How can one explain cross-national differences in innovative activity
across the industrialized democracies? Politics appear to play a strong
causal role here, with case study after case study showing the clear
influence of politics and political institutions on technological
innovation.1
However, this
phenomenon is only sparsely studied by political scientists. Rather,
this area has largely become the purview of a small number of
economists and sociologists who often ignore important political
variables in their analysis. Thus great interest has recently been
generated by a new “varieties of capitalism” (VOC) theory
of innovation which holds that variance in political institutions is
the primary cause of differences in national innovative behavior. In
brief, the central claim of VOC's innovation theory is that the
more a polity allows the market to structure its economic
relationships, the more the polity will direct its inventive activity
toward industries typified by “radical” technological
change. Conversely, the more a polity chooses to coordinate economic
relationships via nonmarket mechanisms, the more it will direct its
inventive activity toward “incremental” technological
change.
This question, of why some countries are more technologically
innovative than others, should interest scholars of international
political economy for several reasons. For example, even among
otherwise friendly nations, economic rivalries between states can often
come to resemble military ones, with competition over trade, jobs, and
markets leading to interstate disputes and strained relations.2
Scholars of security studies will also
appreciate the importance of civilian technological innovation and its
role in production and the general economy as complementary, if not
foundational, to relative military power. See Samuels 1994.
In this competitive
environment, technological innovation is a means not just for wealth
creation but also for economic security; innovation provides the new
products, new processes, and increased efficiencies that are the
driving force behind modern economic growth, relative industrial power,
and competitive advantage.
3 In recognition of this, almost every
industrialized society expends a considerable share of its resources on
the pursuit of technological advance. Yet, despite the random nature of
innovation, and the seemingly clear fiscal and policy requirements for
promoting innovative behavior, some countries are consistently more
successful than others at technological progress, even among the
industrialized democracies. This presents an increasingly nettlesome
puzzle for social scientists.
Furthermore, VOC scholars see innovation theory as a key to resolving
current problems in understanding global trade flows and production
patterns. Classic trade theory holds that free trade will not deplete
national wealth by impelling production abroad but will instead enhance
economic performance and increase each trader's consumption
possibilities. In this basic model, societies specialize production in
their most efficient sectors and then trade the surplus for more goods
than they otherwise could have produced locally. The Heckscher-Ohlin
model improves on this basic theory by arguing that nations'
relative endowments of basic economic factors (land, labor, capital)
should determine the general lines along which international production
and trade are structured. However, VOC proponents point to the rise of
intra-industry trade during the past thirty years that has contradicted
the interindustry trading patterns predicted by the Ricardian or
Heckscher-Ohlin models. Instead of specializing in particular sectors
of production, the industrialized nations have maintained a broad
spectrum of general economic activity and instead have concentrated
their sectoral productive efforts geographically. Recent attempts to
explain these phenomena posit an initially random distribution of
productive activity that is then followed by agglomeration because of
either increasing returns to scale or network externalities.4
VOC scholars generally accept these agglomeration
arguments, but they identify certain nonrandom patterns of
international production and trade that are neither explained nor
predicted by current agglomeration theories.
5 VOC's innovation theory offers a resolution to both of the
anomalies above, suggesting that domestic institutional structures can
account for the different degrees of innovative effort and achievement
between nations, and the production and trade relationships that
subsequently develop. If VOC theory is correct, it would explain why
nations maintain their innovative profiles in spite of strong pressures
to change them, and why certain kinds of innovation-dependent
production might tend to be concentrated in particular countries.
However, the central claim made by VOC's innovation theory has yet
to be proven. The purpose of this article is to use new data on patents
and scholarly publications to test VOC theory's central
assumptions and predictions and to see whether VOC theory properly
describes the empirical world of technological innovation. I
demonstrate that VOC theory does not accurately predict innovative
behavior over time and space, and that VOC's existing empirical
support strongly depends on the inclusion of a major outlier, the
United States, in the set of radically innovative countries. I also
find that some industries are more radically innovative than others in
the short run, as assumed by VOC theory, but that this characterization
cannot be confirmed in the long run as industries age and mature
technologically.
Politics, Economics, and Innovation Theory
For much of the history of political economy, questions about the
causes of national differences in technological innovation have
remained at the periphery of the field.6
Technology is defined as a physical product, or a
process of handling physical materials, which is used as an aid in
problem solving. More precisely, technology is a product or process
that allows social agents to perform entirely new activities or to
perform established activities with increased efficiency.
Innovation is the discovery, introduction, or development of
new technology, or the adaptation of established technology to a new
use or to a new physical or social environment.
One of the
major reasons for this was the apparently random, or at least
inexplicable, nature of innovation itself; even those social scientists
who attempted to deal systematically with technological change
(including Marx, Schumpeter, and Solow) generally regarded it, and the
underlying body of scientific knowledge on which it drew, as a
“black box” proceeding according to its internal processes
largely independent of political or economic forces.
7 This attitude changed
gradually during the Cold War, as vast expenditures by the U.S.
government and industry on research and development (R&D) made it
increasingly clear that technological innovation could be made
responsive to economic and political needs, a fact further punctuated
by the Soviet launch of Sputnik and later by the Japanese and German
economic “miracles.” In response, economists during the
1960s began to investigate whether certain supply-side or demand-side
variables could explain why even developed nations followed different
technological trajectories.
8 This somewhat
inconclusive debate was followed in the late 1970s and 1980s by a plethora
of case and country studies that tended to emphasize the importance of this
or that policy, these or those historical conditions, but failed to produce
any generalizable theory about the rate or direction of national
innovation.
A recurring problem encountered in these debates was the
contradiction between empirical observation and certain fundamental
tenets of the economics of science. Specifically, Arrow had shown that
much productive knowledge takes the form of unpatentable laws of nature
and advances in basic science, and is therefore a nonexcludable public
good available to everyone without charge.9
While
patents and trade secrets act as temporary solutions to this
appropriability problem in the area of applied knowledge, history has
shown that the original inventors of technology often do not capture
most of the benefits of their innovations when these inventions are
transferred across borders, and that these transfers take place even in
spite of considerable efforts to stop them. Theoretically speaking
then, in the long run, developed nations should not display significant
variation in either per capita innovation rates or in the type of
innovative activities that they pursue. Yet differences appear to
abound.
One possible solution to this paradox focuses on institutions.
Institutions are perhaps the only variables that both influence the
incentives for innovative behavior and differ across nations. Indeed,
political scientists and economists have long recognized the capacity
of government, labor, regulatory, and legal institutions to inhibit
free market exchange and thereby hamper innovation. But it was not
until Romer endogenized technological change that social scientists
began to take seriously the ability of institutions to actively enhance
aggregate economic performance through their effects on the rate and
direction of technological progress.10
To date
though, beyond the broadest brushstrokes of political-economic theory,
social scientists have yet to pinpoint the specific mechanisms by which
institutions cause countries to differ technologically.
It is into this environment that VOC theory makes its foray, taking a
radical new approach to explaining cross-national differences in the
direction of technological progress.11
I
am concerned here specifically with those aspects of VOC theory
discussed in Hall and Soskice 2001,
1–44.
VOC theory is broad and foundational; it touches
on multiple aspects of political and economic life, of which innovation
is but one part. At its most basic level, it is a theory of capitalism
by gradation: some countries use markets more than others to coordinate
economic actors and this variation is used to explain a myriad of
comparative and international political-economic behavior. However,
when fully articulated, VOC theory does not divide the world into
“free-trade versus protectionist” or “state-owned
versus privatized” systems of political economy as is
traditionally done. This approach would focus attention on the state,
which VOC scholars wish to avoid. Rather, they view the firm as the
locus of trade and production in the capitalist economy and, therefore,
take the firm, not the state, as their primary unit of analysis. Nor is
the firm a lone or independent actor in VOC's analysis; successful
operation of the firm depends heavily on its relationships with labor,
investors, and other firms. It is these crucial relationships that, in
turn, explain patterns of economic activity and policymaking.
Therefore, the central claims of VOC theory focus on how a given
political-economic institutional structure determines the conduct of
these crucial relationships and how economic actors organize to solve
the classic coordination problems that afflict such relations. At one
end of this relationship spectrum lie the “liberal market
economies” (LMEs), such as the United States, in which firms tend
to coordinate their relations and activities in the manner described by
Williamson: through internal corporate hierarchies and external
competitive market arrangements.
12 At the other end of the spectrum sit
the “coordinated market economies” (CMEs), such as Germany,
in which firms tend to coordinate via nonmarket relationships, with
greater dependency on relational and incomplete contracting, exchanges
of private information within enduring networks, and a high degree of
actor collaboration (as opposed to competition or confrontation). As I
show in the next section, these distinctions have important
implications for explaining and predicting national differences in
innovation.
Varieties of Capitalism's Theory of
Technological Innovation
According to VOC theory, technological innovation comes in two types,
radical and incremental, each of which forms the basis for a different
mode of production. While an exact definition is elusive, VOC scholars
describe radical innovation as that which “entails substantial
shifts in product lines, the development of entirely new goods, or
major changes to the production processes.”13
They argue that radical innovation is therefore
vital to production in high-technology sectors that require rapid and
significant product changes (biotechnology, semiconductors, software)
or in the manufacture of complex systems-based products
(telecommunications, defense, airlines). Incremental innovation, on the
other hand, is that which is “marked by continuous but
small-scale improvements to existing product lines and production
processes.”
14
Unlike production based on radical innovation where speed and
flexibility are crucial, production based on incremental innovation
prioritizes the maintenance of high quality in established goods. This
approach to innovation involves constant improvements in manufacturing
processes to bring down costs and prices, but only occasional minor
improvements in the product line. Incremental innovation is therefore
essential for competitiveness in capital goods production (machine
tools, factory equipment, consumer durables, engines).
VOC theory further predicts that LMEs and CMEs will tend to exert
greater effort toward, and be successful in, different types of
technological innovation. VOC theory interprets innovation as just
another productive activity; therefore, innovation should be sensitive
to the firm's crucial relationships described above and the
institutions that structure them. This does not mean that a given
political-economic structure will result in only one kind of
innovation, but that different institutions will create different types
of comparative advantage for innovators. For example, incremental
innovation requires a workforce that is skilled enough to come up with
innovation, secure enough to risk suggesting it, and autonomous enough
to see it as a part of their job. This in turn requires that firms
provide workers with secure environments, autonomy in the workplace,
opportunities to influence firm decisions, education and training
beyond just task-specific skills (preferably industry-specific
technical skills), and close interfirm collaboration that encourages
clients and suppliers to suggest innovations as well. These are exactly
the kinds of apparatus provided by CME institutions. In fact, CMEs are
defined by the very institutions that provide a comparative advantage
for incremental innovation. These institutions include highly
coordinated industrial-relations systems; corporate structures
characterized by works councils and consensus-style decision making; a
dense network of intercorporate linkages (such as interlocking
corporate directorates and cross-shareholding); systems of corporate
governance that insulate against hostile takeovers and reduce
sensitivity to current profits; and appropriate laws for
relationship-based, incomplete contracting between firms. VOC scholars
argue that this combination of institutions results in long employment
tenures, corporate strategies based on product differentiation rather
than intense product competition, and formal training systems for
employees that focus on high skills and a mix of company-specific and
industry-specific skills; in other words, the very factors that combine
to foster incremental innovation.
On the other hand, VOC scholars argue that these same CME
institutions that provide comparative advantages for incremental
innovation also serve as obstacles to radical innovation. For instance,
worker representation in the corporate leadership combines with
consensus-style decision making to make radical change and
reorganization difficult. Also, long employment tenures make the
acquisition of new skills and rebalancing a company's labor mix
difficult. Dense intercorporate networks also make the diffusion of
disruptive innovations slow and arduous, and technological acquisition
by mergers and acquisitions or takeovers hard. All of these act
against, or reduce the potential rewards of, radical innovation.
In LMEs, the situation is reversed. LMEs are defined by institutions
that provide a comparative advantage for radical innovation, while
creating obstacles to incremental innovation. LMEs have flexible labor
markets with few restrictions on layoffs, which means that companies
can drastically change their product lines and still acquire the proper
labor mix. LMEs also support extensive equity markets with dispersed
shareholders providing innovators of all sizes with relatively
unfettered access to capital. Also, interfirm relations in LMEs allow
for a variety of aggressive asset exchanges with few restrictions on
mergers and acquisition, buyouts, personnel poaching, licensing, and so
on, which permits firms to easily acquire scientific expertise and new
technology. Concentration of power at the top of LME-based firms
augments these institutions, allowing management to force major change
quickly on complex organizations. All of these factors combine to
create large incentives for, and an environment accommodative to,
radical innovation. Conversely, LMEs' capacity for incremental
innovation is limited because of financial arrangements that emphasize
current profitability, corporate structures that concentrate unilateral
control at the top and eliminate workforce security, and antitrust and
contract laws that discourage interfirm collaboration in incremental
innovation. Meanwhile, fluid labor markets and short job tenures
motivate workers to pursue selfish career goals and to acquire mobile
general skills rather than firm-specific or industry-specific skills.
Hence, in VOC's analysis, neither workers nor firms in LMEs tend
to have the incentives or the resources for sustained incremental
innovation.
Testing the Varieties of Capitalism Claims
The purpose of the remainder of this article is not to evaluate the
accuracy of the LME-CME classification system or test a specific causal
mechanism involved in VOC's theory of innovation. Rather, the
question I ask here is whether the international patterns of innovation
that VOC theory predicts actually exist. The VOC causal story outlined
above is both theoretically appealing and dovetails with some widely
held stereotypes about national differences in innovation; however,
little empirical data has yet been produced to support its central
claim. The evidence offered by Hall and Soskice consists of four years
of patent data from the European Patent Office (EPO) that shows that
Germany and the United States concentrate their patents according to
the LME versus CME model discussed above. Specifically, Hall and
Soskice examine patenting activity by Germany and the United States in
thirty technology classes during 1983–84 and 1993–94 (Figure 1). Overall, they found that Germany's
patent specialization was almost equal and opposite that of the United
States in both time periods.15
Hall and
Soskice's methodology will be discussed in greater detail
below.
More specifically, the Germans were found to be more
active innovators in industries that Hall and Soskice characterize as
dominated by incremental innovation (such as mechanical engineering,
product handling, transport, consumer durables, and machine tools);
meanwhile, firms in the United States innovated disproportionately in
industries that the authors perceive as more radically innovative
(including medical engineering, biotechnology, semiconductors, and
telecommunications).
Patent specialization by technology class
I have identified several possible problems with this approach.
First, VOC theory implicitly assumes that some industries are
inherently characterized by radical innovation, others by incremental
innovation, and that these industries have been correctly identified.
Second, in supporting their claims, Hall and Soskice use only four
years' worth of patent data from only two countries, one of which,
the United States, is an outlier by almost any measure. Third, Hall and
Soskice use only simple patent counts as their measure of innovation,
hence frivolous patents are counted the same as highly innovative ones;
nor do Hall and Soskice use any nonpatent measures of innovation.
In the following sections, I will address these issues in turn. In
some instances, I use Hall and Soskice's data and methods to test
the generality of their claims. In others, I take advantage of a new
data set compiled at the National Bureau of Economic Research (NBER) of
more than 2.9 million utility patents granted by the U.S. Patent and
Trademark Office (USPTO) to applicants from the United States and 162
other countries during 1963–99, and the sixteen million citations
made to these patents between 1975 and 1999.16
This new data set allows one to go beyond Hall and Soskice's
empirical investigation to consider some thirty-six years of patenting
activity for all of the LME and CME countries and to use patents
weighted by forward citations in an attempt to control for the quality
of the innovations being patented. Later, I consider data from the
Institute for Scientific Information (ISI) on scholarly and
professional journal publications, also weighted by forward citations,
as an additional measure of innovation.
Independent Variable: LME Versus CME
According to VOC theory, the primary independent variable for
predicting innovation characteristics is the type of national
political-economic institutional structure (LME or CME) within which
innovators operate. The LMEs include Australia, Canada, Great Britain,
Ireland, New Zealand, and the United States. The CMEs include Austria,
Belgium, Denmark, Finland, Germany, Japan, Netherlands, Norway, Sweden,
and Switzerland. In between these two ideal types, and of less
importance to VOC scholars, are a handful of hybrids denoted as
“Mediterranean market economies” (MMEs) that have mixed CME
and LME characteristics. These countries include France, Greece, Italy,
Portugal, Spain, and Turkey.17
Countries
such as Luxembourg and Iceland are eliminated from the VOC typology
because of their small size, while others, such as Mexico, are
disqualified because they are developing nations.
For the
remainder of this article, references to the set of “LME,”
“CME,” or “MME” countries should be understood
to mean only those states listed above, as these are the only ones
explicitly mentioned in the VOC claims tested here. Later, in the
multivariate regressions, “LMEx” will be used to refer to
the set of all LME countries except the United States.
Some critics might question the “LME-ness” or
“CME-ness” of certain states classified above, for example
the Oceanic countries during much of the Cold War. However, I employ
the existing VOC classifications for several reasons. First, in VOC
theory, it is not the amount of protectionism or regulatory burden that
defines an LME or CME and determines its innovative profile, but
whether markets or hierarchies form the context within which economic
actors organize, conduct their relationships, and solve coordination
problems. Therefore when accepting the VOC country classifications, I
privilege the relational aspects of the LME-CME distinction as
discussed by Hall and Soskice, rather than protectionist or
state-interventionist behavior, because the former are the most
relevant and active mechanisms in VOC's theory of innovation.
Second, recall that the LME-CME dichotomy is not definitive but rather
“constitute[s] ideal types at the poles of a
spectrum.”18
All states have
some degree of tariff and nontariff barriers to trade, and no nation is
free from regulation. Therefore, there are shades of LME and CME in
every economy, and these qualities change over time. Hence when
accepting particular classifications, I pay attention not to absolute
qualities but to relative ones. Finally, all classification systems
have debatable aspects, and their acceptance is often based more on
their usefulness rather than their exactitude. Part of the goal of this
article is to test VOC theory as stated, which includes the usefulness
of their typology.
19 Also, although
country size, wealth, and other factors may be important to innovation
theorists, there is no explicit motivation within VOC theory itself for
including these as separate independent variables other than
restricting the data to the advanced capitalist
democracies.
Dependent Variable: Innovation
The most frequently used measure of innovation is patents. The debate
over the proper use of patent data has proceeded vigorously and with
increasing sophistication over the past several decades. The current
consensus holds that patent data are acceptable measures of innovation
when used in the aggregate (for example, as a rough measure of national
levels of innovation across long periods of time), but are not
appropriate when used as a measure of micro-level innovation (to
compare the innovativeness of individual firms or specific industries
from year to year). While this debate is ongoing and is better
recounted elsewhere, this section will address some of the more
pressing issues surrounding patent measures and their use in testing
VOC theory.20
Strictly speaking, a patent is a temporary legal monopoly granted by
the government to an inventor for the commercial use of his or her
invention, where the invention can take the form of a process, machine,
article of manufacture, or compositions of matters, or any new useful
improvement thereof (USPTO).21
Designs
and plant life can also be patented; however, most econometric analysis
of patent data is confined to utility patents granted for inventions
such as those listed above. For a fuller description of patents and
patent laws, classifications, and the application process see
〈http://www.uspto.gov/main/patents.htm〉.
Accessed 24 March 2004.
A patent is a specific property right
that is granted only after formal examination of the invention has
revealed it to be nontrivial (that is, it would not appear obvious to a
skilled user of the relevant technology), useful (that is, it has
potential commercial value), and novel (that is, it is significantly
different than existing technology). As such, patents have
characteristics that make them a potentially useful tool for the
quantification of inventive activity. First, patents are by definition
related to innovation, each representing a “quantum of
invention” that has passed the scrutiny of a trained specialist
and gained the support of investors and researchers who must dedicate
time, effort, and often significant resources for its physical
development and subsequent legal protection. Second, patent data are
widely available and are perhaps the only observable result of
inventive activity that covers almost every field of invention in most
developed countries over long periods of time. Third, the granting of
patents is based on relatively objective and slowly changing standards.
Finally, the USPTO and the European Patent Office provide researchers
with centralized patenting institutions for the two largest markets for
new technology. In practical terms, this allows researchers to get
around the issue of national differences in patenting laws as well as
providing two separate and fairly independent data pools.
Given these qualities, patents have been used as a basis for the
economic analysis of innovative activity for more than thirty-five
years. Current use began with the pioneering work of Scherer and
Schmookler who used patent statistics to investigate the demand-side
determinants of innovation.22
However, the labor intensive nature of patent
analysis, which used to involve the manual location and coding of
thousands of patent documents, severely limited the extent (or at least
the appeal) of their use in political and economic research. These
limitations were eased somewhat during the 1970s when the advent of
machine-readable patent data sparked a wave of econometric
analysis.
23 In the late 1980s, the use of patent
data was further facilitated by computerization, which increased the
practical size of patent data sets into millions of observations. Most
recently, Hall, Jaffe, and Trajtenberg at the NBER have compiled a
statistical database of several million patents complete with
geographic, industry, and citation information, which I use later to
test the VOC claims.
24 However, patents do have significant drawbacks that somewhat
restrict, but by no means eliminate, their usage as an index of
innovation. First, there is the classification problem, in that it is
difficult to assign a particular industry to a patent, especially
because the industry of invention may not be the industry of eventual
production or the industry of use or benefit. I address this issue,
where possible, by using two different patent data sets with assorted
systems and levels of patent classification. Second, it is not yet
clear what fraction of the universe of innovation is represented by
patents, because not all inventions are patentable and not all
patentable inventions are patented. This problem is exacerbated when
attempting comparative research because different industries and
different countries may exhibit significant variance in their
propensity to patent. I address these concerns by using publications
data in addition to patents. In addition, although patents and
publications both may be imprecise measures of innovation, as long as
this measurement error is random and uncorrelated with the explanatory
variables, then regressions using this data should produce unbiased
estimates of the coefficients (and generally with inflated standard
errors).
Finally, some critics point out that patents vary widely in their
technical and economic significance: most are for minor inventions,
while a few represent extremely valuable and far-reaching innovations.
Moreover, it has been found that simple patent counts do not provide a
good measure of the radicalness, importance, or “size” of
an innovation. Simple patents counts correlate well with innovation
inputs such as R&D outlays, but they are too noisy to serve as
anything but a very rough measure of innovation output.25
Therefore, I use patent counts that have been
weighted by forward citations. Forward citations on patents have been
found to be a good indicator of the importance or value of an
innovation, just as scholarly journal articles are often valuated by
the number of times they are cited. The idea here is that minor or
incremental innovations receive few if any citations, and revolutionary
innovations receive tens or hundreds. Empirical support for this
interpretation has arisen in various quarters: citation-weighted
patents have been found to correlate well with market value of the
corporate patent holder, the likelihood of patent renewal and
litigation, inventor perception of value, and other measures of
innovation outputs.
26Testing the VOC Industry Assumption
Armed with a better understanding of patents, I now use them to test
some of the more controversial claims made by VOC scholars. One such
controversy resides in their implicit assumption about the innovative
characteristics of particular industries. VOC theory assumes that some
industries are inherently and statically more radically innovative, and
other industries inherently and statically more incrementally
innovative. However, this assumption is contradicted by a vast
empirical literature that shows that the innovative characteristics of
any given industry are not static but dynamic and depend not so much on
industry type but on the industry's technological maturity.27
More specifically, studies have found that most
industries are typified by two successive waves of innovation: first a
flurry of radical product innovations that eventually converge on a
dominant product design, followed by a flurry of process innovations in
manufacturing the product at lower cost. In each wave, earlier
innovations tend to be more revolutionary than subsequent ones that
build on them. For example, during the first thirty years of automobile
production, more than 100 U.S. firms produced competing models of
automobiles with tremendous variance in features and operability.
During this period, innovation focused on radical product changes:
introduction of enclosed bodies, wheel-based steering, electrical
systems, gasoline-based fuel and engine systems, and so on. These
innovations tended to be revolutionary and dramatically affected the
look and performance of successive versions of the automobile, such
that cars from this period bear little resemblance to the cars of
today. However, as the market converged on a dominant design for
automobiles, product innovations became gradually more incremental, and
the focus of radical innovation shifted to production processes. This
type of innovation dynamic has been observed in almost every industry
that produces assembled products.
If the innovative character of industry changes over time, then Hall
and Soskice's use of snapshots of patent activity in particular
industries may not properly test VOC theory. That is, for the two brief
time periods covered by Hall and Soskice's patent data, do the
researchers correctly identify which industries were more radically or
incrementally innovative? In order to answer this question I rely on
the ability of forward citations to serve as a measure of
“degree” or “value” of an innovation. For my
empirical evidence, I make use of the newly compiled NBER patent data
set described above. Using the USPTO patent classifications, the NBER
scholars have grouped their data into six industry categories, each
consisting of four to seven subcategories (for a total of thirty-six
subcategories), which allows comparison of the average patent citation
rates across different industries.
Table 1 shows the means of the forward
citations per patent by industry category. The industries generally
rank as assumed by VOC theory: information technology and
telecommunications patents receive on average the most forward
citations, followed by drugs and medical, electronic, chemical, others,
and finally mechanical. T-tests reveal that the differences between
these means are significant beyond the 99 percent confidence level.
Even if one sharpens the level of analysis by further subdividing the
industry categories into their smaller subcategories, patent citations
would again behave more or less as assumed by VOC theory.28
Exceptions include patents in the drugs,
biotechnology, food, and organic compounds subcategories that appear to
be relatively poorly cited despite the fact that these are among
VOC's “radically innovative” industries; in the
“incremental” subcategories, patents related to gas, power
systems, resins, and coatings appear to be more highly cited than VOC
theory might assume. These might be partially explained by
classification problems or by differences in the legal or technical
need to cite in these industries.
Patents and forward citations by industry, 1963–99
Of course, analyzing the data in this manner introduces a potential
truncation problem: older patents have had more time to be cited than
younger patents. This problem is exacerbated in the NBER data set
because it only includes citations data from 1975 onwards.29
This is because the citations data were not
computerized before 1975.
Therefore, patents granted before
1975 will suffer from further truncation in that a 1969 patent will
contain the citations received from patents granted during
1975–99, but not from patents granted in 1969–74. I control
for the overall truncation problem by excluding pre-1975 patents from
consideration and by using multivariate regression analysis with a
control for patent age.
30 All regressions
reported here use a patent age based on grant year. Regressions
performed using a patent age based on application year produced similar
results.
The results of these regressions are reported in
Table 2. First, the table shows in all of the
regressions that the coefficient for patent age is significant and
generally positive; note also that the age coefficient increases in
strength when pre-1975 patents are omitted from the data set, and
consistently hugs 0.3 in all regressions conducted using the
1975–99 patent data (see also Tables
5,
6, and
7 below). This
is suggestive of the truncation effects described above. One can
interpret this coefficient as indicating the number of additional
citations received per patent for each year of its existence. The age
coefficient does turn negative in Model 5, where only the very oldest
patents are used. This suggests that patented innovations may have a
“lifespan” of usefulness, generating much subsequent
innovation while young, then slowly fading into obsolescence as either
new innovations come to replace them or their capacity to serve as the
foundation for new innovations is exhausted. Second, Models 1 and 2
show that, even when controlling for patent age (and with the added
understanding that classification errors may exist), the industry
coefficients generally line up as assumed by VOC theory: information
technology and telecommunications (IT/Telecom) patents receive the
most forward citations, followed by drugs and medical, electronic,
chemical, others, and finally mechanical. The coefficients here can be
interpreted as the additional number of citations received per patent
for patents granted to innovations in a particular industry (relative
to the omitted category “Other”
31 “Other” includes innovations in miscellaneous
areas such as house fixtures, furniture, pipes and joints, jewelry,
cutlery, receptacles, undertaking, and amusement devices.
).
The mean citations received per patent in the 1975–99 data set is
4.9 (with a standard deviation of 7.8); therefore, the size of the
innovative differences between industries suggested by the coefficients
is significant, but not immense.
OLS testing of VOC's industry-innovation assumption
As my findings in subsequent sections indicate that VOC's
evidence is sensitive to the United States outlier, I run two
regressions to consider its effects on the industry rankings. In Model
3, I omit the U.S. data entirely, which drastically reduces the
coefficient for the IT/Telecom and Drugs/Medical categories and
increases the coefficients for the Chemicals and Mechanical categories.
When I instead use a U.S. dummy (Model 4), the coefficients change
significantly for only Chemicals and Mechanical patenting. The first
thing to note in both these regressions is that the rankings do not
change in the areas of most concern to VOC theory: chemicals,
mechanical, and “other” patents receive fewer citations
than those in VOC's radically innovative sectors. Second, these
regressions suggest that the United States is in fact a powerful
outlier that affects the nature of global innovation, especially in
frontier sectors.
Given the time dynamics of innovation, it is also important to
confirm that the findings above are not an artifact of averaging across
a long time period. Models 5 and 6 address this concern, revealing that
VOC's industry assumption generally holds even when I limit the
data set to either the earliest or latest five years of patenting
activity. In these regressions, information technology and
telecommunications patents consistently received the most citations,
again followed by drugs and medical and electronics patents. There is
however some shuffling among the remaining categories, especially
mechanical patents that may suggest a recent small surge in innovation
there. But these minor shifts do not create any major problems for the
VOC assumptions. Also, though not shown here, if one were to further
subdivide the six categories above into their thirty-six subcategories,
patent citations would behave more or less as they do at the category
level.32
With the same exceptions at the
subcategory level as those found with the citations averages. See note
28 above.
Finally, given the nonconstant variance in forward
citations across industries (and later, countries), I correct for
heteroscedasticity using Huber-White estimators of standard errors in
all regressions but find no significant differences from the results
generated by the traditional estimator. In sum, patent data generally
support the VOC assumption about industry innovation
characteristics.
Testing VOC's Predictions About National
Innovative Character: Simple Patent Counts
Having confirmed the industry-based innovation assumption above, I
now reconsider the evidence offered by Hall and Soskice (Figure 1). Again, this chart is based on EPO patent
data for the United States and Germany in thirty industries during two
separate two-year periods. For each industry in each time period, Hall
and Soskice calculated a patent specialization index (I) that simply
subtracts a country's fraction of its total patents in a
particular field from the world's fraction of total global patents
in the same field.33
For example, in
biotechnology: IUS biotech =
USbiotech/UStotal −
Worldbiotech/Worldtotal .
Hence a
positive index score means greater specialization in innovation in that
particular type of technology. The chart shows that the United States
specializes its patenting in industries typified by radical innovation,
while Germany's patent specialization is in industries typified by
incremental innovation. The question then is whether this finding holds
true across time and space, or have Hall and Soskice inadvertently
selected outlying countries or years? To test this possibility, I use
the same EPO data set and computational formula used by Hall and
Soskice, but instead I calculate the patent specialization indices
across a much longer time-span (1978–95) and compare the
innovative activities of the entire set of LME and CME countries.
The results of this exercise are summarized in Table 3. Note that rather than requiring an exact
quantitative match, I apply a more lenient qualitative standard for VOC
theory to pass, only testing which country (or set of countries) has a
higher patent specialization index in each of the thirty industries.
Using Hall and Soskice's data and methodology, I was able to
closely reproduce their findings for Germany and the United States in
1983–84 and 1993–94. However, when I extend the time period
to 1978–95, German and U.S. patenting fails to meet VOC
predictions in polymers, new materials, and nuclear engineering. Even
more discrepancies arise when the data set is expanded to compare
patent specialization by the set of all LME countries versus the set of
all CME countries. For example, in the 1983–84 period, the set of
LMEs had higher patent specialization indices than the set of CMEs in
three industries that Hall and Soskice describe as incremental
(mechanical elements, basic materials, polymers), while CME patenting
had higher specialization scores in two radical industries (new
materials, audiovisual technology). But the most striking disparity
occurs when the United States is excluded from the set of LME
countries; under these conditions VOC theory has only marginally more
predictive power than random chance.
Violations of VOC theory for innovation in thirty technology
classes
The NBER patent data provides a second data set with which
to test the patent specialization indices devised by Hall and
Soskice. Such a test adds value in that the NBER data set not only
spans over twice the time period (1963–99) as the EPO data used
by Hall and Soskice, but the NBER data set also consists of USPTO
patents and is therefore completely independent. The NBER data also
uses a completely independent classification scheme that provides
controls for some of the potential classification problems and
idiosyncrasies discussed above. Yet, despite these differences, my test
results are generally the same as those found using Hall and
Soskice's EPO data. I omit a graphic depiction of the results and
instead explain the major findings. Of the eighteen categories of
innovation that I was able to map from Hall and Soskice to the NBER
data, VOC's predictions were borne out relatively well
(approximately 70 to 80 percent of the time, depending on the time
period) when applied to the United States and Germany.34
Agricultural machines (a particularly
difficult category to define in NBER terms) is the only category that
persistently defies the VOC predictions in all time periods; while
patenting in optics, pharmaceuticals, transport, organic chemistry,
weapons, electrical energy, and nuclear engineering (narrowly measured)
each contradicted VOC theory in different time periods.
However, expanding the data set to test all LME countries versus all
CME countries, I find that VOC theory loses a considerable amount of
its predictive power, with a 72 percent success rate in 1983–84,
but only 50 percent in 1993–94, and 56 percent over the entire
1963–99 period. Omitting the United States from the set of LMEs
results in further deterioration, with VOC's success rate ranging
from 44–56 percent. Thus, after analyzing two different data sets
and competing classification methods, it appears that the success of
VOC theory strongly depends on the inclusion of the United States as an
LME.
Testing VOC's Predictions About National
Innovative Character: Patent Citations
So far I have used simple patents counts in my comparisons of LMEs
versus CMEs, yet as I explained above, forward citations of patents are
an even better gauge of radical versus incremental innovation.
Therefore, in this section, I use the forward citations data in the
NBER patent data set to test the VOC country claims directly, retaining
the same techniques that I used above in testing the VOC assumptions
about industries. As my dependent variable in all of the following
regressions I again use citations-received per patent as a proxy for
the radical versus incremental nature of innovation. VOC theory
suggests that country dummies or country-type dummies (LME, CME) are
the primary independent variables of interest, as well as controls for
industry type (again I use industry category or subcategory), and, of
course, a control for patent age should be included to address the
truncation problem. Since the U.S. outlier proved important in the
simple statistical analysis above, I address it in two ways in the
regressions. In some regressions a U.S. dummy is introduced, in others
the United States is simply omitted from the class of LMEs (creating a
new dummy: LMEx). For data, I use the NBER patent data set for all
countries' patenting activity during the period 1975–99.
I begin with regressions using controls only for patent age and
country type, the results of which (Table 4)
reinforce what I found previously: LMEs are more radically innovative
than CMEs (Model 1 versus Model 2), but this finding depends entirely
on the inclusion of the United States as an LME (Model 3). This effect
is apparent even when the CME dummy is run together with that for LMEs
or LMEx's (Models 4 and 5). In each of these regressions, the
coefficients can be interpreted as the additional number of citations
received per patent for patents granted to innovations in a particular
set of nations (LMEs, CMEs, or LMEx's) relative to the rest of the
world. Note how sharply the LME coefficient drops when I introduce a
U.S. dummy variable (Model 6) and, perhaps more interesting, that the
LMEx's appear to be less radically innovative than the CMEs (Model
7). Of equal importance is the small size of the coefficients and the
differences between them. These indicate, for example in Model 4, that
even when I do not control for the U.S. outlier, the innovative
difference between LMEs and CMEs is smaller than a single citation per
patent. Although this may be a statistically significant amount, it is
far smaller than the innovative difference between the most versus
least innovative industries found above and does not suggest a large
innovation gap.
OLS testing of VOC innovation theory, by country type
(1975–99)
VOC theory also includes industry type as a factor in determining
innovative behavior. Hence a second set of regressions are run,
identical to those reported in Table 4 but
with the addition of controls for industry (Table
5). Yet I find no significant differences when the industry
controls are added to the regression models. Again, the LME countries
appear at first to be more radically innovative than the CMEs (Model 1
versus Model 2), but not when the United States is excluded from the
group of LMEs (Model 3). Note also that the industry coefficients in
this regression match those found when I tested the VOC
industry-innovation assumption above (Table
2). To test this finding more directly, I add a U.S. dummy,
which again severely affects the coefficient of the LME dummy (Models 6
and 7). Regressions run at a finer level of analysis using industry
subcategories (not shown) produce similar results.35
An alternate interpretation of VOC theory suggests that in
place of LME/CME/LMEx controls, one might include interaction
terms (LME*industry, CME*industry, and LMEx*industry). I experimented with
such interaction terms but produced the same general results as those
reported above.
OLS testing of VOC innovation theory, by country type and industry
(1975–99)
Given the broad nature of VOC theory and the complex array of causal
mechanisms it hypothesizes, a fixed-effects model is perhaps the most
efficient way to conduct a statistical test of its central predictions.
While the NBER data set affords enough degrees of freedom to use
countries dummies for all 162 nations, computer memory does not. I
therefore run a final set of regressions in which I include dummies for
twenty-three countries with the highest patenting activity.36
As before, all pre-1975 patents are
eliminated to control for truncation effects.
These countries
include the aforementioned LME and CME states in addition to France,
Italy, Spain, Israel, Taiwan, Singapore, and South Korea. Using only
country dummies, controlling for age, and correcting for
heteroscedasticity, I find that the relative strengths of the
coefficients for the remaining dummies do not quite line up along the
lines predicted by VOC theory (
Table
6). Here the coefficients can be interpreted as the additional
number of citations received per patent for patents granted to innovations
in a particular nation relative to those granted to the rest of the world
(ROW). Though not astronomical, the size of the coefficients do indicate
significant innovative differences between states, and that these innovative
differences are comparable to those across different industries. All of the
coefficients are positive, indicating that patents from the rest of the world
generally receive fewer forward citations than patents from these chosen
countries. Patents from the United States receive the most forward citations,
those from Spain, Austria, and New Zealand consistently receive the
least. Interestingly, Australia and New Zealand appear to deserve a
place among the CMEs, while Japan seems to be one of the most radical
innovators (Model 1). While I am not immediately concerned with Hall
and Soskice's hybrid MMEs, the three MMEs that appear in the
regressions (France, Italy, Spain) have major differences between them
and do not appear to form a cohesive group. Also, the high placement of
Israel (arguably a pre-1970s CME, increasingly MME thereafter) and
Taiwan (arguably an MME), not mentioned in VOC theory, further suggest
that there may be more to radical innovation than the variables
captured by Hall and Soskice. Adding controls for industry does not
have a significant impact on the rankings, except for some minor
shuffling (Model 2).
OLS testing of VOC innovation theory, by country and industry
(1975–99)
Finally, if one believes that both quality and quantity of patents
matter, that Ireland with its relative trickle of few but highly cited
patents should not necessarily be considered more radically innovative
than Germany with its slightly less cited ocean of patents, then one
must instead look at total citations received over time. This data is
charted in Figure 2. Here I have merely
multiplied the mean citations received per patent by the total number
of patents for each country. This captures both the number and value of
patents in one measure. The plots are split horizontally into three
groups (LMEs, CMEs, and other countries) for comparison. Again the
figure shows the U.S. outlier, but no strong general differences in
total citations between the different VOC country types.
Total forward citations, 1975–99
In sum, the VOC theory does not appear to explain innovation as
measured by patenting activity. Rather, the success of VOC theory in
predicting innovation appears to depend on the inclusion of the United
States, a major outlier, in the set of liberal market economies. I find
this fact repeated regardless of the source of the patent data, the
type of the industry classification system used, or whether simple
patents or forward citations are used. However, one caveat that bears
repeating is that this finding depends on an assumption of random error
in using patents as a measure of innovation. Social scientists cannot
yet completely describe the correlation between patents (an innovation
output) and total innovation, nor do social scientists fully understand
how propensity to patent varies across industry, across country, and
over time. I therefore, briefly consider the nonpatent evidence for
differences in national innovation in the next section.
Additional Evidence
Patent statistics are by no means the only innovation data that paint
a picture contradictory to the VOC claims. Scholarly journal articles
are another useful measure of innovation that reinforces the
cross-national findings discussed above. Scholarly publications data
offer advantages similar to those of patents, with each journal article
representing a quantum of research innovation that must pass
independent review and that tends to be cited in proportion to its
innovative impact. More importantly, scholarly publications data are
completely independent of patents: they are generally produced by a
different set of innovators, affected by different incentives, and
judged according to different institutional standards.37
Of course, journal articles also suffer many of
the same shortcomings as patents, including difficulties in
classification, problems with valuation, and uncertainty regarding to
what degree journals represent the universe of innovation.
38 The innovative
“representativeness” of journal articles is more of a
problem in the social sciences, and less so in the physical sciences,
see Hicks 1999.
These difficulties
are further complicated by changing journal sets, the lack of a single
standardized referee process, and the relative importance of prestige
and popularity in the publication process. However, just as with
patents, information sciences scholars have found legitimate and
rigorous applications for publications data in measuring innovative
output. While this debate is better summarized elsewhere, the current
consensus is that there is reasonable basis for using journal articles
as a window on innovative activity in the aggregate.
39 VOC theory does not make specific predictions regarding scholarly
publications patterns, and indeed its authors may never have intended
it to do so. Nonetheless, one might infer from VOC theory the following
hypothesis: that scholarly publications by LME researchers should show
specialization in fields associated with revolutionary scientific
advances, while CMEs should show specialization in fields associated
with incremental scientific advances. Although it is not quite clear
what a “radically” versus “incrementally”
innovative field might be, one could simply map the typology used by
Hall and Soskice for industrial sectors as well as academic sectors.
For example, CMEs should excel in publishing in the engineering and
technology journals, LMEs in biology, medicine, and physics. A second
hypothesis might surmise that researchers in the CMEs should excel in
professional journals and applied sciences publications where
incremental research is more prominent, while LME researchers should
publish heavily in the more academic or theoretical sciences journals
where the research tends toward the revolutionary. A third, and less
controversial, hypothesis would be that LME publications should simply
have higher forward citation averages than CME publications.
Yet, none of the patterns hypothesized above can be found in the
cross-national publications data. Consider the ISI's simple
journal publication data compiled in Table 7.
Compare the world publication rates by field with those of the LMEs and
CMEs. As a group, the LMEs tend to consistently specialize in clinical
medicine, biology, earth-space, psychology, social science, health, and
professional journals; CMEs tend to consistently specialize in clinical
medicine, chemistry, and physics. Over time the CMEs have increased
their specialization in biomedical research, physics, and earth-space
but have weakened in clinical medicine, chemistry, and engineering and
technology; while the LMEs have increased their specialization in
biomedical, physics, and earth-space. Using forward citation indices,
Table 8 shows the LMEs beating CMEs in all
fields. When the United States is excluded from the set of LMEs, the
LMEs appear to have higher citations than the CMEs in all fields except
earth and space, engineering, and physics. Relatively speaking,
LMEs' strengths are in chemistry, physics, biomedical research,
and math. CMEs are also strong in chemistry, as well as engineering,
physics, and biology. None of these findings is what one might expect
from VOC theory.
Specialization in scholarly publications
Relative prominence of scientific literature by country/economy
and field (1999)
Finally, despite problems in measuring pre-1960s innovation and
diffusion, history provides researchers with some natural experiments
that deserve further investigation. For example, Japan, during its
first brush with democracy (1910s to 1930s), was distinctively
“LME-like” but does not appear to have followed a
significantly different innovation pattern than did postwar CME Japan.
During this earlier period, labor in Japan was strong and
confrontational, and business did not hesitate to inflict frequent and
severe employment dislocations for the sake of technological advance.
Moreover, the dependence of prewar Japan on external trade and finance
exposed even the powerful zaibatsu to the vicissitudes of
international markets and created many LME-type incentives for economic
actors. Yet the Japanese appear to have been consistent incremental
innovators during this time. On the other hand, the Germans of this
time period rivaled the United States in technological advance,
producing wave after wave of radical innovation in multiple fields
including the gas-powered automobile, the Zeppelin, the Haber-Bosch
process, blood-typing, aspirin, and organic chemicals to name but a
few. Yet, the Germans had many of the same CME-type institutions and
incentives as one finds there today, including a national welfare
system, national health care, and large business cartels negotiating
with each other, and sometimes with workers, in a fairly CME-like
manner. These stylized facts, while not conclusive, do suggest areas
for deeper research and further testing of VOC claims, both as a theory
of innovation and as general theory of political economy.
Conclusions
In this article I have demonstrated that the predictions made by
varieties of capitalism theory regarding national differences in
technological innovation are not supported by the empirical data, and
that the existing evidence depends heavily on the inclusion of a major
outlier, the United States, in the class of liberal-market economies.
My empirical investigation included simple patent counts, patents
weighted by forward citations, and scholarly publications (both simple
counts and weighted). I investigated data covering all of the VOC
countries over the course of several decades, little of which revealed
the patterns predicted by VOC scholars.
These findings carry significant repercussions for both VOC and
innovation theory. First, insofar as patents and scholarly publications
are good indices of innovation, VOC theory clearly fails to provide an
accurate picture of the innovation process, and hence the trade and
production patterns that follow. Whether this is a problem with the
LME-CME classification system or VOC's assumptions and causal
mechanisms is not clear from the evidence presented here. However, I
would suggest that while the firm may be the key actor in capitalist
economies, and the primary producer of goods and services, it is
difficult to ignore the role of the state in innovation as strongly as
VOC's theory and classification system do. Throughout the world,
much useful innovation is the result of state-sponsored and
state-managed R&D, often originating in concerns with national
security. Another stream of innovative R&D in many countries comes
from the public university system, or private universities benefiting
from significant state support. In still other states, innovation takes
the form of incremental improvements on imported technologies, where
the government has had a heavy hand in deciding which technologies will
get imported. Often, the government also plays a key role as a market
maker for, and main diffuser of, new innovations. However, VOC's
innovation theory omits these causal mechanisms entirely. This does not
mean that VOC scholars are wrong to bring the firm onto the center
stage of political economy, but rather that in trying to get away from
a hackneyed focus on government protectionism and state ownership, they
may have overcompensated. Future theorists must find a synthesis
between the corporate-centered relationships emphasized by VOC and the
state-centered mechanisms employed in traditional political economy.
Second, the statistical analyses above consistently point to the
United States as an important factor in explaining global patterns of
innovation. Furthermore, the fixed-effects regressions reported in
Table 6 reveal that many of the world's
most innovative countries are those that also tend to have the
strongest military and economic ties with the United States, including
Japan, Canada, the United Kingdom, Israel, and Taiwan. Together, these
observations suggest that to better understand the political economy of
comparative rates of innovation, future research should perhaps focus
less on domestic institutions and more on international relations. This
is not to argue that domestic institutions are insignificant, but
rather that the scope and depth of a country's relationship with
the lead innovator may also carry significant weight in determining its
technological profile. There are theoretical grounds for this
supposition in that while the basic laws of science may be public
goods, the tacit knowledge required to apply these laws to proper use
and development of new technology is relatively excludable. Therefore,
factors such as foreign direct investment, educational exchanges,
military assistance, and international flows of science and engineering
labor between the lead innovator and other countries should be explored
for their effects on innovation and the agglomeration patterns that
interest VOC and trade theorists.
Of course, the research reported above, while suggestive, does not
necessarily shut the door on a VOC approach to technological
innovation. Innovation is a notoriously difficult phenomenon to measure
quantitatively, and existing measures carry with them considerable
noise, hence further progress needs to be made on method as well as
theory. Nor does the critique here necessarily apply to other aspects
of VOC theory. VOC is a broad approach to social behavior, consisting
of myriad hypotheses regarding almost the whole spectrum of political
economy including corporate governance, monetary policy, welfare
programs, and labor reform. These hypotheses are not necessarily
interdependent and need to be considered and tested each on its own
merits. Finally, as social scientists increasingly turn to institutions
and international relationships to explain various phenomena related to
cross-national variance in innovation, VOC scholars should be applauded
for inserting political science into an area of research from which it
has been all but absent.40
While economists and sociologists have produced excellent studies of
the role of these variables in international technological performance,
the comparative advantage that political scientists bring to the field
in terms of methods and theory make this an area deserving far greater
attention by students of politics. VOC scholars have therefore provided
a valuable and useful starting point for such an endeavor.
