. economy. American Economic Review 91 (March): 1–32.
Jorgenson, Dale W., and Zvi Griliches. 1996. The explanation of productivity change. In Postwar U.S. economic
growth. Cambridge, Mass.: MIT Press.
Jorgenson, Dale W., and Kevin J. Stiroh. 2000. Raising
the speed limit: U.S. economic growth in the information
age. Brookings Papers on Economic Activity 1: 125–211.
———. 2001b. CBO’s method for estimating potential
output: An update. Washington, D.C.: GPO.
Council of Economic Advisers (CEA). 2001. Annual report
of the Council of Economic Advisers. In Economic report
of the president. Washington, D.C.: GPO.
DeLong, J. Bradford, and Lawrence M. Summers. 2001.
The “new economy”: Background, historical perspective,
questions, and speculation. In Economic policy for the
information economy, the proceedings of a symposium
presented by the Federal Reserve Bank of Kansas City,
Jackson Hole, Wyo., August 30.
French, Mark W. 2001. Estimating change in trend growth
of total factor productivity: Kalman and H-P filters versus
a Markov-switching framework. Board of Governors of
the Federal System, Finance and Economics Discussion
Series 2001-44.
Hansen, Bruce E. 2001. The new econometrics of structural change: Dating breaks in U.S. labor productivity.
Journal of Economic Perspectives 15 (Fall): 117–28.
Hobijn, Bart. 2001. Is equipment price deflation a statistical artifact? Federal Reserve Bank of New York Staff
Report 139, November.
Jorgenson, Dale W., Mun S. Ho, and Kevin J. Stiroh. 2002.
Productivity and labor quality in U.S. industries. Prepared
for NBER/CRIW Conference on Measurement of Capital
in the New Economy, April.
———. 2002. Productivity and costs, fourth quarter and
annual averages, 2001. USDL 02-123, March 7.
Gordon, Robert J. 2000. Does the “New Economy” measure up to the great inventions of the past? Journal of
Economic Perspectives 14 (Fall): 49–74.
Ho, Mun, and Dale W. Jorgenson. 1999. The quality of the
U.S. workforce 1948–95. Harvard University, Kennedy
School of Government Paper, February.
McCarthy, Jonathan. 2001. Equipment expenditures since
1995: The boom and the bust. Current Issues in Economics and Finance 7 (October): 1–6.
Oliner, Stephen D., and Daniel E. Sichel. 2000. The
resurgence of growth in the late 1990s: Is information
technology the story? Journal of Economic Perspectives 14 (Fall): 3–22.
———. 2002. Information technology and productivity:
Where are we now and where are we going? Federal
Reserve Bank of Atlanta Economic Review 87 (Third
Quarter): 15–44.
Roberts, John M. 2001. Estimates of the productivity
trend using time-varying parameter techniques. Board
of Governors of the Federal System, Finance and Economics Discussion Series 2001-8.
Stock, James H., and Mark W. Watson. 1998. Median unbiased estimation of coefficient variance in a time-varying
parameter model. Journal of the American Statistical
Association 93 (March): 349–58.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
13
Information Technology and
Productivity: Where Are We
Now and Where Are We Going?
STEPHEN D. OLINER AND DANIEL E. SICHEL
The authors are an associate director and a senior economist, respectively, in the Division of Research
and Statistics at the Federal Reserve Board. They thank Darrel Cohen, John Fernald, Jaime Marquez, David
Stockton, and conference participants for very useful comments and discussions. For extremely valuable help
with data, they are grateful to Charlie Gilbert from the Federal Reserve Board, Bruce Grimm and David
Wasshausen from the Bureau of Economic Analysis, and Michael Harper, Larry Rosenblum, and Steve
Rosenthal from the Bureau of Labor Statistics. This paper draws heavily from the authors’ earlier work,
including text taken directly from Oliner and Sichel (2000a, b) and Sichel (1997).
fter a quarter-century of lackluster
gains, the U.S. economy experienced a
remarkable resurgence in productivity
growth during the second half of the
1990s. From 1995 to 2000, output per
hour in nonfarm business grew at an
average annual rate of about 21/2 percent compared
with increases of only about 11/2 percent per year
from 1973 to 1995.1 Our earlier work, along with other
research, linked this improved performance to the
information technology (IT) revolution that has
spread through the U.S. economy.2 Indeed, by 2000
this emphasis on the role of information technology
had become the consensus view.
However, shortly after this consensus emerged, the
technology sector of the economy went into a tailspin
as demand for IT products fell sharply. Reflecting this
retrenchment, stock prices for many technology firms
collapsed, and financing for the sector dried up. These
developments raised questions about the robustness
of the earlier results that emphasized the role of information technology. They also cast some doubt on the
sustainability of the rapid productivity growth in the
second half of the 1990s. Nonetheless, the recent data
remain encouraging. Productivity gains have continued to be strong, with output per hour rising 2 percent over the four quarters of 2001—a much larger
increase than is typical during a recession.
Against the backdrop of these developments,
much effort has been devoted to estimating the
A
underlying trend in productivity growth (see, in particular, Baily 2002; DeLong 2002; Gordon 2002;
Jorgenson, Ho, and Stiroh 2002; Kiley 2001; Martin
2001; McKinsey Global Institute 2001; and Roberts
2001). For the most part, these papers take a relatively optimistic view of the long-run prospects for
productivity growth.
We add to this literature in two ways. First, to
assess the robustness of the earlier evidence on
the role of information technology, we extend the
growth-accounting results in Oliner and Sichel
(2000a) through 2001. These results continue to
support the basic story in our earlier work; namely,
the data still show a substantial pickup in labor productivity growth and indicate that both the use of
information technology and efficiency gains associated with the production of information technology
were central factors in that resurgence.
Second, to assess whether the pickup in productivity growth since the mid-1990s is sustainable, we
analyze the steady-state properties of a multisector
growth model. This exercise allows us to translate
alternative views about the evolution of the technology sector (and other sectors of the economy)
into “structured guesses” about future growth in
labor productivity. As highlighted by Jorgenson
(2001), the pace of technological progress in hightech industries—especially semiconductors—likely
will be a key driver of productivity growth going forward. Thus, we develop a model that is rich enough
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
15
to trace out the aggregate effects of these driving
influences. We view this steady-state machinery not
as a forecasting model per se but rather as a tool for
generating a likely range of outcomes for labor productivity growth over roughly the next decade.
Beyond that horizon, the uncertainty about the
structure and evolution of the economy is too great
for our steady-state approach to offer much insight.
Our structured guesses of labor productivity
growth range from 2 percent to roughly 23/4 percent
per year. The lower end of the range reflects conservative assumptions for key parameters in our model.
Notably, in this scenario, we assume that the rate of
technological advance in the semiconductor indus-
The pace of technological progress in hightech industries—especially semiconductors—
likely will be a key driver of productivity growth
going forward.
try drops back to its historical average from the
extremely fast pace in the second half of the 1990s
and that the semiconductor and other IT sectors fail
to grow any further as a share of (current-dollar)
economic activity. In contrast, to generate the upper
end of the range, we assume that the pace of technological advance in the semiconductor sector
reverts only halfway to its historical average and
that the various IT sectors continue to grow as a
share of the economy. Of course, much uncertainty
attends this exercise, and we also discuss more
extreme scenarios in which labor productivity
growth in the steady state would fall short of 2 percent or would exceed 3 percent. We believe, however, that these more extreme alternatives are less
likely to occur than the scenarios generating labor
productivity growth in the 2 to 23/4 percent range.
This range, which includes the pace recorded over
the second half of the 1990s, puts us squarely in
the camp of those who believe that a significant
portion—and possibly all—of the mid-1990s’ productivity resurgence is sustainable.
The next section of the paper provides a largely
nontechnical overview of the analytical framework.
Next, we briefly discuss the data we use and then
describe the growth-accounting results extended
through 2001. We continue by laying out the alternative steady-state scenarios that we analyze and
16
then present the steady-state results. Appendix 1
fully describes our multisector model and derives
all the theoretical results that underlie our growthaccounting and steady-state estimates. Appendix 2
provides detailed documentation for each data
series used in the paper.
Analytical Framework
his paper employs the neoclassical growthaccounting framework pioneered by Solow
(1957) and used extensively by researchers ever
since.3 The neoclassical framework decomposes the
growth in labor productivity, measured by output per
hour worked, into the contributions from three broad
factors: increases in the amount of capital per hour
worked (usually referred to as capital deepening),
improvements in the quality of labor and growth in
multifactor productivity (MFP). MFP is the residual
in this framework, capturing improvements in the
way that firms use their capital and labor but also
embedding any errors in the estimated contributions
from capital deepening and labor quality.
The growth-accounting framework can be tailored
to address many different issues. We employ it to
assess the growth contribution from IT capital, taking
account of both the use of this capital throughout
the economy and the efficiency gains realized in its
production. Given this focus, we construct a model
of the nonfarm business economy that highlights
key IT-producing industries. Our model, which
extends the two-sector models developed in Martin
(2001) and Whelan (2001), divides nonfarm business into five sectors. Three sectors produce final
IT goods—computer hardware, software, and communication equipment—and a large non-IT sector
produces all other final goods and services. The
fifth sector in the model produces semiconductors,
which are either consumed as an intermediate input
by the final-output sectors or exported to foreign
firms. To focus on the role of semiconductors in the
economy, the model abstracts from all other intermediate inputs.
Our model relies on several assumptions that are
typically imposed in growth-accounting studies. In
particular, we assume that all markets are perfectly
competitive and that production in every sector is
characterized by constant returns to scale. Labor
and capital are assumed to be completely mobile,
an assumption that implies a single wage rate for
labor across all sectors and a single rental rate for
each type of capital. Within this competitive market
structure, we assume that firms set their investment
and hiring decisions to maximize profits. Moreover,
when firms purchase new capital or hire additional
T
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
workers, we assume they do not incur any adjustment costs that would reduce output while these
new inputs are integrated into the firms’ production
routines. Finally, we do not explicitly model cyclical
changes in the intensity with which firms use their
capital and labor.
These assumptions yield a tractable analytical
framework by abstracting from some notable features
of the actual economy. One could be concerned that
these assumptions are so restrictive as to distort the
empirical results. In consideration of these concerns,
we would not advocate using a framework like ours
to decompose year-to-year changes in productivity
growth because cyclical factors omitted from the
model could substantially affect the results. However,
Basu, Fernald, and Shapiro (2001) showed that the
basic characterization of productivity trends in the
1990s remains intact even after allowing for adjustment costs, nonconstant returns to scale, and cyclical
variations in the use of capital and labor.
With this background, we now discuss the key
analytical results from our model. The rest of this
section presents and interprets these results; formal derivations can be found in Appendix 1.
Growth in aggregate labor productivity. As
shown in the first proposition of Appendix 1, our
model yields a standard decomposition of growth
•
in aggregate labor productivity. Let Z denote the
growth rate of any variable Z. Then, the growth of
output per hour for nonfarm business as a whole
can be written as
•
•
•
•
•
•
•
•
(1) Y − H = α CK ( KC − H ) + α KSW ( KSW − H ) + α KM (KM − H )
•
•
•
•
•
•
•
+ α OK ( K O − H ) + α L q + MFP
4
•
= ∑ α Kj ( K j − H ) + α L q + MFP,
the stocks of computer hardware, software, communication equipment, and all other tangible capital,
respectively; and q denotes labor quality. The α terms
are income shares; under the assumptions of our
model, the income share for each input equals its
output elasticity, and the shares sum to 1. The second
line of equation 1 merely rewrites the decomposition
with more compact notation, where j indexes the
four types of capital.4
Equation 1 shows that growth in labor productivity reflects capital deepening, improvements in
labor quality, and gains in MFP, with the overall
growth contribution from capital deepening constructed as the sum of the contributions from the
four types of capital. Each such contribution equals
the increase in that type of capital per work hour,
weighted by the income share for that capital. This
decomposition is entirely standard and matches the
one used in Oliner and Sichel (2000a, b). Note that
equation 1 does not identify the sectors using capital and labor; all that matters is the aggregate
amount of each input. Under our assumptions, we
need not keep track of the individual sectors
because each type of capital has the same marginal
product regardless of where it is employed, and the
same holds for labor. Hence, transferring capital or
labor from one sector to another has no effect on
labor productivity for nonfarm business as a whole.
Our growth-accounting decomposition depends
importantly on the income shares of the various
types of capital. These income shares are not
directly observable, and we estimate them in accordance with the method used by the Bureau of Labor
Statistics. In this framework, the income share for
capital of type j is
(2) αj = (R + δj – Πj )Tj pj Kj /pY,
j=1
where Y denotes nonfarm business output in real
terms; H denotes hours worked in nonfarm business;
KC, KSW, KM, and KO denote the services provided by
where R is a measure of the nominal net rate of
return on capital, which is the same for all types of
capital under our assumption of profit maximization
1. This paper was already in production at the time of the July 2002 annual revision of the National Income and Product
Accounts (NIPAs), and all numbers in the paper refer to the prerevision data. Had we been able to take the revision into
account, the basic story of the paper would remain intact although productivity growth and high-tech capital deepening in
recent years would be a bit lower than the figures we present.
2. See Oliner and Sichel (2000a), Bosworth and Triplett (2000), Brynjolfsson and Hitt (2000), Jorgenson and Stiroh (2000),
Jorgenson (2001), and Whelan (2000). For a more skeptical view of the role of information technology written at that time,
see Gordon (2000).
3. See Steindel and Stiroh (2001) for an overview of growth accounting, issues related to the measurement of productivity, and
trends in productivity growth in the postwar period.
4. Time subscripts on both the income shares and the various growth rates have been suppressed to simplify the notation. We
use log differences to measure growth rates. The income share applied to a log difference between periods t and t + 1 is measured as the average of the shares in these two periods.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
17
and full capital mobility; δj is the depreciation rate
for capital of type j; Π j measures any expected
change in the value of this capital over and above
that captured in the depreciation rate; Tj is a composite tax parameter; pj Kj is the current-dollar stock
of this capital; and pY is total current-dollar income
in the nonfarm business sector. The intuition behind
equation 2 is straightforward. In a competitive market, each dollar of type j capital must earn a gross
annual return that covers the net return common
to all capital as well as the loss of value that this
capital suffers over the year and the taxes imposed
on the income it generates. The product of this
gross return and the current-dollar stock equals
In addition to explaining the source of the
productivity pickup in the 1990s, we wish to
estimate a plausible range for productivity
growth in the future.
the current-dollar income assumed to be earned by
type j capital, which we divide by total income in
nonfarm business to obtain the desired income
share. Once we calculate each capital share in this
way, the labor share is simply one minus the sum of
the capital shares.
Aggregate and sectoral MFP growth. The
term for aggregate MFP growth in equation 1 can be
decomposed into the contributions from MFP
growth in each sector. In particular, proposition 1 in
the model appendix shows that
•
4
•
•
(3) MFP = ∑ µ i MFPi + µ S MFPS ,
i =1
where i indexes the four final-output sectors, s
denotes the semiconductor sector, and the µ term
for each sector represents its output expressed as
a share of total nonfarm business output in current dollars. This sectoral weighting scheme was
initially proposed by Domar (1961) and formally
justified by Hulten (1978). The Domar weights
sum to more than one, an outcome that may seem
odd at first glance.5 However, this weighting scheme
is needed to account for the production of intermediate inputs. Without this “gross-up” of the
weights, the MFP gains achieved in producing
semiconductors (the only intermediate input in
18
our model) would be omitted from the decomposition of aggregate MFP growth.
To see this point more clearly, note that equation 3
can be rewritten as
•
•
4
•
(4) MFP = ∑ µ i[ MFPi + βiS (1 + θ ) MFPS ],
i =1
where 1 + θ equals the ratio of domestic semiconductor output to domestic use of semiconductors
S
and βi denotes semiconductor purchases by finaloutput sector i as a share of the sector’s total input
costs. This result, derived in proposition 2, shows
that the semiconductor sector, in effect, can be vertically integrated with the final-output sectors that it
supplies. MFP growth in each vertically integrated
sector—the term in brackets—subsumes the MFP
gains at its dedicated semiconductor plants. Thus,
equation 4 shows that the Domar weighting scheme
(in equation 3) can be viewed as aggregating MFP
growth from these vertically integrated sectors.
To make use of equation 3, we need to estimate
MFP growth in each sector of our model. We do this
with the so-called dual method employed by Triplett
(1996) and Whelan (2000), among others. This
method uses data on the prices of output and inputs,
rather than their quantities, to calculate sectoral
MFP growth. We opt for the dual approach because
the required data are more readily available.
The basic intuition behind the dual approach can
be explained with an example involving semiconductors, the prices of which have trended down
sharply over time. To keep the example simple,
assume that input prices for the semiconductor
sector have been stable. Given the steep decline in
semiconductor prices relative to the prices for other
goods and services, MFP growth at semiconductor
producers must be rapid compared to that elsewhere. Were it not, semiconductor producers would
be driven out of business by the ever-lower prices
for their output in the face of stable input costs.
This example illustrates that relative growth rates
of sectoral MFP can be inferred from movements in
relative output prices.6
We rely on this link to estimate sectoral MFP
growth. Proposition 3 provides the details, which
involve some messy algebra. Roughly speaking, each
sectoral MFP growth rate can be written as
•
•
(5) MFPi = MFPo – πi + terms for the relative growth
in sectoral input costs,
•
•
where πi ≡ (pi – po ) denotes the difference in output
price inflation between sector i and the “other final-
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
output” sector, which serves as our benchmark sector. If input costs grew at the same rate in every
sector, the change in relative output prices would
fully characterize the differences in sectoral MFP
growth. However, because semiconductors loom
large in the cost structure of the computer industry,
we know that input costs for that industry are falling
relative to those for other sectors. The additional
terms in equation 5 take account of these differences
in sectoral input costs.
Note that equation 5 determines relative rates
of MFP growth, not the absolute rate in any sector. We pin down the absolute MFP growth rates
in two different ways, the first of which uses equation 3 to force the sectoral MFP growth rates to
reproduce our estimate of aggregate MFP growth.
This case represents the methodology we use to
compute historical growth contributions through
2001. In the second case, which we use for our
steady-state analysis, we condition on an assumed
pace of MFP growth in the “other final-output”
sector, which generates the remaining sectoral MFP
growth rates via equation 5 and aggregate MFP
growth via equation 3.
Analysis of the steady state. In addition to
explaining the source of the productivity pickup in
the 1990s, we wish to estimate a plausible range for
productivity growth in the future. To develop such
a range, we impose additional steady-state conditions on our model, closely following the two-sector
analysis in Martin (2001) and Whelan (2001).
Among the conditions imposed to derive steadystate growth, we assume that output in each sector
grows at a constant rate (which differs across sectors). In addition, we impose conditions that are
sufficient to force investment in each type of capital to grow at the same (constant) rate as the stock
of that capital. Taken together, these conditions can
be shown to imply that production in each final-out-
put sector grows at the same (constant) rate as the
capital stock that consists of investment goods produced by that sector. Two other important conditions
are that labor hours grow at the same (constant) rate
in each sector and that all income shares and sectoral
output shares remain constant.
Under these steady-state conditions, proposition 4
shows that the growth-accounting equation for aggregate labor productivity becomes
•
•
•
4
•
•
•
(6) Y − H = ∑ ( α iK α L )( MFPi + βiS MFPS ) + q + MFP,
i =1
•
where MFP is calculated, as above, from equation 3.
Note that equation 6 contains no explicit terms for
capital deepening, in contrast to its non-steadystate counterpart, equation 1. No such terms appear
because the steady-state pace of capital deepening
is determined endogenously within the model as a
function of the sectoral MFP growth rates. Hence,
the summation on the right side of equation 6 represents the growth contribution from this induced
capital deepening. With this interpretation, it becomes
clear that equations 1 and 6 share a common structure: Both indicate that the growth of labor productivity depends on capital deepening, improvements
in labor quality, and growth in MFP.
To further interpret equation 6, consider the
growth-accounting equation (outside the steady
state) for a simple one-sector model:
•
•
•
•
•
•
(7) Y − H = α K ( K − H ) + α L q + MFP.
Now impose the steady-state condition that output
and capital stock grow at the same rate and substitute
• •
K = Y into equation 7, noting that α K + α L = 1 under
constant returns to scale. The result is
•
•
•
•
•
•
•
(8) Y − H = q + MFP / α L = ( α K / α L ) MFP + q + MFP,
5. It is easy to see that the weights sum to more than one if semiconductor producers sell all of their output to the four finaloutput sectors, with none sold as exports. In this case, semiconductors are strictly an intermediate input, and production by
the four final-output sectors accounts for all nonfarm business output. Hence, the µ terms for these sectors sum to one before
adding in µS. With a little algebra, one can show that the µ terms also sum to more than one in the more general case that
allows for exports of semiconductors.
6. Under perfect competition, the growth rate of MFP in each sector can be inferred exactly from relative price movements.
However, if markets are not perfectly competitive, then the dual methodology would yield an inaccurate reading on MFP
growth to the extent that relative price changes resulted from swings in margins rather than technological developments. Of
course, if there is imperfect competition but margins are constant, then MFP growth rates still can be inferred exactly from
relative price movements because changes in margins would not be a source of changes in relative prices. For the semiconductor sector—where market concentration in microprocessors suggests that this potential problem with the dual methodology could be particularly acute—Aizcorbe (2002) found that a conventional Tornquist index of Intel microprocessor prices
fell 24 1/2 percent per quarter on average from 1993 to 1999; adjusted for movements in Intel’s margins over this period, the
index declined 21 percent per quarter. Thus, swings in margins appear to have had a relatively small average effect on chip
prices over this period.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
19
where the second equality uses the fact that (α K / α L)
= (1/α L) – 1 when α K + α L = 1. Comparing equations 6
and 8 shows that our steady-state growth-accounting
decomposition is the multisector counterpart to the
decomposition in a one-sector model.
Summary. We use equations 1–3 and 5 to decompose the observed growth in labor productivity
through 2001. Equation 1 provides the structure for
the decomposition, while equation 2 shows how we
calculate the income shares, and equations 3 and 5
(implemented with the dual method) show how we
relate aggregate MFP growth to its sectoral components. To estimate the growth of labor productivity
in the steady state, we replace equation 1 with equa-
Are the results from earlier research that
emphasized the role of information technology
still valid given the sharp contraction in the
technology sector?
tion 6, but otherwise we use the same machinery as
for the historical decompositions.
Data
his section provides a brief overview of the data
used for this paper; a detailed description appears
in Appendix 2. To estimate the decomposition of
labor productivity growth, we rely heavily on data
from the Bureau of Economic Analysis (BEA) and
the Bureau of Labor Statistics (BLS). Our starting
point is the data set assembled by the BLS for its
estimates of multifactor productivity. These annual
data cover the private nonfarm business sector in the
United States and provide measures of the growth
of real output, real capital input, labor hours, and
labor quality. At the time we were writing, the BLS
data set ran through 2000, and we extended all
necessary series through 2001.
The income shares in our growth-accounting calculations depend on estimates of the gross rate of
return earned by each asset (R + δj – Πj). To measure
the components of the gross return, we rely again
on data from the BEA and BLS. With just a few
exceptions, the depreciation rates (δj ) for the various types of equipment, software, and structures
are those published by the BEA. Because the BEA
provides only limited information on the depreciation rates for components of computers and periph-
T
20
eral equipment, we follow Whelan (2000) and set
these depreciation rates equal to a geometric approximation calculated from BEA capital stocks and
investment flows. For personal computers, we are
uncomfortable with the BEA’s procedure and instead
set the depreciation rate for PCs equal to the 30 percent annual rate for mainframe computers. (See
Appendix 2 for a discussion of this issue.) To estimate
the capital gain or loss term in the gross return (Πj ),
we use a three-year moving average of the percent
change in the price of each asset. The moving average smooths the often volatile yearly changes in
prices and probably conforms more closely to the
capital gain or loss that asset owners expect to bear
when they make investment decisions. Finally, to
calculate the net return (R), we mimic the BLS procedure, which computes the average realized net
return on the entire stock of equipment, software,
and structures. By using this average net return in
the income share for each asset, we impose the neoclassical assumption that all types of capital earn
the same net return in a given year.
To implement the sectoral model of MFP, we
need data on final sales of computer hardware,
software, and communication equipment as well
as data on the semiconductor sector. Our data on
final sales of computer hardware came from the
NIPAs, and we used unpublished BEA data to calculate final sales of software and communication
equipment. For the semiconductor sector, we used
data from the Semiconductor Industry Association
as well as data constructed by Federal Reserve Board
staff to support the Fed’s published data on U.S.
industrial production.
Decomposition of Labor Productivity Growth
s discussed above, our earlier research documented that information technology was a key
driver behind the resurgence in labor productivity
growth during the second half of the 1990s. Recent
developments—including the bursting of the
Nasdaq bubble and the dramatic retrenchment in
the high-tech sector—have raised questions about
the robustness of those results. By extending our
estimates through 2001, we can assess whether
recent data still support the basic story in our earlier
research. We describe our new numbers and then
compare them to our earlier results.
Results through 2001. Table 1 presents our
decomposition of labor productivity growth in the
nonfarm business sector through 2001. As shown in
the first line of the table, growth in labor productivity picked up from about 1.5 percent per year in the
first half of the 1990s to about 2.4 percent since
A
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 1
Contributions to Growth in Labor Productivity, Using Data as of March 2002
1. Growth of labor productivity1
1974–90
(1)
1991–95
(2)
1996–2001
(3)
Post-1995
change
(3) minus (2)
1.36
1.54
2.43
.89
Contributions from2
2.
3.
4.
5.
6.
7.
Capital deepening
Information technology capital
Computer hardware
Software
Communication equipment
Other capital
.77
.41
.23
.09
.09
.37
.52
.46
.19
.21
.05
.06
1.19
1.02
.54
.35
.13
.17
.67
.56
.35
.14
.08
.11
8.
Labor quality
.22
.45
.25
–.20
9.
10.
11.
12.
13.
14.
Multifactor productivity
Semiconductors
Computer hardware
Software
Communication equipment
Other sectors
.37
.08
.11
.04
.04
.11
.58
.13
.13
.09
.06
.17
.99
.42
.19
.11
.05
.23
.41
.29
.06
.02
–.01
.06
15.
Total IT contribution3
.68
.87
1.79
.92
1
In the nonfarm business sector, measured as the average annual log difference for the years shown multiplied by 100.
Percentage points per year.
3
Equals the sum of lines 3 and 10–13.
Note: Detail may not sum to totals because of rounding.
Source: Authors’ calculations based on BEA and BLS data
2
1995.7 Rapid capital deepening related to information technology capital—the greater use of information technology—accounted for about three-fifths
of this pickup (line 3). Other types of capital (line 7)
made a much smaller contribution to the acceleration in labor productivity, while the contribution
from labor quality actually fell across the two periods.
Multifactor productivity (line 9) is left to account for
a little less than half of the improvement in labor
productivity growth.
Next, we decompose this overall MFP contribution into its sectoral components in order to estimate
the growth contribution from the production of
information technology. Lines 10–14 of Table 1 display
this sectoral decomposition. The results show that
the MFP contribution from semiconductor producers
(line 10) jumped after 1995. Given our use of the dual
methodology, this pickup owes to the more rapid
decline in semiconductor prices in this period, which
the model interprets as a speedup in MFP growth.
7. Note that the figures for output per hour in Table 1 are based on the BLS published series for nonfarm business output. This
series is a product-side measure of output, which reflects spending on goods and services produced by nonfarm businesses.
Alternatively, output could be measured from the “income side” as the sum of payments to capital and labor employed in that
sector. Although the two measures of output differ only slightly on average over long periods of time, a sizable gap has
emerged in recent years. By our estimates, the acceleration in the income-side measure was about one-third percentage point
greater (at an average annual rate) after 1995. We employ the published product-side data to maintain consistency with other
studies; in addition, if an adjustment were made to output and labor productivity growth, it is not clear how that adjustment
should be allocated among the components of capital deepening and MFP growth. Nonetheless, the true pickup in productivity growth after 1995 could be somewhat larger than shown in our table.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
21
CHART 1
Contributions from the Use and Production of Information Technology to
Growth of Labor Productivity in Nonfarm Business
2.0
Production of IT
Use of IT
Pe r c e n t a g e p o in t s
1.6
1.2
0.8
0.4
0
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
Source: Authors’ calculations based on BEA, BLS, and other data
In contrast, the MFP contribution from the other
information technology sectors taken together (lines
11–13) rose only a little after 1995 compared with
the first half of the 1990s.
For computer hardware, the particularly rapid
decline in prices after 1995 might have led one to
believe that MFP growth in this sector had increased
dramatically. However, as indicated earlier, the computer sector—as we define it—excludes the production of the semiconductors embedded in computer
hardware. Thus, MFP in the computer sector represents only efficiency gains in the design and assembly
of computers, not in the production of the embedded
semiconductors. Accordingly, our results indicate
that the faster declines in computer prices after
1995 largely reflected the sharp drop in the cost of
semiconductor inputs rather than independent developments in computer manufacturing.
The MFP contributions from the software and
communication equipment sectors were fairly small
during both the 1991–95 and 1996–2001 periods.
According to the published numbers, the relative
prices of both software and communication equipment fell much less rapidly than did relative computer prices during these periods.8 In addition, for
communication equipment, our numbers indicate
that much of the relative price drop that did occur
22
reflected the plunging costs of semiconductor inputs,
which our sectoral decomposition attributes to MFP
growth in the semiconductor industry, not in communication equipment. Thus, the dual methodology
suggests that the MFP gains in both software and
communication equipment have been smaller than
those in the computer sector.
Putting together the information technology pieces
(line 15), greater use of information technology and
faster efficiency gains in the production of IT capital goods more than accounted for the 0.89 percentage point speedup in labor productivity growth
after 1995. This large contribution can also be seen
in Chart 1; the blue bars show the contribution from
the use of IT and the gray bars show the contribution from the production of IT on a year-by-year
basis. As the chart shows, these contributions surged
after 1994. Although they dropped back in 2001, the
contributions for that year remain well above those
observed before 1995. Based on these results, we
conclude that recent data confirm the main findings
in our earlier work. Namely, the resurgence in labor
productivity is still quite evident in the data, and
information technology appears to have played a
central role in this pickup.
Further comparison to our earlier work.
Table 2 compares our latest numbers to those in
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 2
Acceleration in Labor Productivity between 1991–95 and Post–1995 Period,
Effect of New Data and Revisions
1. Acceleration in labor productivity 1
Oliner and Sichel
(2000a)
through 1999
This paper
through 2000
This paper
through 2001
1.04
1.00
.89
.48
.45
.36
.04
.05
.03
.57
.54
.36
.13
.07
.02
.67
.56
.35
.14
.08
.11
–.13
–.20
–.20
.68
.27
.10
.62
.30
.06
.41
.29
.06
.31
.26
.06
Contributions from2
2. Capital deepening
3. Information technology capital
4.
Computer hardware
5.
Software
6.
Communication equipment
7. Other capital
8. Labor quality
9. Multifactor productivity
10.
Semiconductors
11.
Computer hardware
12.
Other sectors3
1
In the nonfarm business sector, measured as percentage points per year.
Percentage points per year.
3
Includes producers of communication equipment and software.
Note: Detail may not sum to totals due to rounding.
Source: Authors’ calculations based on BEA and BLS data
2
Oliner and Sichel (2000a).9 The first column of the
table shows contributions to the pickup in labor productivity growth from our earlier paper, the second
column presents estimates through 2000 using
recent data, and the third column repeats the contributions through 2001 shown in Table 1. In addition
to the inclusion of data for 2000, the numbers in the
second column differ from those in the first because
of data revisions since our earlier results were completed.10 Clearly, incorporating data for 2000 and
revisions for earlier years changed our results relatively little. The contribution to the productivity
pickup from software capital deepening increased,
but this increase was offset by a more negative
8. Jorgenson and Stiroh (2000) raised the possibility that software prices may have fallen faster than reported in the official
numbers. While this speculation may be correct, software has historically been a craft industry, in which highly skilled professionals write code line by line. In the 1960s and 1970s, several studies examined costs per line of code written. Phister
(1979, 502) estimated a 3.5 percent annual reduction in the labor required to produce one thousand lines of code. Zraket
(1992) argued that the nominal cost per line of code in the early 1990s was little changed from twenty years earlier, a scenario that would yield a real decline similar to Phister’s. Of course, the more recent adoption of software suites, licenses, and
enterprisewide software solutions may well have led to dramatic declines in the effective price of software. All told, we
believe that considerable uncertainty still attends the measurement of software prices.
Jorgenson and Stiroh (2000) also suggested that prices of communication equipment may have fallen faster than reported
in official statistics. Recent work by Doms (2002) provides support for that perspective.
9. Table 2 shows separate MFP contributions only for the semiconductor and computer sectors to maintain comparability with
our earlier work.
10. The most important data revisions that we factored in were the July 2000 and July 2001 NIPA revisions released by the BEA
(which are fully reflected in the latest BLS multifactor productivity data) and the official published estimates of capital stocks
for software. (In our earlier work, we had included our own estimate of software capital stocks.) In addition, we have made
some minor adjustments to our estimation procedures, but these changes had relatively small effects.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
23
contribution from labor quality and a somewhat
smaller contribution from MFP growth.
Extending the results through the recession year
2001 tempers the step-up in labor productivity growth
(line 1), as would be expected given the procyclical
behavior of productivity gains. At the same time,
line 2 indicates that the growth contribution from
capital deepening increased with the inclusion of data
for 2001. The large implied contribution in 2001 may
seem puzzling in light of the recession-related downturn in investment spending. However, recall that
capital deepening reflects the ratio of capital services
to hours worked. The decline in hours in 2001, other
things being equal, boosts the capital-hours ratio. Also,
has contributed significantly to the pickup in labor
productivity growth, quite apart from developments
in IT-producing industries. Second, the muchreduced MFP acceleration in other industries likely
reflected cyclical factors.12 Identifying the magnitude
of such cyclical influences is challenging, and we
believe that the trend cannot be inferred from the
average growth rate between 1995 and 2001. The
first year of that period, 1995, was midway through
the cycle, while the last year, 2001, was a recession
year.13 Thus, taking an average over the 1995–2001
period implicitly draws a line from a point at midcycle
to a point near the bottom of the cycle. Such a line
likely understates the trend over this period.
Labor Productivity Growth in the Steady State
Some observers might argue that the very small
acceleration of MFP outside the IT-producing
sectors indicates that the productivity benefits
of IT have been either narrowly focused or have
been largely reversed over the past year.
note that our growth accounting uses annual-average
data. Because investment spending weakened over
the course of 2001, annual averaging smooths this
decline relative to the change observed over the four
quarters of the year. Similarly, the Tornquist weighting
procedure delays the impact of such changes by using
an average of this year’s and last year’s capital income
shares as aggregation weights for the capital deepening contributions. Thus, some of the effects of the
recession on corporate profits (and hence on the capital income shares) will not show up in our numbers
until 2002. Indeed, a back-of-the-envelope calculation
suggests that the contribution of capital deepening
will drop back in 2002.11
The final effect of folding in data for 2001 is the
noticeably smaller contribution of MFP to the post1995 step-up in labor productivity growth (line 9).
Virtually the entire downward revision is in the large
residual sector consisting of all nonfarm business
except the computer and semiconductor industries
(line 12).
Some observers might argue that the very small
acceleration of MFP outside these IT-producing
sectors indicates that the productivity benefits of IT
have been either narrowly focused or have been
largely reversed over the past year. However, we are
not inclined to accept either interpretation for two
reasons. First, the use of IT throughout the economy
24
ow much of the resurgence in labor productivity growth in the second half of the 1990s is
sustainable? To address this question, we use the
steady-state machinery described earlier to generate
a range of likely outcomes for labor productivity
growth in the future. We do not regard these steadystate results as forecasts of productivity growth for
any particular time period. Rather, this exercise yields
structured guesses of the sustainable growth in labor
productivity consistent with alternative scenarios for
the evolution of key features of the economy.
To construct this range of likely outcomes, we set
lower and upper bounds on steady-state parameters
and then solve for the implied rates of labor productivity growth. We believe that these scenarios encompass the most plausible paths going forward, but there
is substantial uncertainty about future productivity
developments. Hence, as we will discuss, the sustainable pace of labor productivity growth could fall
outside the range that we consider most likely. The
rest of this section describes the lower- and upperbound parameter values that we chose, presents our
steady-state results, and compares our results to
those obtained by other researchers.
Parameter values. Table 3 displays the many
parameters that feed into our model of steady-state
growth. To provide some historical context, the first
three columns of the table show the average value
of each parameter over the 1974–90, 1991–95, and
1996–2001 periods. The next two columns present
our assumed lower-bound and upper-bound values
for each parameter in the steady state, and the final
column briefly indicates the rationale for these
steady-state values.14
Lines 1–15 of the table list the parameters needed
to compute aggregate and sectoral MFP growth in
the steady state. These parameters include each
sector’s current-dollar share of nonfarm business
H
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
output (the µs), outlays for semiconductors as a
share of total input costs in each final-output sector
(the βs), the rate of output price inflation in each
sector relative to that in the other-final-output sector
(the πs), and the growth of MFP in the other-finaloutput sector (MFPO ).
Although the steady-state bounds for some of
these parameters require no discussion beyond the
brief rationale in the table, others need further
explanation.15 Starting with the output shares, we
calibrated the steady-state bounds from the plots in
Chart 2. The short lines in each panel represent the
bounds, which can be compared to the history for
each series. The current-dollar output shares for
producers of computer hardware and communication
equipment have each fluctuated in a fairly narrow
range since the mid-1980s. Our steady-state bounds
largely bracket those ranges. For producers of software and semiconductors, the current-dollar output
shares have trended sharply higher over time, and
our steady-state bounds allow for some additional
increase from the average level in recent years.16
Among the semiconductor cost shares (the βs),
we set the share for computers equal to 0.30, the
middle of the range employed by Triplett (1996).
For software, we set the share to zero. The share for
communication equipment is shown in Chart 3. This
share has risen quite a bit since the early 1990s,
reflecting the increasing amount of computer-like
technology in communication equipment. We set the
steady-state bounds on the assumption that this
trend will persist.
This increase in the semiconductor content of
communication equipment implies that the relative
price for such equipment is likely to fall more rapidly
in the future than it has over history. We built that
expectation into the steady-state bounds for π M,
shown on line 14. These values were chosen to
ensure that the implied MFP growth rate for the
sector, computed by the dual method, remained
close to the average pace over 1996–2001.
This issue does not arise for other sectors, where
the semiconductor cost shares are assumed to change
little, if at all, going forward. For these sectors
(lines 11–13), we set the bounds on relative price
changes (the πs) by reference to historical patterns.
The lower bound for each sector equals the average
rate of relative price change over 1974–2001, while
the upper bound lies midway between that average
How much of the resurgence in labor productivity growth in the second half of the
1990s is sustainable?
and the most rapid rate of relative price decline for
the three subperiods since 1974. Thus, we do not
assume that the extremely rapid declines in computer and semiconductor prices over 1996–2001
will persist in the steady state, even in our optimistic scenario.
Lines 16–28 of the table list the components of
the capital income shares. For the nominal rate of
return on capital and the asset-specific depreciation
11. To show this, we calculated capital deepening for 2002 on the assumption that the growth rate of real investment in hightech equipment snaps back to its robust average pace during 1996–2000 and that hours fall nearly 1 percent in 2002 on an
annual average basis as projected in Macroeconomic Advisers’ January 2002 Economic Outlook. Even under this optimistic
assumption for investment and sluggish forecast for hours, the contribution of capital deepening to labor productivity growth
in 2002 would be below its 2001 value (but still significantly above its pre-1995 value).
12. Even though MFP is often associated with technological change, short-run movements in MFP can be heavily influenced by
cyclical factors that have little relation to technological change. For further discussion of this point, see Basu, Fernald, and
Shapiro (2001).
13. Inferring the trend from the average growth rate between 1995 and 2000 also may be problematic because the average covers a period from midcycle to peak. Moving the initial year back to the prior peak in 1990 is not appealing because we are
interested in what happened to productivity beginning in the mid-1990s.
14. Note that the upper-bound value for each parameter yields a higher rate of productivity growth than the lower-bound value.
For some parameters, such as relative prices, the upper-bound value is numerically smaller than the lower-bound value.
15. In performing similar exercises, DeLong (2002), Kiley (2001), and Martin (2001) start with demand elasticities for high-tech
products to generate output and income shares. In contrast, we set assumptions for output shares and other key parameters directly. Because relatively little is known about high-tech demand elasticities, we prefer the transparency of directly
setting output shares and other parameters based on their historical patterns.
16. The output share for the semiconductor sector plunged in 2001 to the lowest level since 1994 owing to the deep cutbacks
in spending on high-tech equipment during the recession. In setting the steady-state bounds, we assumed that the cyclical
drop would be reversed as the economy recovers from recession.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
25
TABLE 3
Parameter Values for Steady-State Calculations
Historical averages
Steady-state values
1974– 1991– 1996–
lower
upper
1990
1995
2001
bound bound
Parameter
Method for setting
steady-state values
Output shares1 (µ)
1. Computer hardware
2. Software
3. Communication equipment
4. Other final-output sectors
5. Net exports of semiconductors
6. Total semiconductor output
1.06
.84
1.80
96.33
–.04
.30
1.19
1.79
1.68
95.45
–.11
.58
1.32
2.70
1.83
94.14
.00
.91
1.10
3.10
1.60
94.20
.00
1.00
1.40
3.60
2.00
93.00
.00
1.20
See Chart 2.
See Chart 2.
See Chart 2.
Implied by lines 1–3 and 5.
1996–2001 average.
See Chart 2.
Semiconductor cost shares1 (β)
7. Computer hardware
8. Software
9. Communication equipment
10. Other final-output sectors
30.00
.00
1.17
.00
30.00
.00
4.59
.27
30.00
.00
8.88
.37
30.00
.00
13.00
.46
30.00
.00
16.00
.46
Assumed constant value.
Assumed constant value.
See Chart 3.
Implied by lines 1–4, 6–9, and 36.
Relative inflation rates2 (π)
11. Semiconductors
12. Computer hardware
13. Software
14. Communication equipment
–28.90
–19.29
–4.13
–2.44
–21.75
–17.79
–4.83
–4.06
–44.71
–27.15
–3.90
–5.80
–31.01
–20.71
–4.21
–6.00
–37.86
–23.93
–4.52
–7.75
.11
.17
.23
.11
.23
Used historical range.
7.88
4.29
4.55
4.55
4.55
1996–2001 average.
29.74
34.87
13.00
5.87
30.11
37.04
13.00
6.08
30.30
38.46
13.00
6.10
30.30
38.46
13.00
6.10
30.30
38.46
13.00
6.10
1996–2001 average.
1996–2001 average.
1996–2001 average.
1996–2001 average.
Expected capital gains/losses4 (Π)
21. Computer hardware
–12.70
22. Software
3.27
23. Communication equipment
3.65
24. Other business fixed capital
6.31
–11.79
–.56
–.07
2.52
–23.21
–.31
–3.01
2.55
–17.50
–.44
–4.00
2.54
–20.36
–.50
–5.75
2.53
15. Growth of MFPO 3
3
16. Nominal return on capital (R)
Lower bound is 1974–2001 average;
upper bound is midway between that
value and fastest historical decline.
Calibrated to keep the sector’s MFP
growth rate near the 1996–2001 pace.
}{
3
Depreciation rates (δ)
17. Computer hardware
18. Software
19. Communication equipment
20. Other business fixed capital
Capital-output ratios (TpK K /pY )
25. Computer hardware
26. Software
27. Communication equipment
28. Other business fixed capital
Income shares1 (α)
29. Computer hardware
30. Software
31. Communication equipment
32. Other business fixed capital
33. Other capital7
34. Labor
Other parameters
35. Growth of labor quality 3 (q)
36. Ratio of domestic semiconductor
output to domestic use (1 + θ)
1
2
3
4
5
6
7
26
.0192 .0293 .0294
.0191 .0440 .0618
.0876 .1087 .0951
2.4227 2.2648 2.1008
.0300 .0360
.0800 .0900
.0875 .1025
1.9000 2.0500
.92
.75
1.48
18.00
9.81
69.04
1.34
1.85
1.88
17.78
8.90
68.25
1.71
2.67
1.96
17.04
8.93
67.69
1.57
3.48
1.89
15.42
8.93
68.72
1.99
3.92
2.39
16.65
8.93
66.13
.32
.65
.38
.30
.30
.89
.86
1.03
1.03
1.03
See footnote 5.
See footnote 5.
See footnote 6.
See footnote 5.
See Chart 4.
See Chart 4.
See Chart 4.
See Chart 4.
Implied. See Chart 5.
Implied. See Chart 5.
Implied. See Chart 5.
Implied. See Chart 5.
1996–2001 average.
Implied by lines 29–33.
Assumed slower growth.
1996–2001 average.
Current-dollar shares, in percent.
Output price inflation in each sector minus that in the “other final-output” sector, in percentage points.
In percent.
Three-year moving average of price inflation for each asset, in percent.
Lower bound is average over 1991–2001; upper bound is midway between that value and the smaller of the 1991–95 and
1996–2001 values.
The lower and upper bounds equal the corresponding values for the relative inflation rate of communication equipment (line 14), plus 2
percentage points—the assumed rate of inflation in the “other final-output” sector.
Includes land, inventories, and tenant-occupied housing.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 2
Current-Dollar Output Shares
Software
4
1.5
3
Pe r c e n t
Pe r c e n t
Computer hardware
2.0
1.0
2
1
0.5
0
0
1974 1978
1982 1986 1990
1994
1998 2002
1974 1978
Communication equipment
1982 1986 1990
1994
1998 2002
1994
1998
Semiconductors
1.5
2.5
2.0
P e rc e n t
P e rc e n t
1.0
1.5
1.0
0.5
.5
0
0
1974 1978
1982 1986 1990
1994
1998 2002
1974 1978
1982 1986 1990
2002
Source: Authors’ calculations based on BEA, BLS, and other data
rates, we simply project forward the average values
for 1996–2001. These parameters varied only slightly
between the first and second halves of the 1990s;
moreover, the higher nominal return on capital over
1974–90, which was driven in part by the elevated
pace of inflation over that period, is not appropriate
for the current low-inflation environment. For the
next element of the income share, the expected
capital gain or loss on the asset, we set the steadystate bounds in essentially the same way as we did
for the relative inflation rates. For all types of capital
except communication equipment, we chose these
bounds by reference to the historical data, though
we looked back only to 1991 to avoid building in the
higher rates of inflation that prevailed over 1974–90.
The bounds for communication equipment were set
to the analogous bounds for the relative price decline
on line 14, plus 2 percentage points. This add-on
for the assumed rate of inflation in the other-final-
output sector converts the relative price change
into an absolute change.
The final piece of the income share is the (taxadjusted) capital-output ratio, expressed in current
dollars (TpK K/pY ). Chart 4 displays this ratio back
to 1974 for the four types of capital. For computer
hardware and communication equipment, where the
capital-output ratio has not displayed a clear trend
of late, we set the bounds to keep the ratio in its neighborhood of recent years. In contrast, for software and
other fixed capital, we chose the bounds to allow for
a continuation of longer-term trends. Chart 5 shows
the implied bounds for the capital income shares
along with the historical series for these shares. The
one series that bears comment is the share for other
equipment and nonresidential structures, which
plummeted in 2001 as the recession-induced decline
in corporate profits depressed the nominal return
to capital (R).17 The steady-state bounds for this
17. The drop in R had a much greater effect on the income share for this broad capital aggregate than on the income shares for
computers, software, or communication equipment. For these high-tech assets, the rapid trend rate of depreciation is the
dominant piece of the gross return, overwhelming even sizable movements in R.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
27
CHART 3
Semiconductor Cost Share in Communication Equipment
18
16
14
Pe r c e n t
12
10
8
6
4
2
0
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
Source: Authors’ calculations based on BEA, BLS, and other data
CHART 4
Current-Dollar Capital-Output Ratios
Software
.100
.03
.075
Ra tio
Ra tio
Computer hardware
.04
.02
.01
.050
.025
0
0
1974 1978
1982 1986 1990
1994
1974 1978
1998 2002
Communication equipment
1982 1986 1990
1994
1998 2002
Other equipment and nonresidential structures
3.0
.125
.100
Ra t i o
Ra t i o
2.5
.075
2.0
.050
.025
1.5
1974 1978
1982 1986 1990
1994
1998 2002
Note: Each subcomponent within a category is tax-weighted.
Source: Authors’ calculations based on BEA, BLS, and other data
28
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
1974 1978
1982 1986 1990
1994
1998 2002
CHART 5
Current-Dollar Income Shares
Computer hardware
Software
2.5
4
3
Pe r c e n t
Pe r c e n t
2.0
1.5
1.0
2
1
.5
0
0
1974 1978
1982 1986 1990
1994
1974 1978
1998 2002
1994
1998 2002
Other equipment and nonresidential structures
2.5
20
2.0
18
P e rc e n t
P e rc e n t
Communication equipment
1982 1986 1990
1.5
1.0
16
14
12
0
1974 1978
1982 1986 1990
1994
1998 2002
1974 1978
1982 1986 1990
1994
1998 2002
Source: Authors’ calculations based on BEA, BLS, and other data
income share imply at least a partial reversal of this
cyclical decline.
The final parameter of note is the growth of labor
quality (line 35). We assume that labor quality will
increase 0.3 percent per year in the steady state,
noticeably slower than its average annual rise over
recent decades. Jorgenson, Ho, and Stiroh (2002)
suggest a step-down in labor quality growth of similar magnitude while Aaronson and Sullivan (2001)
project a slightly larger drop-off going forward.
Results. Table 4 contains the “structured guesses”
of labor productivity growth in the steady state
using lower-bound and upper-bound parameter
values.18 As shown on line 1, the lower-bound parameter values generate steady-state growth in labor
productivity of about 2 percent while the upperbound values imply growth of slightly more than
23/4 percent.19 This range, which sits well above
the sluggish pace realized from the early 1970s to
the mid-1990s, suggests a relatively optimistic outlook for labor productivity.
To provide intuition for the steady-state range,
note that the lower-bound figure of about 2 percent is
roughly 1/2 percentage point below the pace of labor
18. As noted earlier, our model does not explicitly account for adjustment costs. Nevertheless, we recognize that such costs
could have important implications for labor productivity growth, as emphasized by Kiley (2001) and Basu, Fernald, and
Shapiro (2001). Implicitly, our steady-state estimates of labor productivity growth embed the average historical value of
adjustment costs. Specifically, if adjustment costs have held down labor productivity growth on average historically, our
growth-accounting framework will sweep these effects into the residual—which is MFP growth in “other final output.”
Because our steady-state estimates depend on MFP growth in that residual category, the average historical magnitude of
adjustment costs is implicitly built into these estimates.
19. It is reassuring that the results generated by the steady-state model over historical periods are well aligned with measured
productivity growth. In particular, if we use the steady-state model with the historical average parameter values in Table 3,
it returns an average growth rate for labor productivity of 1.57 percent over 1974–2001, very close to the actual growth rate
of 1.62 percent over this period.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
29
TABLE 4
Steady-State Results
Using lowerbound parameters
Using upperbound parameters
1.98
2.84
.97
.88
.09
.30
.72
1.47
1.31
.16
.30
1.07
1.50
2.17
1. Growth of labor productivity1
Contributions from2
2. Induced capital deepening
3. Information technology capital
4. Other capital
5. Labor quality
6. Multifactor productivity
7. Total IT contribution3
1
In the nonfarm business sector, measured in percent.
Percentage points per year.
3
Equals line 3 plus the contributions included in line 6 from producers of computer hardware, software, communication equipment,
and semiconductors.
Note: Detail may not sum to totals because of rounding.
2
productivity growth during 1996–2001. This slowdown occurs because we assume that the rates of
decline in semiconductor and computer prices revert
to their long-run historical averages from the very
rapid pace realized in the second half of the 1990s.
These assumptions produce a marked slowdown in
MFP growth in the semiconductor sector and, to a
lesser extent, in the computer sector. Nonetheless,
labor productivity growth for nonfarm business as
a whole remains above the 1974–95 average because
the IT sectors, taken together, constitute a larger part
of the economy than they did in this earlier period.
The upper-bound figure of about 2.8 percent in the
steady state is almost 1/2 percentage point above the
1996–2001 pace. The model generates this step-up
even though the price declines for semiconductors
and computers in the steady state (and hence the
rates of MFP growth) are assumed to be less rapid
than those in the second half of the 1990s. The
countervailing factor is that the semiconductor
sector and other IT sectors grow as a share of the
economy compared to that period. The greater
importance of these sectors with relatively fast MFP
growth more than makes up for the slower price
declines for semiconductors and computers.
The remaining lines of Table 4 show the major
factors that contribute to steady-state growth in labor
productivity. These numbers highlight the important role of IT in future labor productivity growth.
In particular, a comparison of lines 2 and 3 indicates that the induced capital deepening in the
steady state is very heavily skewed toward IT cap30
ital in both the lower- and upper-bound scenarios,
just as it was in the latter half of the 1990s. More
broadly, as shown in line 7, the combined contribution of both the induced use and the production of
IT accounts for about three-fourths of overall labor
productivity growth in both the lower- and upperbound scenarios.
As indicated above, our intent is to provide a likely
range for steady-state growth in labor productivity,
not to bound all possible outcomes. For example, the
steady-state model can generate labor productivity
growth above 3 percent per year if we assume that
semiconductor and computer prices continue to fall
at the 1996–2001 pace and allow the semiconductor
output share to rise by the amount seen between the
first and second halves of the 1990s. Conversely, we
can generate numbers for labor productivity growth
between 11/2 and 13/4 percent per year if we assume
that price declines for computers and semiconductors revert to their historical average and that the
computer and semiconductor output shares go back
down to levels seen in the first half of the 1990s. So,
while we are comfortable with a likely range for
steady-state labor productivity growth from 2 percent to 23/4 percent, we are well aware of the uncertainty that attends the exercise we have undertaken.
Comparison to other research. Table 5 compares the steady-state results in this paper to those
obtained by other researchers. There are two points
to take away from this table. First, the range of estimates is very wide, extending from about 11/4 percent
up to 31/4 percent. This range highlights the uncer-
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 5
Alternative Estimates of Steady-State Growth in Labor Productivity, Percent per Year
Point estimate
1. This paper
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
1
2
3
4
5
6
7
Jorgenson, Ho, and Stiroh (2002)1
Congressional Budget Office (2002)2
Economic Report of the President (2002)3
Baily (2002)
Gordon4
Kiley (2001)
Martin5
McKinsey (2001)6
Roberts7
DeLong (2002)
Range
2.0 to 2.8
2.25
2.2
2.1
2.2
≈2.0
2.6
“like the fast-growing late 1990s”
1.3 to 3.0
2.2
2.0
2.6
1.5
1.6
to
to
to
to
to
2.7
2.2
3.2
2.4
2.5
Jorgenson, Ho, and Stiroh measure productivity growth for a broader definition of the economy than do the other papers. To make
their numbers comparable to those in the other studies, add 0.15 percentage point to the point estimate and range shown for
Jorgenson, Ho, and Stiroh in the table.
Table 2–5.
Table 1–2, p. 55.
Based on personal correspondence with Robert Gordon, March 24, 2002.
In personal correspondence of August 2002, Bill Martin reported these numbers for the period 2002–11; these figures are lower than
those in Martin (2001).
Chapter 3, exhibit 13.
Unpublished update to Roberts (2001).
tainty surrounding the future path of productivity
growth. Second, despite the wide band of uncertainty,
most of the point estimates (or range midpoints) fall
within our range of 2 to 23/4 percent per year. Thus,
there is considerable agreement among researchers
that productivity growth likely will remain fairly
strong going forward.
Conclusion
ecent debates about the pickup of productivity
growth in the United States have revolved around
two questions. First, are the results from earlier
research that emphasized the role of information
technology still valid given the sharp contraction in
the technology sector? Second, how much of the
improvement in labor productivity growth since the
mid-1990s could plausibly be sustained? This paper
addressed both questions.
As for the robustness of earlier results, we used
data through 2001 to reassess the role of information technology in the productivity revival since the
mid-1990s. These new growth-accounting results
indicate that the story told in Oliner and Sichel
(2000a) still stands. Namely, output per hour accelerated substantially after 1995, driven in large part
by greater use of IT capital goods by businesses
throughout the economy and by more rapid efficiency gains in the production of IT goods.
R
To address the question of sustainability, we analyzed the steady-state properties of a multisector
growth model. This framework translates alternative
views about the evolution of the technology sector
and other features of the economy into estimates of
labor productivity growth in the steady state. When
we imposed relatively conservative values for key
parameters, this framework generated steady-state
growth in labor productivity of about 2 percent per
year. This estimate rose to roughly 23/4 percent when
we imposed somewhat more optimistic assumptions.
We refer to these estimates as structured guesses
and think of them as identifying a likely range of productivity outcomes over roughly the next decade. Of
course, any such exercise entails substantial uncertainty, and we also discussed scenarios that would
generate a wider range of outcomes.
Our analysis highlights that future increases in
output per hour will depend importantly on the
pace of technological advance in the semiconductor industry and on the extent to which products
embodying these advances diffuse through the
economy. This observation is consistent with the
emphasis in Jorgenson (2001) on semiconductor
technology. Gaining a deeper understanding of
technological developments in this sector should
be a high priority for those attempting to shed light
on trends in productivity.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
31
APPENDIX 1
Model of Sectoral Productivity
4
his appendix presents our model of sectoral
productivity and derives key results for our
analysis of growth in aggregate labor productivity.
The model divides nonfarm business into five
sectors. Four of the sectors produce final output (computer hardware, software, communication
equipment, and all other final output). The fifth
sector produces semiconductors, which are either
consumed as an intermediate input by the finaloutput sectors or exported to foreign firms. To focus
on essential linkages, the model abstracts from all
intermediate inputs besides semiconductors.
T
(A2) Ys = Sd + Sx = ∑ Si+ Sx – Sm
i=1
= Fs (Ls , K1, s , K2, s , K3, s , K4, s , zs).
The next step is to define the relationship
between the sectoral variables and their aggregate
counterparts. Following the guidance of index number theory, we express the growth in aggregate final
output as a superlative index of growth in sectoral
•
final output. Let Z ≡ (∂Z/∂t)/Z denote the growth
in any variable Z. Then the growth of aggregate
nonfarm business output (Y ) in our model is
•
4
•
•
•
(A3) Y = ∑ µ i Yi + µ S ,x Sx − µ S ,m Sm ,
i =1
The Model
Let Yi (i = 1,…, 4) denote the production of the
final-output sectors. Each sector produces investment goods (Ii) and consumption goods (Ci) for
domestic use, where Ii and Ci are identical goods
sold to different agents (firms buy Ii, while households buy Ci). Let Ii,j and Ii,s denote, respectively,
the purchases of Ii by final-output sector j ( j =
1,…, 4) and by semiconductor producers, with Ii =
Σ j Ii,j + Ii,s. Each sector also produces goods for
export (Xi). To produce this output, sector i
employs labor (Li) and various types of capital
(Kj,i, j = 1,…, 4), and it purchases semiconductors (Si) as an intermediate input.1 With this
notation, the production function for each finaloutput sector can be written as
4
+ ∑ Ii, j + Ii, s + Xi
(A1) Yi = Ci =
i=1
= Fi (Li, K1,i, K2,i, K3, i, K4, i, Si, zi)
for i = 1,…, 4,
where zi measures the level of multifactor productivity. Although we do not explicitly model
foreign production, the capital stocks Kj,i should
be regarded as including imported capital goods
of type j. To ease the notational burden, we have
suppressed time subscripts in equation A1 and
will do so throughout this appendix.
The output of the semiconductor sector (Ys) is
either sold as intermediate input to the domestic
final-output sectors (Sd ) or is exported (Sx ). The
semiconductors purchased by each domestic
final-demand sector (Si ) include imported semiconductors (Sm ), implying that the production sold
for domestic use can be written as Sd = Σi Si – Sm.
We assume that semiconductor producers employ
labor and the same set of capital inputs as the
final-output sectors. With these assumptions,
32
where µi ≡ piYi /pY (for i = 1,…, 4), µS,x ≡ ps Sx /pY,
4
piYi + psSx – ps Sm.2
µS,m ≡ ps Sm /pY, and pY ≡ Σi=1
The prices of final output and semiconductors are
denoted by pi and ps, respectively, and pY represents aggregate current-dollar output. Equation A3
expresses the growth in aggregate output as a
share-weighted average of sectoral output growth,
where the shares are in current dollars. Note that
the semiconductors sold to domestic final-output
sectors are an intermediate input for those sectors
and thus do not appear in equation A3; only net
exports of semiconductors enter the equation,
consistent with the treatment of semiconductors
in the NIPAs.
The definition of labor and capital aggregates in
our model is very simple. We assume that labor input
used in a given sector is identical to that used in any
other sector. We also impose this assumption on
each type of capital. Given these assumptions, we can
directly aggregate the sectoral inputs without the
need for superlative aggregation formulas. That is,
4
(A4) L = ∑ Li + Ls;
i=1
4
(A5) Kj = ∑ Kj,i + Kj,s for j = 1,…, 4.
i=1
Moreover, with this setup, there is a common wage
rate (w) for labor in every sector and, likewise, a
common rental rate (rj) for all capital of type j.
Labor input in each sector is the product of
hours worked (Hi) and labor quality (qi), where
quality reflects the characteristics of the workers
employed in that sector. We allow labor quality to
change over time, but given our assumption of
identical labor input across sectors, qi must equal
a common value q in every sector at a given point
in time. Using equation A4, this implies that
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
4
(A6) L = ∑ qHi + qHs = qH,
A11, the growth-accounting equation for aggregate
labor productivity is
i=1
where H represents aggregate hours worked.
To derive the growth-accounting equation for
each sector, we impose the standard neoclassical assumptions of perfect competition and constant returns to scale. We also assume that there
are no adjustment costs. Under these assumptions, profit-maximizing firms will set the marginal revenue product of each input equal to its
one-period cost:
•
•
•
4
•
j=1
•
•
•
•
4
•
•
4
4
•
4
= ∑ α L ( Li L) Li + ∑ ∑ α Kj ( K j,i K j ) K j,i
i =1
j =1 i =1
•
•
4
+ ∑ µ S ( Si YS )Si + ∑ µ i MFPi
(A9) ps = pi(∂Fi /∂Si) for i = 1,…, 4.
i =1
i =1
•
If we totally differentiate equations A1 and A2
and then impose conditions A7 through A9, we
obtain the standard growth-accounting equations:
•
•
+ µ S ,x Sx − µ S ,m Sm
4
•
•
j =1
•
(A8) rj = ps(∂Fs /∂Kj,s ) = pi(∂Fi /∂Kj,i )
for i, j = 1,…, 4;
4
•
4
(A12) Y = ∑ µ i[βiL Li + ∑ β Kj,i K j,i + βiS Si + MFPi ]
i =1
•
•
4
where MFP = Σ i=1
µiMFPi + µsMFPs, αL = wL/pY,
α Kj = rj Kj /pY, µi = piYi /pY, and µS = psYs/pY.
Proof. To begin, substitute the expression for
•
Yi from equation A10 into equation A3:
(A7) w = ps(∂Fs /∂Ls ) = pi(∂Fi /∂Li) for i = 1,…, 4;
•
•
•
Y – H =+ ∑ α Kj (Kj – H) + α Lq + MFP,
•
(A10) Yi = βiL Li + ∑ β Kj,i K j,i + βiS Si + MFPi
•
+ µ S ,x Sx − µ S ,m Sm ,
where the second equality follows (after some
algebra) from the definitions of the αs, βs, and µs.
Next, totally differentiate equation A2 to obtain
j =1
for i = 1,…, 4;
•
•
4
(A11) YS = γ LS + ∑
L
j =1
•
γ Kj
•
•
•
4
•
•
(A13) YS = ∑ ( Si YS )Si +( Sx YS )Sx − ( Sm YS )Sm.
•
K j,S + MFPS ,
i =1
•
where MFPi ≡ (∂Fi /∂zi)/Fi, MFPS ≡ (∂Fs /∂zs)/Fs,
and the βs and γ s represent the following
income shares: βLi ≡ wLi /piYi, the labor share in
sector i; βKj,i ≡ rj Kj,i /piYi, the share for capital of
type j in sector i; βSi ≡ psSi /piYi, the semiconductor share in sector i; γ L ≡ wLs/psYs , the labor
share in the semiconductor sector; and γ Kj ≡
rj Kj,s /psYs, the share for capital of type j in the
semiconductor sector. Given the assumption of
constant returns, the income shares in each sector sum to one.
Multiplying equation A13 by µS and using the definitions of µS, µS,x, and µS,m,
•
4
•
•
•
(A14) ∑ µ S ( Si YS )Si = µ S YS − µ S ,x Sx + µ S ,m Sm.
i =1
Now, substitute equation A14 into equation A12,
which yields
•
•
4
4
4
j=1
i=1
•
(A15) Y = α L ∑( Li L)Li + ∑ α Kj ∑( K j,i K j )K j,i
i=1
4
•
•
+ ∑ µ i MFPi + µ S YS .
i=1
Aggregate Labor Productivity
Proposition 1 derives the expression for growth
in aggregate labor productivity in our model.
Proposition 1. Assume that all markets are
perfectly competitive, that production exhibits
constant returns to scale in every sector, and that
input use is not subject to adjustment costs. Then,
in the model described by equations A1 through
Next, totally differentiate equations A4 and A5:
•
4
•
i=1
4
•
•
(A16) L = ∑( Li L)Li + ( LS L)LS ;
•
•
(A17) K j = ∑ ( K j,i K j ) K j,i + ( K j,S K j ) K j,S ,
i=1
and substitute these equations into A15:
1. When either I or K has a double subscript, the first subscript indicates the sector that produced the investment good
while the second subscript indicates the sector that uses it as an input to production.
2. Equation A3 is just one of several possible superlative indexes of output growth. It differs slightly from the Fisher chain
index used in the NIPAs.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
33
A P P E N D I X 1 (continued)
•
•
•
•
4
(A18) Y = α L [ L − ( LS L) LS ] + ∑ α Kj [K j
(A21)
j=1
•
•
4
4
•
µ S ≡ pSYS / pY = pS [ ∑ Si + Sx − Sm ] / pY
− ( K j,S K j ) K j,S ] + ∑ µ i MFPi + µ S YS
i =1
i=1
•
•
4
4
= ∑ ( piYi / pY )( pS Si / piYi ) + pS ( Sx − Sm )/ pY
•
4
= α L L + ∑ α Kj K j + ∑ µ i MFPi
j=1
i =1
4
i=1
•
= ∑ µ iβ iS + µ S,x − µ S,m .
•
+ µ S [YS − ( α L µ S )( LS L) LS
i =1
•
4
Note that µS,x – µS,m can be written as µS( Sx – Sm)/
YS so that equation A21 becomes
− ∑ ( α Kj µ S )( K j,S K j ) K j,S ]
j=1
•
•
4
•
4
= α L L + ∑ α Kj K j + ∑ µ i MFPi
j=1
•
•
+ µ S [YS − γ LS −
•
= α L+
L
4
(A22)
i=1
L
4
K
∑αj
j=1
•
4
K
∑γ j
j=1
•
K j, S ]
•
4
•
i=1
•
•
4
4
i =1
i =1
i =1
•
K j + ∑ µ i MFPi + µ S MFPS ,
•
i =1
4
= ∑ µ iβ iS / [ ∑ Si / YS ] = ∑ µ iβ Si (1 + θ ),
where the third equality follows from the definitions
of the αs, γs, and µs, and the fourth equality follows
from equation A11. To complete the proof, recall
•
•
•
that L = H + q from equation A6 and that the αs sum
to one under constant returns to scale. Hence,
(A19)
µ S = ∑ µ iβ iS / [1 − ( Sx − Sm )/ YS ]
•
4
where the second equality follows from equation A2 and the third from the definition 1 + θ ≡
4
Si. Finally, substitute equation A22 into
YS / Σ i=1
the expression from Proposition 1 for growth in
aggregate MFP:
(A23)
•
•
4
•
MFP = ∑ µ i MFPi + µ S MFPS
i =1
•
α L L = α L ( H + q ) = H − ∑ α Kj H + α L q.
•
4
•
= ∑ µ i[ MFPi + βiS (1 + θ ) MFPS ].
j =1
i =1
Substitute equation A19 into A18, which produces
(A20)
•
•
Y −H
•
•
= ∑ α Kj ( K j − H )
j =1
4
•
4
•
+ α q + ∑ µ i MFPi
L
i =1
•
+ µ S MFPS .
More on Aggregate MFP
Proposition 1 showed that aggregate MFP
growth in our model equals a share-weighted sum
of MFP growth in each sector. This result can be
rewritten to highlight the input-output connections between semiconductor producers and the
final-output sectors. In effect, we can integrate
semiconductor producers with the final-output
sectors that they supply.
Proposition 2. Under the assumptions of
Proposition 1,
•
4
•
•
MFP = ∑ µ i MFPi + µ S MFPS
i =1
4
•
•
= ∑ µ i[ MFPi + βiS (1 + θ ) MFPS ],
i =1
4
where 1+θ =YS /Σi=1
Si, the ratio of domestic semiconductor output to domestic use of semiconductors.
Proof. Using equation A2 and recalling the
definitions of µi, βSi, µS,x, and µS,m,
34
Measuring Sectoral MFP
To make use of Propositions 1 and 2, we need
to estimate MFP growth in each sector. This estimate can be derived either from the sectoral production functions, as in equations A10 and A11,
or from the sectoral cost functions—the “dual”
approach. We opt for the dual approach because
the required data are more readily available. The
dual counterparts to equations A10 and A11 are:
•
(A24)
•
4
•
•
•
pi = β Li w + ∑ β Kj,i rj + β Si pS − MFPi
j =1
for i = 1,…, 4;
•
(A25)
•
4
•
•
pS = γ L w + ∑ γ Kj rj − MFPS .
j=1
These equations state that the growth in each
sector’s output price equals the growth in the
share-weighted average of its input costs minus
the growth in MFP. MFP growth enters with a
negative sign because efficiency gains hold down
a sector’s output price given its input costs.
To reduce the amount of data needed to estimate MFP growth from equations A24 and A25, we
assume that every sector has the same labor and
capital shares up to a scaling factor that reflects
the intensity of semiconductor use. That is,
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
(A26)
β1L
βL
=…= 4 S = γ L
S
1 − β1
1− β4
β Kj,1
1 − β1S
=…=
β Kj,4
1 − β S4
= γ Kj
Case I: Conditioning on Aggregate MFP Growth
and
•
•
i=1
for j = 1,…, 4.
4
+[β S4 + (1 − β S4 )(1 − ∑ µ i )]π S
•
One can easily verify that the restricted factor
shares sum to one in each sector. Also, given
equation A26, one can show (with some algebra)
that γ L = α L and γ Kj = α Kj; that is, the income shares
for aggregate nonfarm business equal their counterparts in the semiconductor sector. Substituting
equation A26 into A24 and making use of the correspondence between the γs and the αs, we obtain
•
•
4
•
•
•
(A27) pi = (1 − β Si )[α L w + ∑ α Kj rj ] + β Si pS − MFPi
•
•
•
•
Case II: Conditioning on MFP Growth in Sector 4
•
•
MFPS = ( MFP4 − π S )/(1 − β S4 )
•
•
•
MFPi = MFP4 − π i − (β iS − β 4S )MFPS for i = 1,…, 3
•
4
•
i =1
•
•
•
•
•
•
•
•
•
Proof. The proof for Case II is nearly immediate. Subtract equation A29 from equations A28
and A30. After rearranging terms and using equa•
•
•
tion A30 to substitute MFPs for V – Ps, we obtain
•
•
for i = 1,…, 3;
•
(A29) p4 = (1 − β 4S )V + β 4S pS − MFP4 ;
•
MFPi = MFP4 − π i − (βiS − β S4 )MFPS
(A31)
(A28) pi = (1 − β Si )V + β Si pS − MFPi for i = 1,…, 3;
•
•
MFPi = MFP4 − π i − (β iS − β 4S )MFPS for i = 1,…, 3
MFP = ∑ µ i MFPi + µ S MFPS
4
α Kj rj) denote the share-weighted
Let V ≡ (αL w + Σj=1
growth in labor and capital costs for the nonfarm
•
business sector as a whole. Substitute V into the
dual equations A25 and A27, noting that γ L = α L
and γ Kj = α Kj in equation A25. The result is
•
i=1
•
MFPS = ( MFP4 − π S )/(1 − β S4 )
•
j =1
for i = 1,… 4.
3
MFP4 = (1 − β S4 )( MFP + ∑ µ i π i )
•
MFPS = ( MFP4 − π S )/(1 − β S4 ).
(A32)
•
(A30) pS = V − MFPS ,
where we have specifically identified sector 4,
which will serve as the numeraire sector.
We now use the dual equations to derive
expressions for MFP growth in two cases. In the
first case, we infer the rates of sectoral MFP growth
that are consistent with an independent estimate
of aggregate MFP growth (from the Bureau of
Labor Statistics). This case represents the methodology we use to compute growth contributions
through 2001. In the second case, which we use
for our steady-state analysis, we solve for aggregate MFP growth and MFP growth in sectors 1
through 3, conditional on an assumed pace of
MFP growth in sector 4. The next proposition
derives the expressions for sectoral MFP growth
in both cases.
•
•
•
•
Proposition 3. Let πS ≡ ps – p4 and πi ≡ pi – p4
(i = 1,…, 3) denote the rate of change in each
sector’s output price relative to that in sector 4.
Given the dual equations A28–A30, the solutions for sectoral and aggregate MFP growth are
as follows.
Equations A31 and A32, plus the expression for
•
MFP derived in Proposition 1, establish the results
for Case II. Note that the solution is recursive—first
•
solve for MFPs from equation A32, then substitute
the result into equation A31, and finally substitute
all the sectoral MFP growth rates into the expression for aggregate MFP growth.
To prove the result for Case I, substitute equations A31 and A32 into the expression for aggregate MFP growth. After rearranging terms, this
substitution yields
(A33)
•
•
4
•
MFP = ∑ µ i MFPi + µ S MFPS
i=1
4
3
i=1
3
i=1
•
= [ ∑ µ i + [µ S − ∑ µ i(βiS − β S4 )] /(1 − β S4 )] MFP4
3
−∑ µ i π i − [[µ S − ∑ µ i(βiS − β S4 )] /(1 − β S4 )]π S .
i=1
i=1
4
3
µi + [µS – Σ i=1
µi(βSi – β4S )]/(1 – βS4) and
Let B ≡ Σ i=1
•
solve equation A33 for MFP4:
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
35
A P P E N D I X 1 (continued)
•
(A34)
•
•
3
MFP4 = ( MFP + ∑ µ i π i )/ B
4
+ π S ( B − ∑ µ i )/ B.
i=1
With tedious algebra, one can show that B simplifies to be 1/(1 – βS4 ). Using this expression for B,
equation A34 becomes
•
(A35)
•
•
•
•
This reasoning implies that Ij,i = Kj,i and Ij,s = Kj,s
for all i and j. Combining these equalities with
equation A37, we obtain
i=1
3
MFP4 = (1 − β S4 )( MFP + ∑ µ i π i )
i=1
4
+ [β S4 + (1 − β S4 )(1 − ∑ µ i )]π S .
•
•
Proposition 4. Under the steady-state conditions in equations A36–A39 and the restrictions
on the income shares across sectors (equation
A26), the growth-accounting equation for aggregate labor productivity is
i=1
This equation, combined with equations A31 and
A32, completes the proof for Case I. As in Case II,
•
the solution is recursive. First, solve for MFP4 from
equation A35. Then, substitute the result into
equations A31 and A32.
•
(A39) Yj = Kj,i = Kj,s for i, j = 1,…, 4.
•
•
•
4
•
•
•
Y − H = ∑ ( α Ki α L )( MFPi + β Si MFPS ) + q + MFP,
i =1
where
•
•
4
•
MFP = ∑ µ i MFPi + µ S MFPS .
i =1
Analysis of the Steady State
The results presented so far do not require the
economy to have reached a steady state. We now
impose additional conditions to derive the growthaccounting equation for aggregate labor productivity in the steady state.
The first steady-state condition is that labor
input must grow at the same rate in every sector:
(A36)
•
•
Proof. Substitute equations A26, A36, A38,
and A39 into the growth-accounting equations
A10 and A11, and recall that γ L = α L and γ Kj = α Kj
when we impose the cross-sector restrictions on
the income shares. The result is
•
+
•
•
(A41)
(A37)
(A38)
•
•
•
•
•
•
β Si YS
•
+ MFPi for i = 1,…, 4
•
4
•
•
•
•
YS = α L L + ∑ α Kj Yj + MFPS .
j =1
Equations A40 and A41 form a system of five
•
•
•
equations in (Y1,..., Y4, YS ). Solving this system
yields
•
•
•
•
4
(A42) Yi = L + MFPi + ∑ ( α Ki α L )MFPj
j =1
Y j = Cj = Xj = I j,s = I j,i for i, j = 1,…, 4
•
•
4
j =1
L = LS = Li for i = 1,…, 4.
We also require that all components of a given
sector’s output grow at the same rate. Referring
back to equations A1 and A2, this condition
implies the following for the final-output sectors
and the semiconductor sector, respectively:
•
(A40) Yi = (1 − β Si )α L L + ∑ (1 − β Si )α Kj Yj
•
4
+ [β Si + ∑ ( α Ki α L )β Sj ]MFPS
•
j =1
YS = Sx = Sm = Si for i = 1,…, 4.
for i = 1,…, 4;
In addition, we require that all the growth rates in
equations A36–A38 be constant and that the
imported share of each sector’s capital stocks be
constant as well. Because Ij,i grows at a constant
rate over time, the stock of this (domestically
produced) capital will grow at the same constant
rate. Moreover, with the imported share of each
capital stock assumed to be constant, the total
stock, including imported capital, Kj,i, will grow at
the same rate as the domestically produced part.
36
(A43)
•
•
•
4
YS = L + ∑( α iK α L ) MFPj
j=1
4
•
+ [1 + ∑ ( α iK α L )β Sj ] MFPS .
j=1
Now, substitute equations A42 and A43 into equation A3 (the expression for growth in aggregate out•
•
•
put) and rearrange terms, noting that Sx = Sm = Ys
from equation A38:
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
•
•
4
•
•
•
(A44) Y = ∑ µ i Yi + µ S ,x Sx − µ S ,m Sm
•
4
•
= ∑ µ i Yi +( µ S ,x − µ S ,m )YS
i=1
•
•
4
•
•
4
•
4
•
•
4
•
•
+ q + ∑ µ i MFPi + µ S MFPS .
•
+ ∑ µ i MFPi + [ ∑ µ iβiS + µ S ,x − µ S ,m ] MFPS .
i=1
•
i=1
i=1
•
•
(A45) Y − H = ∑ ( α iK α L )( MFPi + βiS MFPS )
= L + ∑ ( α iK α L )( MFPi + βiS MFPS )
4
•
4
Recalling that L = H + q and that µS = Σi=1
µiβiS +
µS,x – µS,m from equation A21, we obtain
i=1
i=1
i=1
APPENDIX 2
Data Sources
T
his appendix describes the data series used
in the paper. All data are annual and cover
the period from 1973 to 2001. Note that we have
not incorporated the July 2002 revision of the
NIPAs, which was released while the paper was
in production.
year, which depends on the level in 1977 and the
level in 1978, straddles the change in methodology.
To account for these effects, we added 0.17 percentage point to the growth rate of the BLS series
for nonfarm business output for each year through
1977 and 0.09 percentage point in 1978.
Real Output in the Nonfarm Business Sector (Y )
Price Index for Nonfarm Business Output (p)
Data through 2000 are from the BLS multifactor productivity data set. (The version we used was
released in March 2002.) In constructing output,
the BLS relies primarily on the BEA real output
series for nonfarm business less housing. Both
the BEA and BLS series are superlative indexes
of output. For 2001, we extended the BLS series
using annual growth rates of the BEA series for
real output in nonfarm business less housing
(NIPA, table 1.8).
Both the BLS and the BEA have incorporated
the effects of technical changes to the consumer
price index (CPI) back to 1978 (specifically, the
introduction of geometric means in the CPI).
However, the output data prior to 1978 must be
adjusted to be methodologically consistent with the
later data. According to the Economic Report
of the President (1999, 94), the introduction of
geometric means prior to 1978 would hold down
CPI inflation by 0.2 percentage point per year. From
1973 to 1977, consumption expenditures accounted
for about 85 percent of nonfarm business output
in current dollars. Thus, the incorporation of geometric means prior to 1978 would reduce inflation
in nonfarm business prices by about 0.17 percentage point per year (0.2 × 0.85) through 1977 and
would boost growth in nonfarm business output by
the same amount each year. In 1978, the adjustment is smaller because the growth rate for that
We measured p as an implicit price deflator,
constructed as the ratio of current-dollar nonfarm business output to real nonfarm business
output from the BLS multifactor productivity
data set. To build in the effects of the CPI revision
described in the previous paragraph, we then
adjusted down the rate of change of this BLS
series by 0.17 percentage point annually for
1974–77 and by 0.09 percentage point for 1978.
For the rate of change in 2001, we extended the
BLS series using the annual growth rate of BEA’s
price index for nonfarm business less housing.
Capital Inputs (KC , KSW , KM , KO )
We constructed these capital inputs in two
steps. The first step develops productive capital
stocks for a detailed set of assets. The second
step aggregates these detailed stocks to the four
capital inputs used in our analysis.
Productive stocks for detailed types of
capital. For each type of capital, we took data
through 2000 directly from the BLS multifactor
productivity data set. The BLS constructs productive stocks for highly disaggregated asset categories, starting with data on real investment for
sixty-one different types of business capital and
then translating these investment flows into productive stocks with the use of hyperbolic ageefficiency profiles.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
37
A P P E N D I X 2 (continued)
We extended these BLS productive stocks to
2001 as follows.1 For nonresidential fixed capital—which constitutes a large majority of all capital used in nonfarm business—we extended the
detailed BLS investment series to 2001 using
NIPA investment data for five broad asset groups:
computers and peripheral equipment, software,
communication equipment, other equipment, and
nonresidential structures. For each group except
computers and peripheral equipment, we used
the growth rate of investment in 2001 for the
group as a whole to extend the investment series
for each asset within the group. For example, we
used the 2001 growth rate for overall NIPA software investment to extend the investment series
for each of the three different types of software.
For computers and peripheral equipment, we
employed a more refined procedure to capture
the differences in trend growth rates across the
assets in this important group. To begin, we used
the BLS data set to calculate the average growth
rate of investment in 1999 and 2000 for each type
of computer and peripheral equipment—mainframes, personal computers, printers, terminals,
integrated systems, and three different types of
storage devices. These growth rates represented
our estimate of “trend” growth in investment for
2001 for each detailed category. Then we scaled
these trend rates so that the resulting individual
investment series would chain aggregate to the
level of total real investment in computers and
peripheral equipment in 2001.2
Given an estimate of real investment in 2001
for each type of nonresidential fixed capital, we
extended the BLS productive capital stocks to 2001
with the perpetual inventory method. Specifically,
for each detailed asset type, we calculated a translation factor (ft) for each year through 2000 from
the following equation:
Kt = ft Kt–1 + (It + It–1)/ 2,
where (following BLS methodology) Kt is measured as the average of the stocks at the end of
years t and t – 1. We used the value of ft in 2000
and the detailed investment data to construct
productive stocks for each type of nonresidential
fixed capital for 2001.
The other assets included in the BLS measure
of nonfarm business capital are tenant-occupied
rental housing, inventories, and land. For tenantoccupied rental housing, we extended the BLS
38
productive stock to 2001 with a simple regression
equation. This equation regressed the BLS productive stock on its own lag and on real investment in multifamily residential structures from
the NIPAs. The coefficients from this equation,
combined with NIPA data on investment in multifamily structures for 2001, generated the estimate
of the stock of tenant-occupied rental housing in
2001. For the stock of inventories, we extended
the BLS series to 2001 using NIPA inventory data.
For the stock of land, we extended the BLS series
to 2001 with the average growth rate of this stock
for the five years through 2000.
Aggregation. The BLS uses the Tornquist formula to aggregate the detailed productive capital stocks into measures of capital services. The
Tornquist aggregate is a weighted average of the
growth rates of the various productive stocks, with
the weight for each asset type equal to its estimated share of total capital income. To construct
our capital aggregate for computer and peripheral
equipment (KC), we applied the Tornquist formula
to the eight components of such equipment. For
software (KSW), we followed a similar procedure
for the three different types of software. For communication equipment (KM), the capital services
aggregate just equals the productive stock; the
Tornquist formula is not needed because we have
no asset detail within this aggregate. Finally, to
construct KO, our first step was to extend the BLS
measure of aggregate capital services to 2001
(using the Tornquist formula). Then, we stripped
out computer and peripheral equipment, software,
and communication equipment from aggregate
capital services to arrive at KO.
Labor Hours (H)
Through 2000, labor hours are from the BLS
multifactor productivity data set. We extended the
data to 2001 using the growth rate in hours of all
persons in the nonfarm business sector from the
BLS Productivity and Cost release.
Labor Quality (q)
The BLS measures labor quality as the difference in the growth rate of labor input and labor
hours. To calculate labor input, the BLS divides the
labor force into a number of age-sex-education
cells and then constructs a weighted average of
growth in hours worked in each cell, with the
weight for each cell equal to its share of total
labor compensation. Through 2000, our measure
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
of labor quality is from the BLS multifactor productivity data set. For 2001, we assumed that
labor quality generated a contribution of 0.25 percentage point to growth in labor productivity, its
average contribution over 1996–2000.
αj )
Income Shares (α
The income share for each detailed type of
nonresidential fixed capital in a given year was
calculated from the following equation:
αj = (R + δj – Πj )pj Kj Tj /pY.
We discuss each component of this equation below.
Note that these income shares vary from year to
year and are not fixed at a period-average value.
For tenant-occupied rental housing, inventories,
and land, the income shares through 2000 were
taken directly from the BLS multifactor productivity data set. For 2001, we extrapolated forward the
BLS year-2000 level of capital income for each asset
using the trend growth rate from 1995 to 2000. We
then divided the estimated 2001 capital income
for each asset by total income in nonfarm business
to obtain income shares.
Once we estimated the income-share series for
each capital asset, the income share for labor equaled
unity minus the total income share for capital.
Depreciation rate (δδj). For the most part,
the depreciation rate for each type of equipment
and structure comes from the BEA (as presented
in Fraumeni 1997, 18–19). However, as indicated
above, the BEA provides very little information
on depreciation rates for the individual types of
computers and peripheral equipment; we followed
Whelan (2000) and set these depreciation rates
equal to a geometric approximation calculated
from the BEA capital stocks and investment flows.
For personal computers, we are uncomfortable
with the BEA’s procedure for depreciation rates,
and instead we set the depreciation rate for PCs
equal to the 30 percent annual rate for mainframe
computers.3 For software, we used the BEA
depreciation rates described by Herman (2000,
19). The BEA assumes that prepackaged software
has a service life of three years and a depreciation
rate of 55 percent per year; own-account and custom software each have service lives of five years
and a depreciation rate of 33 percent per year.
Πj).
Expected nominal capital gain/loss (Π
We calculated Πj as a three-year moving average
of the percent change in the price of asset j (pj).
The moving average serves as a proxy for the
unobserved expectation of price change. Through
2000, the pj series for each asset is the investment
price index from the BLS multifactor productivity
data set. Each pj series was extended to 2001
using the same procedure as that employed for
real investment for each asset. Specifically, we
extended the detailed BLS price series to 2001
1. The BLS actually relies on a two-way disaggregation by type of asset and by industry. For our analysis, we used data by
asset that already have been aggregated across industries.
2. This scaling procedure does not generate sensible results if the estimated trend growth rate of investment for a particular asset differs in sign from the actual 2001 change for the broader group to which it belongs. Because such sign
differences periodically occur for some assets within software, communication equipment, other equipment, and nonresidential structures, we used the simpler procedure described above for extending investment in business fixed
assets other than computer hardware.
3. As described in Herman (2000, 20), the BEA sets the depreciation rate for personal computers so that 10 percent of
the original value remains after five years of service, which implies an annual geometric depreciation rate of 37 percent.
By construction, this depreciation rate captures the full loss of value during each year of the assumed five-year service
life. In contrast, the BEA’s depreciation rates for other types of computer hardware are constructed to capture only the
loss of value over and above the decline in the asset’s constant-quality price index (Πj). This concept of depreciation is
the appropriate one to combine with Πj in order to measure the full loss of asset value in the formula for the income
share. However, for personal computers, the BEA’s depreciation rate, when combined with Πj , double-counts the loss
of value. One fix for this problem would be to drop the Πj term from the income-share formula for PCs. However, doing
so would be appropriate only if the BEA’s depreciation rate of 37 percent accurately measures the full loss of value.
While there is relatively little hard evidence on this subject, our sense is that PCs typically lose more than 37 percent
of their value over the course of a year. Thus, dropping the Πj term from the user cost formula does not seem an adequate solution to this problem. Instead, we set δj for PCs equal to the value for mainframes (30 percent per year) and
plugged this value into the income-share formula, along with the value of Πj for PCs. This may not be an ideal approach,
but given the very limited research on depreciation for PCs, we judged it to be the best choice at present. (For a fuller
discussion of related issues, see Oliner 1994.) Similar problems may affect other assets as well, and we believe that
future research in this area is crucial.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
39
A P P E N D I X 2 (continued)
using NIPA investment prices for five broad categories of nonresidential fixed investment: computers and peripheral equipment, software, communication equipment, other equipment, and nonresidential structures.4 For each individual asset,
the resulting rate of price change was applied to
the year-2000 level of pj to calculate pj for 2001.
Current-dollar productive capital stock
( pj Kj). For each asset, this series is simply the
product of the real productive stock (Kj) and the
asset price index (pj), both of which are discussed above.
Tax adjustment (Tj). For each asset, this
adjustment equals (1 – c – τv)/(1 – τ), where c is
the rate of investment tax credit, τ is the corporate tax rate, and v is the present value of $1 of
tax depreciation allowances. Karl Whelan kindly
provided these series, which are discussed further in Whelan (1999).
Current-dollar nonfarm business output
( pY). Through 2000, this series is from the BLS
multifactor productivity data set. For 2001, we
extended the BLS series using the annual growth
rate of the BEA series for current-dollar output in
the nonfarm business sector less housing.
Nominal net return (R). We calculated R as
the ex post net return earned on the productive
stock of nonresidential equipment and structures. Thus, we obtained R as the solution to the
following equation for each year in our sample:
N4
N
4
j=1
j=1
= ∑ (R + δj – Πj)pj Kj Tj /pY = BLS series for
= ∑ αj,
where the summations are over all N types of nonresidential equipment, software, and structures.
This procedure yielded an annual series for R
through 2000. For 2001, we estimated R from a
regression with the following explanatory variables:
a constant, two lags of R, the rate of price change
for nonfarm business output, the acceleration in
real nonfarm business output, the unemployment
rate, and the share of corporate profits in GNP.
µC , µSW , µM , µS , µO )
Current-Dollar Output Shares (µ
The denominator of each output share is currentdollar nonfarm business output ( pY), the data
source for which was described above. Here, we
focus on the measurement of current-dollar sectoral output, the numerator in each share.
Computer sector. We used NIPA data on final
sales of computers to measure current-dollar
computer output (pCYC). NIPA final sales equals
40
the sum of current-dollar spending on computers
and peripheral equipment in the following categories: private fixed investment, personal consumption expenditures, government expenditures, and
net exports of goods and services. This sum omits
the small portion of final computer output that
ends up in business inventories because the NIPA
inventory data do not break out computing equipment from other inventories.
Software sector. To estimate pSWYSW, we
started with unpublished data from the BEA on
current-dollar final sales of software from 1987 to
2000. We then adjusted this series for software
not produced in the nonfarm business sector by
stripping out the BEA estimate of own-account
software produced by the government.5 Finally,
we extended the 1987 level back to earlier years
and the 2000 level forward to 2001 using NIPA
data on growth in current-dollar software investment by businesses.
Communication equipment sector. To estimate pMYM, we used unpublished data from the
BEA on total current-dollar final sales of communication equipment from 1997 to 2000. We extended
the 1997 level back in time and the 2000 level forward to 2001 using NIPA data on the growth of
current-dollar business investment in communication equipment.
Semiconductor sector. Our series for currentdollar semiconductor output (pSYS) equals currentdollar shipments of products in SIC category 36741
(integrated microcircuits). Federal Reserve Board
staff construct this shipments series as an input to
the Board’s index of industrial production, using
Census Bureau reports through 1999 and trade
data from the Semiconductor Industry Association
(SIA) for 2000 and 2001. Because the shipments
series is not available before 1977, we set the value
of the semiconductor output share (µS) during
1973–76 equal to its 1977 value.
Other final-output sector. We estimated
current-dollar output in this sector ( pOYO) as a
residual after accounting for all other components
of nonfarm business output:
pOYO = pY – pCYC – pSWYSW – pMYM – pS (Sx – Sm),
where the final term is current-dollar net exports of
semiconductors. (This is the only part of semiconductor production that shows up in domestic final
output.) The data sources for pY, pCYC, pSWYSW, and
pMYM were described above. We obtained data on
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
current-dollar net exports of semiconductors as follows. For the period from 1989 to 2001, we started
with series constructed by Federal Reserve Board
staff for current-dollar exports and imports of products in SIC code 3674 (semiconductors and related
devices), which are based on detailed figures from
the International Trade Commission. Because the
3674 category is broader than just semiconductors,
we scaled the export and import series for SIC code
3674 down to 36741 (integrated microcircuits)
using the ratio of domestic shipments in 36741 to
domestic shipments in 3674. Prior to 1989, we did
not have detailed trade data, and we extended the
export and import series back in time using the rate
of change in domestic shipments of semiconductors (the series pSYS described above).
Ratio of Semiconductor Output to
Domestic Semiconductor Use (1 + θ)
Domestic semiconductor use can be expressed
as domestic semiconductor output minus net
exports of semiconductors. Thus,
1 + θ = YS / [YS – (Sx – Sm)]
= pSYS / [pSYS – (pS Sx – pS Sm)],
where the second equality converts each series to
current dollars. The data sources for pSYS and
pS Sx – pS Sm were described above.
πC , πSW , πM , πS )
Rates of Relative Price Change (π
Each πi series (i = C, SW, M, and S) represents
the rate of change in the price ratio pi /pO. Here,
we describe the data source for each price series
that enters these ratios.
Computer sector. pC is measured as an implicit
price deflator for the output of computers in the
NIPAs. We calculated this deflator as the ratio of
current-dollar computer output (defined as the sum
of all final sales of computers and denoted above by
pCYC) to a chain aggregate of real outlays for the
same spending categories, which we denote by YC.
Software sector. pSW is an implicit price
deflator for software produced in the nonfarm
business sector. Using NIPA data, we calculated
this deflator as the ratio of current-dollar software output (the series pSW YSW described above)
to a chain aggregate of real software outlays
denoted by YSW. To construct YSW, we did a “chain
strip-out” of government own-account software
from total final sales of software, parallel to our
calculation of the current-dollar series. The
growth rate of the resulting aggregate series for
real software outlays was about 1 percentage
point per year higher than the growth rate of real
business investment in software over 1987–2000,
the period over which we can construct YSW. To
extend YSW back to years before 1987 and forward
to 2001, we used the annual growth rates of real
business investment in software adjusted by this
1987–2000 wedge.6
Communication equipment sector. pM is an
implicit deflator for the output of communication
equipment in the NIPAs. We calculated this deflator
as the ratio of current-dollar outlays for communication equipment (the series pMYM defined above)
to a chain aggregate of real outlays denoted by YM
and constructed in an analogous manner to pMYM.
To calculate YM we used unpublished data from
the BEA on total real final sales of communication
equipment from 1997 to 2000. We extended the
1997 level back in time and the 2000 level to 2001
using published NIPA data on the growth of real
business investment in communication equipment.
Other final-output sector. Like the other
price series, pO is an implicit deflator, which
equals the ratio of current-dollar output for this
sector (the series pOYO defined above) to a chain
aggregate of the sector’s real output (YO ). We
constructed YO by starting with our series for real
nonfarm business output (Y ) and then chainstripping-out all other components of Y (that is,
real output of computers, software, and communication equipment, along with real exports and
4. Just as for the investment series, the scaling procedure that we used for computer hardware does not generate sensible results if the trend rate of price change for a particular asset differs in sign from the actual 2001 change for the
broader group to which it belongs. Because these sign differences occur for some noncomputer price series, we
employed the simpler extrapolation method described above to extend the price series for nonresidential fixed investment other than computer hardware.
5. Estimates of government own-account software from 1996 to 2000 are available as unpublished data from the BEA. In
addition, Parker and Grimm (2000) provide estimates of government own-account software for 1979 and 1992. Using
these values, we linearly interpolated the government own-account series backward in time.
6. Real software output is the only extrapolated series for which we used a wedge adjustment. For other extrapolated
series, there was not a significant difference between the growth rate of the series in question and the extrapolator series.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
41
A P P E N D I X 2 (continued)
imports of semiconductors). Roughly speaking,
the chain strip-out inverts equation A3 in
Appendix 1 to solve for the growth of YO, and the
resulting growth rates are then linked together to
create a series in index levels. To construct the
series for real exports and imports of semiconductors needed for the chain strip-out, we
assumed that the price of exports and imports of
semiconductors was equal to the semiconductor
price series described in the next paragraph.
Semiconductor sector. For 1977–2001, the
data source for pS is the deflator for SIC 36741
that Federal Reserve staff developed to estimate
industrial production; we used this series to compute the annual percent change in pS for 1978
through 2001. For years before 1978, we calculated
the percent change in pS by extrapolating back in
time using data from Grimm (1998). Specifically,
we calculated the average annual percent change
between 1974 and 1977 in Grimm’s “Summary
price index for MOS memory chips” (p. 12), and
then took the ratio of this average 1974–77 percent change to the percent change for 1978 based
on the Federal Reserve series. We multipied the
1978 percent change in the Federal Reserve
series by this ratio and used the resulting value as
the percent change in pS for each year from 1974
to 1977.
Semiconductors as a Share of Current-Dollar
βCS , βSSW , βMS , βOS )
Input Costs (β
Computer sector. We set βCS equal to 0.3 for
all years. That is, we assumed that semiconductors account for 30 percent of the current-dollar
input cost of computer producers. This value lies
at the middle of the range employed by Triplett
(1996). Although the SIA publishes data on semiconductor usage by the computer industry, these
data are not appropriate for our purpose. As
noted by Flamm (1997, 11), the SIA data cover
only the semiconductors sold by “merchant” producers in the open market; these data exclude
“captive” production by U.S. computer manufacturers, notably IBM. Thus, the SIA-based measure
would greatly understate semiconductor use during the 1970s and 1980s, when IBM was the dominant U.S. computer producer.
S
to zero because
Software sector. We set βSW
semiconductors are not a direct input to software
42
production. (Of course, the software industry uses
computers and communication equipment that
contain semiconductors, but it does not directly
use semiconductors.)
Communication equipment sector. We used
data from the SIA to construct βMS . The SIA provides
data on worldwide shipments of semiconductors
for 1976–2001. The SIA also publishes data for
1985–94 on the share of these worldwide shipments purchased by producers of communication
equipment in the United States. After 1994, the SIA
redefined this latter series to cover “the Americas.”
We linked the series on the U.S.-only share through
1994 with the series on the Americas share from
1995 forward. (Because the share figures are available only back to 1985, we set this share for earlier
years equal to the 1985 value.) We then multiplied
the resulting share series by worldwide semiconductor shipments to calculate the current-dollar
value of semiconductors used by the communication equipment industry in the United States. (To
the extent that semiconductors are used to produce communication equipment elsewhere in
North or South America, the series will overstate
semiconductor use in the United States alone from
1995 forward.) To construct βMS , we divided the
series just described by our estimate of the currentdollar value of communication equipment produced in the United States, pMYM. Prior to 1976 (for
which data on worldwide semiconductor shipments
are not available), we set βMS equal to its 1976 value.
Other final-output sector. To estimate βOS ,
recall the expression for µS in equation A22 of
Appendix 1:
4
µ S = ∑ µ iβ iS (1 + θ ),
i =1
which can be written with explicit sectoral notation as
S
µ S = [µ CβCS + µ SW β SSW + µ M β M
+ µ OβOS ](1 + θ ).
Solving this equation for βOS yields
βOS =
S
µ S − (1 + θ )[µ CβCS + µ SW β SSW + µ M β M
]
.
µ O(1 + θ )
The data sources for all series on the right-hand side
of this expression have already been discussed.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
REFERENCES
Aaronson, Daniel, and Daniel Sullivan. 2001. Growth in
worker quality. Federal Reserve Bank of Chicago
Economic Perspectives 25 (Fourth Quarter): 53–74.
Grimm, Bruce. 1998. Price indexes for selected semiconductors, 1974–96. Survey of Current Business 78
(February): 8–24.
Aizcorbe, Ana. 2002. Why are semiconductor prices falling
so fast? Industry estimates and implications for productivity measurement. Federal Reserve Board Finance and
Economics Discussion Series Paper 2002-20, March.
<www.federalreserve.gov/pubs/feds/2002/200220/
200220pap.pdf> (May 10, 2002).
Herman, Shelby. 2000. Fixed assets and consumer durable
goods. Survey of Current Business 80 (April): 17–30.
Baily, Martin Neil. 2002. The new economy: Post mortem
or second wind? Journal of Economic Perspectives 16
(Spring): 3–22.
Basu, Susanto, John G. Fernald, and Matthew D. Shapiro.
2001. Productivity growth in the 1990s: Technology, utilization, or adjustment? Carnegie-Rochester Conference
Series on Public Policy 55 (December): 117–65.
Bosworth, Barry P., and Jack E. Triplett. 2000. What is
new about the new economy? Information technology,
economic growth, and productivity. The Brookings
Institution. Unpublished paper.
Brynjolfsson, Erik, and Lorin M. Hitt. 2000. Beyond computation: Information technology, organizational transformation, and business performance. Journal of
Economic Perspectives 14 (Fall): 23–48.
Congressional Budget Office. 2002. The budget and economic outlook: Fiscal years 2003–2012. Washington,
D.C.: CBO. <www.cbo.gov> (March 26, 2002).
DeLong, J. Bradford. 2002. Productivity growth in the
2000s. University of California at Berkeley. Unpublished
paper.
Domar, Evsey. 1961. On the measurement of technological
change. Economic Journal 71 (December): 309–29.
Doms, Mark. 2002. Communications equipment: What
has happened to prices? Federal Reserve Board. Unpublished paper.
Economic report of the president. 1999. Washington,
D.C.: U.S. Government Printing Office.
Hulten, Charles R. 1978. Growth accounting with intermediate inputs. Review of Economic Studies 45
(October): 511–18.
Jorgenson, Dale W. 2001. Information technology and the U.S.
economy. American Economic Review 91 (March): 1–32.
Jorgenson, Dale W., Mun S. Ho, and Kevin J. Stiroh.
2002. Projecting productivity growth: Lessons from the
U.S. growth resurgence. Federal Reserve Bank of Atlanta
Economic Review 87 (Third Quarter): 1–13.
Jorgenson, Dale W., and Kevin J. Stiroh. 2000. Raising the
speed limit: U.S. economic growth in the information age.
Brookings Papers on Economic Activity 1, 125–211.
Kiley, Michael T. 2001. Computers and growth with frictions: Aggregate and disaggregate evidence. CarnegieRochester Conference Series on Public Policy 55
(December): 171–215.
Martin, Bill. 2001. American potential. UBS Asset
Management. Research paper, May. <www.ubs.com/e/
globalam_uk/research.Referenz2.pdf> (August 16, 2002).
McKinsey Global Institute. 2001. U.S. productivity
growth 1995–2000: Understanding the contribution
of information technology relative to other factors.
Washington, D.C.: McKinsey Global Institute.
Oliner, Stephen D. 1994. Measuring stocks of computer
peripheral equipment: Theory and application. Federal
Reserve Board. Unpublished paper.
Oliner, Stephen D., and Daniel E. Sichel. 2000a. The
resurgence of growth in the late 1990s: Is information
technology the story? Journal of Economic Perspectives
14 (Fall): 3–22.
———. 2000b. The resurgence of growth in the late 1990s:
Is information technology the story? Federal Reserve
Board Finance and Economics Discussion Series Paper
2000-20, May. <www.federalreserve.gov/pubs/feds/
2000/200020/200020pap.pdf> (March 26, 2002).
———. 2002. Washington, D.C.: U.S. Government
Printing Office.
Flamm, Kenneth. 1997. More for less: The economic
impact of semiconductors. Semiconductor Industry
Association, December.
Fraumeni, Barbara M. 1997. The measurement of depreciation in the U.S. national income and product accounts.
Survey of Current Business 77 (July): 7–23.
Gordon, Robert J. 2000. Does the “new economy” measure up to the great inventions of the past? Journal of
Economic Perspectives 14 (Fall): 49–74.
———. 2002. Technology and economic performance in
the American economy. National Bureau of Economic
Research Working Paper 8771, February.
Parker, Robert, and Bruce Grimm. 2000. Software prices
and real output: Recent developments at the Bureau of
Economic Analysis. BEA. Unpublished paper.
Phister, Montgomery, Jr. 1979. Data processing technology and economics. Santa Monica, Calif.: Santa Monica
Publishing Company.
Roberts, John M. 2001. Estimates of the productivity
trend using time-varying parameter techniques.
Contributions to Macroeconomics 1, no. 1:1–30.
<www.bepress.com/bejm/contributions/vol1/iss1/art3>
(March 26, 2002).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
43
Sichel, Daniel E. 1997. The computer revolution: An
economic perspective. Washington, D.C.: The Brookings
Institution.
Solow, Robert. 1957. Technical change and the aggregate
production function. Review of Economics and Statistics
39 (August): 65–94.
Steindel, Charles, and Kevin J. Stiroh. 2001. Productivity
growth: What is it and why do we care about it? Business
Economics 36 (October): 13–31.
Triplett, Jack E. 1996. High-tech industry productivity and
hedonic price indices. In May 2–3, 1996, OECD workshop
proceedings Industry productivity: International comparison and measurement issues. Paris: Organisation
for Economic Co-operation and Development.
Whelan, Karl. 1999. Tax incentives, material inputs, and
the supply curve for capital equipment. Federal Reserve
Board Finance and Economics Discussion Series Paper
44
1999-21, April. <www.federalreserve.gov/pubs/feds/
1999/199921/199921pap.pdf> (March 26, 2002).
———. 2000. Computers, obsolescence, and productivity.
Federal Reserve Board Finance and Economics Discussion
Series Paper 2000-6, February. <www.federalreserve.gov/
pubs/feds/2000/200006/200006pap.pdf> (March 26, 2002).
———. 2001. A two-sector approach to modeling U.S. NIPA
data. Federal Reserve Board Finance and Economics Discussion Series Paper 2001-4, April. <www.federalreserve.
gov/pubs/feds/2001/200104/200104pap.pdf> (March 26,
2002).
Zraket, Charles A. 1992. Software: Productivity puzzles,
policy challenges. In Beyond spinoff: Military and
commercial technologies in a changing world, edited
by John A. Alic and others. Cambridge, Mass.: Harvard
Business School Press.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Technology and U.S. Wage
Inequality: A Brief Look
DAVID CARD AND JOHN E. D I NARDO
Card is a professor in the Department of Economics at the University of California, Berkeley.
DiNardo is a professor in the Department of Economics and School of Public Policy at the University
of Michigan, Ann Arbor. They thank Ken Chay and David Lee for many helpful discussions, Charles
Nelson of the U.S. Census Bureau and Anne Polivka of the Bureau of Labor Statistics for assistance
in using the data, and Elizabeth Cascio for outstanding research assistance. They also thank David
Autor, Daniel Hamermesh, Lawrence Katz, Paul Beaudry, and participants at the SOLE meeting and
the Royal Statistical Society’s “Explanations for Rising Economic Inequality” conference in November
2001 for comments, suggestions, and criticisms. Card’s research was supported by grants from the
National Science Foundation and the National Institute of Child Health and Development.
he industrial revolution of the eighteenth
and nineteenth centuries left in its wake
a large body of literature, both popular
and scholarly, arguing that technology
had wrought fundamental changes to
the labor market. Some argued that as
important as the steam engine and new machinery
were to this new economy, “mental steam power”
and “intellectual machinery”—the ability of workers
to interact with the new technologies—was of equal
or greater significance. Even as debate about that
earlier period continues more than one hundred
years later, a new debate—with interesting parallels
to that earlier discussion—has ensued about the
effect of computers and other information and communications technology on the labor market.1
Developments in personal computers, for example, led Time magazine to make the device its 1982
“Person of the Year” and argue that “the information
revolution . . . has arrived . . . bringing with it the
promise of dramatic changes in the way people live
and work, perhaps even in the way they think.
America will never be the same.” By the late 1980s
and early 1990s, labor market analysts were finding
it apparent that wage inequality had risen, and a
series of papers argued that these two developments—rapid technological change and rising wage
inequality—were related.2 These papers and the
large literature that followed have paved the way
for the virtually unanimous agreement among econ-
T
omists that developments in computers and related
information technologies in the 1970s, ’80s, and ’90s
have led to increased wage inequality.
In the labor economics literature this consensus
view has become known as the “skill-biased technological change” (SBTC) hypothesis. Specifically, this
hypothesis is the view that a burst of new technologies led to an increased demand by employers for
highly skilled workers (who are more likely to use
computers) and that this increased demand led to
a rise in the wages of the highly skilled relative to
those of the less skilled and therefore an increase in
wage inequality.
In this paper, which is a substantially abridged
version of Card and DiNardo (2002), we reconsider
the evidence for the SBTC hypothesis. We focus considerable attention on changes over time in overall
wage inequality and in the evolution of relative wages
of different groups of workers. In doing so, we conclude that despite the considerable attention this
view has received in the literature, SBTC falls far
short of unicausal explanation of the substantial
changes in the U.S. wage structure of the 1980s and
1990s. Indeed, although there have been substantial
changes in the wage structure in the last thirty
years, many of which are documented here, SBTC
by itself does not prove to be particularly helpful in
organizing or understandings these changes. Based
on the evidence, we conclude that it is time to reevaluate the case that SBTC offers a satisfactory
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
45
explanation for the rise in U.S. wage inequality in
the last quarter of the twentieth century.
An Empirical Framework for Understanding SBTC
here are many theoretical versions of skillbiased technological change. To help fix ideas,
this paper focuses on a simple SBTC formulation,
versions of which have helped guide the large
empirical literature in labor economics.3 Assume
that aggregate labor demand is generated by a constant elasticity of substitution (CES) production
function of the form
T
(1) Y = f (NH, NL ) = A [α( gH NH )(σ–1)/σ
+ (1 – α) (gL NL )(σ–1)/σ ] σ/(σ–1),
where Y represents the value of output; NH represents the labor input (employment or hours) of highskilled workers; HL represents the input of lowskilled labor; σ > 0 is the elasticity of substitution
between the labor inputs; and A, α, gH, and gL are
technological parameters that can vary over time.4 In
many empirical applications NH is measured by the
number of college graduates (or “college-equivalent”
workers), and NL is measured by the number of
high-school graduates (or “high-school-equivalent”
workers.) For given values of the technology parameters, the relative demand for high-skilled labor is
determined by setting the ratio of the marginal
product of the two groups equal to the ratio of their
wages, wH /wL. Taking logarithms of the resulting
expression and first-differencing over time leads to
a simple expression that has been widely used to
discuss the evolution of relative wages:
(2) ∆log[wH /wL] = ∆log[α/(1 – α)]
+ (σ – 1)/σ∆log[gH /gL] – 1/σ∆log[NH /NL].
The equation is assumed to hold true for every
time period (typically a year). If the relative supply
of the two skill groups is taken as exogenous, this
equation completely determines the evolution of
relative wages over time. The technological parameters cannot be observed directly but are often
inferred by making some assumptions about how
they evolve over time. From equation 2, two observations follow directly.
First, changes in relative wages must reflect either
changes in relative supplies or changes in technology.
Other features of the labor market that potentially
affect relative wages (such as the presence of unions,
institutional wage floors, etc.) are essentially ignored.
In the absence of technological change, the relative
wage of high-skilled workers varies directly with
46
their relative supply. Despite some problems of identification, there exists a “consensus” estimate of
σ ≈ 1.5 when the two skill groups are college and
high-school workers.5 This estimate implies, for
example, that a 10 percent increase in the relative
proportion of college-educated workers lowers the
relative wage of college-educated workers by 6.6 percent. Since the relative proportion of highly educated workers has been rising throughout the past
several decades, the only way to explain a rise in
the relative wage of skilled workers (and hence a
rise in wage inequality) is through changes in the
technology parameters α or g.
Second, skill-biased changes in technology lead
to changes in wage inequality. A shift in the parameter A, or an equiproportional shift in gH and gL,
leaves the relative productivity of the two skill groups
unchanged and affects only the general level of
wages. SBTC involves either an increase in α or an
increase in gH relative to gL. A rise in α raises the marginal productivity of skilled workers and at the same
time lowers the marginal productivity of unskilled
workers. This type of technological change has been
referred to as “extensive” SBTC; Johnson (1997)
gives as an example of extensive SBTC the introduction of robotics in manufacturing. The other situation,
sometimes referred to as “intensive” SBTC, arises
when technological change enhances the marginal
productivity of skilled workers without necessarily
lowering the marginal product of unskilled workers.6
Technology or Tautology?
s has been observed, in this framework SBTC
can be defined to exist whenever changes in
relative wages are not inversely related to changes
in relative supply. Indeed, the test for SBTC proposed by Katz and Murphy (1992) is a multifactor
version of this point. Given a priori qualitative or
quantitative evidence on how different skill groups
are affected by changes in technology, however, the
SBTC hypothesis can be tested using data on relative wages and relative labor supplies of different
education/age groups, and we proceed to do so in a
number of ways.
Aggregate trends in technology. A first task in
making the SBTC hypothesis testable is to quantify
the pace of technological change. The most widely
cited source of SBTC in the 1980s and 1990s is the
personal computer (PC) and related technologies,
including the Internet. Chart 1 presents a timeline of
key events associated with the development of personal computers, plotted with two simple measures
of the extent of computer-related technological
change. Although electronic computing devices were
A
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 1
Measures of Technological Change
60
6
50
Computer use at work (right scale)
5
40
4
30
3
Share of IT in GDP (left scale)
20
2
IBM AT
(1984)
Apple II
(1977)
1
1948
1952
1956
1960
1964
1968
1972
1976
Netscape
(1994)
Windows 3.1
(1990)
IBM PC
(1981)
1980
1984
1988
1992
P e r c e n t o f wo r k e r s u s in g c o m p u t e r s
Sh a r e o f IT in d u s t r y o u t p u t in G D P
7
10
1996
Source: Jorgenson (2001) and authors’ analysis of October Current Population Survey, various years
developed during World War II, and the Apple II was
released in 1977, many observers date the beginning
of the computer revolution to the introduction of the
IBM-PC in 1981. This development was followed by
the IBM-XT (the first PC with built-in disk storage) in
1982 and the IBM-AT in 1984. As late as 1989, most
personal computers used Microsoft’s disk operating
system (DOS). More advanced graphical-interface
operating systems gained widespread use only with
the introduction of Windows 3.1 in 1990.
Some analysts have drawn a sharp distinction
between stand-alone computing tasks (such as wordprocessing or database analysis) and organizationrelated tasks (such as inventory control, supply-chain
integration, and internet commerce) and argue that
innovations in the latter domain are the major source
of SBTC.7 This reasoning suggests that the evolution of network technologies is at least as important
as the development of personal computer technology. The first network of mainframe computers (the
ARPANET) began in 1970 and had expanded to
about 1,000 host machines by 1984.8 In the mid1980s the National Science Foundation laid the
backbone for the modern Internet by establishing
NSFNET. Commercial restrictions on the use of the
Internet were lifted in 1991, and the first U.S. site
on the World-Wide Web was launched in December
1991.9 Use of the Internet grew very rapidly after
1. See Berg and Hudson (1992) and Crafts and Harley (1992) for two very different views of the industrial revolution.
2. See, for example, Bound and Johnson (1992); Juhn, Murphy, and Pierce (1993); Levy and Murnane (1992); and Katz and
Murphy (1992).
3. See, for example, Bound and Johnson (1992); Berman, Bound, and Griliches (1994); and Autor, Katz, and Krueger (1998).
For a more complete discussion, see the longer version in Card and DiNardo (2002).
4. This model can be easily extended to include capital or other inputs provided that labor inputs are separable and enter the
aggregate production function through a subproduction function like equation 1.
5. See Katz and Murphy (1992) and Autor and Katz (1999).
6. Note that it is necessary to assume σ > 1 in order for a rise in gH relative to gL to increase the relative wage of skilled workers.
The distinction between the four parameters (A, α, gH, gL) is somewhat artificial because one can always rewrite the production function as Y = [cH NH(σ–1)/σ + cLNL(σ–1)/σ]σ/(σ–1) by suitable definition of the constants cH and cL. The relevant question is
how the pair (cH, cL) evolves over time.
7. This distinction is emphasized by Bresnahan (1999) and Bresnahan, Brynjolfsson, and Hitt (2002).
8. See Hobbes’ Internet Timeline at www.zakon.org/robert/internet/timeline.
9. The World-Wide Web was invented at CERN (the European Laboratory for Particle Physics) in 1989–90.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
47
the introduction of Netscape’s Navigator program in
1994: The number of Internet hosts rose from about
one million in 1992, to twenty million in 1997, and
to one hundred million in 2000.
Qualitative information on the pace of technological change is potentially helpful in drawing connections between specific innovations and changes in
wage inequality. For example, the sharp rise in wage
inequality between 1980 and 1985 (discussed below)
points to technological innovations that occurred
very early in the computer revolution (around the
time of the original IBM-PC) as the key skill-biased
events. By comparison, innovations associated with
the growth of the Internet presumably had very lim-
While some of the early rise in inequality may
have been due to rapid technological change,
we suspect that the increase in the early 1980s
is largely explained by other plausible—albeit
relatively mundane—factors.
ited impact until the mid-1990s. Nevertheless, comparisons of relative timing are subject to substantial
leeway in interpretation, depending on lags in the
adoption of new technologies.
An alternative approach is to attempt to quantify
recent technological changes by measuring the relative size of the information technology (IT) sector
in the overall economy. One such measure, taken
from Jorgenson (2001), is plotted in Chart 1. Notwithstanding the obvious difficulties with the interpretation of such a simplified measure by a fairly
broad measure—IT output as a percentage of total
gross domestic product—information technology
has grown steadily in importance since 1948, with
sustained growth over the past two decades and a
pronounced upsurge in the late 1990s.10 The rapid
expansion of the IT sector in the late 1990s has
attracted much attention, in part because aggregate
productivity growth rates also surged between 1995
and 2000. Many analysts (including Basu, Fernald,
and Shapiro 2001) have argued that this was the
result of an intensive burst of technological change
in the mid- to late 1990s.
A third approach, pioneered by Krueger (1993),
is to measure the pace of computer-related technological change by the fraction of workers who use a
computer on the job. The thin black line in Chart 1
plots the overall fraction of workers who reported
48
using a computer in 1984, 1989, 1993, and 1997.11
Rates of on-the-job computer use, like the IT output
share, show substantial growth over the past two
decades—from 25 percent in 1984, to 37 percent in
1989, and to 50 percent in 1997. Nevertheless, the
fact that one-quarter of workers were using computers on the job in 1984 suggests that some of the
impact of computerization on the workforce preceded the diffusion of personal computers. Indeed,
Bresnahan (1999) has estimated that as early as 1971
one-third of U.S. workers were employed in establishments with mainframe computer access. Specialized
word-processing machines that predated the personal computer were also widely in use in the early
1980s. The absence of systematic data prior to 1984
makes it hard to know whether computer use
expanded more quickly in the early 1980s than in
the late 1970s or the late 1980s, which in turn makes
it difficult to compare changes in the rate of computer
use with changes in wage inequality, especially in
the critical early years of the 1980s.12
While none of the available indicators of technological change is ideal, all of the indicators suggest
that IT-related technological change has been going
on since at least the 1970s and has continued
throughout the 1980s and 1990s. Moreover, some
evidence (based on the size of the IT sector, the
pace of innovations associated with the Internet,
and aggregate productivity growth) suggests that
the rate of technological change accelerated in the
1990s relative to the 1980s.
Whose productivity was raised by recent
changes in technology? The second task in developing an empirically testable version of the SBTC
hypothesis is to specify which skill groups have
their relative productivity raised by SBTC. There
are two main approaches to this issue. The first,
articulated by Autor, Katz, and Krueger (1998), is
to assume that groups that are more likely to use
computers have skills that are more complementary
with computers and experience bigger gains in productivity with continuing innovations in computer
technology.13 We refer to this as the “computeruse/skill-complementarity” view of SBTC. An alternative, advanced by Juhn, Murphy, and Pierce (1991,
1993), is to assume that recent technological changes
have raised the relative productivity of more highly
skilled workers along every dimension of skill,
leading to an expansion of the wage differentials
between groups.14 We refer to this as the “risingskill-price” hypothesis. As it turns out, the two
approaches yield similar implications for comparisons across some dimensions of the wage structure
but different implications for others. Throughout,
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
we will refer to either version or both versions of
SBTC as appropriate.
To set the stage, the table on page 50 shows patterns of relative computer use on the job by different skill groups in 1984, 1989, 1993, and 1997. Rates
of computer use tend to be higher for those with
more schooling. High-school graduates are three to
four times more likely to use computers on the job
than dropouts, and college graduates are about
twice as likely to use computers as those with only a
high-school diploma. Interestingly, although overall
computer use rates have risen, the relative usage
rates of different education groups have remained
fairly stable. Since the wage differentials between
education groups are bigger today than at the start
of the 1970s, this fact would appear to be consistent
with the computer-use/skill-complementarity view.
Moreover, since better-educated workers earn higher
wages, an increase in the wage differential between
the highly and less highly educated is also consistent with the rising-skill-price view of SBTC.
The data in the table also show that women are
more likely to use computers at work than men, and
blacks are less likely to use computers than whites.
Although the gender and race gaps closed slightly in
the early 1990s, the male-female and black-white
gaps remain relatively large. To the extent that
complementarity with computer-based technologies
is measured by computer use rates, these patterns
suggest that recent technological changes should
have led to upward pressure on women’s wages relative to men’s and downward pressure on black
workers’ wages relative to whites’. In the case of the
race differential, the relative wage approach to
gauging the impact of SBTC leads to a similar conclusion.15 In the case of the gender differential,
however, the two methods are inconsistent. Women
earn less than men and, as with the racial wage gap,
part of the gender gap is usually attributed to differences in unobserved skills. Thus, the argument
that recent technological changes have raised the
relative productivity of more highly paid workers—
the rising-skill-price view of SBTC—suggests that
computer technology should have led to a widening
of the male-female wage gap.
Simple tabulations of computer use rates by education and gender hide an important interaction
between these two factors, however. The education
gradient in computer use is much bigger for men
than women while differences in computer use by
gender are much smaller for better-educated workers. Indeed, as shown in the table, college-educated
men are more likely to use a computer than collegeeducated women. To the extent that computer use
indexes the relative degree of complementarity
with new technology, as assumed by the computeruse/skill-complementarity version of SBTC, computer
technology should have widened gender differentials for the most highly educated and narrowed
them for the least educated. By contrast, since men
earn more than women at all educational levels, the
rising-skill-price view of SBTC suggests that the
gender gap should have expanded at all educational
levels. Although the data are not reported in the
table, we have also examined the interactions
between gender and race. Compared to the interaction between education and gender, however, the
race-gender interactions are relatively modest.
Finally, an examination of computer use rates by
age suggests that computer use has expanded
slightly faster for older workers than for younger
workers. As shown in more detail in Card and DiNardo
(2002), computer use rates in the early 1980s were
declining slightly with age. By the late 1990s, however, the age profile of computer use was rising
slightly between the ages of twenty and forty-five
and declining after age fifty. These observations suggest another divergence between the two versions
of SBTC. Based on the age profiles of computer use,
SBTC may have led to a reduction in older workers’
10. See Oliner and Sichel (2000) and Gordon (2000) for interesting discussions of some of these issues.
11. These data are based on responses to questions in the October Current Population Surveys for workers estimated to be out
of school. See Card and DiNardo (2002) for details.
12. Card and DiNardo (2002), using data from the Information Technology Industry (ITI) Council on annual shipments of different types of computers since 1975, find that series constructed from this data show fairly steady growth in shipments
from 1975 to 1984.
13. Note that this hypothesis does not necessarily imply that individuals who use computers will be paid more or less than people in the same skill group who do not.
14. To be slightly more formal, assume that the log of the real wage of individual i in period t (wi t) is a linear function of a single index of individual ability ai = xiβ + ui , where xi is a set of observed characteristics and ui represents unobserved characteristics. Then log(wit) = ptai = xi(ptβ) + ptui, where pt is the economywide “price” of skill. Skill-biased technological
change in the rising-skill-price view is merely an increase over time in pt.
15. Juhn, Murphy, and Pierce (1991) argued that blacks tend to have lower levels of unobserved ability characteristics and that
rising returns to these characteristics held down relative wages for blacks in the 1980s.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
49
TABLE
Use of Computers at Work (Percent)
1984
1989
1993
1997
All workers
24.5
36.8
46.0
49.9
By education
Dropouts
High school
Some college
College (or more)
4.8
19.8
31.9
41.5
7.4
29.2
46.4
57.9
8.9
34.0
53.5
69.1
11.3
36.1
56.3
75.2
47.7
50.5
49.1
48.1
21.1
29.0
31.6
43.2
40.3
52.7
44.1
56.7
73.0
73.2
76.5
77.8
12.9
42.7
27.5
39.6
20.1
58.8
39.2
56.6
24.1
70.5
45.1
67.4
26.8
75.5
46.8
74.7
30.2
69.4
34.2
69.3
34.2
66.9
35.5
62.7
46.9
107.8
51.3
103.9
53.4
104.5
58.3
101.1
25.3
18.2
23.7
37.9
27.2
36.0
47.3
36.2
42.3
51.3
39.9
48.2
Black/white
72.1
71.7
76.7
77.7
By age
Under 30
30–39
40–49
50 and older
24.7
29.5
24.6
17.6
34.9
42.0
40.6
27.6
41.4
50.5
51.3
38.6
44.5
53.8
54.9
45.3
High school/college
By gender
Men
Women
Male/female
By gender and education
High-school men
College men
High-school women
College women
High school/college (men)
High school/college (women)
Male/female (high school)
Male/female (college)
By race
Whites
Blacks
Other
Notes: Entries display percentage of employed individuals who answer that they “directly use a computer at work” in the October Current
Population Survey (CPS) Computer Use Supplements. Samples include all workers with at least one year of potential experience. College
workers include those with a college degree or higher education. All tabulations are weighted by CPS sample weights.
relative wages. On the other hand, since older workers
earn more than younger workers, the rising-skill-price
view of SBTC predicts a rise in age- or experiencerelated wage premiums over the 1980s and 1990s.
In what follows, we briefly review some important changes in wage inequality and in the wage
structure. Throughout we will discuss both problems and puzzles for SBTC. The problems are facts
that seem superficially inconsistent with both (either)
version of the theory; the puzzles are important
50
developments in the wage structure that are potentially consistent with SBTC but appear to be driven
by other causes.
Trends in Overall Wage Inequality
lthough measurement of wage inequality is
substantially more straightforward than the
measurement of technological change, there are a
number of potentially important issues. The Current
Population Survey (CPS) that is the most widely
A
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 2
Alternative Measures of Aggregate Wage Inequality
Sta n d a rd d e via t io n lo g wa g e s ( o r e q u iva le n t )
.65
.60
Standard deviation log hourly
wages, all workers (March)
.55
Standard deviation log annual
earnings, FTFY men (March)
Normalized 90-10 wage
gap, all workers (OGR)
.50
.45
1967
1970
1973
1976
1979
1982
1985
1988
1991
1994
1997
2000
Source: Authors’ analysis of March, May, and Outgoing Rotation Group files from monthly Current Population Survey and EPI
used source for data on individual wages (and the
source we use here), for instance, experienced a substantial redesign in the mid-1990s that appears to
have raised measured inequality. Nonetheless, our
most important findings appear robust to choice of
data sets and a variety of different methodologies
for the measurement of inequality.16
Chart 2 plots three different measures of aggregate wage dispersion. The first is the standard deviation of log annual earnings for full-time full-year
(FTFY) male workers, constructed from March CPS
data from 1968 to 2001.17 The second is the normalized 90-10 log wage gap in hourly earnings, based on
the May CPS files for 1973 to 1978 and the OGR files
from 1979 onward. This series is based on estimates
constructed by the Economic Policy Institute (EPI),
using procedures very similar to ours.18 The third is
the standard deviation of log hourly wages for all
workers in the March CPS files from 1976 to 2001,
weighted by the hours worked in the previous year.
An examination of the chart suggests that the
recent history of U.S. wage inequality can be divided
into three episodes. During the late 1960s and 1970s,
aggregate wage inequality was relatively constant.
The standard deviation of log wages for FTFY men
rose by only 0.01 between 1967 and 1980 (from 0.51
to 0.52).19 Wage inequality measures from the May
CPS/OGR series also show relative stability (or even
a slight decline) between 1973 and 1980 while the
hours-weighted standard deviation of log hourly
wages for all workers in the March CPS was stable
from 1975 to 1980. The 1980s was a period of
expanding inequality, with most of the rise occurring early in the decade. Among FTFY men, for
example, 85 percent of the 10-point rise in the standard deviation of log wages between 1980 and 1989
occurred before 1985. Finally, in the late 1980s
wage inequality appears to have stabilized. Indeed,
none of the three series in Chart 2 shows a noticeable change in inequality between 1988 and 2000.
16. Card and DiNardo (2002) discuss at length issues of measurement and the robustness of the findings to alternative data
sources and measurement methodologies.
17. Here and in what follows, we refer to the data derived from the March supplement to the CPS as the March CPS and the
data from the Outgoing Rotation Group files and the 1973–78 May supplements as the OGR and May CPS data, respectively.
See Card and DiNardo (2002) for more details.
18. See Mishel, Bernstein, and Schmitt (2001, table 2.17). For details on this and all other aspects of the data, see Card and
DiNardo (2002).
19. Similarly, the standard deviation of log wages for all full-time workers (men and women) was slightly lower in 1980 than in 1967.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
51
CHART 3
College/High-School Wage Ratio by Gender, 1975–99
Colle g e /h ig h -s c h o o l lo g wa g e r a t io
.55
.50
Women
.45
.40
.35
Men
.30
.25
1975
1978
1981
1984
1987
1990
1993
1996
1999
Source: Authors’ analysis of March Current Population Survey, various years
The apparent stability of aggregate wage inequality over the 1990s presents a potentially important
puzzle for the SBTC hypothesis, since there were
continuing advances in computer-related technology
throughout the decade that were arguably as skill
biased as the innovations in the early 1980s.
Another interesting feature of the series in Chart 2
is that the rise in wage inequality over the 1980s was
larger for FTFY men than for workers as a whole.
While the reasons for this are unclear, if viewed with
an eye toward SBTC, the relative rise in inequality
for FTFY men is a puzzle. To the extent that SBTC
tends to widen inequality across skill and ability
groups, we would expect to see a larger rise in
inequality for less homogeneous samples (for example, pooled samples of men and women and fulland part-time workers) and a smaller rise for
more homogeneous samples (such as FTFY men).
The data suggest the opposite.20
Components of the Wage Structure
R
eturns to college. We now shift our focus to
specific dimensions of the wage structure. We
begin with wage differences by education, which
are at the core of the SBTC hypothesis. Chart 3 presents estimates of the college/high-school wage gap
by gender for the 1975–99 period, based on average
hourly earnings data from the March CPS.21 Trends
in the college/high-school gap are similar to the
52
trends in overall inequality and suggest three distinct episodes: the 1970s, when the college gap was
declining slightly; the 1980s, when the gap rose
quickly; and the 1990s, when the gap was stable or
rising slightly. For both men and women, the college/
high-school wage gap rose by about 0.15 log points
between 1980 and 1990. The rise for men was concentrated in the 1980–85 period while for women it
was more evenly distributed over the decade. The
similar overall rise in returns to college for men and
women is interesting, however, because as noted
earlier there is a much larger education gradient in
computer use rates for men than women. Based on
this fact, the computer-complementarity version of
the SBTC hypothesis would predict a larger rise in
the college/high-school wage gap for men than for
women during the 1980s and 1990s. On the other
hand, since the college/high-school wage gaps are
similar for men and women, the skill-price version
of SBTC predicts about the same rise in returns for
both. Thus, the similarity of the rise in the college
gap for men and women is a puzzle for one version
of the theory but not for the other.
Some previous authors have argued that variation in the college/high-school wage premium can
be explained by a model like equation 2, with the
added assumption that the effect of changing technology follows a smooth trend (see, for example,
Freeman 1975 and Katz and Murphy 1992). In these
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 4
Relative Supply of College-Educated Labor
–0.2
L o g r e la t ive s u p p ly in d e x
–0.4
–0.6
Log of college/high-school equivalents
–0.8
Fitted with trend break in 1982
–1.0
–1.2
–1.4
1967
1970
1973
1976
1979
1982
1985
1988
1991
1994
1997
2000
Source: Authors’ analysis of the March Current Population Survey, various years. See Card and DiNardo (2002) for details.
studies the relative supply of college workers is estimated by assigning various fractions of “collegeequivalent” and “high-school-equivalent” labor units
to workers in different education categories.22
Using a variant of this method, we derive such a
supply index, which is displayed in Chart 4. A feature of this index—which is revealing about a
potential problem with the SBTC hypothesis—is
that it follows a roughly constant trend between
1967 and 1982 (4.5 percent per year) and a slower
but again nearly constant trend after 1982 (2.0 percent per year).23 Assuming that 1/σ is positive, shifting trends in relative supply can potentially explain
an upward shift in the rate of growth of the college/
high-school wage gap in the early 1980s but not the
slowdown in the 1990s.
The problem is further revealed by comparing
estimates of models based on equation 2 that exclude
or include the 1990s. For example, augmenting the
model with a trend shift term that allows for a possible acceleration in SBTC after 1980, the estimate
of the relative supply term becomes wrong-signed,
and the model substantially overpredicts returns to
college in the late 1990s.24 We conclude that the
slowdown in the rate of growth in the return to
college in the 1990s is a problem for the SBTC
hypothesis that cannot be easily reconciled by shifts
in relative supply.
20. As we explain in the longer version of this paper, although we prefer to measure aggregate wage inequality using the broadest possible sample of workers, the tradition in the inequality literature has been to analyze men and women separately
(although Lee 1999 and Fortin and Lemieux 2000 are important counterexamples). Treating men and women separately,
however, yields substantially the same conclusions for men. For women the trends in inequality are a little different although
they pose essentially the same problems for SBTC. For instance, whether the OGR or March data are used, it is clear that
most of the rise in gender-specific wage inequality, like the rise in overall inequality, was concentrated in the first half of the
1980s, with surprisingly little change in the 1990s.
21. These estimates are obtained from regression models fit separately by gender and year to samples of people with either
twelve or sixteen years of education. The models include a dummy for college education, a cubic in years of potential experience, and a dummy for nonwhite race.
22. For example, a worker with fourteen years of education contributes one-half unit of college labor and one-half unit of highschool labor while a worker with ten years of education contributes something less than one unit of high-school labor.
23. Indeed, a regression of the supply index on a linear trend and post-1982 trend interaction yields an R 2 of 0.997.
24. See Card and DiNardo (2002.) Beaudry and Green (2002) experiment with several variants of equation 2 and report similar findings.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
53
CHART 5
Changing Age Structure of the College/High-School Wage Gap
0.5
C o lle g e /h ig h -s c h o o l wa g e g a p
1994–96
0.4
1989–91
1984–86
0.3
1979–81
0.2
Average of 1959, 1969–70, 1974–76
0.1
26–30
31–35
36–40
41–45
46–50
51–55
56–60
Source: Card and Lemieux (2001)
Education and age. So far we have focused on
the average difference in wages between college and
high-school workers in all age groups. This focus arises
naturally out of a model such as the one described
by equations 1 and 2, where there are only two skill
groups—high and low education—and workers with
different years of labor market experience are treated
as perfect substitutes. In such a model, there is a
unique “return to education” in the economy as a
whole at any point in time. Moreover, the focus on
average returns to college is descriptively adequate
whenever the wage differentials between education
groups are the same for people with different ages or
different years of experience, as in Mincer’s (1974)
human capital earnings function.25
While the rise in the average wage gap between
college and high-school workers has been extensively
documented, the fact that the increases have been
very different for different age groups is less well
known. Specifically, the rise in the college/high-school
wage gap for men is most pronounced among young
workers entering the labor force after the late 1970s.
Moreover, the pattern of this increase does not appear
to be well explained by either the rising-skill-price or
computer-use/skill complementarity versions of SBTC.
One assumption embedded in equation 1 is that
workers with similar educations but different ages are
perfect substitutes in production. Card and Lemieux
(2001) show that one implication of a more general
54
model that allows for imperfect substitution across
age groups is the presence of cohort effects in the
returns structure. Because education is (essentially)
fixed once a cohort enters the labor market, a cohort
with fewer highly educated workers will experience
higher relative returns at each age, leading to cohortspecific deviations from the average pattern. Evidence
of such cohort effects is presented in Chart 5 (taken
from Card and Lemieux 2001), which shows the age
profiles of the college/high-school wage gap for fiveyear age cohorts of men in five periods: 1960–76
(based on pooled data from the 1960 Census and
early CPS surveys), 1979–81, 1984–86, 1989–91, and
1994–96. In the 1960s and early 1970s the college/
high-school wage gap was an increasing and slightly
concave function of age, consistent with the functional form posited by Mincer (1974). Subsequent
changes in the age structure of the college/highschool gap, however, reveal a “twisting” of the age
profile—large increases in the gap at relatively young
ages during the mid-1980s to the mid-1990s and
relatively small changes in the gap for older men.
Based on the data in Chart 5 and a series of formal
statistical tests, Card and Lemieux (2001) argue that
the trends in the college/high-school wage gap for
different age groups reflect systematically higher
college/high-school wage premiums received by successive cohorts that have entered the labor market
since the late 1970s. Moreover, these cohort effects
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 6
Mean Salary Offer Relative to Humanities/Social Sciences
170
R e la t ive m e a n s a la r y o ffe r
Electrical engineering
150
Computer science
Mathematics
130
Chemistry
110
Business/accounting
90
1963
1966
1970
1974
1978
1981
1985
1989
1993
Source: Authors’ analysis of data from National Association of Colleges and Employers
are highly correlated with cohort-specific changes
in the relative supply of college workers. Somewhat
surprisingly, after controlling for cohort-specific supplies, they find that the return to education was about
the same in the mid-1990s as it had been in the mid1970s. This interpretation of the data leaves little or
no room for accelerating technical change; while
one could argue that the spread of computers led to
cohort-specific relative productivity gains for collegeeducated workers, there is no direct evidence of such
a phenomenon. Moreover, the age profiles of the
college/high-school gap in computer use shifted uniformly between 1984 and 1997, rather than twisting
like the returns profiles in Chart 5.
Returns to different college degrees. One
concern with evidence for SBTC based on overall
wage differences between college and high-school
workers is that computer-related technology may
have had different effects on college graduates from
different fields of study. In particular, it seems plausible that the computer revolution would lead to a
rise in the relative demand for college graduates
with more “technical” skills (like engineers and sci-
entists), especially in the early 1980s when microcomputers were first introduced and the college/
high-school wage gap was expanding rapidly. Chart 6
displays mean starting salaries offered to graduating
students with bachelors degrees in various fields,
compiled from a survey of career placement offices
conducted by the National Association of Colleges
and Employers, and brings some evidence to bear
on this possibility.26 For convenience, we have scaled
the data to show mean salaries relative to humanities and social sciences. The most obvious feature
of the data is that the relative salaries in more technical fields rose in the 1970s and fell in the 1980s.
This pattern is particularly true for the relative
salaries in the two fields most closely connected with
computers: computer science and electrical engineering. Paradoxically, the introduction of microcomputers was associated with a fall in the relative
salaries of specialized college graduates with the
strongest computer skills. Although the data in
Chart 6 cover only the period up to 1993, more
recent data suggest that in the late 1990s the relative salaries of electrical engineering and computer
25. The simplest way to justify Mincer’s formulation within the framework of the model in equation 1 is to assume that the relative efficiency units of different age groups depend only on experience (for example, age minus education) and that the
relative efficiency profile is the same for college and high-school labor.
26. An alternative data source on college graduates’ relative salaries in different fields, the Recent College Graduates Survey
(which is available only since 1977), shows similar patterns. See U.S. Department of Education (1998, supp. table 33-1).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
55
CHART 7
Male-Female Wage Gaps
0.55
M a le -fe m a le m e a n lo g wa g e
Annual earnings, FTFY workers (March)
0.45
Hourly earnings,
all workers (March)
0.35
Hourly earnings, all workers (OGR)
0.25
0.15
1967
1970
1973
1976
1979
1985
1982
1988
1991
1994
1997
2000
Source: Authors’ analysis of March and May Current Population Survey, various years, and Outgoing Rotation Group files from monthly
Current Population Survey, various years
science graduates rose back to the levels of the late
1970s. Thus, the IT-sector boom in the late 1990s was
associated with a rise in relative wages of graduates
with computer-related skills.
We regard the trends in the relative salaries of
college graduates in different fields as at least a
puzzle, if not a problem, for the SBTC hypothesis.
While innovations in computer technology do not
necessarily raise the relative demand for workers
with the most specialized computer training, engineers and computer scientists have very high rates
of computer use and also earn higher wages than
other bachelor degree holders. Thus, the decline in
the wage premium for engineers and computer science graduates over the 1980s is inconsistent with
either the computer-use/skill-complementarity or
rising-skill-price versions of the SBTC hypothesis.
Other Changes in the Structure of Wages
T
he male-female wage gap. One of the most
prominent changes in the U.S. wage structure
is the recent closing of the male-female gap. Chart 7
displays three estimates of the gap in wages between
men and women: the difference in mean log annual
earnings of full-time/full-year workers (based on
March CPS data); the difference in mean log average hourly earnings from the March CPS (for 1975
and later); and the difference in mean log average
hourly earnings from the OGR supplements (for
56
1979 and later). Like overall inequality and returns
to college, trends in the male-female wage gap seem
to fall into three distinct episodes. During the 1970s,
the gender gap was relatively stable. During the
1980s and early 1990s the gap fell. Finally, in the midto late-1990s the gap was stable again. Although the
different wage series give somewhat different estimates of the size of the gender gap, all three show a
15 percentage point decline between 1980 and 1992.
Moreover, these trends are very similar for different
age and education groups.
These trends, and their similarity for different
age and education groups, pose a number of problems and puzzles for different versions of the SBTC
hypothesis. As noted earlier, the closing of the gender gap in the 1980s is a particular problem for the
rising-skill-price version of SBTC, which predicts that
technological change raises the return to all different kinds of skills, including the unobserved skills
that are usually hypothesized to explain the gender
gap. Since women use computers on the job more
than men, some observers have argued that the
decline in the gender wage gap is consistent with
the computer-use/skill-complementarity version of
SBTC.27 This theory cannot explain the similarity of
the trends in the gender gap for high-school and
college graduates, however, since college-educated
women are actually less likely to use a computer
than college-educated men. Thus, like Blau and Kahn
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 8
White-Black Wage Gaps
0.30
Wh it e -b la c k m e a n lo g wa g e
0.25
FTFY men (March)
0.20
All men (OGR)
0.15
FTFY women (March)
0.10
0.05
All women (OGR)
0
1967
1970
1973
1976
1979
1982
1985
1988
1991
1994
1997
2000
Source: Authors’ analysis of March and May Current Population Survey, various years, and Outgoing Rotation Group files from monthly
Current Population Survey, various years
(1997), we conclude that the rise in women’s wages
relative to men’s wages over the 1980s must be
attributed to gender-specific factors.
The black-white wage gap. Chart 8 shows the
evolution of another important dimension of wage
inequality—the difference in wages between white
and black workers. The chart shows the wage gaps
for full-time/full-year men and women (derived from
March CPS data) and for all men and women (based
on average hourly earnings from the OGR data).
The gaps for women are similar whether the data
are confined to FTFY workers or not while the gaps
for men are slightly different between FTFY workers
and all workers, at least in the early 1980s. As previous studies have documented, racial wage gaps
are also much smaller for women than for men. More
interesting from our perspective are the trends in
the racial wage gap, which are quite different from
the trends in other dimensions of inequality. During
the 1970s, when the gender gap and overall wage
inequality were relatively stable, the wage advantage
of white workers fell sharply: from 28 to 18 percent
for men and from 18 percent to 4 percent for women.
During the 1980s, when overall wage inequality was
rising and the gender gap was closing, the black-
white wage gap was relatively stable. Finally, over the
1990s, racial wage gaps were roughly constant. The
gaps for high-school and college-educated men and
women are similar to the corresponding gaps for all
education groups and follow roughly similar trends.
Like the gender wage gap, we view the evolution of
racial wage differences as at least a puzzle, and potentially a problem, for SBTC. Both the rising-skill-price
view and the computer-use/skill-complementarity
view suggest that SBTC should have led to a widening of racial wage gaps in the 1980s. The gap in computer use between blacks and whites is about the
same magnitude as the male-female gap, so the same
arguments that have been made about the effect of
computerization on male-female wage differences
would seem to apply to race. Indeed, Hamilton (1997)
argues that a computer skills gap contributed to an
increase in the wage differentials between whites and
blacks. In view of the data in Chart 8, however, it is
clear that other factors must have worked in the
opposite direction to offset any such effects of SBTC.
Work experience. Along with education, gender,
and race, a fourth key dimension of wage inequality
in the U.S. labor market is age. Following Mincer
(1974), most labor market analysts have adopted the
27. For example, Weinberg (2000) argues that “since computer [jobs] are likely to be less physically demanding than the average noncomputer job, the elimination of noncomputer jobs in which men have a comparative advantage and the creation of
computer jobs in which women have a comparative advantage would tend to favor women.”
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
57
assumption that log wages are a separable function
of education and potential labor market experience
(age minus education minus 6): In this framework,
if there is an increase in the return to skill caused
by changes in technology, we should expect the
return to an additional year of experience to rise. As
shown in more detail in Card and DiNardo (2002),
here too the evidence is not favorable for SBTC. For
example, there is little evidence of either a rise or
fall in the average return to experience over the
period 1979 to 1991 for high-school-educated men,
who make up about one-third of all male workers.
Much the same is true for younger college-educated
men (those between the ages of twenty-four and
thirty-seven who have two to fifteen years of potential experience), and for college-educated men in the
middle range of experience, the wage profile actually
became flatter over the 1980s and 1990s.
Given this analysis of the change (or lack of
change) in men’s experience profile, it is difficult to
rationalize the somewhat different evolution of the
potential experience profile for women in an SBTC
framework. During the period 1979–91, experience
profiles did become somewhat steeper for both
high-school- and college-educated women, particularly for women with between two and eighteen
years of potential experience. Manning (2001) has
shown that for women in the United Kingdom, where
the male-female wage gap also closed substantially
over the 1980s, a similar increase in the returns to
experience can be in part explained by a shift across
cohorts in the fraction of time spent working.28 How
far such an analysis could go toward explaining the
these shifts in the experience profile is an interesting question. In any case, we suspect that SBTC has
little to do with the story.
Residual inequality. SBTC has also been proposed as an explanation for the rise in inequality
among workers with similar observable characteristics. To the extent that wage differences between
workers with the same education, age, gender, and
race reflect the labor market’s valuation of unmeasured productivity, the rising-skill-price version
of SBTC predicts a rise in the residual variance
associated with a standard human capital model of
wage determination while the prediction from the
computer-complimentarity version of SBTC is
unclear. As documented in Card and DiNardo (2002),
however, the trends in residual inequality pose
much the same difficulties as the trends in overall
inequality that we have documented here. In particular, we find that most of the modest rise in residual inequality was concentrated in the early 1980s,
which suggests that if SBTC is the cause for this
58
change, it occurred during the earliest years of the
microcomputer revolution.
SBTC and Productivity
final issue worth discussing is the relationship
between SBTC and productivity growth. Many
analysts have noted that the pace of aggregate productivity growth was stable during the 1980s and
early 1990s despite the introduction of computers and
the almost immediate effect that computerization is
presumed to have had on wage inequality.29 To illustrate, Chart 9 plots the log of real output per hour in
the nonfarm business sector of the United States over
the 1947–2000 period, along with a fitted trend line
that allows a productivity slowdown after 1975.30 The
rate of labor productivity growth during the 1980s
and early 1990s was substantially slower than in
1947–75. However, between 1979 and 1986, when
aggregate wage inequality was expanding rapidly,
productivity first fell relative to trend (during the 1980
and 1982–83 recessions), then recovered to its earlier
trend level. There is no indication that developments
in the early 1980s led to an unexpected change in the
productive capacity of the economy.
We regard the absence of a link between SBTC
(as measured by the rate of increase in wage inequality) and aggregate productivity growth as a puzzle,
although not necessarily a problem, for SBTC. While
some theoretical discussions of technological change
assume that any new technology leads to an outward
shift in the economywide production frontier, some
specific versions of SBTC do not. Extensive SBTC—
a rise in the share parameter α in the aggregate
production function given in equation 1 that would
raise the productivity of some workers and lower
that of others—would be consistent with rising wage
inequality but not necessarily raising aggregate
labor productivity. Nonetheless, it is rather surprising that whatever shifts in technology led to the
rapid growth in inequality between 1980 and 1985
appeared to have no effect on the trend in aggregate productivity.
In comparison to the early 1980s, the late 1990s
may turn out to be a better example of a period of
rapid technologically driven output growth. As shown
in Chart 9, aggregate output growth was considerably above trend in the 1998–2000 period. Moreover,
some (but not all) measures of wage inequality show
a rise after 1995 or 1996. Some detailed microlevel
analyses point to specific technology-related changes
in workplace organization that have a significant
impact on productivity (see, for example, Bresnahan,
Brynjolfsson, and Hitt 2002). In view of the confounding effect of the extraordinary business cycle
A
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 9
Trends in Productivity per Hour, Nonfarm Business Sector
1979–86
4.9
L o g o f p r o d u c t ivit y p e r h o u r
4.7
4.5
Productivity
4.3
4.1
Fitted trend, with post-1975 trend break
3.9
3.7
1947
1951
1955
1959
1963
1967
1971
1975
1979
1983
1987
1991
1995
1999
Source: Authors’ analysis of Bureau of Labor Statistics data (see footnote 30)
conditions during the late 1990s, however, it may be
some time before a definitive interpretation of this
period is reached.
Conclusion
hat is one to make of recent trends in wage
inequality and productivity and the links (or
absence of links) to computer-related technology?
From the vantage point of an analyst looking at the
available data in the mid- to late 1980s, there were
many reasons to find skill-biased technological
change a plausible explanation for the large
increase in inequality that began in the early 1980s.
First and foremost, the timing seemed right. During
the 1970s, the college/high-school wage gap narrowed. Richard Freeman’s 1976 book The Overeducated American argued that the U.S. labor
market suffered from an oversupply of educated
workers. By 1985 the situation had clearly reversed,
and education-related wage gaps and other dimen-
W
sions of wage inequality were on the rise. At the
same time, the personal computer was making dramatic inroads into the workplace, the stock market
valuation of technology firms was rising, and articles in the business press were expounding the
effects of the new technology. Analysts in the late
1980s had no way of knowing that, although computer use would continue to expand over the next
decade and the stock market value of technology
firms would rise, the increase in wage inequality
was largely over.
Viewed from 2002, the rise in wage inequality
now appears to have been an episodic event. Of the
17 percent rise in the 90-10 wage gap between 1979
and 1999 for all workers in the OGR wage series
(see Chart 2), 13 percentage points (or 76 percent)
occurred by 1984, the year that the IBM-AT was
introduced. While some of the early rise in inequality
may have been due to rapid technological change,
we suspect that the increase in the early 1980s is
28. To the extent that the measured potential experience profile reflects the relationship between wages and actual experience,
for example, such a shift in the labor force participation rates of women would cause a steepening of the wage/potential
experience profile.
29. For example, in a 1996 statement Alan Greenspan observed that “the advent of the semiconductor, the microprocessor, the
computer, and the satellite . . . has puzzled many of us in that the growth of output as customarily measured has not evidenced a corresponding pickup” (quoted in McGuckin, Stiroh, and van Ark 1997).
30. The productivity series is series PRS85006093, from the U.S. Bureau of Labor Statistics, uploaded December 2001. The fitted
trend in the log of output per hour is 0.0262 in the period 1947–75 and 0.0139 in the period 1976–2000.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
59
CHART 10
Real Minimum Wage, 1973–2000
7.0
R e a l m in im u m wa g e ( 2000 $)
6.5
6.0
5.5
5.0
4.5
4.0
1973
1976
1979
1982
1985
1988
1991
1994
1997
2000
Source: Authors’ analysis. See also <www.dol.gov/esa/minwage/chart.htm>.
largely explained by other plausible—albeit relatively
mundane—factors. A primary candidate is the fall
in the real value of the minimum wage. In 1979 the
modal wage for women with a high-school education
was $2.90 an hour—the level of the federal minimum wage (DiNardo, Fortin, and Lemieux 1996).
Over the next five years the consumer price index
rose by 48 percent while the minimum wage increased
by only 15 percent, leading to a steep decline in the
influence of the minimum wage on the lower tail of
the wage distribution. Chart 10 plots the real value
of the federal minimum wage between 1973 and
2000. Examination of this figure suggests that it is
nearly a mirror image of the inequality series in
Chart 2. Indeed, as shown in Chart 11, predictions
from a simple regression of the normalized 90-10
wage gap (from the May CPS and OGR data) on the
log of the real minimum wage track the actual wage
gap very closely. This simple model explains over
90 percent of the variation in the 90-10 wage gap
and captures many of the key turning points.
Of course, neither this informal analysis nor the
more exhaustive study by Lee (1999) imply that the
minimum wage can explain all the changes in the
wage structure that occurred in the 1980s and
1990s. Indeed, we have documented several important changes that cannot be explained by the minimum wage, including the closing of the gender gap.31
Nevertheless, we suspect that trends in the minimum wage and other factors such as declining unionization and the reallocation of labor caused by the
1982 recession can help to explain the rapid rise in
overall wage inequality in the early 1980s.
Overall, the evidence linking rising wage inequality to skill-biased technological change is surprisingly
weak. Moreover, we conjecture that a narrow focus
on technology has diverted attention away from many
interesting developments in the wage structure that
cannot be easily explained by skill-biased technological change. Perhaps the perspective of a new
decade will help to open the field of unexplained
variance to all players.
31. Lee (1999) presents a detailed cross-state evaluation of the effect of the minimum wage on overall wage inequality and concludes that the fall in the real minimum wage can explain nearly all the rise in aggregate inequality in the 1980s. That the
minimum wage explains most of the change in overall inequality, but cannot explain specific changes in the wage structure,
is not as puzzling as it might first appear. Fortin and Lemieux (1998, 2000) show that although the 1980s saw very large
increases in gender-specific wage inequality, changes in the overall distribution of wages were much smaller. Lee’s analysis
suggests that these are largely explainable by the minimum wage.
60
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 11
Wage Inequality and the Minimum Wage
No r m a lize d 90-10 wa g e d iffe r e n t ia l
0.59
0.57
Normalized 90-10 wage
gap (May/OGR data)
0.55
Predicted wage gap from regression
on log of real minimum wage
0.53
0.51
0.49
0.47
0.45
1973
1976
1979
1982
1985
1988
1991
1994
1997
2000
Source: Authors’ analysis of various May and Outgoing Rotation Group files from monthly Current Population Survey and EPI. See also
<www.dol.gov/esa/minwage/chart.htm>.
REFERENCES
Autor, David, and Lawrence Katz. 1999. Changes in the
wage structure and earnings inequality. In Handbook of
labor economics, vol. 3A, edited by Orley Ashenfelter
and David Card. Amsterdam: Elsevier.
Bound, John, and George Johnson. 1992. Changes in the
structure of wages in the 1980s: An evaluation of alternative explanations. American Economic Review 83
(June): 371–92.
Autor, David, Lawrence Katz, and Alan B. Krueger. 1998.
Computing inequality: Have computers changed the
labor market? Quarterly Journal of Economics 113
(November): 1169–1214.
Bresnahan, Timothy F. 1999. Computerisation and wage
dispersion: An analytical reinterpretation. Economic
Journal 109 (June): 390–415.
Basu, Susanto, John G. Fernald, and Matthew D. Shapiro.
2001. Productivity growth in the 1990s: Technology, utilization, or adjustment? NBER Working Paper 8359, July.
Bresnahan, Timothy F., Erik Brynjolfsson, and Lorin M.
Hitt. 2002. Information technology, workplace organization
and the demand for skilled labor: Firm-level evidence.
Quarterly Journal of Economics 117 (February): 339–76.
Beaudry, Paul, and David A. Green. 2002. Changes in U.S.
wages 1976–2000: Ongoing skill bias or major technological change? NBER Working Paper 8787, February.
Card, David, and John E. DiNardo. 2002. Skill-biased technological change and rising wage inequality: Some problems and puzzles. Journal of Labor Economics 20, no. 4.
Berg, Maxine, and Pat Hudson. 1992. Rehabilitating
the Industrial Revolution. Economic History Review
45 (February): 24–50.
Card, David, and Thomas Lemieux. 2001. Can falling
supply explain the rising return to college for younger
men? A cohort-based analysis. Quarterly Journal of
Economics 116 (May): 705–46.
Berman, Eli, John Bound, and Zvi Griliches. 1994. Changes
in the demand for skilled labor within U.S. manufacturing
industries. Quarterly Journal of Economics 109 (May):
367–98.
Blau, Francine, and Lawrence M. Kahn. 1997. Swimming
upstream: Trends in the gender wage differential in the
1980s. Journal of Labor Economics 15 (January): 1–42.
Crafts, N.F.R., and C.K. Harley. 1992. Output growth and
the British Industrial Revolution: A restatement of the
Crafts-Harley view. Economic History Review 45
(November): 703–30.
DiNardo, John E., Nicole Fortin, and Thomas Lemieux.
1996. Labor market institutions and the distribution of
wages, 1973–1992: A semi-parametric approach. Econometrica 64 (September): 1001–44.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
61
Fortin, Nicole, and Thomas Lemieux. 1998. Rank regressions, wage distributions, and the gender gap. Journal
of Human Resources 33 (Winter): 610–43.
Krueger, Alan B. 1993. How computers have changed the
wage structure: Evidence from microdata, 1984–1989.
Quarterly Journal of Economics 108 (February): 33–60.
———. 2000. Are women’s wage gains men’s losses? A
distributional test. American Economic Review 90
(May): 456–60.
Lee, David S. 1999. Wage inequality during the 1980s:
Rising dispersion or falling minimum wage? Quarterly
Journal of Economics 114 (August): 977–1023.
Freeman, Richard B. 1975. Overinvestment in college
training? Journal of Human Resources 10 (Summer):
287–311.
Levy, Frank, and Richard J. Murnane. 1992. U.S. earnings and earnings inequality: A review of recent trends
and proposed explanations. Journal of Economic
Literature 30 (September): 1333–81.
———. 1976. The Overeducated American. New York:
Academic Press.
Gordon, Robert J. 2000. Does the “New Economy” measure up to the great inventions of the past? Journal of
Economic Perspectives 14 (Fall): 49–74.
Hamilton, Barton Hughes. 1997. Returns to computer
skills and black-white wage differentials. Washington
University in St. Louis, Department of Economics.
Unpublished manuscript.
Johnson, George. 1997. Changes in earnings inequality:
The role of demand shifts. Journal of Economic
Perspectives 11 (Spring): 41–54.
Jorgenson, Dale W. 2001. Information technology and
the U.S. economy. American Economic Review 91
(March): 1–32.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce.
1991. Accounting for the slowdown in black-white wage
convergence. In Workers and Their Wages. Edited by
Marvin H. Kosters. Washington, D.C.: AEI Press.
———. 1993. Wage inequality and the rise in returns to
skill. Journal of Political Economy 101 (June): 410–42.
Katz, Lawrence, and Kevin M. Murphy. 1992. Changes in
relative wages, 1963–1987: Supply and demand factors.
Quarterly Journal of Economics 107 (February): 35–78.
62
Manning, Alan. 2001. Can rising labor market attachment
explain the closing of the UK gender gap? London School
of Economics, Department of Economics. Unpublished
manuscript.
McGuckin, Robert H., Kevin J. Stiroh, and Bart van Ark.
1997. Computers, productivity, and economic growth.
Report No. 1206-97-RR. New York: Conference Board.
Mincer, Jacob. 1974. Education, experience, and earnings. New York: Columbia University Press.
Mishel, Lawrence, Jared Bernstein, and John Schmitt.
2001. The state of working America: 2000/2001.
Ithaca, N.Y.: Cornell University Press. Additional tables
at <www.epinet.org/datazone/>.
Oliner, Stephen D., and Daniel E. Sichel. 2000. The
resurgence of growth in the late 1990s: Is information
technology the story? Journal of Economic Perspectives 14 (Fall): 3–22.
U.S. Department of Education. National Center for
Education Statistics. 1998. The condition of education.
Washington, D.C.: GPO. Tabulations at <nces.ed.gov/
pubs98/condition98>.
Weinberg, Bruce. 2000. Computer use and the demand
for female workers. Industrial and Labor Relations
Review 53 (January): 290–308.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Productivity, Computerization,
and Skill Change
EDWARD N. WOLFF
Wolff is a professor of economics at New York
University. He thanks the Century Foundation
for financial support for the research.
obert Solow was, perhaps, the first to
point out the anomaly between productivity growth and computerization.
Indeed, he quipped that we see computers everywhere except in the productivity statistics. As we shall see, industries
that have had the greatest investment in computers
(namely, financial services) have ranked among the
lowest in terms of conventionally measured productivity growth. Moreover, at least until recently, there
has been little evidence of a payoff to computer
investment in terms of productivity growth.
However, another recent phenomenon of considerable visibility has been the rapid degree of industrial restructuring among U.S. corporations. This
paper argues that standard measures of productivity
growth are only one indicator of structural change.
There are others, such as changes in direct input and
capital coefficients. Changes in occupational mix and
the composition of inputs were greater in the 1980s
than in the preceding two decades. This pattern coincides with the sharp rise in computerization.
Though most of the literature has focused on the
connection between information technology (IT) or
information and communications technology (ICT)
and productivity, little work has been conducted on
the linkage between IT and broader indicators of
structural change (with a few exceptions noted
below). One purpose of this paper is to help fill this
gap. Indeed, this study finds evidence from regres-
R
sion analysis that the degree of computerization has
had a statistically significant effect on changes in
industry input coefficients and other dimensions of
structural change.
Another apparent anomaly arises when we consider the relationship between schooling and skills
on the one hand and productivity growth on the
other hand. Human capital theory predicts that rising educational attainment and skills will lead to
increasing productivity. Considerable policy discussion has also focused on the importance of education
and skill upgrading as an ingredient in promoting
productivity growth. Yet, as this discussion will show,
while overall productivity growth in the United States
slowed after 1973, the growth of schooling levels and
skills continued unabated. Indeed, college completion rates accelerated after 1970. In the time series
data, from 1947 to 1997, there is virtually no correlation between the growth of total factor productivity on the one hand and that of skills or educational
attainment on the other. Likewise, on the industry
level, sectors with the highest skills—namely services—have had the lowest productivity growth.
This paper will concentrate on the relation of
skills, education, and computerization to productivity growth and other indicators of technological
change on the industry level. I find no evidence
that the growth of educational attainment has any
statistically measured effect on industry productivity growth. The growth in cognitive skills, on the
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
63
other hand, is significantly related to industry productivity growth though the effect is very modest.
Moreover, the degree of computerization is not significant. In contrast, computerization has had a statistically significant effect on changes in industry
input coefficients.
The paper begins with a review of some of the
pertinent literature on the role of skill change and
computerization on productivity changes in the U.S.
economy. The next two sections introduce the
accounting framework and model and present
descriptive statistics on postwar productivity
trends. Descriptive statistics are also presented for
key variables that have shaped the pattern of pro-
This study finds evidence from regression
analysis that the degree of computerization
has had a statistically significant effect on
changes in industry input coefficients and
other dimensions of structural change.
ductivity growth over the postwar period, and multivariate analysis is conducted on the industry level
to assess their influence.
Review of Previous Literature
uman capital theory views schooling as an investment in skills and hence as a way of augmenting worker productivity (see, for example, Schultz
1960 and Becker 1975). This line of reasoning leads
to growth accounting models in which productivity
or output growth is derived as a function of the
change in educational attainment. The early studies
on this subject showed very powerful effects of
educational change on economic growth. Griliches
(1970) estimated that the increased educational
attainment of the U.S. labor force accounted for
one-third of the aggregate technical change between
1940 and 1967. Denison (1979) estimated that
about one-fifth of the growth in U.S. national income
per person employed (NIPPE) between 1948 and
1973 could be attributed to increases in educational
levels of the labor force. Jorgenson and Fraumeni
(1993) calculated that improvements in labor quality
accounted for one-fourth of U.S. economic growth
between 1948 and 1986.
Yet some anomalies have appeared in this line of
inquiry. Denison (1983), in his analysis of the productivity slowdown in the United States between
H
64
1973 and 1981, reported that the growth in NIPPE
fell by 0.2 percentage points whereas increases in
educational attainment contributed 0.6 percentage
points to the growth in NIPPE. Maddison (1982)
reported similar results for other OECD countries
for the 1970–79 period. Wolff (2001), using various
series on educational attainment, found no statistically significant effect of the growth in mean years
of schooling on GDP growth per capita among
OECD countries over the 1950–90 period.
A substantial number of studies, perhaps
inspired by Solow’s quip, have now examined the
linkage between computerization or information
technology (IT) in general and productivity gains.
The evidence is mixed. Most of the earlier studies
failed to find any excess returns to IT over and
above the fact that these investments are normally
in the form of equipment investment. These studies
include Franke (1987), who found that the installation of automated teller machines was associated
with a lowered real return on equity; Bailey and
Gordon (1988), who examined aggregate productivity growth in the United States and found no significant contribution of computerization; Loveman
(1988), who reported no productivity gains from IT
investment; Parsons, Gotlieb, and Denny (1993),
who estimated very low returns on computer
investments in Canadian banks; and Berndt and
Morrison (1995), who found negative correlations
between labor productivity growth and high-tech
capital investment in U.S. manufacturing industries.
Wolff (1991) found that the insurance industry had
a negative rate of total factor productivity growth
during the 1948–86 period in the United States even
though it ranked fourth among sixty-four industries
in terms of computer investment.
The later studies generally tend to be more positive. Both Siegel and Griliches (1992) and Steindel
(1992) estimated a positive and significant relationship between computer investment and industrylevel productivity growth. Oliner and Sichel (1994)
reported a significant contribution of computers to
aggregate U.S. output growth. Lichtenberg (1995)
estimated firm-level production functions and
found an excess return to IT equipment and labor.
Siegel (1997), using detailed industry-level manufacturing data for the United States, found that
computers are an important source of quality
change and that, once correcting output measures
for quality change, computerization had a significant positive effect on productivity growth.
Brynjolfsson and Hitt (1996, 1998) found, over
the 1987–94 time period, a positive correlation
between firm-level productivity growth and IT
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
investment when accompanied by organizational
changes. Lehr and Lichtenberg (1998) used data for
U.S. federal government agencies for the 1987–92
period and found a significant positive relation
between productivity growth and computer intensity.
Lehr and Lichtenberg (1999) investigated firm-level
data among service industries for the 1977–93 period
and also reported evidence that computers, particularly personal computers, contributed positively
and significantly to productivity growth. Ten Raa
and Wolff (2001), developing a new measure of
direct and indirect productivity gains, found that
the computer sector was the leading sector in the
U.S. economy during the 1980s as a source of economywide productivity growth. They also found very
high productivity spillovers between the computerproducing sector and sectors using computers. In
their imputation procedure, these large spillovers
were attributable to the high rate of productivity
growth within the computer industry.
Stiroh (1998) and Jorgenson and Stiroh (1999,
2000) used a growth accounting framework to
assess the impact of computers on output growth.
Jorgenson and Stiroh (1999) calculated that onesixth of the 2.4 percent annual growth in output can
be attributed to computer outputs compared to
about 0 percent for the 1948–73 period. The effect
came from capital deepening rather than from
enhanced productivity growth. A study by Oliner
and Sichel (2000) provides strong evidence for a
substantial role of IT in the recent spurt of productivity growth during the second half of the 1990s.
Using aggregate time-series data for the United
States, they found that both the use of IT in sectors
purchasing computers and other forms of information technology and the production of computers
appear to have made an important contribution to
the speedup of productivity growth in the latter
part of the 1990s. Hubbard (2001) investigated how
on-board computer adoption affected capacity utilization in the U.S. trucking industry between 1992
and 1997. He found that the use of computers
improved communications and resource allocation
decisions and led to a 3 percent increase in capacity
utilization within the industry.
One other factor that will be used in the data
analysis is research and development (R&D). A large
literature, beginning with Mansfield (1965), has now
almost universally established a positive and significant effect of R&D expenditures on productivity
growth (see Griliches 1979 and 1992 and Mohnen
1992 for reviews of the literature).
Modeling Framework
I
begin with a standard neoclassical production
function fj for sector j:
(1) Xj = Zj fj (KCj, KEj, KSj, L j, Nj, R j),
where Xj is the (gross) output of sector j, KCj is the
input of IT-related capital, KE j is the input of other
machinery and equipment capital goods, KS j is the
input of plant and other structures, Lj is the total
labor input, Nj is total intermediate input, Rj is the
stock of R&D capital, and Zj is a (Hicks-neutral)
total factor productivity (TFP) index that shifts the
production function of sector j over time.1 For convenience, the time subscript has been suppressed.
Moreover, capacity utilization and adjustment costs
are ignored. It then follows that
(2) dlnXj = dlnZj + εCjdlnKCj + εE jdlnKEj + εSjdlnKSj
+ εLj dlnLj + εNj dlnNj + εR jdlnRj,
where ε represents the output elasticity of each
input and dlnZj is the rate of Hicks-neutral TFP
growth. If the assumption of competitive input markets and constant returns to scale is imposed, it follows that an input’s factor share (α j ) will equal its
output elasticity. Employing the standard measure
of TFP growth, πj, for sector j,
(3) πj ≡ dlnXj /dt – α C j dlnKCj /dt – αE j dlnKE j /dt
– αSj dlnKSj /dt – αLj dlnL j /dt – α N j dlnNj /dt.
It then follows that
(4) πj = dlnZj /dt + αR j dlnRj /dt.
In particular, in the standard neoclassical model,
there is no special place reserved for IT capital in
terms of its effect on TFP growth.
As Stiroh (2002) argues, there are several reasons that the standard necoclassical model might be
expected to fail in the case of the introduction of a
radically new technology that might be captured by
IT investment. These include the presence of productivity spillovers from IT, problems of omitted
variables, the presence of embodied technological
change, measurement error in variables, and reverse
causality. If for one of these reasons the output elasticity of IT, εCj, exceeds its measured input share,
αCj, say, by uCj, then
(5) πj = dlnZj /dt + αRj dlnRj /dt + uCj dlnKCj /dt.
1. This equation is a modified form of the production function used by Stiroh (2002).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
65
In other words, conventionally measured TFP growth,
πj, will be positively correlated with the growth in
ICT capital.
A similar argument applies to labor productivity
growth, LP, defined as
(6) LPj ≡ dlnXj /dt – dlnLj /dt.
(BEA) worksheets. Deflators for 1982, 1987, 1992,
and 1996 are calculated from the Bureau of Labor
Statistics’ Historical Output Data Series (obtained
on computer diskette) on the basis of the currentand constant-dollar series. See the appendix for
details on sources and methods and a listing of the
forty-five industries.
If the assumption of competitive input markets and
constant returns to scale is again imposed, it follows that
C = forty-five-order matrix of capital coefficients,
where ci j is the net stock of capital of type i (in
1992 dollars) used per constant dollar of output j.
(7) LPj = dlnZj /dt + αCj dlnkCj /dt + αE j dlnkE j /dt
+ αSj dlnkSj /dt + αNj dlnnj /dt + αRj dlnRj /dt,
The capital matrix in constant dollars was provided by the BEA (see the appendix for sources)
and is based on price deflators for individual components of the capital stock (such as computers,
industrial machinery, buildings, etc.).
Measures of structural change may provide
a more direct and robust test of the effects
of computerization on changes in technology
than standard measures of productivity
growth do.
where lowercase symbols indicate the amount of
the input per worker.2 If for the reasons cited above
there is a special productivity “kick” from IT investment, then the estimated coefficient of kCj /dt should
exceeds its factor input share.
However, as indicated in the literature survey
in the previous section, very few studies, with the
exception of Siegel and Griliches (1992), have found
a direct positive correlation between industry TFP
growth and IT investment. As a result, this study
considers other indicators of the degree of structural change in an industry. These include changes
in the occupational composition of employment and
in the input and capital composition within an
industry. Productivity growth and changes in input
composition usually go hand in hand. To illustrate,
three new matrices are introduced:
A = forty-five-order matrix of technical interindustry
input-output coefficients, where ai j is the amount
of input i used per constant dollar of output j.
The technical coefficient (A) matrices are constructed on the basis of current-dollar matrices and
sector-specific price deflators. Sectoral price indices
for years 1958, 1963, and 1967 were provided by the
Brandeis Economic Research Center and those for
1972 and 1977 from the Bureau of Economic Analysis
66
M = occupation-by-industry employment coefficient
matrix, where mi j shows the employment of occupation i in industry j as a share of total employment in industry j.
The employment data are for 267 occupations and
64 industries and were obtained from the decennial
Census of Population for the years 1950, 1960, 1970,
1980, and 1990 (see Wolff 1996 for details).
Then, since for any input I in sector j, αI j =
pI Ij /pj Xj, where p is the price, equation 3 can be
rewritten as
(8) πj = –[Σi pi dai j + Σi pi,cdci j + Σi wi dbi j]/pj,
where pi is the price of intermediate input i, pi,c
is the price of capital input i, bi j = mij Lj /Xj is the
total employment of occupation i per unit of output in industry j, and wi is the wage paid to workers in occupation i. In this formulation, it is clear
that measured TFP growth reflects changes in
the composition of intermediate inputs, capital
inputs, and occupational employment. Using the
multiplication rule for derivatives, equation 8 can
be rewritten as
(9) πj = –[Σi pi daij + Σ i pi,cdci j + Σ iwi λj dmi j
+ Σ i wi mi j dλj ]/pj,
where λj = Lj /Xj . From equation 5 it follows that, in
the circumstances enumerated above, there may
be a positive correlation between measures of coefficient changes (such as dai j, dci j, and dmi j) and
IT investment.
Though productivity growth and changes in
input composition are algebraically related, there
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
are several reasons they may deviate. First, there
are costs of adjustments associated with radical
restructuring of technology, so there may be a
considerable time lag between the two (see David
1991, for example). Second, while new technology
is generally used to lower costs and hence increase
measured output per unit of input, new technology
might be used for other purposes such as product
differentiation or differential pricing. Third, in the
case of services in particular, output measurement problems might prevent one from correctly
assessing industry productivity growth. This problem could, of course, be partly a consequence of
product differentiation and price discrimination.
Measures of structural change may therefore provide a more direct and robust test of the effects
of computerization on changes in technology
than standard measures of productivity growth
do, particularly when a radically new technology
is introduced and the consequent adjustment
period is lengthy.
Finally, the change in average worker skills is
included in the production function. There are two
possible approaches. Let the effective labor input E =
QL, where Q is a measure of average worker quality
(or skills). Then equation 1 can be rewritten as
(10) Xj = Zj f *j (KCj, KE j, KS j, E j, Nj, Rj).
Again assuming competitive input markets and constant returns to scale (to the traditional factors of
production) and still using equation 6 to define labor
productivity growth, one obtains
(11) LPj = dlnZj/dt + αCj dlnkCj /dt + αE j dlnkE j /dt
+ αSj dlnkSj /dt + αNj dlnnj /dt + αLj dlnQj /dt
+ αR j dlnRj /dt.
In this formulation, the rate of labor productivity
growth should increase directly with the rate of
growth of average worker quality or skills.
The second approach derives from the standard
human capital earnings function. From Mincer (1974),
lnw = a0 + a1S,
where w is the wage, S is the worker’s level of
schooling (or skills), and a0 and a1 are constants. It
follows that
(dlnw)/dt = ai(dS/dt).
By definition, the wage share in sector j is αLj =
wj Lj /Xj. Under the assumptions of competitive
input markets and constant returns to scale, αL j =
εLj, a constant. Therefore, Xj /Lj = wj /εL j. In this case,
effective labor input E is given by the equation: Ln E
= Q + ln L. It follows from equation 6 that
(12) LPj = dlnZj /dt + α Cj dlnkCj /dt + α E j dlnkE j /dt
+ αSj dlnkSj /dt + α Nj dlnnj /dt + αLj dQj /dt
+ αR j dlnRj /dt.
In other words, the rate of labor productivity growth
should be proportional to the change in the level of
average worker quality or skills over the period.
Descriptive Statistics
T
echnological change. Table 1 shows the annual
rate of TFP growth for twelve major sectors over
the decades of the 1950s, 1960s, 1970s, and 1980s.
The periods are chosen to correspond to the employment by occupation and industry matrices. Factor
shares are based on period averages (the TornqvistDivisia index). The labor input is based on persons
engaged in production (PEP), the number of fulltime and part-time employees plus the number of
self-employed persons, and the capital input is
measured by fixed nonresidential net capital stock
(1992 dollars).3 (See the appendix.)
As shown in Table 1 (and Chart 1), the annual
rate of TFP growth for the entire economy fell from
1.4 percent per year in the 1950s to 1 percent per
year in the 1960s, plummeted to 0.4 percent per
year in the 1970s (the “productivity slowdown”
period), but subsequently rose to 0.8 percent in the
1980s.4 In the goods-producing industries (including communications, transportation, and utilities),
there was generally a modest slowdown in TFP productivity growth from the 1950–60 period to the
1960–70 periods, followed by a sharp decline in the
2. Technically, the assumption of constant returns to scale of the traditional factors of production is imposed, so that αCj + αEj
+ αSj + αNj + αLj = 1.
3. A second index of TFP growth was also used, with full-time equivalent employees (FTE) as the measure of labor input.
Results are very similar on the basis of this measure and are not reported below.
4. In November 1999, the BEA released a major revision of the U.S. national accounts. The new BEA data showed a faster rise in real
GDP and hence labor productivity during the 1990s than the older data indicated. One major element of the revision is the treatment of software expenses as a capital good rather than as an intermediate purchase. However, the BEA has not released the corresponding revised capital stock data. As a result, the statistics in this paper are based on the older BEA national accounts data.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
67
TABLE 1
Total Factor Productivity (TFP) Growth by Major Sector, 1950–90
1950–60
1960–70
1970–80
1.54
2.22
4.00
1.95
0.40
1.10
2.99
5.35
1.05
3.19
–2.36
1.72
1.59
2.97
2.55
3.47
–2.33
–3.41
–4.48
2.19
1.07
0.13
2.94
2.66
5.52
3.06
0.49
3.12
2.23
0.88
1.46
0.62
1.45
1.27
–0.59
2.25
1.32
1.27
2.49
3.03
1.08
1.41
0.12
0.60
0.14
–0.05
–1.01
0.37
0.25
0.86
–1.53
–0.35
0.38
0.10
–0.07
0.59
–0.66
0.15
–0.03
–0.28
2.12
0.70
1.39
1.50
0.58
0.96
0.25
0.58
0.38
2.04
0.07
0.77
1.48
0.48
0.88
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
Government and
government enterprises
Total goods
Total services
Total economy (GDP)
1980–90
1950–90
Note: Average annual growth in percentage points.
CHART 1
Annual TFP Growth, Mean Substantive Complexity, Mean Education,
and Percent of Adults with a College Education, 1952–97
25
Pe rc e n t ag e p oi nt s , pe r c en t, o r sc a la r
23
21
Percent of adults with college degree
19
17
15
13
Mean education (years)
11
9
7
Mean substantive complexity (SC)
5
3
TFP growth
1
–1
1952
1956
1960
1964
1968
1972
Note: Annual TFP growth is a five-year running average in percent per year.
Source: See Appendix.
68
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
1976
1980
1984
1988
1992
1996
TABLE 2
Dissimilarity Index (DIOCCUP) of the Distribution of Occupational
Employment by Major Sector, 1950–90
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
Government and
government enterprises
Total goods
Total services
All industries
1950–60
1960–70
1970–80
1980–90
Average
1950–90
0.000
0.022
0.040
0.100
0.077
0.030
0.032
0.078
0.001
0.025
0.025
0.039
0.050
0.024
0.061
0.169
0.001
0.020
0.005
0.014
0.023
0.014
0.043
0.053
0.017
0.045
0.053
0.096
0.088
0.048
0.128
0.105
0.005
0.028
0.031
0.062
0.060
0.029
0.066
0.101
0.026
0.043
0.061
0.019
0.117
0.091
0.029
0.033
0.029
0.078
0.080
0.047
0.038
0.068
0.057
0.046
0.054
0.042
0.045
0.047
0.063
0.022
0.050
0.061
0.056
0.056
0.014
0.026
0.019
0.110
0.077
0.095
0.062
0.045
0.055
Note: Computations are based on employment by occupation aggregated for each of the major sectors.
1970s (with agriculture, mining, and construction
recording negative productivity growth) and then a
substantial recovery in the 1980s. The major exceptions are durable manufacturing and communications, whose TFP growth rate rose from the 1960s
to the 1970s. TFP growth in the goods-producing
industries as a whole averaged 2.1 percent per year
in the 1950s, fell to 1.5 percent per year in the
1960s, and then collapsed to 0.3 percent in the
1970s before climbing back to 2 percent per year in
the 1980s.
TFP growth has been much lower in the service
sector than among goods-producing industries—
0.48 percent per year over the 1950–90 period for
the former compared to 1.48 percent per year for
the latter. The pattern over time is also generally different for the service industries. TFP growth in
wholesale and retail trade had a similar pattern to
that in goods industries—strong in the 1950–60
period (1.1 percent per year) before falling to 0.6
percent in the 1960s, turning negative in the next
decade, and then rebounding to 0.9 percent per year
in the 1980s. However, in finance, insurance, and
real estate (FIRE) general services, and the government sector, TFP growth dropped from the 1950s to
the 1960s, recovered somewhat in the 1970s, and
then slipped once again in the 1980s, turning negative in each case. Overall, annual TFP growth among
all services fell monotonically between the 1950s
and the 1980s, from 0.7 to 0.1 percent.
As noted above, I use three measures of structural change. The first measure is the degree to which
the occupational structure shifts over time. For this,
I employ an index of similarity. The similarity index
for industry j between two time periods 1 and 2 is
given by
(13) SI12 = (Σi m1i j m 2i j )/[Σi (m1i j ) 2 Σi(m 2i j ) 2 ]1/ 2.
The index SI is the cosine between the two vectors
st1 and st2 and varies from 0 (the two vectors are
orthogonal) to 1 (the two vectors are identical). The
index of occupational dissimilarity, DI, is defined as
(14) DIOCCUP 12 = 1 – SI 12.
Descriptive statistics for DIOCCUP are shown
in Table 2. The DIOCCUP index for the total economy, after rising slightly from 0.050 in the 1950s
to 0.056 in the 1960s dropped to 0.019 in the
1970s but then surged to 0.095 in the 1980s, its
highest level of the four decades. These results
confirm anecdotal evidence about the substantial
degree of industrial restructuring during the 1980s.
Similar patterns are evident for the major sectors
as well. In fact, seven of the twelve major sectors
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
69
TABLE 3
Dissimilarity Index (DIACOEFF) for Technical Interindustry Coefficients by Major Sector, 1950–90
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
Government and
government enterprises
Total goods
Total services
All industries
1950–60
1960–70
1970–80
1980–90
Average
1950–90
0.008
0.041
0.012
0.013
0.022
0.043
0.270
0.048
0.006
0.065
0.004
0.007
0.012
0.067
0.024
0.087
0.004
0.070
0.028
0.009
0.027
0.016
0.051
0.020
0.009
0.092
0.008
0.014
0.025
0.017
0.170
0.147
0.007
0.067
0.013
0.011
0.021
0.036
0.129
0.075
0.015
0.015
0.034
0.049
0.033
0.047
0.017
0.010
0.066
0.010
0.010
0.027
0.023
0.017
0.043
0.054
0.046
0.026
0.061
0.047
0.020
0.057
0.036
0.017
0.046
0.027
0.024
0.043
0.030
0.029
0.045
0.033
0.023
0.048
0.031
Note: Sectoral figures are based on unweighted averages of industries within the sector.
experienced their most rapid degree of occupational change during the 1980s. The three sectors
that experienced the greatest occupational restructuring over the four decades were utilities (0.101),
FIRE (0.068), and communications (0.066). Occupational change was particularly low in agriculture
(0.005), mining (0.028), transportation (0.029), and
construction (0.031).
It is also apparent that the association between
the DIOCCUP index and industry TFP growth is quite
loose. Though the degree of occupational restructuring has been somewhat greater in the goodsproducing industries than in services (average scores
of 0.062 and 0.045, respectively, for the 1950–90
period), the difference is not nearly as marked as for
TFP growth (annual rates of 1.5 percent and 0.5 percent, respectively, over the same period). Moreover,
while FIRE ranks second-highest in terms of occupational change, it is the fourth-lowest in terms of
TFP growth. In contrast, while agriculture ranks
fourth-highest in terms of TFP growth, it ranks
lowest in terms of occupational restructuring. The
DIOCCUP index provides a separate and relatively
independent dimension of the degree of technological change occurring in an industry.
A second index reflects changes in the technical
interindustry coefficients within an industry:
70
(15) DIACOEFF12 = 1 – (Σi a1i j a 2i j )/
[Σi(a1i j )2 Σi(a 2i j )2 ]1/ 2.
Figures in Table 3 indicate that the DIACOEFF
index for the total economy, after falling from 0.036
in the 1950–60 period to 0.027 in the 1960s, rose to
0.030 in the 1970s and again to 0.033 in the 1980s.
Eight of the twelve major sectors also recorded an
increase in the degree of change in their interindustry coefficients between the 1960s and the 1980s.
The sectors with the greatest interindustry coefficient change over the four decades were communications (0.129), utilities (0.075), and mining (0.067),
and the two with the least were agriculture (0.007)
and durable manufacturing (0.011).
The correlation between the DIACOEFF index
and industry TFP growth is again quite small. While
TFP growth was much higher in goods-producing
industries than in services, DIACOEFF was higher for
services than the goods sector. While agriculture,
durable manufacturing, and nondurable manufacturing all ranked high in terms of TFP growth, they were
the three lowest in terms of coefficient changes. The
DIACOEFF index provides another independent indicator of the degree of industry technological change.
A third index measures the change in capital
coefficients within an industry:
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 4
Dissimilarity Index (DIKCOEFF) for Capital Coefficients, 1950–90
1950–60
1960–70
1970–80
1980–90
Average
1950–90
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
0.002
0.016
0.011
0.005
0.009
0.002
0.015
0.003
0.000
0.008
0.016
0.007
0.006
0.009
0.028
0.001
0.001
0.025
0.032
0.009
0.006
0.011
0.045
0.002
0.005
0.038
0.061
0.007
0.009
0.008
0.087
0.003
0.002
0.022
0.030
0.007
0.008
0.007
0.044
0.002
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
0.045
0.020
0.057
0.019
0.014
0.033
0.014
0.027
0.035
0.024
0.043
0.062
0.026
0.026
0.047
Total goods
Total services (except government)
Total economy (except government)
0.008
0.038
0.020
0.007
0.024
0.014
0.011
0.029
0.018
0.014
0.050
0.028
0.010
0.035
0.020
Note: Sectoral figures are based on unweighted averages of industries within the sector. Data on investment by type are not available for
the government and government enterprises sectors.
(16) DIKCOEFF12 = 1 – (Σic1i j c2i j )/
[Σi(c i j ) Σi(c i j ) ] .
1
2
2
2 1/ 2
Table 4 shows that the DIKCOEFF index for the
total economy, after declining from 0.020 in the
1950s to 0.014 in the 1960s, increased to 0.018 in the
1970s and to 0.028 in the 1980s. DIKCOEFF rose in
nine of the eleven major sectors (capital stock by
type is not available for the government sector)
between the 1960s and the 1980s. General services
and communications showed the greatest change in
capital coefficients over the 1950–90 period and
agriculture and utilities the least. Here, again, while
TFP growth was much higher in goods than in service industries, DIKCOEFF was higher for the latter
than the former. Moreover, while agriculture,
durable manufacturing, and nondurable manufacturing were all among the top industries in terms of
TFP growth, they were among the lowest in terms
of capital coefficient changes.
Changes in skills and educational attainment.
As discussed in the previous two sections, the human
capital model predicts a positive relation between
changes in average education or average skill levels
and productivity growth. Figures on mean years of
schooling by industry are derived directly from
decennial Census of Population data for 1950, 1960,
1970, 1980, and 1990.
Educational attainment has been widely employed
to measure the skills supplied in the workplace.
However, the usefulness of schooling measures is
limited by such problems as variations in the quality
of schooling both over time and among areas, the
use of credentials as a screening mechanism, and
inflationary trends in credential and certification
requirements. Indeed, evidence presented in Wolff
(1996) suggests that years of schooling may not
closely correspond to the technical skill requirements of the jobs.
As a result, I also make use of the fourth (1977)
edition of the Dictionary of Occupational Titles
(DOT) for direct measures of workplace skills. For
some 12,000 job titles, it provides a variety of alternative measures of job-skill requirements based
upon data collected between 1966 and 1974. It probably provides the best source of detailed measures
of skill requirements covering the period 1950 to
1990. Three measures of workplace skills, described
below, are developed from this source for each of
267 occupations (see Wolff 1996 for more details).
Substantive complexity (SC). Substantive complexity is a composite measure of skills derived from
a factor analytic test of DOT variables. It was found
to be correlated with general educational development, specific vocational preparation (training
time requirements), data (synthesizing, coordinating,
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
71
analyzing), and three worker aptitudes—intelligence (general learning and reasoning ability), verbal, and numerical.
Interactive skills (IS). Interactive skills can be
measured, at least roughly, by the DOT “people”
variable, which, on a scale of 0 to 8, identifies
whether the job requires mentoring (0), negotiating
(1), instructing (2), supervising (3), diverting (4),
persuading (5), speaking-signaling (6), serving (7),
or taking instructions (8). For comparability with
the other measures, this variable is rescaled so that
its value ranges from 0 to 10 and reversed so that
mentoring is now scored 10 and taking instructions
is scored 0.
The human capital model predicts a positive
relation between changes in average education
or average skill levels and productivity growth.
Motor skills (MS). Motor skills is another DOT
factor-based variable. Also scaled from 0 to 10, this
measure reflects occupational scores on motor
coordination, manual dexterity, and “things”—job
requirements that range from setting up machines
and precision working to feeding machines and handling materials.
Composite skills (CS). I also introduce a measure
of composite skill, CS, which is based on a regression
of hourly wages in 1970 on SC, MS, and IS scores
across the 267 occupations. The resulting formula is
CS = 0.454SC + 0.093MS + 0.028IS
SC is the dominant factor in determining relative
wages in 1970, followed by MS and then IS.5
Average industry skill scores are computed as a
weighted average of the skill scores of each occupation, with the occupational employment mix of the
industry as weights. Computations are performed
for 1950, 1960, 1970, 1980, and 1990 on the basis of
consistent occupation by industry employment
matrices for each of these years constructed from
decennial census data. There are 267 occupations
and 64 industries.
Chart 1 provides some evidence on trends in both
cognitive skills (substantive complexity), mean education of the workforce, and the percentage of adults
72
with a college degree or more. Cognitive skills do not
appear to be closely correlated with TFP growth.
The average annual change in the SC index between
1947 and 1973 was .0156 points while TFP growth
averaged 1.4 percent per year and .0170 points
between 1973 and 1997, when TFP grew at only
0.6 percent per year. Moreover, the growth of college
graduates in the adult population was much greater
in the later period, averaging 0.45 percentage points
per year, than in the earlier period, averaging only
0.28 percentage points per year. Mean schooling, on
the other hand, tracks TFP more closely. The average annual change in mean education was 0.096
years over the 1948–73 period and 0.053 years over
the 1973–97 period.
There is also very little cross-industry association between skill levels and productivity growth.
As Table 5 shows, cognitive skill levels (SC) were,
on average, higher in the service sector than the
goods sector. In the 1980s, employees in FIRE had
the highest average SC score (5.25), followed by
general services (4.85), communications (4.74), and
the government sector (4.61). On the other hand, the
growth in mean SC was somewhat higher in goods
industries (0.53 points) than in services (0.43 points)
between 1950 and 1990.
The pattern is very similar for the mean education of the workforce. Average schooling was higher
in services than in the goods sector and was led
by general services (13.7 in 1980–90), followed by
FIRE (13.5), government (13.4), and communications (13.3). The change in mean education over the
four decades was also larger in the goods sector
(3.4 years) than in the service sector (2.6 years).
Investment in OCA. My measure of IT capital is
the stock of office, computing, and accounting equipment (OCA) in 1992 dollars, which is provided in the
BEA’s capital data (see the appendix for sources).
These figures are based on the BEA’s hedonic price
deflator for computers and computer-related equipment. As shown in Table 6 (and Chart 2), investment
in OCA per person engaged in production (PEP)
grew more than ninefold between the 1950s and the
1990s, from $28 (in 1992 dollars) per PEP to $263.
Indeed, by 1997 it had reached $2,178 per worker. By
the 1980s, the most OCA-intensive sector by far was
FIRE, at $1,211 per employee, followed by utilities
($628), mining ($393), durables manufacturing
($345), and communications ($285). On the whole,
the overall service sector has been investing more
intensively in computer equipment than the goods
sector has, but this pattern was largely due to the
very heavy investments made by FIRE. The trade and
general service sectors were actually below average in
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 5
Average Skill Level by Period and Major Sector, 1950–90
1950–60
1960–70
1970–80
1980–90
Change
1950–90
1. Mean years of education (in years)
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
Government and
government enterprises
Total goods
Total services
Total economy
8.05
9.19
9.53
10.28
9.75
9.78
11.42
10.69
9.06
10.41
10.25
11.00
10.48
10.55
11.98
11.19
10.45
11.56
11.21
11.67
11.34
11.44
12.62
11.78
11.45
12.45
12.04
12.39
12.13
12.27
13.31
12.68
4.02
4.21
3.11
2.90
3.05
3.21
2.52
2.79
10.62
11.82
11.56
11.18
12.40
12.34
11.89
12.95
13.08
12.51
13.53
13.66
2.33
2.29
2.72
11.50
12.02
12.69
13.37
2.42
9.59
11.20
10.36
10.51
11.88
11.25
11.43
12.62
12.13
12.23
13.23
12.86
3.43
2.60
3.23
2. Mean substantive complexity
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
Government and
government enterprises
Total goods
Total services
Total economy
3.67
3.35
3.67
3.50
2.98
3.16
4.02
3.85
3.64
3.71
4.02
3.71
3.12
3.25
4.26
3.87
3.61
3.98
4.16
3.84
3.34
3.35
4.51
4.07
3.64
4.13
4.22
3.96
3.49
3.32
4.74
4.33
0.01
1.02
0.80
0.65
0.58
0.11
0.93
0.56
3.91
4.63
4.32
3.84
4.96
4.46
3.88
5.13
4.73
3.98
5.25
4.85
0.04
0.90
0.52
4.24
4.30
4.46
4.61
0.42
3.41
4.18
3.78
3.57
4.26
3.94
3.73
4.44
4.15
3.83
4.57
4.30
0.53
0.43
0.62
Note: Figures are weighted averages of individual industries within each major sector.
5. The regression results for 1970 hourly wages (HOURWAGE) are as follows: HOURWAGE = 1.145 + 0.454SC + 0.093MS +
0.028IS, N = 267, R 2 = 0.535(4.78) (12.1) (2.37) (0.70), with t-ratios shown in parentheses. See the DOT, chapter 3, section 2,
for more discussion and analysis and for corresponding regression results for other years.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
73
TABLE 6
Annual Investment in Office, Computing, and Accounting Equipment (OCA)
per Persons Engaged in Production (PEP), 1950–90 (1992$, Period Averages)
1950–60
1960–70
1970–80
1980–90
Ratio of
1980–90 to
1950–60
A. Goods-producing industries
Agriculture, forestry, and fisheries
Mining
Construction
Manufacturing, durables
Manufacturing, nondurables
Transportation
Communications
Electric, gas, and sanitary services
0.1
14.3
6.8
24.5
49.2
43.7
49.1
47.2
0.3
28.6
6.9
21.5
54.5
36.5
43.6
41.8
2.1
53.3
5.8
30.2
98.3
29.6
51.1
54.5
4.9
392.9
7.7
119.9
345.3
72.7
285.2
628.3
67.4
27.5
1.1
4.9
7.0
1.7
5.8
13.3
B. Service industries
Wholesale and retail trade
Finance, insurance, and real estate
General services
14.0
140.0
22.9
20.3
162.7
23.4
42.5
339.4
23.0
279.8
1211.0
148.0
20.0
8.7
6.5
Total goods
Total services (except government)
Total economy (except government)
26.4
30.4
28.2
27.7
37.8
32.6
42.0
70.0
57.0
162.1
329.4
262.7
6.1
10.8
9.3
Note: Data on investment in OCA are not available for the government and government enterprises sectors.
CHART 2
Annual TFP Growth and OCA Investment per Worker, 1947–97
Pe r ce n t an d h u nd r e ds o f 19 92$ p e r P EP
26
22
18
OCA investment per PEP
14
10
6
TFP growth
2
–2
1947
1951
1955
1959
1963
1967
1971
1975
1979
1983
1987
1991
1995
Note: Annual TFP growth is a five-year running average in percent per year. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
74
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 3
TFP Growth and OCA Investment per Worker, 1950–90
5
P e rc e n t a n d h u n d r e d s o f 1992$ p e r PEP
OCA investment per PEP
TFP growth
4
3
2
1
0
–1
Agri.
Mining
Construct.
Durables Nondurables Transport. Communic.
Utilities
Trade
FIRE
Services
Note: Annual TFP growth is a five-year running average in percent per year. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
terms of OCA investment per PEP. Total investment
in equipment, machinery, and instruments (including
OCA) per PEP was more than fourteen times greater
than OCA investment even in the 1980s though by
1997 it accounted for almost exactly one-third of total
equipment investment.
On the surface, at least, there does not appear to
be much relation between OCA intensity and TFP
growth. While investment in OCA per worker rose
almost continuously over the postwar period, TFP
growth tracked downward, at least until the early
1980s (see Chart 2). Moreover, the sector with the
highest amount of OCA investment per worker,
FIRE, averaged close to zero in terms of TFP
growth over the postwar period (see Chart 3).
On the other hand, OCA investment seems to
line up well with measures of structural change. As
shown in Chart 4, the sectors with two highest rates
of investment in OCA per PEP over the 1950–90
period are FIRE and utilities, which also rank in the
top two in terms of the average value of DIOCCUP
over the same period. The sector with the lowest
investment in OCA per worker is agriculture, which
also ranks lowest in terms of DIOCCUP. Utilities
ranks highest in terms of DIACOEFF over the
1950–90 period and second-highest in terms of OCA
investment per employee while agriculture ranks
lowest in both dimensions (see Chart 5). The asso-
ciation is not quite as tight between OCA investment and DIKCOEFF (see Chart 6). However, here
again agriculture ranks lowest in both dimensions.
R&D. As shown in Chart 7, the ratio of R&D
expenditures to total GDP has remained relatively
constant over time, at least in comparison to the
wide fluctuations in TFP growth. It averaged 2 percent in the 1960s, fell to 1.5 percent in the 1970s,
recovered to 1.9 percent in the 1980s, and remained
at this level in the 1990–97 period. The pattern is
very similar for individual industries, with the
notable exceptions of industrial machinery (including OCA) and instruments, which show a continuous rise over the three periods. The ratio of R&D
to sales was considerably higher—by almost a factor of three—in durable manufacturing than in
nondurables. In the 1980–90 period, it ranged from
a low of 0.4 percent in food products to a high of
18.3 percent in other transportation (including
aircraft). The other major R&D-intensive industries,
in rank order, are instruments, electric and electronic equipment, industrial machinery, chemicals,
and motor vehicles.
An alternative indicator of R&D activity is the
number of full-time-equivalent scientists and engineers engaged in R&D per 1,000 full-time-equivalent
employees. Like the ratio of R&D expenditures to
GDP, this series shows a drop between the 1960s and
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
75
CHART 4
DIOCCUP and OCA Investment per Worker, 1950–90
12
P e rc en t a n d h u n d r e d s o f 1992$ p e r PEP
OCA investment per PEP
DIOCCUP
10
8
6
4
2
0
Agri.
Mining
Construct.
Durables Nondurables Transport.
Communic.
Utilities
Trade
FIRE
Services
FIRE
Services
Note: DIOCCUP is an average for the period in percent. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
CHART 5
DIACOEFF and OCA Investment per Worker, 1950–90
14
Pe r ce n t a n d h u nd r e ds o f 19 92$ p e r P EP
OCA investment per PEP
DIACOEFF
12
10
8
6
4
2
0
Agri.
Mining
Construct.
Durables Nondurables Transport. Communic.
Utilities
Trade
Note: DIACOEFF is an average for the period in percent. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
76
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 6
DIKCOEFF and OCA Investment per Worker, 1950–90
P e rc e n t a n d h u n d r e d s o f 1992$ p e r PEP
5
OCA investment per PEP
DIKCOEFF
4
3
2
1
0
Agri.
Mining
Construct.
Durables Nondurables Transport.
Communic.
Utilities
Trade
FIRE
Services
Note: DIKCOEFF is an average for the period in percent. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
1970s, from 5.4 to 4.8, and a recovery in the 1980s
to 6.4 (see Chart 7). However, it shows a further
increase to 7.3 in the 1990–96 period. This indicator also gives a very similar industry ranking. The
leading industries in the 1980s, in rank order, are
other transportation, chemicals, electric and electronic equipment, industrial machinery, instruments, and motor vehicles.
R&D expenditures does a much better job in lining up with TFP growth than either OCA or equipment investment. Both R&D intensity and TFP
growth fell from the 1960s to the 1970s and then
recovered in the 1980s. Moreover, there is a strong
cross-industry correlation between TFP growth and
R&D intensity—for example, both R&D intensity
and TFP growth are higher in durable manufacturing than in nondurable manufacturing.
Regression Analysis
n the first regression, the dependent variable is
the rate of industry TFP growth. The independent variables are R&D expenditures as a percent of
net sales and the growth in the stock of OCA capital. The statistical technique is based on pooled
I
cross-section time-series regressions on industries
and for the decades that correspond with the
decennial census data. The sample consists of fortyfive industries and three time periods (1960–70,
1970–80, and 1980–90).6 The estimating equation is
(17) TFPGRTHj = β0 + β1RDSALESj
+ β2OCAGRTH j + vj,
where TFPGRTHj is the rate of TFP growth in sector
j, RDSALESj is the ratio of R&D expenditures to net
sales in sector j, OCAGRTH is the rate of growth of
the stock of OCA capital, vj is a stochastic error term,
and the time subscript has been suppressed for
notational convenience. It is assumed that the vj t are
independently distributed but may not be identically
distributed. The regression results reported below
use the White procedure for a heteroscedasticityconsistent covariance matrix.
From equation 4 it follows that the constant β0
is the pure rate of (Hicks-neutral) technological
progress. From Griliches (1980) and Mansfield
(1980), the coefficient of RDSALES is interpreted
as the rate of return of R&D under the assumption
6. The 1950–60 period cannot be included in the regression analysis because the R&D series begins fully only in 1958.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
77
CHART 7
DIKCOEFF and OCA Investment per Worker, 1950–90
8
7
Pe r c e n t a n d n u m b e r p e r 1,000
Scientists and engineers per 1,000 employees
6
5
4
3
Ratio of R&D expenditures to GDP
2
1
0
TFP growth
–1
1957
1961
1965
1969
1973
1977
1981
1985
1989
1993
1997
Note: DIKCOEFF is an average for the period in percent. OCA investment is in hundreds of 1992 dollars per PEP.
Source: See Appendix.
that the (average) rate of return to R&D is equalized across sectors.7 Time dummies for the periods
1970–80 and 1980–90 are introduced to allow for
period-specific effects on productivity growth not
attributable to R&D or OCA investment. A dummy
variable identifying the ten service industries is also
included to partially control for measurement problems in service sector output.
Basic Regression Results
egression results for the full sample are shown
in columns 1 and 2 of Table 7. The constant
term ranges from 0.015 to 0.016. These estimates
are comparable to previous estimates of the Hicksneutral rate of technological change (see Griliches
1979, for example). The coefficient of the ratio of
R&D expenditures to net sales is significant at the
5 percent level. The estimated rate of return to R&D
ranges from 0.20 to 0.21. These estimates are about
average compared to previous work on the subject
(see Mohnen 1992, for example, for a review of previous studies).8
The coefficient of the growth of OCA is negative
but not statistically significant. The same result holds
for two alternative measures of IT, the growth in the
stock of computers and the stock of OCA plus communications equipment (OCACM). As noted above,
these specifications really measure the excess returns
R
78
to computer investment over and above that to capital in general since TFP growth already controls for
the growth of total capital stock per worker. The
coefficient of the dummy variable for service industries is significant at the 1 percent level; its value is
–0.017. The coefficient of the dummy variable for
the 1970–80 period is negative (significant in one of
the two cases), and that for the 1980–90 period is
positive (but not significant).
Because of difficulties in measuring output in
many service industries, regressions were also performed separately for the thirty-one goods-producing
industries (see the appendix table).9 The coefficient
values and significance levels of the constant term,
R&D intensity, the dummy variable for services, and
the two time period dummy variables are strikingly
similar to those for the all-industry regressions (see
specifications 3 and 4 of Table 7). The coefficient of
the growth in computer stock remains negative but
insignificant (specification 4).10
The next two regressions, focus on the “computer
age,” the period from 1970 onward. Does the effect
of computerization on productivity growth now show
up for this restricted sample? The answer is still negative, as shown in specifications 5 and 6 of Table 7.
The coefficients of the other two computerization
variables, the rate of growth in the stock of computers and that of OCACM are also insignificant
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 7
Cross-Industry Regressions of Industry TFP Growth (TFPGRTH) on R&D Intensity and OCA Investment
Independent
variables
Specification
(4)
(5)
(1)
(2)
(3)
Constant
0.015**
(3.45)
0.016**
(3.59)
0.014*
(2.59)
0.014**
(2.63)
Ratio of R&D
expenditures to sales
0.203*
(2.17)
0.212*
(2.24)
0.199#
(1.89)
Annual growth
in OCA
–0.039
(1.36)
(6)
(7)
(8)
0.011
(1.38)
0.020
(1.53)
0.010
(1.24)
0.005
(0.35)
0.205#
(1.93)
0.338*
(2.28)
0.348#
(2.00)
0.171*
(2.26)
0.131#
(1.86)
–0.024
(0.62)
–0.053
(1.27)
–0.102
(1.21)
–0.060
(1.29)
–0.016
(0.19)
–0.032**
(3.08)
–0.023*
(2.10)
Dummy variable
for services
–0.017**
(3.47)
–0.017**
(3.34)
Dummy variable for
1970–80
–0.010#
(1.89)
–0.006
(0.95)
–0.012#
(1.74)
–0.009
(1.05)
Dummy variable for
1980–90 (or 1987–97)
0.003
(0.59)
0.007
(1.13)
0.009
(1.22)
0.011
(1.37)
0.012#
(1.95)
0.008
(0.80)
0.005
(0.81)
0.195
0.171
0.0249
132
All
1960–90
0.205
0.174
0.0251
132
All
1960–90
0.127
0.098
0.0280
93
Goods
1960–90
0.131
0.092
0.0281
93
Goods
1960–90
0.216
0.178
0.0286
88
All
1970–90
0.145
0.078
0.0289
42
Goods
1970–90
0.232
0.201
0.0267
88
All
1977–97
R2
Adjusted R 2
Standard error
Sample size
Sample
Period
–0.018*
(2.47)
0.187
0.129
0.0292
44
All
1987–97
Note: Significance levels: #, 10%; *, 5%; **, 1%. The full sample consists of pooled cross-section time-series data, with observations on
each of 44 industries in 1960–70, 1970–80, and 1980–90 or in 1977–87 and 1987–97 (sector 45, public administration, is excluded
because of a lack of appropriate capital stock data). The goods sample consists of 31 industries (industries 1 to 31 in the Appendix table).
The coefficients are estimated using the White procedure for a heteroscedasticity-consistent covariance matrix. The absolute value of the tstatistic is in parentheses below the coefficient. See the Appendix for sources and methods.
(results not shown). R&D intensity remains significant in these regressions, and the estimated return
to R&D is higher, between 34 and 35 percent. The
same results for computerization (and R&D investment) are found when the sample is further restricted
to the 1980–90 period.
Specification 7 in Table 7 is based on a pooled
sample of observations for the 1977–87 and 1987–97
periods, while specification 8 is restricted to the
1987–97 period. As before, the coefficient of the
growth of OCA per worker is negative but not significant. Likewise, the coefficients of the rate of
growth in the stock of OCACM per employee and
the rate of growth of computers per employee are
insignificant (results not shown). In these regressions, the coefficient of R&D intensity remains significant but is somewhat lower (a range of 0.13 to
0.17) while the coefficient of the service dummy
variable also stays significant but is higher in
absolute value (a range of –0.23 to –0.032).
Regression results with worker skills. Table 8
shows the regression results for the various measures of worker skills and for the two alternative
formulations. Following equations 11 and 12, I use
labor productivity growth as the dependent variable. The first specification does not include skill
7. The proof is that RDSALES = dR/X. From equations 2 and 4 it follows that π = εR(dR/R) = εR(dR/X)(X/R) = (εR X/R)(dR/X).
Therefore, β1 = (εR X/R) = (dX/X)(X/R)/(dR/R) = dX/dR. The term dX/dR is the marginal productivity of R&D capital, which
is equivalent to the rate of return to R&D.
8. The coefficient of the number of full-time-equivalent scientists and engineers engaged in R&D per employee is also significant in
every case, typically at the 1 percent level. The tables present results using R&D expenditures because it is more conventional.
9. Since output measurement problems are less likely to affect transportation, communications, and utilities, they are classified as goods-producing industries here.
10. Results are again similar when the sample of industries is further restricted to the twenty manufacturing industries (results
not shown).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
79
TABLE 8
Cross-Industry Regressions of Industry Labor Productivity Growth on R&D Intensity,
Capital Investment, and Skill Change, 1960–90
Independent
variables
Specification
(4)
(5)
(6)
(7)
0.031**
(3.23)
0.030*
(3.39)
0.038*
(2.00)
0.017**
(2.81)
0.014#
(1.74)
0.182#
(1.86)
0.174#
(1.84)
0.184#
(1.95)
0.178#
(1.86)
0.174#
(1.77)
0.170#
(1.77)
Growth in total capital
per worker
0.235*
(2.27)
0.237*
(2.31)
0.239*
(2.34)
0.252*
(2.45)
0.244*
(2.31)
0.251*
(2.43)
Growth in substantive
complexity (SC)
0.181
(1.19)
0.125#
(1.78)
Growth in interactive
skills (IS)
–0.055
(0.44)
Growth in motor
skills (MS)
–0.015
(0.09)
(1)
(2)
Constant
0.017**
(2.96)
0.033
(1.47)
Ratio of R&D
expenditures to sales
0.164#
(1.73)
Growth in
OCA per worker
–0.006
(0.20)
Growth in total capital
less OCA per worker
0.262*
(2.50)
(3)
Growth in composite
skills (CS)
0.202#
(1.89)
Growth in mean
education
0.110
(1.14)
Change in substantive
complexity (SC)
0.224
(0.90)
Change in interactive
skills (IS)
–0.346
(1.04)
Change in motor
skills (MS)
0.006
(0.02)
Change in mean
education
0.056
(0.66)
Dummy variable
for services
–0.014**
(2.66)
–0.013#
(1.93)
–0.011*
(2.14)
–0.011*
(2.05)
–0.012*
(2.13)
–0.015**
(2.92)
–0.013*
(2.47)
Dummy variable for
1970–80
–0.009#
(1.47)
–0.009
(1.60)
–0.009
(1.65)
–0.009*
(1.59)
–0.012*
(2.12)
–0.009
(1.59)
–0.012*
(1.98)
Dummy variable for
1980–90
0.005
(0.82)
0.006
(0.96)
0.006
(1.00)
0.006
(1.08)
0.009
(0.99)
0.008
(1.23)
0.004
(0.77)
R2
Adjusted R 2
Standard error
Sample size
0.217
0.179
0.0252
132
0.236
0.186
0.0251
132
0.234
0.197
0.0249
132
0.237
0.200
0.0249
132
0.223
0.186
0.0251
132
0.226
0.176
0.0253
132
0.218
0.180
0.0252
132
Note: Significance levels: #, 10%; *, 5%; **, 1%.The sample consists of pooled cross-section time-series data, with observations on each
of 44 industries in 1960–70, 1970–80, and 1980–90 (sector 45, public administration, is excluded because of a lack of appropriate capital
stock data).The coefficients are estimated using the White procedure for a heteroscedasticity-consistent covariance matrix. The absolute value
of the t-statistic is in parentheses below the coefficient. See the Appendix for sources and methods.
80
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
change but splits total capital into OCA and other
capital. The coefficient of the growth of OCA per
worker is virtually zero, and the t-statistic is close to
zero. This result provides further corroboration of a
lack of a special effect of OCA investment on productivity growth.
In the second specification, I include the annual
change of the three measures of workplace skill:
substantive complexity (SC), interactive skills (IS),
and motor skills (MS). I also include the growth of
total capital per worker. None of the skill variables
is statistically significant in this regression. The
coefficients of the growth of IS and MS are, in fact,
negative. However, when the growth in cognitive
skills is included by itself, its coefficient becomes
marginally significant (at the 10 percent level). Its
elasticity is 0.13. The growth in the composite skill
index (CS) is also significant at the 10 percent level
(with a higher t-ratio) and its elasticity is 0.20
(specification 4). The best fit (highest adjusted R 2 )
occurs with the use of the CS variable. The coefficient
of the growth in mean schooling is also positive, with
an elasticity of 0.11, but not statistically significant
(specification 5).
Estimated coefficients for the change in mean
skills and mean schooling are not as significant as
those for the corresponding growth rates (specifications 6 and 7). None of the coefficients is even close
to significance. These results suggest that the labor
productivity growth is more closely related to the
growth in worker schools rather than to their absolute
change. This set of results remains robust among
alternative samples—goods-producing industries
only and for the 1970–90 period.11
In the set of regressions shown in Table 8, R&D
intensity is significant at the 10 percent level and its
estimated value is somewhat lower than in the corresponding TFP regressions (Table 7). The coefficient of the dummy variable for services is also
slightly lower (in absolute value) than in the TFP
regressions. The coefficient of the growth of total
capital per worker is in the range of 0.24 to 0.25,
somewhat lower than its income share, and is significant at the 5 percent level in all cases.
As discussed in the introduction, Brynjolfsson
and Hitt (1996, 1998) found a positive correlation
between firm-level productivity growth and IT
investment when the introduction of IT was accompanied by organizational changes. This finding suggests that interaction effects may exist between
OCA investment and changes in occupational composition. This was investigated by adding an interaction term between the growth of OCA per worker
and DIOCCUP to the labor productivity regression
equation derived from equation 11. The regression
was estimated for the full sample of industries over
both the 1960–90 and the 1970–90 periods and for
goods industries only over the two sets of periods.
The coefficient of the interaction term is statistically
insignificant in all cases and actually negative in
about half the cases.12
Computerization is found to be strongly linked
to occupational restructuring and changes in
material usage and weakly linked to changes
in the composition of capital.
Other indicators of technological activity.
In the last set of regressions, shown in Table 9, measures of structural change are used as dependent
variables. As before, the statistical technique is
based on pooled cross-section time-series regressions on industries and for the decades that correspond with the decennial Census data. The sample
consists of forty-four industries and two time periods (1970–80 and 1980–90).13 The basic estimating
equation is of the same form as equation 17, with
R&D intensity and the growth of OCA stock as independent variables. Dummy variables are also
included for the service sector and the 1970–80
period. Moreover, following equation 11, I also use
the growth of OCA per worker and OCA investment
per worker as independent variables in place of the
growth of total OCA stock.
The first of the dependent variables is the change
in occupational composition (DIOCCUP). In contrast
to the TFP regressions, the coefficient of investment
in OCA per worker is positive and significant at the
11. Results remain almost unchanged when an alternative measure of labor productivity growth, based on full-time-equivalent
employees (FTE) instead of persons engaged in production, is used as the dependent variable.
12. Regressions were also estimated with interaction terms between the growth of OCA per worker and the growth or change
in SC < CS and mean education. None of these interaction terms was found to be statistically significant.
13. The 1950–60 and 1960–70 periods are not included in the regression analysis because OCA investment was very small during
these periods. The government sector, moreover, cannot be included because of a lack of data on OCA investment.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
81
TABLE 9
Cross-Industry Regressions of Indicators of Structural Change on Computer Investment
Independent
variables
Constant
DIOCCUP
DIOCCUP
Dependent variable
DIACOEFF
DIACOEFF
DIKCOEFF
DIKCOEFF
0.048**
7.29
0.055**
(8.00)
0.001
(0.13)
–0.02*
(2.24)
0.016**
(2.98)
0.008
(1.02)
0.206
(1.17)
0.129
(0.71)
0.032#
(1.81)
0.031#
(1.66)
Ratio of R&D
expenditures to sales
0.251
(1.10)
0.214
(0.97)
0.136
(0.59)
0.309
(1.57)
Investment in OCA
per worker
0.060**
(3.07)
0.048*
(2.23)
0.043**
(5.24)
0.024**
(2.98)
Initial level of
OCA per worker
Dummy variable
for services
0.008
(0.08)
0.017
(1.51)
0.026**
(2.83)
Dummy variable for
1970–980
–0.021*
(2.30)
-0.001
(0.12)
–0.007
(0.89)
R2
Adjusted R 2
Standard error
Sample size
Industries
0.112
0.091
0.0470
88
All
0.145
0.104
0.0457
88
All
0.250
0.223
0.0429
88
All
0.271
0.227
0.0410
88
All
0.135
0.104
0.0339
88
All
0.165
0.114
0.0341
88
All
Note: Significance level: #, 10%; *, 5%; **, 1%. DIOCCUP is dissimilarity index for occupational coefficients; DIACOEFF is dissimilarity index
for technical interindustry coefficients; DIKCOEFF is dissimilarity index for capital coefficients. The sample consists of pooled cross-section
time-series data, with observations on each of forty-four industries (excluding the government sector) in 1970–80 and 1980–90. The coefficients are estimated using the White procedure for a heteroscedasticity-consistent covariance matrix. The absolute value of the t-statistic
is shown in parentheses below the coefficient estimate.
1 percent level in the regression without the service
and time period dummy variables and positive and
significant at the 5 percent level when the dummy
variables are included. The coefficients of the alternative computerization measures, the growth in OCA
per employee, investment in OCACM per worker,
and the rate of growth in the stock of OCACM per
employee are also significant at the 1 or 5 percent
level (results not shown). However, the best fit is
provided by investment in OCA per worker. The
results also show that R&D intensity is not a significant explanatory factor in accounting for changes in
occupational composition, nor is the dummy variable
for services. However, the time period dummy variable is significant at the 5 percent level.14
The second variable is DIACOEFF, a measure of
the degree of change in interindustry technical coefficients. In this case too, computerization is significant
at the 1 percent level with the predicted positive coefficient. The best fit is provided by investment in OCA
per worker. The coefficient of R&D intensity is positive
but not statistically significant, as is the coefficient of
the dummy variable for services. The coefficient of
the time dummy variable is virtually zero.
82
The third index of structural change is DIKCOEFF,
a measure of how much the composition of capital
has changed over the period. In this case, it is not
possible to use investment in OCA as an independent variable since, by construction, it will be correlated with changes in the capital coefficients.
Instead, I use the initial level of OCA per worker.
The computerization variable has the predicted
positive sign and is significant, though only at the 10
percent level. The coefficient of R&D is positive but
insignificant. However, the dummy variable for services is positive and significant at the 1 percent
level. The coefficient of the dummy variable for
1970–80 is negative but not significant.
In sum, computerization is found to be strongly
linked to occupational restructuring and changes in
material usage and weakly linked to changes in the
composition of capital. For the first result, it might
be appropriate to look at the construction of industry OCA by the BEA. The allocation of investment
in OCA is based partly on the occupational composition of an industry. As a result, a spurious correlation may be introduced between industry-level OCA
investment and the skill mix of an industry. The
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
cross-industry correlation between OCA per worker
and the mean SC level is 0.48 in 1970, 0.39 in 1980,
and 0.56 in 1990 while that between OCA per worker
and the mean schooling level of an industry is 0.46
in 1970, 0.29 in 1980, and 0.37 in 1990.
However, there is no indication that this allocation procedure should affect the change in occupational composition and hence introduce a spurious
correlation between OCA investment and the DIOCCUP variable. Moreover, the time-series evidence
shows a marked acceleration in the degree of occupational change between the 1970s and 1980s, when
OCA investment rose substantially. Regressions of
the change in occupational composition (DIOCCUP)
on both the growth of equipment per worker and
the growth of total capital per worker fail to yield
significant coefficients. As a result, we can surmise
that this finding is on solid ground.
Conclusion and Interpretation of Results
hree sets of findings emerge from the regression
analysis. First, the regression results provide
some modest evidence that skill growth is positively
linked with productivity growth. The coefficients
of the growth in both cognitive skills (SC) and the
composite skill (CS) index are marginally significant (at the 10 percent level). The effects are not
large—elasticities of 0.125 and 0.202, respectively.
Between 1947 and 1997, cognitive skills have grown
at an average annual rate of 0.41 percent, and composite skills by 0.33 percent. The growth of cognitive skills over this period would have added 0.05
percentage points to the growth of annual labor
productivity, while the growth of composite skills
would have added 0.07 percentage points. On the
other hand, the coefficient of the growth of the
mean education of the workforce, while positive, is
not statistically significant. Its estimated elasticity
is 0.110. Since mean education grew, on average, by
0.69 percent per year over the 1947–97 period, its
growth would have added 0.07 percentage points to
annual labor productivity growth.
These findings appear to be inconsistent with
growth accounting models, which have attributed a
substantial portion of the growth in U.S. productivity to increases in schooling levels. The conflict
stems from methodological differences in the two
techniques. Growth accounting simply assigns to
schooling (or measures of labor quality) a (positive)
role in productivity growth based on the share of
labor in total income. In contrast, in regression
T
analysis an estimation procedure is used to determine whether a variable such as education is a significant factor in productivity growth.
The findings on the role of education in productivity growth also appear to be at variance with the standard human capital model. There are several possible
reasons. First, the causal relation between productivity and schooling may be the reverse of what is normally assumed. In particular, as per capita income
rises within a country, schooling opportunities
increase, and more and more students may seek a college education (see Griliches 1996 for a discussion of
the endogeneity of education). Second, the skills
acquired in formal education, particularly at the university level, may not be relevant to the workplace.
Rather, higher education may perform a screening
function, and a university degree may serve employers mainly as a signal of potential productive ability
(see Arrow 1973 or Spence 1973). As enrollment
rates rise, screening or educational credentials may
gain in importance, and a higher proportion of university graduates may become overeducated relative
to the actual skills required in the workplace.
A third possibility is that university education
may be associated with rent-seeking activities
rather than lead directly to productive ones. This
pattern may be true for many professional workers,
such as lawyers, accountants, advertising personnel, and brokers. A fourth possible explanation is
the increasing absorption of university graduates by
“cost disease” sectors characterized by low productivity growth, such as health, teaching, law, and
business (see Baumol, Blackman, and Wolff 1989).
These are essentially labor activities and, as such,
are not subject to the types of automation and
mechanization that occur in manufacturing and
other goods-producing industries. Moreover, these
industries may be subject to output measurement
problems, particularly in regard to quality change.
Second, there is no evidence that computer
investment is positively linked to TFP growth. In
other words, there is no residual correlation
between computer investment and TFP growth over
and above the inclusion of OCA as normal capital
equipment in the TFP calculation. This result holds
not only for the 1960–90 period but also for the
1970–90, 1980–90, 1977–97, and 1987–97 periods.
The result also holds among exclusively goodsproducing industries and among exclusively manufacturing industries. This finding is not inconsistent
with recent work on the subject. Oliner and Sichel
14. It is not possible to use changes in skill levels or education as independent variables since, by definition, they would be associated with shifts in occupational composition.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
83
(2000), for example, found a strong effect of computers on productivity growth only beginning in the
mid-1990s, which is beyond my period of analysis.
Third, in contrast, computerization is strongly and
positively associated with other dimensions of structural change. These include occupational restructuring and changes in the composition of intermediate
inputs. The evidence is a bit weaker for its effects on
changes in the composition of industry capital stock.
The bottom line is that the diffusion of IT
appears to have shaken up the U.S. economy, beginning in the 1970s. However, it is a technological revolution that shows up more strongly in measures of
structural change rather than in terms of productivity, if the previous literature is a good guide on the
latter issue. In particular, the strongest results of
the effects of OCA on productivity growth are found
for the late 1990s in the United States. My results
seem to indicate that OCA has had strong effects on
changes in occupational composition and input
structure dating from the early 1970s.
These two sets of results might reflect the high
adjustment costs associated with the introduction of
new technology. The paradigmatic shift from electromechanical automation to information technologies
might require major changes in the organizational
structure of companies before the new technology
can be realized in the form of measured productivity gains (see David 1991 for greater elaboration of
this argument). The results of computerization are
also consistent with an alternative interpretation of
its role in modern industry. The argument is that a
substantial amount of new technology (particularly,
information technology) may be used for product
differentiation rather than productivity enhancement. Computers allow for greater diversification
of products, which in turn also allows for greater
price discrimination (for example, airline pricing
systems) and the ability to extract a large portion of
consumer surplus. Greater product diversity might
increase a firm’s profits, though not necessarily its
productivity. Some evidence on the production differentiation effects of computers is provided by
Chakraborty and Kazarosian (1999) for the U.S.
trucking industry (for example, speed of delivery
versus average load).
APPENDIX
Data Sources and Methods
C
apital stock figures. Figures are based on
chain-type quantity indexes for net stock of
fixed capital in 1992$, year-end estimates. OCA
investment data are available for the private
(nongovernment) sector only. Source: U.S. Bureau
of Economic Analysis, CD-ROM NCN-0229,
“Fixed Reproducible Tangible Wealth of the
United States, 1925–97.”
Educational attainment: (a) Median years of
schooling, adult population; (b) percent of adults
with four years of high school or more; and (c) percent of adults with four years of college or more.
Source: U.S. Bureau of the Census, Current Population Reports Reports <www.census.gov/hhes/
income/histinc/incperdet.html>. “Adults” refers to
persons twenty-five years of age and over in the
noninstitutional population (excluding members of
the armed forces living in barracks). (d) Mean (or
median) schooling of workers by industry for 1950,
1960, 1970, 1980, and 1990 is derived from the
decennial U.S. Census of Population Public Use
Samples for the corresponding years.
Input-output data: The original input-output
data are eighty-five-sector U.S. input-output tables
84
for 1947, 1958, 1963, 1967, 1972, 1977, 1982, 1987,
1992, and 1996 (see, for example, Lawson 1997 for
details on the sectoring). The 1947, 1958, and 1963
tables are available only in single-table format. The
1967, 1972, 1977, 1982, 1987, 1992, and 1996 data
are available in separate make and use tables.
These tables have been aggregated to forty-five
sectors for conformity with the other data sources.
The 1950, 1960, 1970, 1980, and 1990 input-output
tables are interpolated from the benchmark U.S.
input-output tables.
NIPA employee compensation: Figures are from
the National Income and Product Accounts (NIPA)
<www.bea.gov/bea/dn/nipaweb/>. Employee compensation includes wages and salaries and
employee benefits.
NIPA employment data: Full-time-equivalent
employees (FTE) equals the number of employees
on full-time schedules plus the number of employees on part-time schedules converted to a full-time
basis. FTE is computed as the product of the total
number of employees and the ratio of average
weekly hours per employee for all employees to
average weekly hours per employee on full-time
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
schedules. Persons engaged in production (PEP)
equals the number of full-time-equivalent employees plus the number of self-employed persons.
Unpaid family workers are not included.
Research and development expenditures:
R&D expenditures performed by industry include
company, federal, and other sources of funds.
Company-financed R&D performed outside the
company is excluded. Industry series on R&D and
full-time equivalent scientists and engineers
engaged in R&D per full-time equivalent employee
run from 1957 to 1997. Source: National Science
Foundation <www.nsf.gov/sbe/srs/nsf02312/>. For
technical details, see National Science Foundation, Research and Development in Industry
(Arlington, Va.: National Science Foundation)
NSF96-304, 1996.
TABLE
45-Sector Industry Classification Scheme
Industry
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
1987 SIC codes
Agriculture, forestry, and fishing
Metal mining
Coal mining
Oil and gas extraction
Mining of nonmetallic minerals, except fuels
Construction
Food and kindred products
Tobacco products
Textile mill products
Apparel and other textile products
Lumber and wood products
Furniture and fixtures
Paper and allied products
Printing and publishing
Chemicals and allied products
Petroleum and coal products
Rubber and miscellaneous plastic products
Leather and leather products
Stone, clay, and glass products
Primary metal products
Fabricated metal products, including ordnance
Industrial machinery and equipment, exc. electrical
Electric and electronic equipment
Motor vehicles and equipment
Other transportation equipment
Instruments and related products
Miscellaneous manufactures
Transportation
Telephone and telegraph
Radio and TV broadcasting
Electric, gas, and sanitary services
Wholesale trade
Retail trade
Banking; credit and investment companies
Insurance
Real estate
Hotels, motels, and lodging places
Personal services
Business and repair services except auto
Auto services and repair
Amusement and recreation services
Health services, including hospitals
Educational services
Legal and other professional services and nonprofit organizations
Public administration
01–09
10
11–12
13
14
15–17
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
371
37 [exc. 371]
38
39
40–42, 44–47
481, 482, 484, 489
483
49
50–51
52–59
60–62, 67
63–64
65–66
70
72
73, 76
75
78–79
80
82
81, 83, 84, 86, 87, 89
—
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
85
REFERENCES
Arrow, Kenneth. 1973. Higher education as a filter.
Journal of Public Economics 2, no. 3:193–216.
———. 1980. R&D and the productivity slowdown.
American Economic Review 70 (May): 343–47.
Bailey, Martin, and Robert Gordon. 1988. The productivity
slowdown, measurement issues, and the explosion in computer power. Brookings Papers on Economic Activity,
no. 2: 347–432.
———. 1992. The search for R&D spillovers. Scandinavian Journal of Economics 94 (Supplement): 29–47.
Baumol, William J., Sue Anne Batey Blackman, and Edward
N. Wolff. 1989. Productivity and American Leadership:
The Long View. Cambridge, Mass.: MIT Press.
Becker, Gary S. 1975. Human Capital: A Theoretical
and Empirical Analysis. 2d ed. New York: Columbia
University Press and National Bureau of Economic
Research.
Berndt, Ernst R., and Catherine J. Morrison. 1995. Hightech capital formation and economic performance in U.S.
manufacturing industries. Journal of Econometrics 65
(January): 9–43.
Brynjolfsson, Eric, and Lorin Hitt. 1996. Computers and
productivity growth: Firm-level evidence. MIT Sloan
School of Management. Photocopy.
———. 1998. Information technology and organizational
design: Evidence from micro data. MIT Sloan School of
Management. Photocopy.
Chakraborty, Arreya, and Mark Kazarosian. 1999. Product
differentiation and the use of information technology:
Evidence from the trucking industry. NBER Working
Paper 7222, July.
David, Paul A. 1991. Computer and dynamo: The modern productivity paradox in a not-too-distant mirror.
In Technology and productivity: The challenge for
economic policy. Paris: Organisation for Economic
Co-operation and Development.
Denison, Edward F. 1979. Accounting for slower economic growth: The United States in the 1970s.
Washington, D.C.: Brookings Institution.
———. 1983. The interruption of productivity growth in
the United States. Economic Journal 93 (March): 56–77.
Franke, R.H. 1987. Technological revolutions and productivity decline: The case of U.S. banks. In Computers in the
Human Context: Information Technology, Productivity,
and People, edited by T. Forester. Cambridge, Mass.:
Massachusetts Institute of Technology Press.
Griliches, Zvi. 1970. Notes on the role of education in production functions and growth accounting. in Education,
income, and human capital. Studies in Income and
Wealth 35, edited by W. Lee Hansen, 71–115. New York:
Columbia University Press and National Bureau of
Economic Research.
———. 1979. Issues in assessing the contribution of
research and development to productivity growth. Bell
Journal of Economics 10 (Spring): 92–116.
86
———. 1996. Education, human capital, and growth: A
personal perspective. NBER Working Paper 5426, January.
Hubbard, Thomas N. 2001. Information, decisions, and
productivity: On-board computers and capacity utilization
in trucking. NBER Working Paper No. 8525, October.
Jorgenson, Dale W., and Barbara M. Fraumeni. 1993.
Education and productivity growth in a market economy.
Atlantic Economic Journal 21 (June): 1–25.
Jorgenson, Dale W., and Kevin J. Stiroh. 1999. Information
technology and growth. American Economic Review 89
(May): 109–15.
———. 2000. Raising the speed limit: U.S. economic
growth in the information age. Brooking Papers on
Economic Activity, no. 1: 125–211.
Lawson, Anne M. 1997. Benchmark input-output accounts
for the U.S. economy, 1992. Survey of Current Business
77 (November): 36–83.
Lehr, William, and Frank Lichtenberg. 1998. Computer
use and productivity growth in U.S. federal government
agencies, 1987–92. Journal of Industrial Economics 44
(June): 257–79.
———. 1999. Information technology and its impact on
productivity: Firm-level evidence from government and
private data sources, 1977–1993. Canadian Journal of
Economics 32 (April): 335–62.
Lichtenberg, Frank R. 1995. The output contribution of
computer equipment and personnel: A firm-level analysis.
Economics of Innovation and New Technology 3, no.
4:201–17.
Loveman, G.W. 1988. An assessment of the productivity
impact of information technologies. In Management in
the 1990s. Cambridge, Mass.: Massachusetts Institute of
Technology Press.
Maddison, Angus. 1982. Phases of capitalist development. Oxford, England: Oxford University Press.
Mansfield, Edwin. 1965. Rates of return from industrial
research and development. American Economic Review
55 (May): 310–22.
———. 1980 Basic research and productivity increase
in manufacturing. American Economic Review 70
(December): 863–73.
Mincer, Jacob. 1974. Schooling, experience, and earnings. New York: National Bureau of Economic Research.
Mohnen, Pierre. 1992. The relationship between R&D
and productivity growth in Canada and other major
industrialized countries. Ottawa, Canada: Canada
Communications Group.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Oliner, Stephen, and Daniel Sichel. 1994. Computers and
output growth revisited: How big is the puzzle? Brookings
Papers on Economic Activity, no. 2:273–317.
Steindel, Charles. 1992. Manufacturing productivity and
high-tech investment. Federal Reserve Bank of New York
Quarterly Review 17 (December): 39–47.
———. 2000. The resurgence of growth in the late 1990s:
Is information technology the story? Federal Reserve
Board of Governors. Photocopy.
Stiroh, Kevin J. 1998. Computers, productivity, and input
substitution. Economic Inquiry 36 (April): 175–91.
Parsons, Daniel J., Calvin C. Gotlieb, and Michael Denny.
1993. Productivity and computers in Canadian banking.
Journal of Productivity Analysis 4 (June): 91–110.
———. 2002. Are ICT spillovers driving the New Economy?
Review of Income and Wealth 48 (March): 33-57.
Schultz, Theodore W. 1960. Capital formation by education. Journal of Political Economy 68, no. 6:571–83.
ten Raa, Thijs, and Edward N. Wolff. 2001. Outsourcing
of services and the productivity recovery in U.S. manufacturing in the 1980s and 1990s. Journal of Productivity
Analysis 16, no. 2:149–65.
Siegel, Donald. 1997. The impact of computers on manufacturing productivity growth: A multiple-indicators,
multiple-causes approach. Review of Economics and
Statistics 79 (February): 68–78.
Wolff, Edward N. 1991. Productivity growth, capital
intensity, and skill levels in the U.S. insurance industry,
1948–86. Geneva Papers on Risk and Insurance 16
(April): 173–90.
Siegel, Donald, and Zvi Griliches. 1992. Purchased services, outsourcing, computers, and productivity in manufacturing. In Output measurement in the service
sector, edited by Zvi Griliches. Chicago: University of
Chicago Press.
———. 1996. Technology and the demand for skills.
OECD Science, Technology and Industry Review 18:
96–123.
Spence, Michael. 1973. Job market signaling. Quarterly
Journal of Economics 87, no. 3:355–74.
———. 2001. Human capital investment and economic
growth: Exploring the cross-country evidence. Structural
Change and Economic Dynamics 11 (December):
433–72.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
87
Global Banks, Local Crises:
Bad News from Argentina
MARCO DEL NEGRO AND STEPHEN J. KAY
Del Negro is an economist on the macropolicy team, and Kay is a senior
economic analyst in the Latin America Research Group in the Atlanta Fed’s
research department. They thank Mike Malone, Ann Misback, and John
Atkinson for help with the legal issues and Virgina DeJesus-Rueff for help
with the data. They are also grateful to Dave Gordon, Jose-Antonio Meade,
Carlos Provencio, Myriam Quispe-Agnoli, Will Roberds, Ellis Tallman, and
especially Larry Wall for very helpful comments and conversations.
Finally, they thank Elena Casal for research assistance.
magine a world where your favorite bank is
like Starbucks: you can find a branch at every
corner, in every city on the globe. Imagine a
world where in emerging markets all banks
are international because local banks have
either disappeared or been bought out. The
world just described is not as distant from reality as
one might think, especially in the Americas. Nearly
everywhere you go in Latin America, from San Luis
Potosí in Mexico to Santiago in Chile, Citibank has
an office (see Chart 1). In the last few years, large
U.S. and European banks have expanded their
presence in several Latin American countries at a
staggering pace to the extent that today in some
countries they own or control the majority of the
domestic banking system.
In the past few decades, banking crises have
been a recurrent phenomenon in Latin America.1
Some have argued that the internationalization of
the banking sector has ushered in a new era. A
November 2001 report by Salomon Smith Barney
states that “One of the main benefits that the presence of foreign banks in Latin America should
produce is the overall decline in systemic risk. . . .
We believe systemic risk in the [Argentine] banking
system (one that caused the collapse of the system
of payments) is low, as 43% of its equity is controlled by foreigners”(23). The rationale for this
optimism is as follows. When an intermediation
sector is purely domestic, any financial crisis, major
I
currency depreciation, or government bankruptcy is
a systemic shock that could cause the collapse of
the entire system. The fact that international banks
now own or control a sizable fraction of local banking systems, the reasoning goes, has changed the
picture considerably. Some international banks hold
such a large and internationally diversified portfolio
of assets that a country-specific shock in a small
economy, like Argentina, should not be able to
endanger their financial health. Hence, what used
to be systemic risk from the perspective of local
banks with undiversified portfolios might no longer
be systemic from the standpoint of large international banks. In economic terms, Argentina is about
the size of Connecticut. Given the size and resources
of a typical large international bank, a crisis in a
country like Argentina could be overcome by such a
bank—or so the reasoning went.
This scenario, if true, would be very good news
for depositors in emerging markets. While in the
United States deposit insurance shields depositors
from the risk of bank insolvency, in some emerging
markets there is no deposit insurance at all.2 In others, like Argentina, its scope and resources are limited.3 This lack or limitation of deposit insurance in
emerging markets means that a shock to the asset
side of a bank often translates into a shock to the
liability side: Depositors bear at least some of the
brunt of bank insolvency, especially when it is systemic. In this light, the international diversification
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
1
CHART 1
Citibank Branches in the Western Hemisphere
US
446
Mexico
249
2 Bahamas
5 Dominican Republic
15 Puerto Rico
Panama 5
Colombia 17
Ecuador 5
7 Venezuela
34 Brazil
Chile 19
54 Argentina
Source: Citibank <www.citibank.com> (March 2002)
of foreign banks’ assets is an attractive feature for
depositors in emerging markets because it reduces
the portfolio exposure to country-specific shocks
and hence makes deposit safer.
Yet Argentina’s experience shows that the presence of international banks was not enough to prevent local banking crises and sizable losses to
depositors. Specifically, the point of this article is as
follows: The “bad news” from Argentina is that
depositors in emerging markets may not reap the
full benefits of international portfolio diversification. The article argues that depositors may not
reap the full benefits because international banks
have limited liability, at least under some circumstances—for instance, whenever the local government heavily intervenes in the banking system.
Hence, there is a key difference between a crisis in,
say, Connecticut and a crisis in Argentina. If the
branch of any bank in Connecticut is producing
heavy losses, for example, the U.S. regulator will
not simply liquidate the branch and let the parent
2
company—that is, the bank—forfeit its obligations
to the depositors at that branch. The parent company has no choice but to face its liabilities, at least
to the extent that the bank as a whole is solvent.
If the same events occur at the bank’s branch in
Argentina, however, the bank can conceivably
refuse to shore up the local branch—or at least
threaten to do so—even if the parent company as a
whole has enough liquidity to withstand the crisis.4
Because of this limited-liability feature, the Argentine branches or subsidiaries of international banks
may face the crisis as stand-alone entities. And while
the parent company’s portfolio is highly diversified
internationally, the branch’s or subsidiary’s portfolio
often is not.5
Given the sensitive nature of this topic, it is
important that the message of this article is not misunderstood. The article does not argue that the
presence of international banks is detrimental to
emerging markets. On the contrary, there is substantial evidence that opening the banking system to
foreign banks is beneficial to emerging markets from
all points of view, including macroeconomic stability.
Also, the article does not argue that the limitedliability feature itself is detrimental for emerging
markets. While the limited-liability feature of international banks may seem bad ex post—and, of
course, it is from the perspective of Argentine
depositors—this feature may well be desirable, perhaps even necessary, ex ante. Indeed, the earlier
analogy comparing a crisis in Connecticut and one in
Argentina needs at least one important qualification.
In the unlikely event that the State of Connecticut
were to implement some of the actions taken by the
Argentine government—such as forced conversion
of dollar-denominated bank assets into pesos at a
less-than-market rate or limitations on holdings of
dollar-denominated assets—the banks could certainly challenge those actions in a federal court.6
International banks do not have this option in the
Argentine case. Hence, the limited-liability feature is
needed to protect banks from foreign governments’
actions; in the absence of limited liability, the incentive for foreign government to (implicitly or explicitly)
expropriate the assets of international banks would
be too high. Although this article does not study the
welfare implications of this limited-liability feature,
the concluding section offers some further thoughts
on the issue. In particular, it argues that the limitedliability feature may also create perverse incentives
for international banks to the extent that local
depositors are not fully aware of it.
This article first presents some evidence of the
globalization of the banking sector in Latin America,
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
CHART 2
Foreign Banks’ Participation and Control as a Percentage of Total Sector Assets, November 2001
80
Assets participation
Assets control
70
60
Pe r c e n t
50
40
30
20
10
0
Argentina
Brazil
Chile
Mexico
Source: Salomon Smith Barney (2001)
documenting the dramatic increase of the phenomenon in the late nineties. The discussion also
demonstrates that the involvement of international
banks in any country has often been very large relative to the size of the banking sector in that specific country but relatively small in comparison to
the overall size of the international bank. The article
then reviews the literature on the pros and cons of
international banks in emerging markets, specifically
in Latin America. The discussion focuses on the
literature that addresses the question, Does the
presence of foreign financial institutions enhance or
reduce the stability of the domestic banking system? The study then examines the legal issues that
are behind the limited-liability feature. Indeed, the
institutional information in this section, which is
sometimes neglected by the literature, is the main
value added of this article. Finally, the article
addresses the “bad news” from Argentina and discusses some implications of this phenomenon.
International Banks in Latin America:
Some Facts
n the largest Latin American countries a sizable
portion of the banking sector is, directly or indirectly, in the hands of international financial institutions. Chart 2 shows the percentage of assets
controlled by foreign banks in the four largest Latin
American banking systems. The definition of “control” is the same used in the report by Salomon
I
1. Since 1980, Argentina alone has suffered two banking crises, in 1980–82 and 1989–90 (see Caprio and Klingebiel 1996).
2. In the United States the Federal Deposit Insurance Corporation (FDIC) covers deposits up to $100,000. Before the FDIC
Improvement Act (FDICIA) of 1991, essentially all creditors of large banks were covered by the FDIC. FDICIA substantially
limits this coverage (see Wall 1993).
3. Kane and Demirgüc-Kunt (2001) document that deposit insurance has become very popular of late in emerging markets: In
the past fifteen years the fraction of countries offering deposit insurance has increased from about 30 percent to 70 percent.
The remainder of the paper provides further details on the deposit insurance scheme in Argentina.
4. Of course, international banks can close their operations in emerging markets at will, but the point addressed in the article is
the circumstances under which international banks have limited liability.
5. Some argue that this lack of diversification was partly due to the Argentine government’s forcing banks to hold government paper.
6. In 1933 the federal government actually implemented both actions: It suspended the gold clauses (which tied the value of
certain assets to gold) and forced all private parties to hand all gold (coins, bullions, and certificates) to the federal government. Those actions were challenged in federal courts and finally in the Supreme Court. In all (four) cases, the Supreme
Court sided with the federal government (Kroszner 1999).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
3
TABLE 1
Foreign Participation and Control of Loans, Deposits, and Equity, November 2001
Loans
control
Loans
participation
Deposits
control
Deposits
participation
Equity
control
Equity
participation
46.5
24.4
44.9
72.9
44.2
24.6
36.0
57.7
46.4
16.3
44.5
76.2
43.3
16.5
36.4
60.7
43.1
30.1
54.1
74.8
40.5
29.8
46.4
61.1
Argentina
Brazil
Chile
Mexico
Note: All figures are percentages. November participation is applied to June 2001 figures.
Source: Salomon Smith Barney (2001)
CHART 3
Foreign Control of Total Loans
80
December 1996
June 1999
March 1998
December 2000
November 2001
70
60
P e rc e n t
50
40
30
20
10
0
Argentina
Brazil
Chile
Colombia
Mexico
Peru
Venezuela
Source: Salomon Smith Barney
Smith Barney from which the data were taken: An
international bank controls a domestic bank if its
stake in the domestic bank is at least 40 percent.7
The chart shows that foreign banks control almost a
third of banking sector assets in Brazil, the largest
Latin American economy. In the second-largest
economy, Mexico, the figure rises to a staggering
three-quarters. In the third- and fourth-largest
economies (in financial terms), Argentina and Chile,
foreign banks control 53 percent and 59 percent of
total assets, respectively. The numbers for the share
of assets owned by international banks (“participation”) are lower, but not very much so, suggesting
that international financial institutions usually own
large stakes in the banks they control.
4
Table 1 looks at other measures of international
banks’ involvement in Latin America, particularly the
share of loans, deposits, and equity either controlled
or owned by foreign financial institutions in the four
largest Latin American countries. In Brazil international banks control a quarter of loans, 16 percent of
deposits, and 30 percent of equity. The corresponding figures for Mexico are 73 percent, 76 percent,
and 75 percent. For Argentina and Chile these figures are approximately 40 percent to 50 percent.
The table clearly shows that, no matter how one
measures it, the presence of international banks in
Latin America is large.8
The picture just described would have been
almost unthinkable a decade ago. Chart 3 shows the
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
TABLE 2
dramatic expansion of foreign control of total loans
in the banking sector from December 1996 to
November 2001. Foreign control over loans increased
by 30 percent in Argentina, more than doubled in
both Brazil and Chile, and increased sixfold in
Mexico. Table 2, which lists foreign control of total
assets in the banking sector in 1994, 1999, and 2001,
also shows how rapidly foreign control evolved in
the 1990s. Foreign control of assets in Mexico
evolved from 1 percent in 1994 to 45 percent in
2001.9 In Argentina, Brazil, and Chile, foreign control of total assets tripled during that period.
Explaining this dramatic increase in foreign
banks’ presence in Latin America goes beyond the
scope of this article. According to the literature
(Clarke and others 2000; Clarke and others 2001;
Barajas, Steiner, and Salazar 2000; Demirgüc-Kunt
and Huizinga 2000), one reason for this increase
seems to be that domestic banks were not very efficient, at least relative to foreign banks. Since competition from local banks in emerging markets is
often not as stiff as competition at home, for many
U.S. and European banks the Latin American market opens profit opportunities in the provision of
financial services. In some countries, the increase in
economic integration between the home country
and the host country also prompted those international banks that wanted to “follow their clients” to
expand their role in Latin America. For instance,
since the beginning of the North American Free
Trade Agreement (NAFTA) in 1994, economic integration between the United States and Mexico has
increased dramatically. Changes in regulations have
also played a major role. In Mexico, before NAFTA,
Citibank was the only international bank permitted
to conduct (limited) banking operations. Until
December 1998 regulations prohibited foreign control of Mexico’s three largest banks, which account
for about 60 percent of loan market share. The lifting of those restrictions prompted a dramatic
expansion of foreign banks’ role in Mexico.
Foreign Control of Total Assets, 1994–2001
Argentina
Brazil
Chile
Mexico
1994
1999
2001
17.9
8.4
16.3
1.0
48.6
16.8
53.6
18.8
53.1
27.0
48.0
45.4
Note: Control is defined as a 50 percent stake.
Source: IMF, Salomon Smith Barney, authors’ calculations
Finally, financial crises themselves contributed
to the increasing presence of international financial institutions in Latin America (see Peek and
Rosengren 2000b). In the aftermath of the Mexican
crisis, for instance, the government was very eager
to sell the banks it had just rescued. International
banks were an important source of new capital for
a banking sector that desperately needed a capital
infusion. The same situation occurred in Argentina
in the aftermath of the Tequila Crisis.
Which international banks are the biggest players in the Latin American arena? For each of the
largest eight financial institutions involved in Latin
America, Table 3 shows the amount of loans made
by banks controlled by these institutions and these
loans as a percentage of total loans. Three of the
banks shown in the table are a notch above all others in terms of involvement in Latin America: two
Spanish banks, BBVA and Santander Central
Hispano (SCH), and a U.S. financial institution,
Citigroup.
How large a stake do international banks have in
Latin America? Table 4 lists the amount of loans
that the three major players control in the four
largest banking sectors in the region.10 The table
indicates that loans made to these four countries
represent a sizable portion of the loan portfolio of
these banks. For Citigroup this share is roughly 9 percent. For the two Spanish banks the figure is even
7. If a 50 percent threshold is used the figures do not change substantially, with the exception of Mexico, where the Spanish
bank Banco Bilbao Vizcaya Argentaria (BBVA) owns 49 percent of BBVA Bancomer.
8. Peek and Rosengren (2000b) argue that all these measures grossly underestimate the importance of international banks for
lending to Latin America. The asset and loan measures include subsidiaries and branches of international banks that operate in the host countries but neglect offshore lending. Peek and Rosengren show that until 1997 the latter component was
more important than the former for Argentina, Brazil, and Mexico.
9. In this case, control is defined as at least a 50 percent stake. This definition excludes Mexico’s largest bank, BBVA Bancomer,
of which the Spanish bank BBVA owns a 49 percent stake. If the lower 40 percent threshold is used, foreign control of total
assets would be 73 percent.
10. Again, these figures actually underestimate the exposure of international banks because they do not include offshore lending (Peek and Rosengren 2000b). One should also be careful in interpreting these figures as appropriate measures of risk,
which is more properly computed from the exposure in relation to the parent bank’s capital or equity rather than the overall asset position (see Goldberg 2001).
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
5
TABLE 3
Top Eight Foreign Banks in Argentina, Brazil, Chile, and Mexico, 2001a
Bank
In U.S.$ billions
As a percent
of total loans
36.6
34.5
34.8
9.2
5.1
4.5
4.1
3.7
11.5
10.8
10.9
2.9
1.6
1.4
1.3
1.2
Banco Bilbao Vizcaya Argentaria (Spain)
Santander Central Hispano (Spain)
Citibank (U.S.)
FleetBoston (U.S.)
HSBC (U.K.)
ABN Amro (Netherlands)
Scotiabank (Canada)
Sudameris (France/Italy)
a
As a percentage of total loans controlled
Source: Salomon Smith Barney
TABLE 4
The Largest Three Foreign Banks’ Loans to Argentina, Brazil, Mexico, and Chile
Bank
Loans to
top four
banking As % of
Net loans sectors net loans
SCH
154.9
Citigroup 381.8
BBVA
133.9
34.3
34.6
36.7
Loans to As % of Loans to As % of Loans to As % of Loans to As % of
Argentina net loans Brazil net loans Mexico net loans Chile net loans
22.1
9.1
27.4
6.5
4.7
5.0
4.2
1.2
3.7
2.4
3.3
1.2
1.6
0.9
0.9
14
25
28
9.1
6.6
20.9
11.4
1.6
2.5
7.4
0.4
1.9
Note: In billions of U.S. dollars and as a percentage of total loans. Loans are shown in billions of U.S. dollars. Net loans are total loans
less loan loss reserves as of year-end 2001. Country loans are as of November 2001.
Source: Dow Jones Interactive, Salomon Smith Barney
larger, about 22 percent for SCH and 29 percent for
BBVA. On the one hand, these numbers suggest
that these banks, especially SCH and BBVA, could
be severely affected by a systemic crisis in Latin
America as a whole. On the other hand, if one
focuses on any specific country, one finds that, with
the exception of Mexico, the exposure of these
banks is relatively small. For Argentina the share of
net loans is around 4 percent for the two Spanish
banks and a mere 1.2 percent for Citigroup. In
Mexico, by contrast, international banks have quite
a bit at stake: 6.6 percent of Citigroup’s net loans,
9.1 percent of SCH’s net loans, and 20.9 percent of
BBVA’s net loans.
Foreign Banks and Domestic Crises:
What Do We Know?
he literature on international banks in Latin
America (and elsewhere) is relatively recent,
just like the phenomenon it studies. The literature
can best be understood within the context of the
T
6
policy questions faced by decision makers in Latin
America and other developing countries: Should
we allow entry to foreign banks? What are the
gains? What are the potential risks? Box 1 offers a
brief review of the literature. This review focuses
on the evidence the literature has gathered on the
narrower questions, Does the presence of foreign
financial institutions enhance or reduce the stability of the domestic banking system? What are foreign banks going to do in time of crisis? Given the
apparently endemic instability of Latin American
economies, these are key questions policymakers
face. Opponents of foreign bank entry claim that in
a crisis international banks will abandon—“Vive
les rats!”—the domestic economy to its destiny.
Proponents argue that, on the contrary, global
banks will provide stability to the domestic financial sector since they are less affected by shocks
that are idiosyncratic to the host country.
In principle, the presence of foreign banks has two
contrasting effects on the stability of domestic bank-
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
BOX 1
International Banks in Latin America: The Literature
he literature on international banks in Latin
America can be best understood within the
context of the policy questions faced by decision
makers in Latin America and other developing
countries: Should we allow entry to foreign
banks? What are the gains? What are the potential risks?1
One of the main benefits related to the entry of
foreign banks is the increased efficiency of the
financial system. On this point, the literature
strongly suggests that efficiency increases following foreign banks’ entry into developing countries.2 For one thing, banks that expand abroad
are typically the “best of the crop” in the country
of origin (Focarelli and Pozzolo 2000). Hence,
they are likely to export improved management
and information technology practices to the host
country. Second, the literature finds that foreign
banks are generally more efficient than domestic
competitors (Barajas, Steiner, and Salazar 2000;
Clarke and others 2000). Third, a number of studies find that foreign bank entry has been associated with increased efficiency of domestic financial
intermediaries (see Claessens, Demirgüc-Kunt, and
Huizinga 1998; Clarke and others 2000).
The payoff from increased efficiency can be
very large. Levine (2001) argues that there is
substantial empirical evidence supporting the following causal chain: first, foreign bank entry
enhances the efficiency of the banking sector;
second, efficiency in the intermediation sector
spurs growth by boosting productivity.
Opponents of financial openness, however,
emphasize the other side if the coin. By squeezing the interest margins and profitability of
domestic banks, the entry of foreign banks may
push local intermediaries out of the market. This
reasoning implies that entire sectors that were
previously dependent on local banks—small
firms, for instance—may find themselves without
T
access to credit, with detrimental consequences
for the economy. The evidence on whether these
consequences actually ensue in countries with
extensive foreign bank presence is inconclusive.
The literature finds that small businesses are
indeed less likely than larger ones to receive
credit from foreign banks (Berger, Klapper, and
Udell 2000; Clarke and others 2002). After the
size of the banks in the sample is controlled for,
however, the negative relationship between foreign ownership and lending to small businesses
tends to disappear, if not to be reversed. A different but related argument brought forward by
opponents of foreign banks’ entry is that these
banks tend to “cherry-pick” their customers, leaving domestic banks with a worse pool of potential
creditors than before. There is little evidence
supporting this point, however, and the existing
evidence points in the opposite direction (Crystal,
Dages, and Goldberg 2001).
The multitude of banking crises during the last
two decades point to the weaknesses of the regulatory and supervisory environment in many
emerging markets. Disclosure standards are also
inferior in developed countries, especially to
standards in the United States. Proponents of
financial openness argue that, by allowing foreign
bank entry, emerging markets indirectly benefit
from the more advanced supervisory and disclosure environment in the country of origin (see
Peek and Rosengren 2000b). Opponents of financial openness counter that foreign bank entry
leaves the domestic regulator in a weaker position than before. For one thing, the regulator’s
ability to exercise moral suasion is lessened. In
addition, foreign banks may be more responsive
to changes in regulations at home than in the host
country (Peek and Rosengren 2000b). Specifically,
regulatory changes in the country of origin may
affect lending in the host country.
1. An exhaustive review of the literature so far can be found in Clarke and others (2001).
2. This evidence for developing countries is in contrast to that found for developed countries, in particular for the United
States (see, for instance, Hasan and Hunter 1996).
ing systems (leaving aside the issue of limitedliability, which will be discussed later). On the one
hand, the portfolio diversification of global banks
makes the domestic financial system less fragile with
respect to domestic shocks. On the other hand, their
presence means that the host country may become
more exposed to external shocks—more specifically
to shocks that affect the country of origin of the banks.
Global banks are generally larger, and have a
more diversified portfolio of assets, than local
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
7
banks.11 The international portfolio diversification
of global banks is advantageous for the host country’s financial system both ex post, in the event of a
crisis, and ex ante. If a crisis occurs, global banks
are likely to have both less portfolio exposure to the
domestic economy and greater access to liquidity
than local banks do. Ex ante, according to standard
portfolio theory, the presence of international banks
may imply that the interest rate paid on loans by
domestic firms is lower, other things being equal,
than when only local banks are present. Since local
banks have all their eggs in the same basket, they
are willing to add one more egg to that same basket
only if the price is high enough to compensate them
Foreign banks’ entry can in principle have two
contrasting effects on the domestic financial
system—diminished sensitivity to domestic
shocks and higher exposure to shocks in the
international banks’ country of origin.
for the additional risk they are undertaking. Global
banks have their eggs in many baskets. Hence, the
additional risk undertaken by international banks of
putting one more egg in the Argentine basket is
lower than that undertaken by an Argentine bank,
so the international banks might be willing to
demand a lower return. In equilibrium, depending
on the market structure of the banking system, this
willingness to demand a lower return may translate
into a lower cost of capital for domestic firms.
In a nutshell, whenever the banking system is
closed, country-specific shocks are necessarily systemic and hence may threaten the stability of the
system. From the perspective of international
banks, however, those very same shocks are idiosyncratic. Hence, the entry of international banks
makes the domestic financial system less fragile
with respect to domestic shocks.12 An important
corollary of this point is the following: The high
volatility of Latin American economies is not at all
an obstacle to the expansion of international banks,
at least to the extent that this volatility is idiosyncratic.13 On the contrary, the higher the volatility is,
the higher the relative advantage of foreign versus
domestic banks.
The discussion now turns to the second question: Does the presence of foreign banks mean that
the host country may inherit global shocks? Even in
8
the absence of foreign banks, emerging markets are
certainly not isolated from global financial shocks,
as shown very clearly by the Asian crises. Yet some
argue that the presence of foreign banks exacerbates the host country’s exposure to global shocks.
For some global banks, idiosyncratic shocks in the
country of origin—Spain, for instance—may affect
the lending behavior of their subsidiaries abroad. In
addition, Kyle and Xiong (2001) have shown that
“contagion” may be the rational outcome of international financial integration via a wealth effect. While
international banks are not the focus of Kyle and
Xiong’s study, the logic of their argument may apply
to international banks as well. In summary, a country that opens its banking system to foreign banks
may become less sensitive to its own shocks but at
the same time increase its exposure with respect to
shocks generated elsewhere.
Empirically, there is some evidence that both
effects are at work—that foreign banks’ entry
makes the banking system (1) less sensitive to
domestic shocks and (2) more sensitive to external
ones. On the first point, the evidence suggests that
lending by global banks is stronger and more stable
than lending by domestic financial institutions even
in the face of crises in the host country. Dages,
Goldberg, and Kinney (2000) show that during the
Tequila Crisis foreign banks in both Mexico and
Argentina did not “cut and run.” The authors find
that foreign banks had both the highest loan growth
and the lowest volatility in lending growth before,
during, and after the crisis for both Argentina and
Mexico. Dages, Goldberg, and Kinney also find that
lending by foreign banks is less sensitive to changes
in domestic real GDP growth than is lending by
domestic banks although their research cannot statistically reject the hypothesis that private domestic
and foreign banks have the same proportionate
response to cyclical forces.14 Goldberg (2001) also
shows that U.S. banks’ claims on emerging markets
are not highly sensitive to fluctuations in localcountry GDP. A study by Demirgüc-Kunt, Levine,
and Min (1999), based on the work of DemirgücKunt and Detragiache (1997), finds that the presence of foreign banks reduced the likelihood of a
banking crisis in the host country.15
A related issue, investigated by Crystal, Dages,
and Goldberg (2001), is whether foreign-owned
banks are any sounder in terms of lending practices
than domestically owned ones. Crystal, Dages, and
Goldberg find that in the seven largest Latin
American economies, foreign-owned banks fare
marginally better than local ones in terms of financial strength ratings (Moody’s Bank Financial
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Strength Ratings) although there are no significant
differences between foreign and private domestic
banks. For Argentina, Chile, and Colombia, the
authors examine banks’ balance sheet data and find
that foreign banks tend to have more aggressive
loan provisioning and higher loan recovery rates
than domestically owned banks do. In summary, the
findings of Crystal, Dages, and Goldberg suggest
that foreign banks’ entry may lead to a sounder
banking system in the host country.
The fact that international banks are perceived
to be sounder than local banks in times of crises has
led some to argue that foreign banks’ presence
opens the possibility of a “capital flight at home.”
Before the appearance of foreign banks, investing
abroad was the only safe haven for domestic depositors, given the lack of credible deposit insurance.
Now, under the assumption that foreign banks are
strong enough to withstand a crisis, all depositors
need to do is transfer their savings from local to foreign financial institutions. There is indeed some evidence of such a “flight to quality” during the Asian
crisis and during the Tequila Crisis in Argentina
(IMF 2000; also, see Kane 2000 for a discussion of
the policy implications of the “flight to quality”).16
There is also evidence that the presence of foreign banks may increase the host country’s exposure to home country shocks. Specifically, evidence
shows that lending by international banks responds
to economic fluctuations in the country of origin.
Peek and Rosengren (2000a) have widely documented that the lending behavior of Japanese banks
in the United States was heavily conditioned by
events at home and that these changes in the lending pattern had real effects in the host country.
Goldberg (2001) studies the determinants of U.S.
banks’ claims to emerging markets. She finds that
the relationship between claims to Latin America
and movements in U.S. real GDP growth is significantly procyclical even after controlling for fluctuations in local GDP and local and U.S. interest rates.
In summary, foreign banks’ entry can in principle have two contrasting effects on the domestic
financial system—diminished sensitivity to domestic shocks and higher exposure to shocks in the
international banks’ country of origin. In addition,
both effects are empirically relevant, raising the
question of which of the two is the most important
quantitatively. Although no study to our knowledge
directly addresses this question (except perhaps
Demirgüc-Kunt, Levine, and Min 1999), the first
effect is likely to be more important than the second for Latin American countries. Latin American
economies have historically been very volatile, and
these fluctuations have had a disrupting impact on
the local banking system. Therefore, it is likely that
the gains from a diminished sensitivity of lending to
local shocks outweigh the costs of higher sensitivity
to shocks originated elsewhere. The results from
the existing literature suggest that foreign banks’
entry is likely to make the banking system more
stable. To what extent do the recent events in
Argentina lead us to reassess this conclusion, if at
all? This question is addressed later in the article,
but the next section takes a brief detour into some
relevant legal issues.
11. Goldberg (2001) shows that 60 percent of the exposure of large U.S. banks engaged in international lending is in industrialized countries.
12. The next two sections of the article argue, however, that the limited-liability feature of international banks undermines some
of the benefits from international portfolio diversification.
13. This point is forcefully made in Stockman (2001). Stockman discusses a related issue, namely the idea of an “optimum central bank area,” in opposition to the standard “optimum currency area.” The “optimum currency area” literature emphasizes
the supposed disadvantages of having asymmetric (that is, uncorrelated) shocks. The idea of an optimum central bank area
emphasizes the advantages of uncorrelated shocks from the perspective of a central bank.
14. For Mexico, the above statements hold true for banks with similar impaired loan ratios. For developed countries some of the
evidence suggests otherwise. Tallman and Bharucha (2000) find that in Australia during the 1986–93 credit crunch foreign
banks cut lending more than domestic ones did.
15. Interestingly, these authors find that the significant variable in reducing the likelihood of a crisis is not so much the share of
foreign banks but rather the number of foreign banks.
16. Some authors further argue that “in countries that allow foreign currency deposits, depositors may be more comfortable
placing such deposits in foreign banks that have ready access to foreign currency during a banking crisis, with the lender of
last resort for the bank being the central bank in the banks’ home country rather than that of the host country” (Peek and
Rosengren 2000b, 49). In essence, these authors argue that in the absence of limited liability the parent company may have
to shore up local branches or subsidiaries. To the extent that this operation affects the solvency of the parent company at
home, the home regulator may end up implicitly bailing out the host country banking system. However, because of the limitedliability feature of international banks, it is unlikely that the home country central bank would end up acting as a lender of
last resort, particularly if the banking crisis is accompanied by interventions on the part of the foreign government, as was
the case in Argentina. The next section directly addresses this issue.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
9
Legal Niceties
hat is the relationship between the foreign
subsidiary of an international bank and the
parent company? If the foreign subsidiary or branch
is insolvent, to what extent can depositors or other
creditors successfully seek payment from the parent company? If a U.S. bank decides to close down
a branch in, say, Connecticut, depositors can withdraw their money at any other branch in the country.
Does the same apply to depositors of a U.S. bank’s
branch in a Latin American country, say, Argentina?
If not, why not? We are not experts in international
law and hence do not pretend to give a definite
answer to these questions. Rather, the goal of this
W
“A member bank shall not be required to
repay any deposit made at a foreign branch . . .
if the branch cannot repay . . . due to . . . an
action by a foreign government. . . .”
Section 326, Riegle-Neal Act
section is to raise these questions—arguing that
they are relevant for the issues discussed here—
and provide some guidelines for addressing them.
The questions posed above have a clear practical relevance for Argentine depositors. They are
also relevant for the larger issues discussed in this
article, namely, To what extent do depositors reap
the benefits of the fact that global banks have an
internationally diversified portfolio? To the extent
that a global bank can walk away from a country in
crisis without being held accountable for the subsidiary’s or the branch’s liabilities, an incentive
arises to pull out if these liabilities exceed the
expected profit from remaining in the country.
Hence, at least under some circumstances, the
presence of international banks may be no safety
net for local depositors during a crisis. These questions are also relevant for home and host country
regulators.17 To the extent that foreign banks have
a limited liability, the home country regulator may
not be as concerned about the repercussions of
foreign banking crises on the financial health of
the parent company as it would be otherwise.
This discussion has argued that the above questions are relevant. To address them, let us first
consider the case in which the parent company’s
subsidiary, or the branch, operates in the United
States. In the case of the subsidiary the key notion
10
is the one of “corporate veil” (see Cox, Hazen, and
O’Neal 1997 and Hamilton 1991). A subsidiary’s
creditors cannot go after the parent company’s
assets in case of default only if the corporate veil
is in place. Loosely speaking, the corporate veil is
in place when the following two conditions are
satisfied. First, the subsidiary must present itself
to creditors as a clearly separate entity from the
parent company. Second, it must act as such—that
is, the subsidiary must be independently managed,
and the parent company must have no more clout
than the majority shareholder in any other corporation. If the subsidiary is a bank, a regulator in
the United States is particularly keen on enforcing
the corporate veil.18 To prevent claims on deposit
insurance, the regulator wants to avoid a situation
in which the subsidiary endangers its financial
health by making transfers (sweetheart loans, etc.)
to the parent company. Just like any other shareholder, the parent company can profit from the
subsidiary only via the dividends it receives. In the
case of a domestic branch there is of course no corporate veil. Hence, a bank is fully liable for all of its
branches, at least those within the United States.
When the subsidiary operates abroad, the corporate veil argument suggests that the parent company
is in general not liable for obligations undertaken by
its subsidiaries. In order to obtain payment from the
head office, creditors would have to show that the
corporate veil has been pierced. In recent court
cases—such as the one filed in Spain against BBVA
(Reuters Business Briefing, June 18, 2002)—Argentine
depositors are arguing that the corporate veil
between local subsidiaries and the parent company
was thin. As discussed above, in the United States
the corporate veil is in place to the extent that the
subsidiary presents itself to creditors as a clearly
separate entity from the parent company. Some of
the success of global banks in attracting deposits,
Argentine depositors argue, derived precisely from
the fact that they marketed themselves as being
“safer” than local banks because they have the
backing of the parent company. In times of crises
this backing is the main motivation behind the flight
to quality. Bank advertising tended to stress the
reliability of the corporate name, which further
reassures depositors that their money is secure.19
In the case of foreign branches the distinction
between a branch and a subsidiary is often more
blurred than in the United States. In several countries, such as Argentina, branches of international
banks are essentially treated as separate entities
from the head office by the domestic regulator. For
instance, foreign branches have to meet capital
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
requirements as a separate entity, that is, without
relying on the parent company’s capital.
Most importantly for branches of U.S. banks, section 25C of the Federal Reserve Act (section 326 of
the Riegle-Neal Interstate Banking and Branching
Efficiency Act, codified at 12 U.S. Code section
633) establishes that foreign branches have limited
liability under some circumstances:
A member bank shall not be required to repay
any deposit made at a foreign branch of the bank
if the branch cannot repay the deposit due to
(1) an act of war, insurrection, or civil strife; or
(2) an action by a foreign government or instrumentality (whether de jure or de facto) in the
country in which the branch is located; unless
the member bank has expressly agreed in writing
to repay the deposit under those circumstances.
This law was added in 1994 after Citibank was
sued by depositors at foreign branches in Vietnam
and the Philippines and lost the cases. The Philippine
case is particularly instructive. In 1983 the Philippine
government had confiscated all foreign exchange,
making it impossible for the Manila branch of
Citibank to repay Wells Fargo’s local subsidiary, Wells
Fargo Asia Limited, out of local branch assets. The
court ruled that “Citibank’s worldwide assets were
available for satisfaction of Wells Fargo Asia Limited’s
claims” in spite of the fact that the original contract
did not explicitly state so (see Wells Fargo 1991).
After this and a similar ruling in the Vietnamese case,
U.S. lawmakers sought to protect U.S. banks with
foreign branches from actions by host-country governments. The 1994 law makes it clear that world-
wide assets of a global bank are not in peril if the
foreign branch’s failure to honor its obligations is
the result of a foreign government’s intervention.
Most analysts regard the Argentine case as falling
into this category: The asymmetric conversion of
dollar-denominated banks’ assets and liabilities (see
the next section) and the restrictions on foreign
exchange appear to be clear examples of government interventions.
Hence, the chances of Argentine depositors of
U.S. banks recovering their funds in the United
States are dim. Also, since government intervention
of this sort is not at all rare in the event of a banking crisis, the “news” from Argentina may well be
relevant for other emerging markets as well.20
What would happen in the absence of a foreign
government’s intervention—that is, if section 326 is
not applicable? Consider the following hypothetical
scenario: The Argentine government defaults on its
debt but refrains from the actions discussed
above.21 Under this scenario, for some Argentine
branches or subsidiaries of international banks,
locally held assets would still not suffice to cover
their deposits. If the parent company refuses to
shore up the local branch, can local creditors successfully seek payment from the head office? While
this scenario is only hypothetical, one can argue
that the question is still relevant to the case of
future crises in emerging markets. At least for
branches of U.S. banks, deposit contracts generally
state that depositors can collect their funds only
locally.22 The contracts also state that the bankdepositor relationship is governed by the local
jurisdiction; hence, a U.S. court may refuse to even
consider the case (although such a refusal did not
17. This article does not delve into the issue of cross-border supervision. IMF (2000) summarizes the principles and practices
of cross-border supervision, with particular reference to the Basel Concordat.
18. According to U.S. law, if a bank holding company owns more than one bank subsidiary, each subsidiary is responsible for losses
of other bank subsidiaries owned by the same holding company regardless of whether the corporate veil is in place or not.
19. In the opinion of some analysts, in Argentina “the foreign owners created the illusion that Argentines were depositing their
money into a global financial network. Argentines were told that their money was just as safe as if it was deposited in New
York, Madrid, or Hong Kong” (Molano 2002).
20. Many previous banking crises in Latin America—for instance, the 1989–90 crisis in Argentina—were also characterized by
similar government interventions. One does not need to look far to find evidence of government interventions following large
shocks to the economy. Roosevelt’s actions in the aftermath of the Great Depression—the abandoning of the gold standard,
the Bank Holidays, and the repudiation of the gold clauses—have close parallels with the Argentine government’s actions
during the current crisis (although the Argentine government imposed a different conversion rate for banks’ assets and liabilities). Kroszner (1999) argues that the repudiation of the gold clauses—which is the equivalent of the Argentine’s government conversions of all dollar loans into pesos—was actually perceived as a beneficial action by financial markets.
Needless to say, the Argentine government was not as successful.
21. The government debt’s default could also be considered a form of government intervention. Whether this is the case from a legal
perspective, from an economic point of view it is a very different action from, say, a forced conversion of assets: debt holders are
fully aware of the possibility that the debt issuer might default and ask for a risk premium as a compensation for the possibility.
22. Of course, this stipulation applies to the extent that the local branch has enough funds to meet its liabilities. To our knowledge, the contracts do not explicitly state what would happen in case of liquidation of the branch.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
11
occur in the court cases mentioned above).23 A local
court may well demand that the parent company
honor its obligations, but the court may have little
power of enforcement. In conclusion, it is not clear
that creditors of branches or subsidiaries can successfully attach the parent company’s assets even in
the absence of outright government intervention.
Bad News from Argentina
efore the current crisis, the Argentine banking
system was hailed as a success story for Latin
America (see Kiguel 2002): A 1998 World Bank study
rated Argentina’s regulatory regime among the top
three in emerging markets (see Calomiris and Powell
B
Before the current crisis, the Argentine
banking system was hailed as a success
story for Latin America.
2000). Because Argentina was under a currency
board (so-called convertibility) regime, the central
bank was, by law, severely restricted in its role as a
lender of last resort.24 Hence, the regulator had to
make sure that the banking system could stand on
its own. To achieve this goal, policymakers pursued
a two-tier strategy. The first tier consisted of
strengthening prudential regulation (see Calomiris
and Powell 2000 for an insightful description of
Argentine regulatory approach). Capital requirements were stricter than those imposed by the
Basel Committee: The capital asset ratio was set
at 11.5 percent as opposed to the 8 percent level recommended by the Basel Committee (Kiguel 2002).
Furthermore, capital requirements were adjusted
depending on the CAMEL rating of the bank.25
To better assess the riskiness of financial institutions, the regulator also required banks to issue subordinated debt for an amount equivalent to 2 percent
of deposits (although foreign banks with good credit
ratings did not have to comply) and to be monitored
by international credit rating agencies. Banks were
also subject to liability requirements—that is, reserve
requirements for all liabilities (not only for deposits),
depending on their maturity. Liability requirements
amounted to about 30 percent of the system deposits
(Caprio and Honohan 1999). Indicative of the regulator’s faith in foreign financial institutions is the fact
12
that as much as 80 percent of the liquidity requirement could be fulfilled by holding balances at qualifying foreign banks, possibly abroad.
Deposit insurance, which had been abolished in
1992, was reinstated in 1995 during the Tequila
Crisis—albeit with a limited scope—with the purpose of strengthening depositors’ confidence in the
banking system. Deposit insurance was funded via a
premium on banks that varied from 0.015 to 0.06
percent of deposits and was implemented via an
entity (Seguro de Depositos Sociedad Anonima,
or SEDESA) that by law could not rely on resources
from either the central bank or the Treasury. The
scheme covered only deposits up to $30,000 and in
principle should have been endowed with enough
resources to cover 5 percent of deposits. By the end
of 2001, however, the fund had only $270 million,
which covered about 0.4 percent of all deposits. A
key feature of the scheme, particularly in light of
what was to follow, was that it could invest up to 50
percent of its assets in government bonds (Sistema
de Seguro 2002).
The second tier of the strategy consisted in welcoming foreign banks’ entry, especially in the aftermath of the Tequila Crisis. Argentina quickly became
one of the first countries in Latin America with
substantial foreign bank presence. Finally, the
central bank set up a contingent credit line with
international banks—a partial substitute for the
lack of a lender of last resort. The Argentine financial system’s ability to withstand the Tequila Crisis
without major losses, in spite of large shocks to
deposits (Kiguel 2002), and to weather successfully
the East Asian, Russian, and Brazilian crises seemed
to suggest that Argentina had found the avenue to
banking system stability.26
Of course, the current crisis changes the picture
considerably. The Argentine economy unraveled in
2001, culminating with the collapse of the convertibility plan that had linked the peso to the dollar at
parity (see Box 2 for a brief chronology of the
Argentine crisis). The default on government debt
in December 2001 had devastating consequences
for the banking system as a sizable portion of bank
assets (about 21 percent in October 2001) was in
government liabilities.27 In November 2001 the government induced the banks to “voluntarily” swap
government bonds for illiquid government liabilities, prompting large deposit withdrawals: Deposits
fell 24 percent by the end of the year. In the final
days of the De la Rua government only a freeze on
deposits could prevent a widespread bank run.28 In
January 2002 convertibility ended and the peso
underwent a large devaluation. By government
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
decree, in February 2002 all dollar-denominated
loans were converted to pesos at one to one while
dollar-denominated deposits were converted at 1.4
pesos per dollar. According to Moody’s, the banking
system’s losses as a result of the crisis could reach
$54 billion. Deposit insurance quickly ran out of
funds (Para Scotia 2002): In February 2002 a presidential decree revised the deposit guarantee law to
allow for compensation of depositors via nontradable government securities.29 Given that a sizable
fraction of deposits (72 percent by December 2001)
was dollar-denominated, the central bank could
hardly intervene as a lender of last resort. By early
2002 international banks were ready to leave the
country or at least threatening to do so (Ryst 2002).
Given the Argentine government’s heavy-handed
intervention in the banking system, the analysis of
the previous section suggests that at least for U.S.
banks the parent company may not be liable for the
local branches.
As the crisis unraveled, some of the supposed
benefits of international banks did not quite materialize as expected. As mentioned above, one of the
main advantages of international banks is that their
portfolios are well-diversified and hence can withstand a localized crisis. This advantage was indeed
true for many of the international banks involved in
the Argentine crisis. However, banks’ local branches
and subsidiaries, when considered as stand-alone
entities, had portfolios that were by and large just as
vulnerable as those of domestic banks to the shocks
that hit the economy, like the government’s default.30
To the extent that international banks could walk
away from the subsidiaries’ liability, from the deposi-
tors’ perspective the local branches or subsidiaries
of international banks were indeed stand-alone
entities. Interestingly, the data suggest that this fact
seemed to be understood by Argentine depositors—
although this specific question certainly deserves a
much deeper analysis than the one undertaken
here. Chart 4 seems to indicate that little or no
flight to quality took place as the crisis developed
during 2001 except in the very last months. The
observed flight to quality was specifically toward
branches of foreign banks perhaps because of their
lower exposure to government liabilities.31
In summary, the bad news from Argentina is that
even a sizable presence of global banks may not be
enough to protect depositors from the occurrence
of a banking crisis. This article argues that one of
the reasons why this is the case is that under some
circumstances—and most likely under the circumstances that developed in Argentina following
heavy government intervention in the banking system—international banks are shielded from their
liabilities. In other words, they may not be legally
compelled to recapitalize Argentine branches or
subsidiaries. As we write, only a few foreign banks
(Credit Agricole, Scotiabank) have explicitly abandoned their Argentine branches or subsidiaries. To
the extent that Argentine taxpayers will assume at
least part of those liabilities or that depositors will
be forced into accepting a subpar compensation for
their funds, some foreign banks may decide to stay
in the end.32 Negotiations are under way. In these
negotiations, a key factor affecting foreign banks’
bargaining power has to do with reputation. On the
one hand, a default in Argentina may harm the
23. Interestingly, the court’s motivation was as follows: “If the goal is to promote certainty in international financial markets,
it makes sense to apply New York law uniformly, rather than conditioning the deposit obligations to the vagaries of local
law.…” (Wells Fargo 1991).
24. The 1992 central bank charter barred the central bank from offering either implicit or explicit guarantees for bank liabilities
to the extent that these guarantees were backed by fiscal funds (see Schumacher 2000). The central bank was, however,
able to extend repos and rediscounts to financial intermediaries, albeit under restrictions, and to change the reserve requirements. During the 1995 Tequila Crisis the central bank used both instruments in order to weather the crisis (see Calomiris
and Powell 2000).
25. The CAMEL score is a measure of the financial health of a bank.
26. Schumacher (2000) reports that by December 1995 nine banks had failed, and thirty had been acquired or merged, out of
a total of 137 private banks.
27. For subsidiaries of foreign banks the exposure to the government was also around 20 percent. For branches of foreign banks,
however, the corresponding figure was much lower—around 10 percent.
28. The freeze on deposits is still in place as this article is written.
29. The deposit law guarantee now states that these securities cannot be endorsed; depositors would have to hold them to maturity. See Sistema de Seguro (2002).
30. However, as discussed above, Crystal, Dages, and Goldberg (2001) find that foreign banks’ portfolios were in general marginally sounder than those of domestic banks. See also footnote 27.
31. Martinez Peria and Schmukler (2001) study the extent to which depositors discipline banks in Latin American countries.
32. In the 1989 crisis bank deposits were replaced with bonds that traded at a large discount; the swap was known as the Bonex
plan. A similar plan, known as Bonex II, is currently been considered by the authorities.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
13
BOX 2
Chronicle of the Argentine Crisis
Events
14
EMBI+ bond spread
over U.S. Treasuries
(at end of period)1
October 2000
Confidence erodes after Vice President Carlos
Alvarez resigns.
815
December 2000
The IMF leads a $39.7 billion three-year rescue
package.
773
January 2001
Capital returns to the country, central bank
reserves increase $1.3 billion, and deposits
increase $1.2 billion.
663
February 2001
Allegations of malfeasance are made against central bank President Pedro Pou. The Turkish crisis
begins.
803
March–April 2001
Economy Minister Jose Luis Machinea resigns.
His replacement, Ricardo López-Murphy, holds
office less than two weeks. Domingo Cavallo
takes over. Devaluation fears grow after the
Convertibility Law is altered to eventually link the
peso with the dollar and the euro.
1,042
June 2001
Argentina completes a $29.5 billion debt swap.
1,025
July 2001
Sharp falls in deposits occur, and bond spreads
widen. Congress approves a zero deficit law calling for the immediate cut of the fiscal deficit
through budget cuts and tax hikes. Salaries and
pensions over $500 are cut by 13 percent.
1,599
August–September 2001
New fiscal austerity measures are enacted. The
announcement of an IMF assistance package calms
default fears. Unemployment is at 17.2 percent.
The IMF announces up to $8 billion of additional
loans ($5 billion available immediately and $3 billion available later depending on future reforms).
1,595
October 2001
Opposition Peronists win in legislative elections.
2,136
November 2001
The government announces a new, ostensibly voluntary, debt swap of as much as $16 billion in
high-yield government bonds held by local banks
and pension funds for securities that pay lower
interest but are guaranteed by tax revenue. The
IMF endorses the swap. Sovereign bond spreads
widen. A sharp decline occurs in deposits. Tax
revenue drops, and the zero fiscal deficit plan
becomes clearly unsustainable.
3,340
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
Events
EMBI+ bond spread
over U.S. Treasuries
(at end of period)1
December 2001
2001 deposits fall from $85 billion to $64.6 billion. 2001 GDP falls 3.9 percent. Restrictions on
deposits are imposed in the wake of the run on
deposits. Withdrawals are limited to 250 pesos
per week (later raised to 1,200 pesos per month).
Violent protests occur. Domingo Cavallo and
Fernando De la Rua resign. Interim President
Rodriguez Saá announces a moratorium on foreign debt.
4,404
January–February 2002
President Eduardo Duhalde is sworn in, and the
convertibility law ends. A dual exchange rate is
announced in January, and a floating exchange
rate is introduced in February. Bank assets are converted to pesos at 1 to 1; liabilities are converted at
1.4 to 1. The banking system is in crisis because of
a currency mismatch, a decline in value of asset
portfolios, and losses from holdings of $30 billion in
government debt. Converting dollar loans at parity
could generate losses up to $18 billion for the banking sector. The government announces that bank
losses will be partially compensated by issuing
bonds and indexing loans to inflation.
4,098
March–April 2002
The largest private bank, Banco Galicia, receives
an $800 million bailout from the central bank and
fifteen local banks. Foreign banks postpone decisions on recapitalization. The central bank intervenes in a Scotiabank subsidiary. The government
continues to negotiate with the IMF. Economy
Minister Remes Lenicov resigns after mandatory
bonds-for-deposit swaps are rejected. He is
replaced by Roberto Lavagna.
4,831
May 2002
Scotiabank (Canada) and Credit Agricole (France)
plan to sell or close their Argentine units. Societe
Generale (France) agrees to recapitalize its
Argentine unit.
6,123
June 2002
Voluntary deposit-for-bond swaps are announced.
$9.5 billion in ten-year dollar bonds is to be provided to banks to compensate for losses associated
with the devaluation and currency mismatch.
Negotiations with the IMF are set to resume.
6,791
1. Bond spreads are from JP Morgan’s Emerging Market Bond Index (EMBI+) for Argentina.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
15
CHART 4
Deposit Evolution of Foreign and Domestic Banks in 2001
60
Total foreign banks
50
Total domestic banks
40
Pe r c e n t
Foreign-owned banks
State-owned banks
30
20
Local private banks
Foreign bank branches
10
Jan.
Feb.
March
April
May
June
July
August
Sept.
Oct.
Nov.
Dec.
Source: Banco Central de la República Argentina
position of these international banks in other
emerging markets. On the other hand, an unconditional recapitalization of local branches could
induce other emerging-market governments to
believe that foreign banks may always pick up the
bill for their lack of fiscal discipline.
Conclusions
here are interesting similarities between the
policy debate that took place in the United
States in the 1980s and early 1990s with regard to
interstate branching and the current debate in
emerging markets on international banks (see IMF
2000). Proponents of interstate branching in the
United States saw the gains in efficiency from
increased competition and the increased stability
due to wider portfolio diversification as the two
main benefits from lifting restrictions. Opponents
claimed that out-of-state branches would draw
funds away from local markets and neglect local
small businesses.33 Likewise, opponents of foreign
banks’ entry into developing markets claim that
these banks neglect lending to small enterprises
and may amplify credit rationing in times of crisis.
In contrast, proponents of foreign banks emphasize
the benefits to be gained from efficiency and portfolio diversification. This article documents that the
empirical literature by and large sides with the pro-
T
16
ponents of global banks’ entry. Many of the arguments against international banks do not seem to
find empirical support.
This article focuses mainly on the issue of banking systems’ stability during a crisis, specifically on
the following claim, as summarized in an IMF
report: “It has been suggested that foreign banks
can provide a more stable source of credit and can
make the banking system more robust to shocks.
The greater stability is said to reflect the fact that
the branches and subsidiaries of large international
banks can draw on their parent for additional funding and capital when needed. In turn, the parent
may be able to provide such funding because it will
typically hold a more internationally diversified
portfolio than domestic banks, which means that its
income stream will be less correlated with purely
domestic shocks” (IMF 2000, 163).34
The discussion points out that, at least under some
circumstances, international banks may not be fully
liable for the obligations of their foreign branches or
subsidiaries. Because of this limited-liability feature,
local depositors may not reap the full benefits from
portfolio diversification offered by the presence of
foreign banks. During crises, and especially in cases of
crises-cum-government-intervention, the branch or
subsidiary may default and depositors may not be
able to make claims against the parent company’s
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
assets. Hence, under such circumstances, the greater
portfolio diversification of international banks is of no
avail to local depositors.
These arguments, especially in light of recent
events in Argentina, suggest that international
banks’ presence is not a panacea against banking
crises. But it is important to note that this argument should not be taken as an argument against
foreign banks’ entry. First, it is not clear that a
financial system closed to foreign banks would be
any better. Past crises in Latin America strongly
suggest that it would not. Second, the literature
has pointed out a number of other important benefits from foreign banks’ entry. Third, it is not clear
that a priori the limited-liability feature of foreign
banks reduces welfare. One may argue that the
limited-liability feature of foreign banks increases
the cost of financial crises for governments and
thus may induce governments to pursue policies
that avoid crises. Finally, in the absence of this feature, the expansion of international banks might
not have occurred in the first place. At the same
time, however, it is not clear that all the incentives
generated by the limited liability feature are in the
right direction. To the extent that local depositors
are unaware of international banks’ limited liability,
these banks have an incentive to borrow locally and
invest in high-yield government securities: The
limited-liability feature, if it applies, covers international banks from the risk of government default.
More work at both the theoretical and empirical
level is needed to investigate these issues.
33. See Jackson and Eisenbeis (1997) for an empirical refutation of the first point.
34. Note that the IMF report does not necessarily endorse these views.
REFERENCES
Barajas, A., R. Steiner, and N. Salazar. 2000. Foreign
investment in Colombia’s financial sector. In The internationalization of financial services: Issues and lessons
for developing countries, edited by S. Claessens and
Marion Jansen. Boston: Kluwer Academic Press.
agenda for further research. World Bank Policy Research
Working Paper 2698, October.
———. 2002. Bank lending to small business in Latin
America: Does bank origin matter? World Bank Policy
Research Working Paper 2760, January.
Berger, A.N., L.F. Klapper, and G.F. Udell. 2000. The
ability of banks to lend to informationally opaque businesses. Photocopy.
Cox, James D., Thomas L. Hazen, and F. Hodge O’Neal.
1997. Corporations. New York: Aspen Publishers.
Calomiris, Charles W., and Andrew Powell. 2000. Can
emerging market bank regulators establish credible discipline? The case of Argentina, 1992–1999. National
Bureau of Economic Research Working Paper 7715, May.
Crystal, Jennifer S., B. Gerard Dages, and Linda S.
Goldberg. 2001. Does foreign ownership contribute to
sounder banks in emerging markets? The Latin American experience. Federal Reserve Bank of New York.
Photocopy.
Caprio, Gerald, Jr., and Patrick Honohan. 1999. Beyond
capital ideals: Restoring banking stability. World Bank
Policy Research Working Paper 2235, November.
Dages, B. Gerard, Linda Goldberg, and Daniel Kinney.
2000. Foreign and domestic bank participation in emerging markets: Lessons from Mexico and Argentina. Federal
Reserve Bank of New York Economic Policy Review 6
(September): 17–36.
Caprio, Gerald, Jr., and Daniela Klingebiel. 1996. Bank
insolvencies: Cross-country experience. World Bank
Policy Research Working Paper 1620, July.
Demirgüc-Kunt, Asli, and Enrica Detragiache. 1997. The
determinants of banking crises: Evidence from developing
and developed countries. International Monetary Fund
Working Paper 97/106, September.
Claessens, Stijn, Asli Demirgüc-Kunt, and Harry
Huizinga. 1998. How does foreign bank entry affect the
domestic banking market? World Bank Policy Research
Working Paper 1918, June.
Clarke, George, Robert Cull, Laura D’Amato, and Andrea
Molinari. 2000. On the kindness of strangers? The impact
of foreign entry on domestic banks in Argentina. In The
internationalization of financial services: Issues and
lessons for developing countries, edited by S. Claessens
and Marion Jansen. Boston: Kluwer Academic Press.
Clarke, George, Robert Cull, Maria Soledad Martinez
Peria, and Susana M. Sanchez. 2001. Foreign bank entry:
Experience, implications for developing countries, and
Demirgüc-Kunt, Asli, and H. Huizinga. 2000. Determinants
of commercial bank interest margins and profitability:
Some international evidence. World Bank Economic
Review 13 (May): 379–408.
Demirgüc-Kunt, Asli, Ross Levine, and H.G. Min. 1999.
Opening to foreign banks: Issues of stability, efficiency,
and growth. In Proceedings of the Bank of Korea 1998
Conference on the Implications of Globalization of
World Financial Markets. Seoul: Bank of Korea.
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
17
Focarelli, Dario, and Alberto Pozzolo. 2000. The determinants of cross-border shareholding: An analysis with
bank-level data from OECD countries. Banca d’Italia
Temi di Discussione del Sevizio Studi no. 381.
Levine, Ross. 2001. International financial liberalization
and economic growth. Review of International
Economics 9 (November): 688–702.
Goldberg, Linda. 2001. When is U.S. bank lending to
emerging markets volatile? Federal Reserve Bank of
New York. Photocopy.
Martinez Peria, Maria Soledad, and Sergio Schmukler.
2001. Do depositors punish banks for bad behavior?
Market discipline, deposit insurance, and banking crises.
Journal of Finance 56 (June): 1029–49.
Hamilton, Robert W. 1991. The law of corporations in a
nutshell. 3d ed. St. Paul, Minn.: West Publishing Co.
Molano, Walter T. 2002. The Latin American advisor.
BCP Securities, February 6.
Hasan, I., and William Hunter. 1996. Efficiency of Japanese
multinational banks in the United States. In Research
in finance, vol. 14, edited by Andrew H. Chen, 157–73.
Greenwich, Conn., and London: JAI Press.
Para Scotia, el sur no existe (Para Scotia). 2002. Página
12, May 8. <www.pagina12.com.ar> (May 8, 2002).
International Monetary Fund Staff Team (IMF), led by
Donald J. Mathieson and Garry J. Schinasi. 2000. International capital markets: Developments, prospects, and
key policy issues. Washington, D.C.: IMF Publication
Services.
Kane, Edward J. 2000. Capital movements, banking
insolvency, and silent runs in the Asian financial crisis.
Pacific-Basin Finance Journal 8 (May): 153–75.
Jackson, William E., III, and Robert A. Eisenbeis. 1997.
Geographic integration of bank deposit markets and
restrictions on interstate banking: A cointegration
approach. Journal of Economics and Business 49,
no. 4:335–46.
Kane, Edward J., and Asli Demirgüc-Kunt. 2001. Deposit
insurance around the globe: Where does it work?
National Bureau of Economic Research Working Paper
8493, September.
Kiguel, Miguel A. 2002. The Argentine financial system
in the last decade. In Proceedings of the Federal
Reserve Bank of Atlanta 2001 Conference on
Domestic Finance and Global Capital in Latin
America. Atlanta: Federal Reserve Bank of Atlanta.
<www.frbatlanta.org/econ_rd/larg/events/conf2001/
kiguel_paper.pdf> (July 24, 2002).
Peek, Joe, and Eric S. Rosengren. 2000a. Collateral damage: Effects of the Japanese bank crisis on real activity
in the United States. American Economic Review 90,
no. 1:30–45.
———. 2000b. Implications of globalization of the banking sector: The Latin American experience. New England
Economic Review (September/October): 45–62.
Ryst, Sonja. 2002. Argentine banks facing grim outlook
amid economic plans. Dow Jones Newswires, February 5.
Salomon Smith Barney. 2001. Foreign financial institutions in Latin America. 2001 Update. November 28.
Schumacher, Liliana. 2000. Bank runs and currency runs
in a system without a safety net: Argentina and the
“tequila” shock. Journal of Monetary Economics 46
(August): 257–77.
Sistema de Seguro de Garantía de los Depósitos
Bancarios (Sistema de Seguro). 2002. Ley 24.485.
<www.bcra.gov.ar/pdfs/marco/sedesa.pdf> (July 16, 2002).
Stockman, Alan C. 2001. Optimal central bank areas, financial intermediation, and Mexican dollarization. Journal of
Money, Credit, and Banking 33 (May): 648–68.
Tallman, Ellis W., and Nargis Bharucha. 2000. Credit
crunch or what? Australian banks during the 1986–93
credit cycle. Federal Reserve Bank of Atlanta Economic
Review 85 (Third Quarter): 13–33.
Kroszner, Randall S. 1999. Is it better to forgive than
receive? Repudiation of the gold indexation clause in longterm debt during the Great Depression. Paper presented
September 2 at the Federal Reserve Bank of Atlanta.
Wall, Larry D. 1993. Too-big-to-fail after FDICIA. Federal
Reserve Bank of Atlanta Economic Review 78 (January/
February): 1–14.
Kyle, Albert S., and Wei Xiong. 2001. Contagion as a
wealth effect. Journal of Finance 56 (August): 1401–43.
Wells Fargo Asia Ltd. v. Citibank N.A. 1991. 936 F.
2nd 723 (2nd Cir.).
18
Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W Third Quarter 2002
@FedFRASER