WPS3739
Patenting and Research and Development: A Global View*
Mariano Bosch
Daniel Lederman
William F. Maloney
The World Bank
Abstract: Using a new global data base on patents and innovation inputs, this paper
employs several recently developed econometric techniques to examine the dynamic
relationship between research and development and U.S. patents granted. We confirm at
the country level the recurrent micro-level finding of a strong relationship between the
two and estimate the OECD elasticity to be effectively unity. This conflicts with the
frequent micro level finding of strong diminishing returns in knowledge generation and
suggests the importance of spillover effects measurable only at the aggregate level.
Developing countries, however, do show diminishing returns and implicit rates of return
to innovative effort in these countries appear to be a fraction of those found in the OECD.
We then seek to explain the variance in elasticities by introducing proxies for the
functioning of the national innovation system, the set of institutions and agents that create
and diffuse knowledge. Across the entire sample, several variables, in particular
education, security of intellectual property rights, and in some specifications the quality
of research institutions and their interaction with the private sector impact the elasticity of
transformation of R&D into patents.
World Bank Policy Research Working Paper 3739, October 2005
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the
exchange of ideas about development issues. An objective of the series is to get the findings out quickly,
even if the presentations are less than fully polished. The papers carry the names of the authors and should
be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely
those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors,
or the countries they represent. Policy Research Working Papers are available online at
http://econ.worldbank.org.
*We are grateful for helpful comments from Eli Berman, Bronwyn Hall, Andrés Rodriguez Clare and
especially Frank Windmeijer. This research was funded by the regional studies program of the Office of
the Chief Economist for Latin America and the Caribbean at the World Bank.
I. Introduction
The idea that technological progress is driven by research effort is the central
insight of the endogenous growth literature.1 In several of the canonical models (Romer
1990, Grossman and Helpman 1991 and Aghion and Howitt 1992), the assumption of
constant returns to knowledge generation plus large spillovers to other researchers
implies that more resources invested in R&D leads to increasing growth rates.2
However, there is disagreement over the technology generating new knowledge.
Jones (1995) critiqued the relevance of such models arguing that in the US, both growth
and the intermediate measure of knowledge generation, patents, appear to be constant
over time despite large increases in R&D employment, implying decreasing, rather than
the increasing returns implicit in previous models.3 In fact, a long and expanding
literature offers support for this view. While it may well be the case that future
researchers may "stand on the shoulders" of those past and find innovation easier (see,
for instance, Romer 1990, Aghion and Howitt 1992), this may not be sufficient to offset
increasing costs of generating knowledge as technological opportunity in a particular
field diminishes (Evenson 1984, Evenson and Kislev 1976, Kortum 1997, Segerstrom
1998). Further, the fact that patenting gives monopoly rights over the rents from new
ideas can give rise to increasing redundancy of R&D efforts when multiple firms pursue
the same discovery (see Jones 1995, Kortum 1993, Reinganum 1984,1989). Not only can
these factors enter directly as determinants of the rate of growth, but as Jones and
Williams (2000) among others note, they dictate whether the market determined level of
R&D is above or below socially optimal.4
1See for example, Barro and Sala-I-Martin 1995; Romer 1987, 1990; Aghion and Howitt 1992; and
Grossman and Helpman 1991.
2See, for example, Romer (1990), Segerstrom et al (1990).
3Aghion and Howitt (1998) counter, arguing that, measured as a share of GDP, R&D effort has been
relatively constant, consistent with their extended model.
4 As Jones and Williams note (2000) the "the standing on shoulders" effect implies socially sub-optimal
levels of R&D investment while the is "step on toes effect" raise the private return above the social and
imply above socially optimal levels of investment in R&D
1
An increasingly sophisticated empirical literature attempts to directly estimate the
knowledge production function using patents as a proxy for new knowledge. Jaffe and
Trajtenberg (2002) and Griliches (1990) offer as stylized facts that in cross section
patents are probably roughly proportional to R&D implying an elasticity of unity, while
the "within" panel dimension suggests strongly decreasing returns to scale.5 The few
studies in the second category (Hausman, Hall and Griliches 1984,1986; Blundell,
Griffith and Windmeijer 2002), along with Blundell, Griffith and Van Reenen (1995) that
looks at the impact of market structure on innovation, have contributed important
advances in the theory of count data estimators in a panel context. The latter two focus
particularly on modeling dynamics and controlling for the unobserved heterogeneity that
renders cross sectional estimates suspect. In this sense, they parallel closely the recent
literature on dynamic panel modeling with continuous dependent variables (Arellano and
Bond 1991, Blundell and Bond 1998). Both issues are found to be of central importance
in modeling the innovation process and generating consistent parameter estimates.
All the panel studies confirm the stylized fact of decreasing returns to scale.
Hausman, Hall and Griliches's (1984) non-dynamic estimates of the elasticity of patents
with respect to R&D are in the range of 0.3-0.6, depending on the technique employed.
Hall, Griliches and Hausman's (1986) estimates hover around .35 and are similar to those
estimated in a dynamic context by Blundell, Griffith and Windmeijer (2002) of around
0.5. Using industry level panel data Kortum and Lerner (2000) find an elasticity of .48-
.52. There thus seems to be support in the micro level within estimates for the more
pessimistic view of increasing difficulty and congestion.
Nevertheless, the firm level estimates, and to a lesser degree those at the industry
level, may miss the spillover effects that one firm's R&D may contribute to another firm
or industries knowledge generation effort. The literature on estimating the returns to
R&D, for example, finds differences of several multiples between the private returns to
5In cross section, Bound et. al (1984) introduce a wide variety of estimates, the most comparable being
around .32-.38 and Pakes and Griliches (1980) offer an estimate of .61 although Griliches (1990) suggests
that when accounting for reporting error, there is little evidence of diminishing return in cross section. See
also Klette and Kortum (2002) for a discussion of the stylized facts around the patents and R&D.
2
R&D, estimated at the firm level, and those at the national level suggesting substantial
spillovers.6 Jaffe (1986) also finds that firms whose research is in areas where there is
much research activity by other firms generate, on average, more patents per dollar of
R&D and he finds the magnitudes of the spillovers to be substantial.
To date, there are very few investigations at the aggregate level that would more
completely capture such spillovers and none are comparable to the micro studies. Bottazi
and Peri (2003) have estimated the R&D/patent relationship at the aggregate level using a
cross section of European regions. As in the other cross-sectional firm literature, they
find an elasticity close to 0.9 and in addition, significant, although small, regional
spillover effects within 300 kilometers. Using an OECD panel on 17 countries, Furman,
Porter and Stern (2002) more generally explore the relationship between a range of
innovation inputs and institutions and patents but also do not offer parameters estimates
comparable to the panel studies discussed above.
In this paper we construct a large panel of advanced and developing countries and
generate estimates for the R&D-patent relationship at the global level that are comparable
to those at the micro level discussed above. Including a large sample of emerging
countries puts us somewhere between the count and continuous dependent variable
literatures. Only 6.5% of the sample has observations of zero value, just under half the
number found in Blundell, Griffith and Windmeijer (2002), so clearly the issue of
integers and zero values is reduced, yet not so small as to be dismissed altogether. We,
therefore employ the estimators recently developed in both literatures. This allows tests
for robustness to the assumptions underlying the various estimators. We find that
regardless of the technique employed, there is a strong relationship between R&D effort
at the global level and innovation with a long run elasticity surprisingly close to 1.
We contribute to the understanding of the dynamics of knowledge creation. Hall,
Griliches and Hausman (1986) argue that, at the firm level, patenting occurs virtually
6One strand of the empirical literature working broadly in the growth tradition has measured the returns to
R&D using both micro and cross-country data.6
3
instantaneously, suggesting that much of R&D expenditures are dedicated to
development as opposed to basic or applied research. Blundell, Griffith and Windmeijer
(2002), however find the long run elasticity substantially larger than the short run. Since
spillovers may occur only with some lag from the initial discovery, we may expect larger
lags at the aggregate level. In using both annual and quinquennial data, we find that in
fact, the lags are quite long.
We then divide the sample into OECD and the developed countries and identify
significant and important differences in scale economies and very large differences in
implicit returns to R&D between the two samples. To examine the determinants of the
differences in observed elasticities, we introduce interactively several variables likely to
capture, especially, the degree of spillovers. In this sense we broadly parallel the
literature on National Innovation Systems (NIS)7 that focuses on how the deployment of
human capital and financial resources in the various national institutions universities,
public think tanks, firms- and the interactions among these institutions affect the capacity
of a country to generate knowledge. Furman, Porter and Stern's work on the OECD
established that such factors often enter additively in determining the number of patents.
Given the greater range of endowments, institutions, and institutional quality across our
sample, we expect, and find, significant and important effects of these factors on the
patenting elasticity.
The remainder of the paper is organized as follows. Section 2 describes the data,
section 3 outlines the econometric approach, and section 4 presents the main results
regarding the relationship between patents and R&D expenditures. Section 5 examines
elements of the NIS that may influence the corresponding elasticity. Section 6 concludes.
7See Nelson 1993; OECD 1998, 2001; Lundvall et al. 2002, Furman et al 2002
4
II Estimating the Innovation Function
Data
Our dataset consists of a panel of 49 developed and developing countries from the
1960s to the present. We first review the measures of innovation output and input.8
Innovation Output: A large literature uses patents as an imperfect measure of innovation
output although the limitations of the measure are well known (see for example, Griliches
1990, Trajtenberg 2001, Jaffe and Trajtenberg 2002). Of perhaps greater concern for the
present work is that patents granted by national agencies are not comparable due to
differences in national standards, costs of applying for patents, levels of intellectual
property protection, pecuniary benefits from patenting, and other country-specific
institutional features. To ensure comparability we follow Jaffe and Trajtenberg (2002),
Branstetter (2001), Furman, Porter and Stern (2002) and use the number of patents
granted by the United States Patent and Trademark Office (USPTO).9 In the absence of a
global patenting agency, the US remains the principal locus of patenting activity and the
USPTO offers reliable panel data for the period 1963 to 2000 for a large number of
countries. Granted patents are assigned by country of origin based on the country of
residence of the first inventor.
That said, countries may differ systematically in their propensities to apply for
patents in the US. Those with a large volume of exports to the US have a greater interest
in patenting any invention that may be embodied in their exports.10 Similarly, country
endowments and economic structure may also impact the ratio of innovations to patents:
manufacturing in general may generate higher patents per innovative dollar than natural
resource based sectors. To control for these effects we include in our specifications the
volume of merchandise exports to the US, drawn from the IMF Direction of Trade
Statistics, and Leamer's (1984) index of natural resources endowments, net exports of
8See Lederman and Saenz (2003) for a complete discussion of the data.
9 The U.S. PTO demands that the invention be "novel and nontrivial, and has to have commercial
application" (Jaffe and Trajtenberg 2002, 3-4).
10See Trajtenberg (2001).
5
resource intensive products over labor force, constructed from the UN Statistics
Department's Commodity Trade Statistics (COMTRADE).
Innovation Effort: Consistent with much of the microeconomic literature we employ real
R&D expenditures as a measure of innovation effort. The data are derived ultimately
from national surveys that use as a common definition of expenditures that include
"fundamental and applied research as well as experimental development." 11 The data
thus include not only the basic science expected in the more advanced countries, but also
investments in the adoption and adaptation of existing technologies often thought more
germane to developing countries. The series are constructed based on underlying data
published by UNESCO, the OECD, the Ibero American Science and Technology
Indicators Network (RICYT) and the Taiwan Statistical Data Book.
Though it would be desirable to study the evolution and efficiency of both private
and public R&D, we work with aggregate R&D for several reasons. First, the data
sources divide R&D not into private and public R&D, but into productive and non-
productive sectors, the latter accounting for roughly 20% of the total.12 The definition of
"productive sector" includes both public and private for profit and not-for profit firms
while "non-productive sector" includes R&D financed or undertaken by the executive
branch of government.
Second, the productive/non productive split seems to occasionally lead to some
critical issues in categorization. For instance, if a public company finances its R&D from
retained earnings, this will count as productive sector R&D. If instead the same R&D is
financed by a transfer from the treasury to the firm, it counts as "non-productive" R&D.
For several countries in our sample, there were striking shifts in composition from one
year to the next suggesting sensitivity to these accounting conventions. By contrast, the
total R&D series were reasonably stable. The final consideration is more prosaic: many
11UNESCO Statistical Yearbook (1980) pg. 742. Definitions are common to the OECD, Ibero American
Science and Technology Indicators Network (RICYT), World Bank ,and Taiwan Statistical Yearbook and
all are based on the Frascatti manual definition.
12The median for countries with both series is 21%.
6
developing countries tabulate only the aggregate values and hence to maintain the largest
sample we use the most common definition.
We describe the data in tables 1 and 2 and in figures 1, 2 and 3. Table 1 tabulates
the basic descriptive statistics of all variables and table 2 the correlations among them13.
Several facts merit attention. First, the patents variable ranges from 0 (Bolivia, Jamaica,
Togo) to around 30,000 (Japan). 14 Similarly there are great differences in terms of R&D
investment across countries. Figure 1 plots the overall relationship between patenting and
R&D effort. Looking more closely, Figure 2 shows the same scatter for 2 groups of
observations, those with less and more than 1000 patents. Although some OECD
countries have had bad years of patenting below 1000, the graphs effectively capture the
differences in the OECD/non-OECD samples. Both depict a clear relationship between
patents and R&D expenditures. In addition to the strong relationship with log R&D, table
2 shows patents to be positively correlated with the log of US trade and negatively
correlated with natural resource abundance.
Methodology
The estimation of the links between innovative capacity and its outputs, measured
as the number of patents, is not straightforward for several reasons. First, the nature of the
patent variable as discrete and non negative but with a non negligible probability of being
zero has, in the firm level literature, dictated the use of count-data models with an
exponential specification under the assumption of either Poisson or negative binomial
distributions (see Hausman, Hall and Griliches's 1984 and Blundell, Griffith and
Windmeijer 2002). However, the smaller number of zero-patent observations and
generally greater patenting rates reduce somewhat the integer concerns underlying count
models. Several recent studies at the country level treat the patent variable as continuous
and estimate a standard dynamic log linear specification (see Botazzi and Peri 2003;
Furman, Porter and Stern 2002). In our sample, roughly 6.5 % of the observations are
zeros- not negligible particularly in the LDC sample where they are concentrated, but
13Appendix table A.1 tabulates in detail the means of the variables by country as well as the number of
observations per country
14Note that the US has been left out of the sample due to the difficulty of controlling for trade with itself.
7
low enough to justify alternative approaches. Both approaches seem to fit the data
reasonable well, although for low levels of patents there seems to be an increase in
variance in the log linear specifications.15 We explore both approaches here.
The second issue arises from the desirability of some dynamic structure in the
estimation. There is likely to be lag between the commitment of the R&D expenditures
and the actual output derived from them. For this reason, Hall, Griliches and Hausman,
(1986) investigate the existence of a lag structure in the patents-R&D relationship.
Although they argue that there is no evidence of a long lag between R&D and patenting,
past R&D history seem to add to the current year's patent applications, although the
effects are small. Additionally, Blundell Griffith and Windmeijer (2002) argue that there
may be a feed-back effect between R&D expenditures and the patents generated by those
expenditures. They develop an estimator which explicitly models the dynamics of the
count process in the panel data taking into account that feed-back between patents and
R&D.
Each of the techniques employed in the literature offers advantages and
drawbacks. Our strategy is to explore the data with several of them to assess the
robustness of our findings to their underlying assumptions.
Static Modeling
Within the context of static models we estimate variants of the standard
specification used in the R&D-Patents literature (see Hausman, Hall, and Griliches 1984).
As Blundell, Griffith and Van Reenen (1995) note, convenient specifications can be
derived from a Cobb-Douglas technology production function such as
Pit = R & Dit i
15Countries with 0 patents were assigned ln Patents=0. Regressions using the linear approach in following
sections include a dummy variable for these observations.
8
where Pit is the number of patents of country i in period t, R & Dit is the total expenditure
in R&D of country i at time t, is the elasticity relating the two, and i is a fixed
individual effect for country i. In the count-data context where zero values are of central
concern we estimate
Pit = exp( ln R & Dit + 1 lnUSit + 2 ln LIit )i + it
US (1a)
In the context where zeros are less problematic, we estimate a log linear specification:
ln Pit = ln R & Dit + 1 lnUSit + 2 ln LIit +i + it
US (1b)
In both cases, Pit US is the patents count of country i at time t , R&D is the innovation
input, US, the value of real merchandise exports towards the United States, LI is the
Leamer index of natural resource abundance, and i captures country specific fixed
effects.
Dynamic modeling
The introduction of dynamics and more generally of pre-determined regressors in
both 1a and 1b adds complications. First, the inclusion of a lag dependent variable in
equations 1a and 1b yields the usual inconsistent estimates using within group estimators
(Nickel 1981).16 In the context of count data models, Chamberlain (1992) and
Wooldridge (1997) develop a GMM estimator to deal with individual fixed effects when
the regressors are pre-determined. However, as Blundell, Griffith and Windmeijer (2002)
argue, the GMM estimators show small sample bias when the time series show a high
degree of persistence as is the case with expenditures with R&D. In order to jointly deal
with the existence of country fixed effects and predetermined regressors more generally,
they suggest an estimator that measures the fixed effect component i directly. In the
context of innovation functions at the firm level, the key unobserved individual
16Fixed effects estimators create a correlation between the de-meaned variables and the regression error
implying a bias of order 1/t.
9
heterogeneity comes from the large differences in the stock of knowledge among firms so
the capacity to innovate should be captured by the past history of innovations defined as
the pre-sample mean of the dependent variable, Pip = US 1 TP-1PiUS where TP is the
TP r=0 0-r
number of pre-sample observations. The approach is also valid at the country level and
has the additional advantage that, since normally the patents series are more complete
than the R&D series, a substantial number of observations can be dedicated to calculating
the pre-sample mean without restricting the effective sample size used for estimation.
An additional complication specific to count-data models is the way the lagged
dependent variable enters the specification. Blundell, Griffith and Windmeijer (2002)
argue that including the LDV as a logged term within the exponential leads to problems
transforming observations with value zero, and can also lead to greater than unity
coefficients on the lagged dependent variable that imply an explosive time path for
patents. They therefore propose introducing the lag of the dependent variable linearly,
hence giving the name "linear feedback model," which corresponds to the following
specification:
Pit = Pit + exp( ln R & Dit + 1 lnUSit + 2 ln LIit + ln Pip )i + it
US US US
-1 (2a)
Estimating 1b in a dynamic linear dynamic panel context has been extensively
treated elsewhere (see Arellano 2003). From the empirical point of view the recent
growth literature provides numerous examples of dynamic panels in the presence of very
persistent regressors in order to estimate an specification such as
ln Pit = i + ln Pit + ln R & Dit + 1 lnUSit + 2 ln LIit + it
US US
-1 (2b)
Very briefly and closely following Anderson and Hsiao (1982), Arellano and Bond
(1991) and Caselli et. al. (1996) in the growth literature, we difference the data to
eliminate unobserved fixed effects i yielding
10
lnPit = lnR & Dit + 1lnUSit + 2lnLIit + it,
US (3)
However, unless the idiosyncratic error follows a random walk, this differencing
necessarily gives the transformed error a moving-average, MA, structure that is
correlated with the differenced lagged dependent variable. This can be overcome by
using instruments dated t-n and earlier. Arellano and Bond (1991) employ lagged levels
as a proxy for differences in a Generalized Method of Moments (GMM) context.
However, in the context where explanatory variables show little variation over time, as is
the case with R&D here and with schooling or natural resource endowments in growth
regressions, levels are often poor instruments. Bond, Hoeffler and Temple (2001) show
that this "weak instruments" problem can be severe in cross-country growth regressions
with panel data. For this reason, we follow Blundell and Bond (1998), Arellano and
Bover (1995), and Levine, Loayza and Beck (2000) in employing a system estimator that
combines (3) with equation (2b) in levels, using lagged differences of the endogenous
variables as instruments. Finally we make use of the procedure proposed by Windmeijer
(2000) to minimize the small sample bias in the standard errors from the efficient two
step estimation.
Each estimating technique implies a different approach to the data. The static
estimates of equations 1a and 1b allow us to maximize the number of observations used;
however they neglect the possible dynamics of the patents-R&D process and the
endogeneity of the R&D variable. Estimation of dynamic equation 2a is also undertaken
using the same data periodicity although the sample is restricted somewhat since we now
require that each country have at least two consecutive observations. Finally the GMM
estimations require additional lags for use as instruments which, for most of our sample,
are not available due to gaps in the data. For this reason we use quinquennial averages
since 1975.
A priori, our preferred estimates are the exponential linear feedback model and
the log linear GMM estimation on the grounds that they control for individual
heterogeneity and possible endogeneity of the explanatory variables. However, whenever
11
appropriate we present alternative techniques using the same time span and sample for
the sake of comparability.
III. Results
Static Estimates
We begin following the existing firm level literature by estimating static
specifications. Because we want to be broadly comparable with the dynamic estimates
that follow, we introduce a static version of the PSM estimator as well.
Table 3 presents estimates of equations 1a and 1b using yearly data from 1976-
2000 under both the Negative Binomial and Poisson distributions, using the naïve levels,
the Within Group (WG), and the Pre Sample Mean (PSM) specifications.17 For
comparison, we begin in the first panel with a linear estimator under the assumption that
the aggregate data does not suffer "too much" from zeros and integer issues: the
dependent variable is log of patents and a dummy variable is included for those
observations with 0 patents.
The first point to highlight is that regardless of technique, strong evidence
emerges for an aggregate innovation production function mapping the inputs to outputs of
innovation. Further, the controls for US exports are generally of expected (positive)sign
but the control for natural resources is highly unstable in sign and intermittently
significant.
The test for over dispersion- that the conditional variance of patent counts in fact
exceeds the mean- is significant in all of the Negbin regressions suggesting that of the
two distributions, it is preferred over the Poisson model. Nonetheless, with a few notable
exceptions, the results tell a broadly consistent story. The levels specification yields
estimates of the elasticity of patents with respect to R&D ranging from 1.00 to 1.14
across distributional assumptions. Controlling for heterogeneity through either the WG
17We use data on patents between 1963 and 1975 to construct the pre sample mean.
12
or PSM lowers the coefficient, with the exception of the WG Poisson which, at 1.65,
jumps to 50% above the highest of the levels coefficients. With this exception, the
pattern of estimates-Negbin falling in between the levels and Poisson- corresponds
closely to that predicted by Blundell, Griffith, and Windmeijer (2002) in the context of
endogenous regressors: the levels estimator is biased upwards and the WG estimator is
biased downwards and the superior PSM lying in between. In this scenario, the true
value is likely to be bracketed on the high end by the similar Negbin and Poisson
estimates (.78 and .86. respectively) and the linear PSM of .44 is at the low end. The
substantially lower linear estimates under the WG and PSM estimators cast doubt on the
idea that aggregate data obviates the issues of integers and zeros. The instability of the
WG estimates, and the fact the WG Negbin is the only specification to yield a
counterintuitive sign on US exports call the robustness of the estimator somewhat into
question in this context.
Dynamic Estimates
Table 4 presents the dynamic estimates of the innovation function. In the first
two panels, we present the linear and exponential models, analogous to the previous
exercise although, to recall, the linear feedback model has only been developed in the
exponential case for the Poisson distribution. It is important to note that in the linear
models the ln R&D coefficient retrieves the short run elasticity which needs to be
transformed in standard fashion to a long run elasticity whereas in the exponential models
the coefficient already gives the long run elasticity. This implies a lack of direct
comparability of the ln R&D coefficients across specifications and hence we focus on the
short and long run elasticities. Since including lagged variables implies a loss of a
quarter of the observations, the actual numerical results between the dynamic long run
coefficients and the ln R&D coefficients in the static case are not strictly comparable.
Nonetheless, it is striking is that the same patterns of relative magnitudes of the long run
coefficients emerge again with the ordering of levels, PSM and WG holding in the linear
case and the pattern not holding in the Poisson case due to, again, an unusually high
coefficient on the WG estimator. With the exception of this estimate, the long run
elasticities are broadly similar in magnitude to those of the static case.
13
Of more interest are the comparisons within the dynamic specifications. In
particular, the second two panels work with five year averages necessary to generate the
instruments for the system GMM estimator. The specification is clear of second order
serial correlation and passes the Sargan test for over identifying restrictions. We also
replicate the exponential exercises at quinquennial frequency, for comparability with the
linear specifications and with the estimates at annual frequency.
Though the long run elasticity differs across models, there are also impressive
commonalities. The linear and exponential estimates at 5 year frequency are strikingly
similar to the OLS, GMM, levels Poisson and PSM ranging between .86 and 1. The
exponential levels and PSM estimates at annual frequency also track closely their
quinquennial analogues. The annual levels and PSM linear specifications are
substantially lower at .71 and .54 and the WG specifications are especially low, with the
exception of the 2.64 on the annual exponential estimator, showing the same highly
erratic behavior across estimation technique seen in the static regressions. Again, in the
annual exponential and quinquennial linear cases, they generate a counterintuitive
negative sign on US trade which also casts some doubt on their reliability. If, for these
reasons and for the usual bias reasons associated with fixed effects in a dynamic context,
we discount the WG estimator, and hold aside the linear PSM at annual frequency, all
estimates are found between .71 and 1.06 with the bulk clustered around .88-1.06.
These results strongly support a relationship between R&D and patents at the
economy wide level. Further, they suggest elasticities that are substantially higher than
those at the firm level and overall, consistent with constant returns to scale in the
production of patents.
Advanced vs. developing countries
Tables 5 and 6 present the summary statistics for the variables that may influence
the magnitudes of the patenting elasticities in the high income and low income samples.
14
As mentioned earlier, the vast majority of emerging countries have far below 1000
patents and the mean is a mere fraction of that found in the advanced countries which
never shows any zeros.
Table 7 breaks the sample into OECD and developing countries.18 We again
report the results for our two preferred estimators although we now have clearer reasons
to choose one over the other. Of the 6.5 % of zero patent values in the sample, all are
found among the developing countries. Hence, in theory, the LFM estimator should be
preferable for this sample and the GMM for the advanced countries. In practice, there is
little difference in the estimates for the OECD-.95 for the LFM and 1.04 for the GMM-
offering somewhat stronger evidence for constant returns in the advanced countries.
The suspicion that, for some reason, the higher returns to scale arise from using
aggregate as opposed to private sector R&D spending turns out not to be justified. Using
a reduced sample for the OECD of countries with complete data on R&D done by the
productive sector versus the non productive sector (including government) still generates
an elasticity very close to unity for both aggregate- and productive-sector investment.
Since our aggregate sample spans a wide variety of sectors and countries which
cannot be seen as the aggregate counterpart to the US manufacturing firm level data,
asserting that the gap between this finding and the bulk of the micro findings captures
spillovers is probably too heroic. Nor does a finding of constant returns to scale in the
production of patents necessarily imply the same is true of knowledge generally. Still,
the robustness to technique, sample and proxy for variable of the CRS finding for the
OECD sample is suggestive that there we cannot rule out scale economies emerging at
the aggregate level that may offset the increasing difficulty and/or patent race effects that
might be found at the level of the firm.
The results from the developing countries are strikingly different, however, with
the LFM for the developing countries, yielding a coefficient of .7 and the probably
18OECD includes Korea but excludes Mexico which looks far more like the poor country sample.
15
inappropriate GMM, .9. Given the relatively small standard errors, the former estimate
suggests substantial decreasing returns to scale.
The final row of table 8 provides some "back of the envelope" calculations of the
"return to R&D measured in patents" by multiplying the elasticities by the ratio of the
average levels of patents to average level of R&D. If we crudely assume that patents
have equal average value across samples, then we can calculate the relative rate of return
to R&D. For the OECD sample, we find only a slight increase above those emerging
from either the Hall et al. or Griffith et al. estimates at the micro level. This strongly
contrasts with direct estimates of high social rates of return not appropriated privately
summarized by Jones and Williams (1998) which could well reflect the lack of
comparability among data sets. Staying within our sample, however, it can be said that
the OECD countries show a rate of return that, in the more reliable LFM estimates, is
higher by a factor of 5. If we assume that interest rates are probably higher in the
developing world than in the OECD and hence the private rate of return should also be
higher, developing countries exhibit decreasing returns to scale and substantially fewer
spillovers than those enjoyed in the advanced countries.
IV. Assessing and Explaining the Efficiency of Innovation
The striking differential in elasticities and implicit rates of return raises the
question as to what elements of LDC national innovation systems are giving rise to this
result. The greater variance in the nature and quality of the institutions comprising the
NIS across poor and rich countries offers the possibility of estimating their impact on
how an economy uses R&D to produce knowledge.
Data
We experiment with four variables, interacting them with R&D to see which may
raise or lower the elasticity. Again, we describe the data in tables 1,2 and figure 3. As
tables 5 and 6 show, the means of all of the variables are also substantially lower for the
16
non -OECD than is the case with the OECD although the amount of overlap is quite
large.
Years of Education: Recent literature stresses the complementarities between
technological progress and education. In Lucas (1988), for instance, knowledge spillovers
involve interactions among smart people (see also more recently Acemoglu and Zilibotti
2001). We employ the standard measure of the quality of human capital, from the Barro
and Lee (2000) data base, namely the average years of schooling at the beginning of each
five year period.
Perceived Quality of Academic Institutions: Low quality researchers, or even good
quality researchers housed in institutions with poor incentive systems are less likely to
produce patentable innovations for a given allocation of resources and would be less
likely to detect possibilities for building on other individuals' discoveries. The World
Economic Forum's World Competitiveness Report (2000) provides the average of the
subjective rating by entrepreneurs on a 7 point scale of the perceived quality of their
research institutions.
Level of Collaboration between the Research Institutions and the Private Sector: The
literature increasingly stresses the importance of networks and collaboration across the
various elements of the NIS as essential to the efficiency of generating new ideas, and to
building on existing ones. Again, the World Competitiveness Report asks entrepreneurs
to rank on a seven-point scale the level of collaboration among the private sector and
research institutions.
Intellectual Property Rights (IPRs): IPRs, in theory, can reduce the flow of spillovers by
making it more costly to create new inventions building on others' discoveries. On the
other hand, weak intellectual property protection within a country may lead firms to
simply hide their discoveries preventing spillovers all together. Thus, the protection and
revelation aspects of IPRs have countervailing effects with unclear ex ante impact. The
measure is taken from Park (2001) who combines indexes on patents, copyrights and
17
trademarks into one IPR measure. These sub-indices in turn are ranked on coverage of
protection, duration of protection, restrictions on exclusiveness of property rights,
membership in international treaties, and enforcement of rights on a scale from zero to
one, and then aggregated to form an index from 1 to 5. The variable is available only at
quinquennial frequency, but among proxies available, it is the only one with temporal
variation.
As is clear, the quality of these proxies is mixed, ranging from standard measures
to very subjective evaluations of difficult to measure phenomena. Further, the subjective
measures of quality of institutions and collaboration between the research institutions and
the private sector lack a time dimension and hence offer limited variation. Both the years
of education and the intellectual property indexes are reported every five years. Since all
of these variables appear correlated with level of development, we add to our core control
variables and interactive term with GDP per capita. Thus, the interpretation of the
interactions of our proxies is net of any effect due to the level of development.
Results
Given the lack of temporal variation and the potential difficulty of introducing
several interactive variables simultaneously, we begin with the non-dynamic version of
our preferred PSM specification. We then add back the dynamics and, despite the few
degrees of freedom, we run the exercise again with the GMM for completeness.
The results with the static specification suggest that our variables do have the
power to explain substantial differences in the elasticity. In the first column of table 9 the
significance of the interactive with log GDP and its positive sign confirm the results of
the previous section that the elasticity is higher for OECD countries than for developing
countries.
Columns 2-5 add the four NIS variables sequentially. All enter strongly
significantly with level of education and the quality of research institutions emerging as
18
the most important. In all cases, the coefficient on the log GDP interactive variable drops
by at least a third. Since, as table 2 suggests there is a high degree of correlation even
within the NIS variables, Column 6 includes them all simultaneously. All preserve their
significance although the quality of academic institutions and the level of education
emerge as the most quantitatively important. The negative sign and marginal
significance of the collaboration variable may arise because of the high correlation of the
two subjective variables and their lack of variance over time. In the next two columns,
we introduce them separately and confirm their significance and positive influence
individually.
In the full LFM specification (table 10), all variables retain their signs and overall
level of magnitude although the two subjective variables lose substantial significance. In
the specifications with all NIS variables included, only education emerges as significant
with IPRs retaining its sign but being significant only at the 15% level. The subjective
interactive dummies maintain sign but not significance. Dropping them, the IPR variable
appears again significant. Together, IPRs and education levels reduce the coefficient on
the GDP/capita interactive term by over half from .029 to .013 suggesting that they
account for much of the difference in elastic observed across income groups. The results
for the GMM dynamic specification (table 11) are strikingly similar with, again,
education and IPR entering significantly separately but only education maintaining its
level of significance when combined. Again, given the few degrees of freedom, this is
perhaps not surprising. Though admittedly imperfect, the proxies for elements of the NIS
appear to be significant determinants of the rate at which R&D is converted into patents.
V. Conclusions
Using a new global data base on patents and innovation inputs and drawing on
recent advances in dynamic estimation techniques, this paper confirms at the country
level the recurrent micro-level finding of a strong relationship between expenditures in
R&D and innovation output measured by US patents granted. The results are broadly
robust to estimation technique and imply a unitary elasticity for OECD countries or
19
constant returns to scale. This contrasts with firm level estimates using the same
techniques that imply decreasing returns to scale. One interpretation is that the increasing
difficulty of invention and redundant innovation efforts brought on by patent races
observed at the micro level are offset by spillover effects only observable at higher levels
of aggregation. The lack of comparability of the aggregate and micro samples leaves this
as a conjecture only.
What is very striking is both the lower elasticity found among developing
countries (suggesting strong decreasing returns to scale) and implicit rates of return that
are perhaps 20% of those found in the OECD. We find that several elements of the
national innovation system contribute to the differing elasticities observed. Education,
and the security of intellectual property rights emerge robustly as important factors and in
non dynamic specifications, the quality of research institutions and their interaction with
the private sector enter as well.
20
References:
Acemoglu, Daron, and Fabrizio Zilbotti. 2001. "Productivity Differences." The Quarterly
Journal of Economics 2:563-606.
Aghion, Philippe, and Peter Howitt. 1992. "A Model of Growth through Creative
Destruction." Econometrica 60: 323-51.
Aghion, Philippe, and Peter Howitt. 1998. Endogenous Growth Theory. The MIT Press.
Cambridge, Massachusetts.
Anderson, T.W., and Cheng Hsiao. 1982. "Formulation and Estimation of Dynamic
Models Using Panel Data." Journal of Econometrics 18: 47-82.
Arellano, Manuel, and Stephen Bond. 1991. "Some Tests of Specification for Panel Data:
Montecarlo Evidence and an Application to Employment Equations." Review of
Economic Studies 58(2): 277-297.
Arellano, Manuel, and Olympia Bover. 1995. "Another Look at the Instrumental-
Variable Estimation of Error Component Models." Journal of Econometrics 68(1): 29-52.
Arellano, Manuel. 2003. Panel Data Econometrics. Oxford and New York :Oxford
University Press.
Barro, Robert J., and Jong-Wha Lee. 2000. "International Data on Educational
Attainment: Updates and Implications." NBER Working Paper 7911, Cambridge,
Massachusetts.
Barro, Robert J., and Xavier Sala-i-Martin. 1995. Economic Growth. McGraw-Hill.
Blundell, Richard and Stephen Bond. 1998. "Initial Conditions and Moment Restrictions
in Dynamic Panel Data Model" Journal of Econometrics 87:115-143.
Blundell, Richard, Rachel Griffith, and John Van Reenen. 1995. "Dynamic Count Data
Models of Technological Innovation." Economic Journal 105: 333-344.
Blundell, Richard, Rachel Griffith, and Frank Windmeijer. 2002. "Individual Effects and
Dynamics in Count Data Models." Journal of Econometrics 108: 113-131.
Bond, Stephen R., Anke Hoeffler, and Jonathan Temple: 2001. "GMM Estimation of
Empirical Growth Models." CEPR Discussion Paper Series No. 3048.
Bottazzi, Laura, and Giovanni Peri (2003) - "Innovation and spillovers in regions:
Evidence from European patent data." European Economic Review (forthcoming).
21
Bound, John, Clint Cummings, Zvi Griliches, Bronwyn H. Hall, and Adam B. Jaffe.
1984. "Who Does R&D and Who Patents?" National Bureau of Economic Research
Working Paper 908, in Z. Griliches ed., R&D, Patents, and Productivity, (Chicago:
University of Chicago Press, 1984), 21-54.
Branstetter, Lee G. 2001. "Are Knowledge Spillovers International or Intranational in
Scope? Microeconomic Evidence from the U.S. and Japan." Journal of International
Economics 53: 53-79.
Caselli, Francesco, Gerardo Esquivel, and Fernando Lefort. 1996. "Reopening the
Convergence Debate: A New Look at Cross-Country Growth Empirics." Journal of
Economic Growth 1:363-390.
Chamberlain, Gary. 1992. "Comment: Sequential Moment Restrictions in Panel Data."
Journal of Business and Economic Statistics 10: 20-26.
Evenson, Robert E.. 1984. "International Invention: Implications for Technology Market
Analysis." In Zvi Griliches, ed., R&D, patents and productivity. University of Chicago
Press. Chicago.
Evenson, Robert E., and Yoav Kisleve. 1976. "A Stochastic Model of Applied Research."
Journal of Political Economy 84:265-281.
Furman, Jeffrey L., Michael Porter, and Scott Stern. 2002. "The Determinants of National
Innovative Capacity." Research Policy 31:899-933.
Griliches, Zvi. 1990. "Patent Statistics as Economic Indicators: A Survey." Journal of
Economic Literature 28(4):1661-1707.
Grossman, Genne E., and Elhanan Helpman. 1991 "Quality Ladders in the Theory of
Growth." Review of Economic Studies, 58(1) 43-61.
Hall, Bronwyn H, Jerry Hausman and Z. Griliches. 1986. "Patents and R&D: Is There a
Lag?" International Economic Review 27(2):265-83.
Hausman, Jerry, Bronwyn H. Hall, and Zvi Griliches. 1984. "Econometric Models for
Count Data with an Application to the Patents-R&D Relationship." Econometrica 52(4):
909-938.
Jaffe, Adam B.. 1986. "Technological Opportunity and Spillovers of R&D: Evidence
from Firms' Patents, Profits and Market Value." American Economic Review 76: 984
1001.
Jaffe, Adam B., and Manuel Trajtenberg. 2002. Patents, Citations & Innovations: A
Window on the Knowledge Economy. Cambridge, Massachusetts: The MIT Press.
22
Jones, Charles I., and John C. Williams. 1998. "Measuring the Social Return to R&D."
The Quarterly Journal of Economics 113: 1119-1135.
Jones, Charles I., and John C. Williams. 2000. "Too Much of a Good Thing? The
Economics of Investment in R&D." Journal of Economic Growth 5(1): 65-85
Jones, Charles I.. 1995. "R&D Based Models of Economic Growth." Journal of
Political Economy 103(4):759-84.
Klette, Tor J. and Samuel Kortum. 2002. "Innovating Firms and Aggregate Innovation."
NBER Working Paper 8819.
Kortum, Samuel. 1997. "Research, Patenting, and Technological Change." Econometrica
65(6): 1389-1419.
Kortum, Samuel. 1993. "Equilibrium R&D and the Patent--R&D Ratio: U.S. Evidence"
American Economic Review 83(2):450-457.
Kortum, Samuel, and Josh Lerner. 2000. "Assessing the Contribution of Venture Capital
to Innovation." RAND Journal of Economics 31(4):674-692.
Leamer, Edward. 1984. Sources of Comparative Advantage: Theory and Evidence.
Cambridge Massachusetts: MIT Press.
Lederman, Daniel, and Laura Saenz. 2003. "Innovation around the World: A Cross-
Country Data Base of Innovation Indicators." Mimeographed. Office of the Chief
Economist for LCR, The World Bank, Washington, DC.
Levine, Ross, Norman Loayza, and Thorsten Beck. 2000. "Financial Intermediation and
Growth: Causality and Causes." Journal of Monetary Economics 46: 31-77.
Lucas, Robert E., Jr.. 1988. "On the Mechanics of Economic Development." Journal of
Monetary Economics 22:3-42.
Lundvall, Bengt-Ake, Bjorn Johnson, Esben S. Andersen, and Bent Dalum. 2002.
"National Systems of Production, Innovation and Competence Building." Research
Policy 31:213-31.
Nelson, Richard R.. 1993. National Innovation Systems: A Comparative Analysis. New
York: Oxford University Press.
Nickell, Stephen J.. 1981. "Biases in Dynamic Models with Fixed Effects."
Econometrica, 49: 1417-1426.
Pakes, Ariel, and Zvi Griliches. 1980. "Patents and R and D at the Firm Level: A First
Report." Economics Letter 5: 377-381.
23
Park, Walter. 2001. "Intellectual Property and Patent Regimes." Chapter 4 in Economic
Freedom of the World: 2001 Annual Report. Freedom House: New York.
OECD (Organization for Economic Co-Operation and Development). 1998. Technology,
Productivity and Job Creation: Best Policy Practices. Paris: OECD Publications.
OECD. 2001. Innovative Clusters: Drivers of National Innovation Systems. Paris: OECD
Publications.
Reinganum, Jennifer F.. 1984. "Practical Implications of Game Theoretic Models of
R&D." American Economic Review 74(2): 61-66.
Reinganum, Jennifer F.. 1989. "The Timing of Innovation: Research, Development and
Diffusion" Chapter 14, Richard Schmalensee and Robert D. Willig eds. Handbook of
Industrial Organization, Volume 1. North Holland New York. 850-908.
Romer, Paul M.. 1987. "Growth Based on Increasing Returns Due to Specialization."
American Economic Review Papers and Proceedings 77: 56-62.
Romer, Paul M.. 1990. "Endogenous Technological Change." Journal of Political
Economy 98: 71-102.
Segerstrom, Paul S. 1998. "Endogenous Growth without Scale Effects" American
Economic Review 88 (5) 1290-1310.
Segerstrom, Paul S., T.C.A. Anant, and Elias Dinpoulos. 1990. "A Schumpeterian Model
of the Product Life Cycle." American Economic Review 80(5): 1077-92.
Trajtenberg, Manuel. 2001. "Innovation in Israel 1968-1997: A Comparative Analysis
Using Patent Data." Research Policy 30(3): 363-389.
UNESCO. 1980. Statistical Yearbook. United Nations, Paris.
Windmeijer, Frank. 2000. "A finite sample correction for the variance of linear two-step
GMM estimators." IFS Working Papers W00/19, Institute for Fiscal Studies.
Wooldridge, Jeffrey M..1997. "Multiplicative Panel Data Models without the Strict
Exogeneity Assumption," Econometric Theory 13, 667-678.
World Economic Forum. 2000. World Competitiveness Report. Lausanne, Switzerland.
24
Table 1: Descriptive statistics
Std. % of
Variable Mean Dev. Min Max zeros
Patents 1132 3509 0 31104 6.45
ln R&D 20.90 2.33 11.25 25.82
ln US trade 15.20 1.68 7.82 19.05
Natural Resources 0.21 1.66 -4.74 14.88
Quality of Academic Institutions 4.93 0.97 2.80 6.30
Collaboration with the Private Sector 3.97 0.95 2.54 5.78
Education 7.05 2.23 2.21 11.84
IPR 2.93 0.94 0.33 4.57
ln (Mean Patents) 3.48 2.70 -2.56 8.39
25
Table 2: Correlations
Patents ln R&D UStrade Natural ln (Mean Quality Coll. Educ. IPR
Res. Patents) Acad. Private
Inst. Sector
Patents 1.00
ln R&D 0.51 1.00
ln US trade 0.47 0.69 1.00
Natural Resources -0.27 -0.16 -0.07 1.00
ln (Mean Patents) 0.27 0.66 0.35 0.01 1.00
Quality of Academic Institutions 0.32 0.59 0.34 0.08 0.90 1.00
Collaboration with the Private Sector 0.27 0.51 0.34 0.27 0.64 0.71 1.00
Education 0.30 0.63 0.30 -0.07 0.64 0.64 0.70 1.00
IPR 0.47 0.84 0.61 -0.07 0.69 0.61 0.60 0.64 1.00
26
Table 3: Static Estimates Patents on R&D
Observations 636
Countries 49
Period 1976-99
Linear Exponential
Negbin Poisson
Levels WG PSM Levels WG PSM Levels WG PSM
(1) (2) (3) (4) (5) (6) (7) (8) (9)
ln R&D 1.01 *** 0.36 *** 0.44 *** 1.14 *** 0.56 *** 0.78 *** 1.00 *** 1.65 *** 0.86 ***
(0.05) (0.04) (0.05) (0.03) (0.04) (0.06) (0.03) (0.01) (0.04)
ln US trade 0.17 *** 0.25 *** 0.14 *** 0.11 *** -0.29 *** 0.14 *** 0.19 *** 0.04 *** 0.22 ***
(0.05) (0.07) (0.03) (0.03) (0.05) (0.03) (0.03) (0.01) (0.02)
Natural Resou 0.05 *** -0.10 *** -0.02 0.06 *** 0.00 0.00 0.05 *** -0.01 *** 0.01
(0.02) (0.02) (0.03) (0.02) (0.02) (0.01) (0.01) (0.00) (0.01)
ln (Mean Patents) 0.61 *** 0.33 *** 0.11 ***
(0.04) (0.04) (0.04)
Over-Dispersion 0.65 *** 0.28 *** 0.50 ***
(0.04) (0.01) (0.06) (0.04)
Note: The dependent variable in the linear specifications is log Patents. We also include an dummy for those
countries with no Patents. The dependent variable in the exponential specification is the count of Patents. Robust
standard errors were estimated in all the specifications. All specifications include time dummies. WG and PMS
refer to the within-group and the pre-sample mean estimations respectively.
*Significant at 10%, ** Significant at 5%, *** Significant at 1%
27
Table 4: Dynamic Estimates: Patents on R&D
Annual Five Years
Observations 482 105
Countries 46 43
Period 1978-99 1985-99
Annual Five Years
Linear Exponential Linear Exponential
Levels WG PSM Levels WG PSM OLS WG GMM Levels WG PSM
Patents(-1) 0.95 *** 0.62 *** 0.91 *** 0.32 *** 0.41 *** 0.45 *** 0.82 *** 0.26 *** 0.62 *** 0.46 *** 0.47 0.54 ***
(0.02) (0.03) (0.24) (0.09) (0.08) (0.06) (0.13) (0.05) (0.12) (0.09) (0.30) (0.11)
ln R&D 0.03 0.12 *** 0.03 1.06 *** 2.64 *** 0.91 *** 0.18 *** 0.23 *** 0.37 ** 0.91 *** 0.16 0.88 ***
(0.02) (0.03) (0.02) (0.04) (0.37) (0.07) (0.07) (0.06) (0.15) (0.16) (0.10) (0.12)
ln US trade 0.02 0.13 ** 0.03 * 0.19 ** -0.24 0.22 ** 0.11 *** -0.06 *** 0.14 ** 0.29 *** 1.38 *** 0.28 ***
(0.01) (0.06) (0.02) (0.10) (0.19) (0.12) (0.02) (0.02) (0.05) (0.06) (0.25) (0.08)
Natural Resources 0.00 -0.04 * -0.01 0.09 *** 0.03 * 0.07 ** 0.15 -0.19 0.01 0.01 -0.28 *** 0.01
(0.01) (0.02) (0.02) (0.02) (0.02) (0.04) (0.11) (0.13) (0.01) (0.01) (0.06) (0.01)
ln (Mean Patents) 0.04 *** 0.10 0.06
(0.02) (0.07) (0.05)
Long-Run Elast. of R&D 0.71 0.33 0.37 1.06 2.64 0.91 1.00 0.31 0.96 0.91 0.16 0.88
Short-Run Elast. of R&D 0.03 0.12 0.03 0.71 1.56 0.50 0.18 0.23 0.21 0.50 0.09 0.40
Sargan Test (p-value) 0.63
First Order (p-value) 0.24
Second Order (p-value) 0.23
Note: The dependent variable in the linear specifications is log Patents. We also include an dummy for those countries with no Patents. Robust standard errors
were estimated in all the specifications. The dependent variable in the exponential specification is the count of Patents. The exponential models were computed
using EXPEND (see Windmeijer 2000), whereas the linear models were estimated in Stata. For the PSM models we include two lags of the dependent variable in
set of instruments. The GMM system estimator used two lags of the variables and first differences to instrument. All the specifications include time dummies.
*Significant at 10%, ** Significant at 5%, *** Significant at 1%
28
Table 5: OECD Statistics
Observations=346 Mean Std. Dev. Min Max % of zeros
Patents 1975 4560 1 31104 0
ln R&D 22.3 1.58 18.88 25.8
ln US trade 15.7 1.53 12.73 19.0
Natural Resources 0.28 2.08 -3.69 14.9
Quality of Academic Institutions 5.53 0.62 4.10 6.3
Collaboration with the Private Sector 4.53 0.75 2.92 5.8
Education 8.32 1.71 3.27 11.8
IPR 3.53 0.56 1.98 4.6
ln (Mean Patents) 5.21 2.15 1.26 8.4
Table 6: non-OECD
Observations=290 Mean Std. Dev. Min Max % of zeros
Patents 34 91 0 754 14.40%
ln R&D 19.23 1.95 11.25 22.80
ln US trade 14.64 1.67 7.82 18.46
Natural Resources 0.13 0.91 -4.74 5.07
Quality of Academic Institutions 4.23 0.82 2.80 6.20
Collaboration with the Private Sector 3.30 0.71 2.54 5.30
Education 5.53 1.79 2.21 9.82
IPR 2.22 0.80 0.33 3.90
ln (Mean Patents) 1.41 1.60 -2.56 4.13
29
Table 7: Elasticities: OECD vs. non-OECD
Model Periodicity Total OECD Non-OECD
Static
Linear PSM Annual 0.44 *** 0.82 *** 0.14 ***
(0.05) (0.05) (0.02)
Exponential PSM Annual 0.78 *** 0.83 *** 0.22 ***
(0.06) (0.04) (0.03)
Dynamic
Linear GMM Quiquenial 0.96 *** 1.04 *** 0.89 ***
(0.09) (0.06) (0.11)
Exponential PSM Annual 0.91 *** 0.95 *** 0.70 ***
(0.06) (0.04) (0.12)
Table 8: Returns to R&D
Hall et al Griffit et al LFM GMM
(1986) (2002)
Total OECD NON-OECD Total OECD NON-OECD
Firms 642 407
Patents 35 35 1132 1975 34 1097 1831 119
R&D Millions of Dollars of 1976 21.48 35.00
R&D $M of 1995 49.91 81.30 9010 12050 722 8320 13700 1110
R&D/Patents 1.43 2.32 7.96 6.10 21.24 7.58 7.48 9.33
Correction Applied/Granted 0.65 0.65
R&D/Patents (Corrected) 2.19 3.57
Elasticity 0.3 0.506 0.91 0.95 0.7 0.98 1.04 0.9
Implied Returns to R&D 13.68% 14.16% 11.43% 15.57% 3.30% 12.92% 13.90% 9.54%
30
Table 9: NIS Efficiency Static Negative Binomial with Pre-sample mean
Annual
Observations 636
Countries 49
Period 1976-99
ln R&D 0.366 *** 0.155 *** 0.314*** 0.372*** 0.301*** 0.135*** 0.188*** 0.300***
(0.03) (0.01) (0.03) (0.03) (0.03) (0.01) (0.04) (0.05)
ln US trade 0.262 *** 0.324 *** 0.280*** 0.231*** 0.316*** 0.331** 0.320*** 0.288***
(0.03) (0.04) (0.04) (0.02) (0.03) (0.03) (0.03) (0.05)
Natural Resources -0.033 ** -0.061 ** -0.056* -0.075*** -0.004* -0.060** -0.060** -0.055**
(0.02) (0.03) (0.03) (0.02) (0.00) (0.03) (0.03) (0.03)
ln GDPcap 0.028 *** 0.024 *** 0.021*** 0.013** 0.018** 0.011** 0.008** 0.006**
(0.01) (0.01) (0.05) (0.01) (0.09) (0.01) (0.00) (0.00)
ln (Mean Patents) 0.210 *** 0.174 *** 0.216*** 0.229*** 0.225** 0.184*** 0.207*** 0.233***
(0.04) (0.04) (0.05) (0.04) (0.05) (0.04) (0.04) (0.04)
Quality 0.097 *** 0.115*** 0.065***
(0.03) (0.03) (0.01)
Colaboration 0.054*** -0.049* 0.025***
(0.02) (0.03) (0.00)
Education 0.068*** 0.053*** 0.045*** 0.050***
(0.02) (0.02) (0.01) (0.00)
IPR 0.021*** 0.014** 0.015*** 0.016***
(0.01) (0.01) (0.01) (0.00)
Note: The dependent variable is the count of Patents. Robust standard errors were estimated in all the
specifications. All specifications include time dummies.
31
Table 10: NIS Efficiency LFM Estimator
Annual
Observations 482
Countries 46
Period 1976-99
Lagged Patents 0.419 *** 0.400*** 0.440*** 0.479*** 0.409*** 0.478 ***
(0.07) (0.06) (0.07) (0.08) (0.07) (0.08)
ln R&D 0.260 ** 0.182* 0.220* 0.376*** 0.260** 0.350 ***
(0.13) (0.11) (0.12) (0.12) (0.13) (0.11)
ln US trade 0.290 *** 0.330*** 0.300*** 0.222*** 0.315*** 0.260 ***
(0.08) (0.08) (0.11) (0.07) (0.11) (0.08)
Natural Resources -0.014 -0.037 -0.036 -0.046* -0.008 -0.044
(0.02) (0.04) (0.03) (0.03) (0.01) (0.05)
ln GDPcap 0.029 *** 0.029*** 0.026*** 0.018*** 0.027** 0.013 **
(0.01) (0.01) (0.01) (0.00) (0.01) (0.01)
ln (Mean Patents) 0.154 ** 0.120** 0.162*** 0.097** 0.147** 0.098 **
(0.07) (0.06) (0.06) (0.05) (0.07) (0.04)
Quality 0.034
(0.02)
Colaboration 0.023
(0.02)
Education 0.049** 0.058 ***
(0.02) (0.02)
IPR 0.014 ** 0.010 *
(0.07) (0.01)
Note: The dependent variable is the count of Patents. Robust standard errors were estimated in all
the specifications. All specifications include time dummies.
32
Table 11: NIS Efficiency GMM Estimator
Observations 105
Countries 43
Period 1985-99
Lagged Patents 0.580 *** 0.634 *** 0.583 *** 0.603*** 0.586*** 0.534 ***
(0.14) (0.17) (0.18) (0.14) (0.12) (0.16)
ln R&D 0.147 * 0.071 0.179 0.057 0.127 0.118
(0.08) (0.06) (0.13) (0.05) (0.11) (0.10)
ln US trade 0.229 *** 0.274 *** 0.204 *** 0.322*** 0.201*** 0.237 ***
(0.06) (0.08) (0.07) (0.08) (0.07) (0.07)
Natural Resources -0.011 0.001 -0.012 -0.026 0.013 -0.020
(0.03) (0.00) (0.01) (0.02) (0.01) (0.02)
ln GDPcap 0.013 *** 0.012 * 0.009 * 0.007** 0.010*** 0.010
(0.00) (0.01) (0.01) (0.00) (0.00) (0.01)
Quality 0.018
(0.01)
Colaboration 0.025
(0.02)
Education 0.047*** 0.037 ***
(0.00) (0.01)
IPR 0.010*** 0.003
(0.00) (0.00)
Sargan Test (p) 0.88 0.85 0.86 0.90 0.99 1
First Order (p) 0.51 0.35 0.24 0.15 0.18 0.11
Second Order (p) 0.22 0.27 0.23 0.30 0.29 0.32
Note: The dependent variable is log Patents. Robust standard errors were estimated in all the
specifications. All specifications include time dummies.The GMM system estimator used two lags of
the variables and first differences to instrument.
33
Annex 1: Countries in the sample
Quality Coll. Ln(Mean
Code N Patents R&D US trade Natural Acad. Private
Res. Edu. IPR Patents
Inst. sector 60-75)
ARG 14 26.36 20.58 14.43 0.64 3.40 2.71 7.64 2.59 3.09
AUS 11 420.73 22.11 15.26 2.26 5.90 4.84 10.14 3.37 5.04
AUT 20 336.45 21.72 14.14 -0.76 5.60 4.81 8.30 4.09 5.26
BOL 8 0.38 16.97 12.23 0.21 2.80 2.66 5.13 2.19 0.96
BRA 10 53.40 22.27 16.01 0.09 4.40 3.03 3.81 2.21 2.74
CAN 22 1723.91 22.67 18.44 2.29 5.70 4.81 10.51 2.95 6.95
CHE 10 1231.50 22.57 15.22 -1.70 6.30 5.15 9.32 3.55 6.98
CHL 16 5.50 19.26 14.31 1.10 4.30 3.77 6.79 2.57 1.31
CHN 5 66.40 22.45 17.95 -0.01 4.50 4.00 5.61 1.55 1.98
COL 7 5.71 18.88 15.07 0.25 3.70 2.98 4.59 2.25 1.59
CRI 10 1.70 16.88 14.08 0.54 4.80 3.66 5.44 1.61 -0.08
DEU 10 7508.50 24.78 17.43 -1.36 5.90 5.04 9.57 3.79 8.39
DNK 21 205.67 21.52 14.46 0.77 5.30 4.52 9.57 3.69 4.83
ECU 7 1.14 16.50 14.47 0.76 3.10 2.60 6.19 2.21 -0.62
EGY 10 1.80 20.84 13.36 -0.13 4.30 3.12 3.97 1.99 0.48
ESP 19 128.42 21.98 15.13 -0.77 4.80 3.24 5.91 3.57 4.06
FIN 20 292.55 21.54 14.21 1.85 6.30 5.78 8.83 3.18 3.92
FRA 22 2642.77 24.11 16.46 -0.73 6.20 4.59 7.32 3.88 7.47
GBR 18 2647.83 23.77 16.98 -0.63 6.10 4.73 8.68 3.54 7.93
GRC 14 9.43 19.53 13.09 -0.54 4.10 2.92 7.04 2.48 2.18
HUN 5 41.40 19.62 13.85 -0.10 5.20 3.59 8.67 3.37 3.50
IDN 12 2.08 19.17 15.74 0.22 3.70 3.14 3.35 0.41 1.55
IND 17 18.59 21.21 15.13 -0.01 5.20 2.76 3.31 1.56 2.53
IRL 21 45.52 19.98 14.50 0.92 5.60 4.77 8.10 3.07 2.55
ISR 16 346.25 21.43 15.24 -1.29 6.20 4.80 9.08 3.57 3.98
ITA 23 1053.87 23.00 16.38 -1.52 4.60 3.30 5.91 3.93 6.35
JAM 5 0.60 14.31 13.02 -0.09 4.50 3.10 3.81 2.86 -0.08
JOR 9 0.11 16.16 9.17 -1.23 4.50 3.12 3.49 1.86 -2.56
JPN 24 16017.38 25.41 18.28 -1.77 5.70 5.08 8.63 3.88 7.95
KOR 23 640.22 22.03 16.52 -0.92 4.90 4.07 8.13 3.62 1.26
LKA 4 0.25 16.48 13.16 0.00 4.00 3.00 5.36 2.95 -1.47
MEX 13 45.31 20.58 17.61 0.24 3.70 2.86 5.46 2.00 4.13
MYS 6 9.00 18.81 16.15 1.29 4.20 3.67 6.06 2.71 -0.26
NLD 22 821.00 22.66 15.52 1.21 6.20 5.14 8.44 4.20 6.31
NOR 17 107.12 21.22 14.75 5.66 5.40 4.62 9.14 3.31 4.20
NZL 7 51.71 20.12 14.23 3.65 5.60 4.47 11.28 3.48 2.85
PER 16 2.25 16.84 14.06 0.22 3.60 2.81 5.98 1.57 1.33
PHL 10 5.00 18.50 15.09 -0.02 4.00 3.00 6.35 2.67 1.67
POL 5 13.60 20.74 13.51 -0.15 4.50 3.13 9.82 2.90 3.01
PRT 10 3.90 19.85 13.62 -0.84 4.40 3.15 3.98 2.28 1.48
SGP 13 51.23 20.32 16.31 -1.96 5.60 5.30 6.10 3.08 0.69
SLV 9 0.33 17.95 13.42 -0.05 2.90 2.54 3.84 2.49 0.27
SWE 12 824.33 22.50 15.38 0.29 6.00 5.37 9.71 3.62 6.49
THA 14 3.21 19.03 15.17 0.04 4.20 3.52 4.77 1.91 -0.96
TTO 2 0.50 15.84 13.94 3.13 3.90 3.40 7.40 3.35 0.57
TUR 13 2.15 20.31 14.14 -0.14 3.50 3.11 4.01 1.80 0.65
URY 10 1.80 17.49 12.22 0.27 3.90 3.30 6.88 2.43 0.43
VEN 19 21.42 19.29 16.04 1.66 4.00 2.67 5.16 1.76 2.03
ZAF 5 99.80 20.75 14.58 0.32 5.10 4.42 5.11 3.57 4.06
34
Figure 1: Patents vs. R&D Expenditures
000
30
0
stn 0002
teaPl
taoT
00
100
0
10 15 20 25
Ln of Total R&D Expenditure
35
Figure 2: Patents vs. R&D Expenditures
1000
30000
800
stn 20000 stn 600
teaPl teaPl
taoT taoT 400
10000
200
0 0
22 23 24 25 26 10 15 20 25
Ln of Total R&D Expenditure Ln of Total R&D Expenditure
36
Figure 3: ln Patents vs NIS variables
10 10
8 8
s s
ent 6 ent 6
Pat Pat
alotT 4 alotT 4
Ln Ln
2 2
0 0
3 4 5 6 7 2 3 4 5 6
Quality of Research Institutions Collaboration
10 10
8 8
s s
ent 6 ent 6
Pat Pat
alotT 4 alotT 4
Ln Ln
2 2
0 0
2 4 6 8 10 12 0 1 2 3 4 5
Years of Education Property Rights
37
38