Productivity, trade and the R&D content of intermediate inputs
Shuichiro Nishioka�
and
Marla Ripolly
May, 2012
Abstract
This paper explores a novel way to evaluate the extent to which R&D knowledge embodied inintermediate inputs correlates with productivity at the industry level. We propose the conceptof R&D content of intermediates, which represents the knowledge stock embodied in intermedi-ate inputs used in production. Using a sample of 32 countries and 13 manufacturing industrieswe compute the elasticity of industry-level TFP with respect to the R&D content of intermedi-ates. We �nd that among high-R&D industries, the R&D embodied in inputs purchased fromthe own industry is signi�cantly associated with industry-level TFP. In this case, both own-industry domestic inputs as well as those imported from G5 countries are relevant. In contrast,intermediate input trade does not appear to be a signi�cant channel of R&D di¤usion amonglow-R&D industries.
Keywords: international transfer of technology, intermediate input trade, R&D stock, industry-level TFPJEL Classi�cation: F14, O33, O5.
�Department of Economics, West Virginia University, 1601 University Avenue, Morgantown WV 26506-6025,e-mail: [email protected].
yDepartment of Economics, University of Pittsburgh, 4933 W.W. Posvar Hall, Pittsburgh PA 15260, e-mail:[email protected].
1 Introduction
The idea that technology transfers across borders are a powerful engine to increase productivity
levels around the world continues to resonate among academics and policy makers. Technological
knowledge embodied in intermediate inputs traded across industries and countries constitutes one
of the channels of technology di¤usion. Intermediate input trade has been at the center of both
theoretical and empirical studies of productivity. On the theory side, intermediate input trade has
been modeled as a mechanism of knowledge di¤usion in the endogenous growth literature introduced
by Romer (1990), Grossman and Helpman (1991) and Aghion and Howitt (1992). According to
these models, the use of domestic and foreign intermediate inputs is associated with increases
in total factor productivity. On the empirical side, the idea that intermediate inputs embody
R&D knowledge and their use is correlated with higher productivity at the industry level was �rst
introduced by Schmookler (1966), and was subsequently examined by Terleckyj (1974), Scherer
(1982), Griliches and Lichtenberg (1984), Goto and Suzuki (1989), and Keller (2002a, b), among
others. While earlier empirical papers were restricted to the US or individual countries, papers in
the last decade or so have examined intermediate input trade and productivity in a larger sample
of countries. However, the analysis of intermediate input trade across industries and countries has
been limited due to data availability on comparable and reliable international input-output tables.
For instance, an assumption in Keller (2002a) is that the input-output structure in all eight OECD
countries in his sample, including the split between domestic and foreign inputs, is the same as in
the US. Newly available data suggest important di¤erences in the input-output structure across a
larger set of countries.
This paper explores a novel way to evaluate the extent to which R&D knowledge embodied in
intermediate inputs a¤ects productivity at the industry level. We propose the concept of R&D
content of intermediates, which represents the R&D stock embodied in intermediate goods used
in production. This concept parallels the one of factor content of trade in the international trade
literature. Our concept of R&D content of intermediate inputs has two distinguishing features.
The �rst one is that for any speci�c industry one can identify the R&D knowledge originated
from each of the other industries, and can also identify the speci�c source countries. The second
distinguishing feature is that our concept of R&D content of intermediate inputs embeds both
the direct and the indirect �ows of inputs across industries as in Leontief�s input-output model.
1
Constructing measures of the R&D content of intermediates requires the use of a global input-output
matrix that speci�es all industries and countries involved in the intermediate purchases. We are
able to construct these measures by using recently available input-output data for 32 countries
and 13 manufacturing industries in 1995, 2000, and 2005. The data are the most comprehensive
set available to date. We use our R&D content measure together with industry-level measures
of total factor productivity (TFP) in order to examine the correlation between productivity and
the R&D content of intermediate inputs. A distinct feature of our analysis is that we compute
not only the elasticity of industry-level TFP with respect to the R&D content of intermediates
for the manufacturing sector as a whole, but we also allow the elasticity to vary among high and
low-R&D intensive industries. Speci�cally, we classify chemicals, electrical and optical equipment,
motor vehicles and other transport equipment as high-R&D industries, while low-R&D intensive
industries correspond to food products, textiles, wood and paper products, rubber and plastics,
mineral products, basic and fabricated metals, and machinery.
The analysis yields four main �ndings. First, our proposed measure of R&D content of inter-
mediates provides a reasonable characterization of how industries and countries di¤er in terms of
their exposure to the R&D knowledge produced by other industries and countries. For instance,
consider the case of G5 countries (US, Japan, France, Germany and the UK) which are the R&D
leaders in the world. Our R&D content measure suggests that while high-R&D industries in G5
countries have a domestic R&D stock that is much larger than the total R&D embodied in the
intermediate inputs they purchase (domestic and foreign), the opposite tends to occur in non-G5
countries. In addition, regardless of the country, low-R&D industries tend to have own-industry
domestic R&D stocks smaller than the total R&D content of the intermediate inputs they pur-
chase. Our R&D content measure provides an interesting and useful representation of the e¤ective
embodied knowledge �ows among industries and countries.
The second �nding is that the elasticity of industry-level TFP with respect to R&D content
of intermediates is positive and signi�cant among all 13 manufacturing and high-R&D intensive
industries. In our benchmark equation we estimate an elasticity of 0:082 for high-R&D industries,
twice the estimated elasticity for all manufacturing industries of 0:041 (both statistically signi�cant
at the 5% con�dence level). These results are quite robust to a number of alternative speci�cations.
The elasticities we estimate are within the range of magnitude of others estimated in the literature
2
(e.g., Keller, 2002b), with the important distinction that we use a novel measure of the foreign
R&D industries can acquire, and we also distinguish low and high-R&D industries.
The third �nding is that among high-R&D industries, own-industry inputs as well as foreign
inputs are the ones that have a signi�cantly positive e¤ect on industry-level TFP. We estimate a
TFP elasticity to own-industry R&D content of 0:078, three times the elasticity obtained for the
whole set of 13 manufacturing industries. The TFP elasticity to foreign R&D content is estimated
at 0:075. But more interestingly, we �nd that G5 countries are the key sources of R&D among high-
R&D industries: the TFP elasticity to G5-foreign R&D content is estimated to be 0:075. These
results suggest that knowledge spillovers are largely industry speci�c: the most important sources
of R&D embodied in inputs for high-R&D industries are high-R&D industries in G5 countries.
Things are quite di¤erent for low-R&D industries. Our fourth �nding is that intermediate
input trade does not appear to be the relevant channel by which foreign R&D di¤uses to low-R&D
industries. We do not �nd the TFP elasticity to R&D content to be statistically signi�cant among
low-R&D industries under any speci�cation. In addition, G5 countries do not appear to be key
sources of R&D embodied in inputs for these industries. Our proposed concept of R&D content of
intermediates is useful to identify important di¤erences among low and high-R&D industries.
Our paper relates to a number of literatures. First, as pointed out before, it relates to the
theoretical literature on R&D knowledge and productivity. Second, our paper relates to the more
general empirical literature pioneered by Coe and Helpman (1995), and followed by Lichtenberg
and van Pottelsberghe de la Potterie (1998, 2001), Xu and Wang (1999), Madsen (2007), and Coe,
Helpman and Ho¤maister (2009), among others. This literature documents the extent to which
domestic and foreign knowledge a¤ect productivity at the aggregate level. While some of these
papers measure knowledge using R&D stocks, others use patent counts (e.g., Madsen, 2007). A
feature of these papers is that they construct measures of foreign R&D by using shares of imports
on GDP as weights for the foreign R&D stocks from source countries.1 Our paper di¤ers from these
studies in that we focus on industry-level rather than aggregate data, and on intermediate input
�ows at the domestic and international levels. In addition, rather than using total bilateral imports
as a fraction of GDP as weights for foreign R&D stocks, we use our measure of R&D content of
intermediate inputs as a more informative way to capture how foreign knowledge has an impact on
1Xu and Wang (1999) use shares of capital-goods imports as weights in constructing the foreign R&D stock.
3
industry-level productivity.
Third, our paper relates to other studies that examine the relationship between productivity
and R&D at the industry-level, especially to Keller (2002a, b). The main �nding in Keller (2002a) is
that the use of own-industry domestic inputs accounts for 50% of the e¤ect of R&D on productivity.
However, his sample includes only eight OECD countries, and country-speci�c input-output tables
are not used in the analysis. Ours is the �rst paper to systematically use reliable international input-
output tables for a sample of 32 countries in order to capture accurate patterns of intermediate input
purchases in domestic and international markets at the industry level. The use of this data set allows
us to construct a measure of R&D content of intermediates and to estimate its correlation with
industry-level productivity. Last, our paper also relates to an emerging literature that examines
the e¤ect of imported inputs on plant productivity, including Kasahara and Rodrigue (2008) and
Halper, Koren and Szeidl (2011). The former uses Chilean data, while the latter Hungarian data,
but both document that using foreign intermediates increases productivity at the plant level. Our
data are available at the two-digit (ISIC Rev.3) industry level rather than at the plant level, but
we �nd evidence of increased productivity from the use of foreign inputs only among high-R&D
industries.
The remainder of the paper is organized as follows. Our proposed R&D content of interme-
diates measure is presented in Section 2. Section 3 discusses the data construction and provides
a description of the main features of the R&D content measure. Section 4 contains the empirical
framework and results. We discuss our results and perform a robustness analysis in Section 5.
Section 6 concludes.
2 R&D content of intermediate inputs
The relationship between productivity and R&D investments has been the focus of both theoretical
and empirical models. The typical equation commonly estimated in the aggregate, country-level
empirical literature pioneered by Coe and Helpman (1995) takes the form
lnTFPit = �i + � lnSit + � lnSfit + "it (1)
4
where TFPit is productivity in country i at time t; �i is a country-speci�c �xed e¤ect; Sit is the
domestic R&D stock, and Sfit is a measure of the foreign R&D stock. The issue is how to construct
measures of Sfit. The approach generally taken is to use shares of imports among pairs of countries
as weights for the respective R&D stocks as follows
Sfit =Xj
!ijtSjt =Xj
importsijtGDPjt
Sjt
where "importsijt" correspond to those between recipient country i and source country j.
The industry-level version of equation (1) can be written as
lnTFPiht = �it + �h + � lnSiht + � lnSfiht + "iht (2)
where h is an industry index and �h an industry �xed e¤ect. Keller (2002b) adopts a non-linear
version of the equation above as follows2
lnTFPiht = �it + �h + � lnhSiht + �S
fiht
i+ "iht (3)
and uses distance-weighted R&D stocks to measure Sfiht
Sfiht =Xj2G5
Sjht exp(��Dij)
where Dij is the geographic distance between recipient country i and source country j.3 Notice
that in the equation above source countries are restricted to be G5 countries, and industry h in
the recipient country is only a¤ected by the R&D stock of the same industry h in G5 countries.
In another paper, Keller (2002a) uses US input-output tables to derive inter-industry trade of
intermediate goods across industries and countries. Again using his non-linear speci�cation, the
foreign R&D stock Sfiht is divided in two components: same or own-industry input trade and
other-industry inputs. The former is measured as
Sf�owniht =Xj 6=i
mijhtSjht
2Keller (2002b) uses a country-by-industry dummy �ih and a time dummy �t.3We discuss the implications of the linear versus the non-linear speci�cation below in Section 4.3.
5
where mijht is the share of imports by recipient country i from source country j in own-industry
h. The foreign other-industry component is measured as
Sf�otheriht =Xg 6=h
hgSf�ownjgt
where hg is the share of a country�s imports of the g intermediate good that go to industry h:
We propose the concept of R&D content of intermediates in order to measure both the domestic
and foreign R&D stocks in equation (2). The idea is to capture the R&D knowledge embodied in the
domestic and foreign intermediate inputs that any given industry uses. The concept we propose
parallels the well-known concept of factor content of trade. In a recent paper, Tre�er and Zhu
(2010) revisit the de�nition of factor content of trade and in the process construct a global country-
by-industry input-output matrix. Our purpose here is to use this global input-output matrix in
order to de�ne R&D content of intermediates.
Our concept of R&D content of intermediates represents an improvement over the measures
of domestic and foreign R&D stocks presented above. First, our R&D content measure captures
in a uni�ed framework all the R&D embodied in domestic and foreign inputs. Thus, for any
speci�c industry one can identify the R&D knowledge originated from each of the other industries,
and can also identify the speci�c source countries. Second, it properly weights R&D stocks by
using intermediate input purchases across industries and countries. Since we analyze industry-level
productivity, this is the most accurate way to capture the relevant R&D knowledge. Intermediate
input �ows are then e¤ective ways of embodied technology di¤usion across sectors and countries.
Third, our concept of R&D content of intermediate inputs embeds both the direct and the indirect
�ows of inputs across industries as in Leontief�s input-output model.
As in Tre�er and Zhu (2010), let g and h = 1; : : : ; G be indexes for goods (industries), and
i and j = 1; : : : ; N for countries. Assume every good is consumed as a �nal product or used as
an intermediate input. Let Qi be a G � G diagonal matrix for country i�s gross output, and let
Bij(g; h) is the amount of intermediate input g from country i used to produce one unit of gross
output of country j�s good h.4 If i = j, Bii(g; h) is the typical (g; h) element of the G�G matrix
of domestic intermediate requirements Bii. If i 6= j, Bij(g; h) is the typical (g; h) element of the
G � G matrix of foreign intermediate requirements Bij . Based on these de�nitions, the following
4 In Tre�er and Zhu (2010) Qi is a G� 1 matrix but we �nd it more convenient to have a diagonal G�G matrix.
6
global matrixes can be de�ned:
Q =
266666664
Q1 0 ::: 0
0 Q2 ::: 0
......
. . ....
0 0 ::: QN
377777775; B =
266666664
B11 B12 ::: B1N
B21 B22 ::: B2N...
.... . .
...
BN1 BN2 ::: BNN
377777775;
where Q is an NG�NG diagonal matrix of gross output, and B is an NG�NG matrix of global
intermediate techniques whose typical element is Bij(g; h). We will be using the global country-
by-industry input-output matrix BQ to de�ne the R&D content of intermediates. Notice the
distinction between gross output, net output, and intermediate usage: Q is gross output; Q�BQ,
or (I � B)Q; is net output (�nal demand); and BQ is intermediate demand for production. In
addition, (I � B)�1Q represents the total requirements (direct and indirect inputs) needed to
produce Q.
We now use the framework above to de�ne the R&D content of intermediate inputs. Let Si be
a 1 � G row vector whose g-th element is the total business R&D stock used directly to produce
good g in country i.5 Since we assume all factors are fully employed, we can construct Di, a 1�G
row vector whose g-th element is the R&D stock per unit of good g. Then, Di satis�es:
DiQi = Si
so we can de�ne D to be a 1�NG global vector of direct R&D requirements
D =
�D1 D2 ::: DN
�;
and S to be a 1�NG global vector of domestic R&D stocks
S =
�S1 S2 ::: SN
�5 In our empirical application if industry g is non-manufacturing, we set Sig = 0 even though transactions between
manufacturing and non-manufacturing industries are all recorded in matrix B. We do not construct R&D stocks fornon-manufacturing sectors because the data for these industries are quite noisy. In addition, most of the businessR&D is concentrated in manufacturing industries. As we do include all 27 manufacturing and non-manufacturingindustries in matrix B, then the R&D content of an intermediate input produced by a manufacturing industry, butpurchased indirectly through a non-manufacturing industry will be taken into account.
7
where S = DQ.
Consider the recursive inputs demands for the production of intermediate goods BQ. For exam-
ple, the production of intermediate input textiles requires thread or yarn while producing interme-
diate input buttons requires plastic products. Matrix B�BQ represents the inputs directly needed
to produce BQ. Following this recursive algorithm to compute intermediate demands we have
that the direct and indirect input requirement needed to produce BQ is given byP1n=1B
n(BQ).
Next, in order to de�ne the R&D content of intermediates as the total R&D stock embodied in
intermediate inputs BQ, we need to compute the total requirements needed to deliver BQ, which
are given by
BQ+1Xn=1
Bn(BQ) =1Xn=0
Bn(BQ) = (I �B)�1BQ:
We de�ne the R&D content of intermediate inputs F as
F = D(I �B)�1BQ (4)
so F is a 1 � NG global vector of total R&D stock embodied in the production of intermediate
inputs. A typical element Fih from vector F corresponds to the measure of the domestic and
foreign R&D embodied in the intermediate inputs sector h in country i purchases from all sectors
and countries. Notice that a measure including only the �direct�R&D content of intermediates F d
would be given by:
F d = DBQ: (5)
In sum, relative to available ways of establishing the relationship between productivity and
R&D investments, our concept of R&D content of intermediates represents a more accurate way
to capture the knowledge embodied in intermediate inputs used in the production process. As we
turn now to discuss, we are able to use recently available input-output data from the OECD in
combination with bilateral imports data in order to construct matrix B. Our analysis improves
over Keller�s (2002a) paper in a number of ways: we are able to use country-speci�c input-output
data to properly weight industry-level R&D stocks; we introduce the idea of R&D content to
capture both direct and indirect �ows of inputs; we use a much bigger sample of countries; and we
report elasticities of TFP with respect to R&D content for all manufacturing, but also for high and
low-R&D industries separately.
8
3 Data description
In this section we discuss the most relevant details of the data construction and we illustrate the
main features of our measure of R&D content of intermediate inputs. Our data set includes 32
countries and 27 industries (13 of which are manufacturing) in years 1995, 2000, and 2005.6 ;7 We
restrict the analysis to these three years due to data availability to construct matrix B. There
are four main data modules that we discuss in turn: the structure of input trade across industries
and countries, industry-level R&D stocks, the global country-by-industry matrix of intermediate
transactions B, and the industry-level TFP indexes.
3.1 Input trade across industries and countries
Table 1 summarizes the average input-output structure of 13 manufacturing industries in 32 coun-
tries in year 2000. The table compares G5 versus non-G5 countries, as well as high (panel 1.1)
versus low-R&D industries (panel 1.2).8 Table 1 suggests several interesting patterns. Intermediate
input spending (of both manufacturing and non-manufacturing goods) as a share of gross output
(�rst row in both panels) is quite similar among high and low-R&D industries, as well as among
G5 and non-G5 countries: it is between 61 and 68% (slightly higher for high-R&D industries in
non-G5 countries). The second row in both panels reports intermediate purchases of manufacturing
inputs as a fraction of gross output. Some di¤erences emerge here: while this fraction is around
45% among high-R&D industries, it is about 35% among low-R&D industries. Similar di¤erences
can be observed from looking at own industry intermediate spending as a fraction of gross output
(third row): it is around 27% among high-R&D and 20% among low-R&D industries. Finally, inter-
mediate spending from other manufacturing industries as a fraction of output (sixth row) displays
6Countries in the sample are (* indicates non-OECD): Australia (AUS), Austria (AUT), Belgium (BEL), Canada(CAN), China* (CHN), the Czech Republic (CZE), Germany (GER), Denmark (DNK), Spain (ESP), Finland (FIN),France (FRA), the United Kingdom (GBR), Greece (GRC), Hungary (HUN), Ireland (IRL), Israel (ISR), Italy (ITA),Japan (JPN), South Korea (KOR), Mexico (MEX), the Netherlands (NLD), Norway (NOR), New Zealand (NZL),Poland (POL), Portugal (PRT), the Slovak Republic (SVK), Slovenia (SVN), Sweden (SWE), Turkey (TUR), ChineseTaipei* (TWN), the United States (USA), and South Africa* (ZAF).
7The 13 manufacturing industries are listed in Table 2. The 14 non-manufacturing sectors are agriculture, min-ing, re�ned petroleum products, other manufacturing (including recycling), electricity, construction, wholesale andretail trade, hotels and restaurants, transportation, �nancial intermediation, real estate and business services, publicadministration, education, and health and social work.
8Details on the sources used to construct Table 1 are provided in the Appendix. In this paper we focus onmanufacturing industries, as they are the main producers of business R&D knowledge in the world. High-R&Dindustries correspond to those with the highest R&D capital to output ratio. Table 1 as well as all results presentedin the paper are robust to alternative de�nitions of high-R&D industries. Results using alternative de�nitions areavailable upon request.
9
only slight di¤erences: it is between 18 to 20% among high-R&D and 15 to 18% among low-R&D
industries. The picture that emerges so far from Table 1 is that high-R&D industries tend to use
proportionally more manufacturing inputs, specially own-industry inputs, than low-R&D indus-
tries. In addition, at this level of aggregation di¤erences between G5 and non-G5 countries are not
sizeable.
Di¤erences between G5 and non-G5 countries emerge when intermediate goods are divided
into domestic and foreign inputs. Table 1 suggests that while non-G5 countries buy relatively less
domestic own-industry inputs than G5 countries, the opposite is true in the case of other-industry
inputs. For instance, while high-R&D industries in non-G5 countries buy 31% of their own-industry
inputs domestically, the corresponding amount in G5 countries is 57%. In contrast, again among
high-R&D industries, non-G5 countries buy 60% of other-industry inputs domestically, while in G5
countries it is 48%. This observation suggests that foreign knowledge embodied in own-industry
inputs is the one most likely to matter for productivity among non-G5 countries. There are also
di¤erences in the domestic / foreign composition of inputs among high and low-R&D industries.
Table 1 indicates that while in low-R&D industries G5 countries purchase 76% of own-industry
inputs domestically, in high-R&D industries non-G5 countries purchase 69% of own-industry inputs
from abroad. This suggests that domestic own-industry R&D content is likely to be correlated with
productivity in low-R&D industries among G5 countries, while foreign own-industry R&D content
would correlate with productivity in high-R&D industries among non-G5 countries. In sum, Table
1 suggests that it is more likely to �nd signi�cant correlations between TFP and foreign own-R&D
content in high-R&D industries among non-G5 countries. As we show below, our results con�rm
this idea.
3.2 Industry-level R&D stocks
The domestic R&D stock or knowledge capital Sigt for country i industry g at time t is computed
using the perpetual inventory methods as follows:
Sigt = (1� �)Sig;t�1 +Rigt
where � is the depreciation rate of knowledge obsolescence, and Rigt is the real business R&D
expenditure. Details on the sources of real business R&D expenditure data are in the Appendix.
10
Although we use R&D data from as early as 1987, the R&D stocks used in the empirical analysis
are only those for 1995, 2000 and 2005. To obtain the initial value of the real business R&D stock,
we compute:
Sig0 =Rig0� + gig
where gig is the average growth rate of real business R&D expenditures for industry g in country
i over the whole period for which data are available. Following Coe and Helpman (1995) we use
� = 0:15.9
Figures 1.1 and 1.2 illustrate how concentrated the R&D stock is in a few manufacturing
industries and a handful of source countries by using data from year 2000. Figure 1.1 shows that
about 40% of the business R&D stock in manufacturing in our sample of 32 countries is accounted
for by electrical machinery, followed by 20% in chemicals, about 15% in motor vehicles and 12% in
other transport. In turn, Figure 1.2 suggests that at least 60% but mostly about 75% of the R&D
stock in the sample is produced by G5 countries.
3.3 Global matrix B and R&D content of intermediates
To construct the global intermediate techniques matrix B, we derive domestic BjjQj and foreign
BijQj input-output tables from OECD Input-Output Database (2010) and industry-level bilateral
imports from OECD STAN Bilateral Trade Database (2009).10 More details on input-output tables
are in the Appendix. Following Tre�er and Zhu (2010), we allocate country j�s purchases of foreign
intermediates BijQj from each trading partner i according to its industry-level import shares. For
example, suppose China imports $20 million from Japan�s machinery industry, and that China�s
total foreign demand (intermediate plus �nal) in the machinery industry is $100 million. Then,
the allocation procedure assumes that each industry in China imports 20% of its total foreign
demand for machinery products from Japan. This same proportion applies regardless of whether
the product is used as intermediate input or a �nal consumption good. This imputation implicitly
assumes that international �ows of goods respond to trade determinants regardless of whether they
9Coe and Helpman (1995) consider depreciation rates ranging from 0:05 to 0:15. Since there are negative growthrates for several industry-level real R&D expenditures from 1987 to 2006, we use a depreciation rate of 0:15 to avoidnegative or large starting values for the R&D stock.10B is an NG � NG matrix with N = 32 and G = 27. We do not consider transactions of intermediate inputs
outside of the 32 countries. For example, we do not account for US imports of intermediate inputs from Brazil, whichis not included in our data set. Below we present robustness checks for the case in which we include an additional"rest-of-the-world" country constructed as an aggregate of Brazil, India and Indonesia.
11
are used as intermediate inputs or �nal goods, a �proportionality hypothesis�. Using Asian input-
output tables that provide direct information on the sources countries of foreign inputs, Puzzello
(2012) examines whether this proportionality hypothesis causes biases and �nds that although it
understates the use of foreign inputs, specially in sectors where they are most used, net biases are
small.11
With global matrix B at hand, as well as with Q and D we can construct our measure of
R&D content of intermediates F according to equation (4). Table 2 displays summary statistics
for both the R&D content of domestic (panel 2.1) and foreign intermediates (panel 2.2) in 2000,
as well as growth rates between 1995 and 2005. A number of facts emerge from these statistics.
First, in high R&D producer countries such as the US and Japan, all industries appear to absorb
high R&D content through the purchase of domestic intermediate inputs. However, China exhibits
the largest growth rates in domestic R&D content in all but one industry between 1995 and 2005.
This is an interesting fact as China certainly absorbs foreign knowledge, but as Table 2 indicates,
has dramatically increased domestic R&D in the last decades across all manufacturing industries.
Turning now to panel 2.2, notice that although the US is one of the leading R&D producers in the
world, a number of industries in the US appear to absorb the maximum foreign R&D content in the
sample (food, wood, paper, chemicals, rubber and plastics, fabricated metals and motor vehicles).
Other countries such as China, Mexico, Germany and France record the maximum foreign R&D
content exposure in other industries. China again exhibits the largest growth rates in foreign R&D
content of intermediates in a number of industries (food, wood, paper, chemicals, basic metals,
electrical machinery and other transport) in the 1995-2005 period.
To get a more precise sense of the magnitudes of R&D content in own-industry, other-industry,
as well as the domestic and G5-foreign components, Figures 2.1 through 2.4 display these values for
the 13 manufacturing industries for China, the Czech Republic, the UK and the US for year 2000
(respectively). In China, chemicals and electrical machinery exhibit both the largest own-industry
domestic R&D content levels and the largest own-industry G5-foreign R&D content (Figure 2.1).
In relative terms, G5-foreign R&D content is larger than the domestic component, specially in
electrical machinery. In the Czech Republic (Figure 2.2) G5-foreign R&D content is also largest
11According to Table 1, the largest di¤erence in the use of foreign inputs between low and high-R&D industriesoccurs for the case of own-industry inputs. Speci�cally, foreign own-industry inputs are relatively more used in high-R&D industries both among G5 and non-G5 countries. In this respect, if a bias were to be present, it would onlymake our results stronger.
12
in electrical machinery, although about 10 times smaller than the levels observed in China. Motor
vehicles in the Czech Republic have the highest domestic R&D content of intermediates, and the
second largest G5-foreign R&D content. Turning to Figure 2.3, the UK has the highest domestic
R&D content in high-R&D industries, specially chemicals. Although the UK is part of the G5
countries, other G5 countries are important sources of foreign R&D content, specially in electrical
machinery and other transport. Finally, like in the UK, the US has the highest domestic R&D
content in high-R&D industries, but this time the largest content is in electrical machinery. In
contrast with the UK though, other G5 countries do not appear to be signi�cant sources of R&D
content for the US.
3.4 Industry-level TFP
We construct industry-level TFP indexes using real value added (Yigt), labor compensation shares
(�igt), physical capital (Kigt), and labor (Ligt) for each country i, industry g and period t. Details
on the speci�c sources of these data are in the Appendix. We follow the methodology of Caves,
Christensen, and Diewert (1982) and compute
lnTFPigt =
"lnYigt �
1
N
Xi
lnYigt
#� 12
�igt +
1
N
Xi
�igt
!"lnLigt �
1
N
Xi
lnLigt
#
�"1� 1
2
�igt +
1
N
Xi
�igt
!#"lnKigt �
1
N
Xi
lnKigt
#
where N = 32 is the number of countries in the data set. This index has been used by Harrigan
(1997) and Keller (2002b). A few comments on this TFP index are in order. First, the TFP index
is normalized because as show in the equation above, it is constructed by expressing value added,
capital and labor relative to their respective averages among the countries in the sample for each
industry and year. This normalization is without loss of generality and partially eliminates trends
in absolute TFP levels. In addition, since due to data limitations we only include years 1995, 2000
and 2005, issues pertaining to the presence of unit roots, cointegration and serial correlation are
not likely to apply to our panel data. In this context, the elasticities we estimate may capture
only short-term e¤ects of R&D on TFP. Second, the equation above uses an average labor share
computed over two terms: labor share in country i, �igt, and cross-country average labor share
in the whole sample, (1=N)Pi �igt. The normalization of the TFP index and this particular
13
computation of labor shares impose a good amount of structure in the construction of TFP, which
should help reducing simultaneity issues in the estimation (Keller, 2002b). Finally, factor shares
�igt are cost-based since they are more robust than revenue-based shares in the absence of constant
returns to scale.
It is important to notice that we use real value added to compute industry-level TFP, rather
than the theoretically preferred gross output. As van Ark and Timmer (2006) point out, "....for
comparisons of levels of multi factor productivity (MFP) at the industry level, gross output is
the preferable concept, as it simultaneously treats intermediate inputs and primary factor inputs.
Nevertheless, due to lack of data on intermediate inputs, value added based MFP measures are
much more common" (p. 19). Although we are using the best available data, these do not include
industry-level real gross output, so as it is common practice, we are using real value added. However,
we have value added de�ators based on the double-de�ation method for most countries in our
sample.12 Theory indicates �see Bruno (1978) and Basu and Fernald (1995)�that starting from
double-de�ated value added measures, there is no bias in constructing TFP growth rates using
valued added or gross output data, as long as intermediate inputs are used in �xed proportions
(Theorem 2 in Bruno). Fortunately, although the input-output matrix changes across the periods
we are examining, these changes are not so dramatic to suspect systematic large biases from using
real valued added in computing TFP measures in our sample.
Panel 2.3 on Table 2 displays summary statistics for the TFP index in our sample. Roughly
speaking, TFP indexes above one re�ect higher than average productivity, while the opposite is
true for indexes below one. A variety of countries appear to lead productivity in di¤erent industries,
although they all correspond to richer countries such as the US, France, Canada, the Netherlands,
Australia, Norway and Finland. Meanwhile China appears to have the lowest TFP index in all but
one industry in year 2000.13 In terms of growth rates between 1995 and 2005, Korea is leading in
a number of industries such as food, textiles, wood and paper products, while China is leading in
chemicals.12Out of our sample of 32 countries, the only countries for which we do not have industry-level value added de�ators
based on the double-de�ation method are the non-OECD countries (China, Taiwan and South Africa), Australia,Ireland, Turkey and the UK. We perform robustness checks for our results by excluding non-OECD countries fromthe sample (see Tables 4 and 5 below).13 It is worth noting that this is not the case in 2005, as China�s relative productivity increased during this 5-year
period.
14
4 Empirical framework and results
In this section we present the empirical implementation of our exercise and the main results.
Speci�cally, we estimate the elasticity of industry-level TFP with respect to the R&D content of
intermediates. We compute elasticities for all 13 manufacturing industries, but we also estimate
separate elasticities for high and low-R&D industries. In addition, using the di¤erent components
of matrix B, we separately estimate the e¤ects of R&D embodied in domestic, foreign, own-industry
or other-industry intermediates.
4.1 Benchmark estimation
The benchmark equation we estimate links (log) industry-level TFP with aggregate R&D content
as follows:
lnTFPiht = �it + �h + � lnFiht + "iht (A)
where i indexes country, t time (1995, 2000 and 2005), and h industry.14 We use country and time-
speci�c �xed e¤ects for the following two reasons. First, they capture measurement errors associated
with the input-output table, which is country and year-speci�c. Second, other potentially important
variables to determine productivity such as accumulation of human capital and institutions are likely
to be country and year-speci�c (Coe, Helpman and Ho¤maister, 2009).15
Estimating � consistently requires to deal with standard econometric issues, particularly serial
correlation, simultaneity, and endogeneity. Serial correlation is unlikely to be an issue in our data
set, as our panel data spans more across industries and countries than across years (only 1995, 2000
and 2005). This is in contrast with studies that focus on long-run time-series data to study the
relationship between TFP and foreign R&D capital (e.g., Coe, Helpman, and Ho¤maister, 2009),
where the use of econometric techniques on panel cointegration is particularly important. Next, we
avoid simultaneity issues in three di¤erent ways. First, we follow Keller (2002b) and employ the
TFP index of Caves, Christensen, and Diewert (1982). Unlike the TFP index in Coe, Helpman,
and Ho¤maister (2009), a considerable amount of structure is imposed in developing our TFP
14While equation (1) in Coe and Helpman (1995) and equation (3) in Keller (2002b) include the domestic R&Dstock Siht in their estimations, we do not have lnSiht in equation (A). The reason is that the R&D content ofintermediates lnFiht already accounts for the own-industry domestic R&D embodied in the intermediate inputs.Including lnSiht in our speci�cation would imply "double accounting."15All results we present are not sensitive to the inclusion of standard human capital measures and are available
upon request.
15
index, which, as discussed before, should reduce simultaneity problems. Second, we use country
and industry-speci�c de�ators for value added and gross �xed capital formations for most countries.
Since the real R&D expenditures provided by OECD are de�ated by country-level GDP de�ators,
we reduce the possibility that our results are induced by common de�ators. Third, below we
estimate alternatives to equation (A) (equations D and E), which treat R&D source and recipient
countries separately. Under these alternative speci�cations, the possibility that R&D content and
TFP are subject to common shocks is diminished. Regarding endogeneity issues, the ideal situation
would be to have instruments for R&D measures. There are no obvious instruments available. One
possibility would be to use lagged variables as instruments. Below we check the robustness of our
estimates by using 5-year lagged values of embodied R&D stocks.
Table 3 presents the results of estimating equation (A) for all 13 manufacturing industries, for
high-R&D intensive industries (chemicals, electrical machinery, motor vehicles and other transport
equipment), and for low-R&D intensive industries. The table suggests important di¤erences be-
tween high and low-R&D industries. The R&D content of intermediates is positively associated
with industry-level TFP among all industries. We estimate � = 0:041 for all 13 manufacturing
industries (5% con�dence level). This coe¢ cient doubles among high-R&D industries with and
estimated � = 0:082. In contrast, the R&D content of intermediates does not appear to be sig-
ni�cantly correlated with productivity among low-R&D industries. While equation (A) is quite
aggregated, our formulation of the R&D content measure allows for a decomposition of source
countries and industries, as we now turn to analyze.
4.2 Disaggregation
We now use the global country-by-industry aspect of our measure of R&D content of intermediates
F in order to further investigate the groups of countries and industries that drive the results from
equation (A). For this purpose we estimate four additional equations. The �rst equation separates
the total e¤ects of R&D content (domestic and foreign) by industry source, i.e., it separates own
16
industry intermediates, from intermediates from other industries as follows
lnTFPiht = �it + �h + �1 ln
24Xj
Djt(h)Ajit(h; h)Qit(h)
35 (B)
+�2 ln
24Xj
Xg 6=h
Djt(g)Ajit(g; h)Qit(h)
35+ "ihtwhere matrix A = (I �B)�1B so that R&D content of intermediates can be de�ned as F = DAQ.
The �rst term in brackets represents the R&D stock embodied in intermediate goods industry h in
country i buys from its own industry in all countries (including itself) at time t, while the second
term in brackets is the R&D stock embodied in intermediate goods industry h in country i buys
from all other industries in all countries.
Second, we estimate an equation to separate the e¤ect of R&D content from domestic and
foreign source countries as follows:
lnTFPiht = �it + �h + �3 ln
"Xg
Dit(g)Aiit(g; h)Qit(h)
#(C)
+�4 ln
24Xj 6=i
Xg
Djt(g)Ajit(g; h)Qit(h)
35+ "ihtwhere the �rst term in brackets represents the domestic R&D content of intermediates for industry
h in country i, and the second term is brackets is the foreign R&D content of intermediates for
industry h in country i.
Third, we estimate an equation to single out the role of G5 countries as foreign sources of
own-industry R&D content as follows:
lnTFPiht = �it + �h + �2 ln
24Xj
Xg 6=h
Djt(g)Ajit(g; h)Qit(h)
35 (D)
+�5 ln[Dit(h)Aiit(h; h)Qit(h)] + �6 ln
24 Xj 6=i;j2G5
Djt(h)Ajit(h; h)Qit(h)
35+ "ihtwhere the �rst term in brackets is the other-industry R&D content (domestic and foreign), the
second term is the domestic own-industry R&D content, and the third term the G5-foreign own-
17
industry R&D content of intermediates.
Last, we estimate a corresponding equation to single out the role of G5 countries as source of
other-industry R&D content
lnTFPiht = �it + �h + �1 ln
24Xj
Djt(h)Ajit(h; h)Qit(h)
35 (E)
+�7 ln
24Xg 6=h
Dit(g)Aiit(g; h)Qit(h)
35+ �8 ln24 Xj 6=i;j2G5
Xg 6=h
Djt(g)Ajit(g; h)Qit(h)
35+ "ihtwhere the �rst term in brackets is the own-industry R&D content (domestic and foreign), the second
term is the domestic other-industry R&D content, and the third term the G5-foreign other-industry
R&D content of intermediates. As discussed before, equations (D) and (E) serve as a robustness
check to benchmark equation (A) to the extent that they should ameliorate simultaneity issues.
Speci�cally, R&D content in G5 countries should be subject to shocks di¤erent than those a¤ecting
TFP in non-G5 recipient countries.
We estimate equations (B) through (E) for all 13 manufacturing industries, as well as separately
for high and low-R&D intensive industries. While the left-hand side of equations (B) and (C)
includes industry-level TFP measures for all 32 countries, we estimate equations (D) and (E) for
27 non-G5 recipient countries only. Results are presented on the last four columns in Table 3.
The table makes apparent a sharp di¤erence between high and low-R&D industries. While none of
the TFP elasticities estimated for low-R&D industries are statistically signi�cant (panel 3.3), the
R&D content of intermediates is positively associated with TFP among high-R&D industries in all
equations (panel 3.2). Estimates of equation (B) indicate that among high R&D industries, the own-
industry R&D content is signi�cantly correlated with TFP with �1 = 0:078 (5% con�dence level).
According to the estimates of equation (C), the TFP elasticity with respect to R&D embodied in
foreign inputs is estimated to be �4 = 0:075: Moreover, we estimate �6 = 0:075 in equation (D),
so inputs produced in G5 countries are critical in enhancing TFP among high-R&D industries. In
addition, the own domestic R&D content is signi�cantly associated with TFP with �5 estimated to
be 0:038 in equation (D). Interestingly, we do not �nd evidence under equations (D) and (E) that
the other-industry R&D content is correlated with TFP among high-R&D industries.
When all 13 manufacturing industries are included (panel 3.1) TFP is signi�cantly correlated
18
with own-industry R&D content with �1 estimated to be 0:025, as well as with foreign R&D
content with �4 estimated to be 0:062 (both with 5% con�dence). Table 3 indicates that both these
results are driven by the high-R&D industries. In addition, estimation of equation (E) among all
13 manufacturing industries implies a marginally signi�cant (10% con�dence) correlation between
G5-foreign other-industry R&D content and TFP with �8 estimated to be 0:051. This e¤ect is
not driven by high-R&D industries, but it is also not signi�cant among low-R&D industries when
considered in isolation (equation (E), panel 3.3). In sum, Table 3 suggests that the extent to which
intermediate input trade is a channel of R&D di¤usion is quite limited to high-R&D intensive
industries for which both domestic and foreign inputs matter. A distinct role for G5 countries is
identi�ed among high R&D industries. In contrast, the data suggest intermediate input trade is
not a relevant channel of knowledge di¤usion among low-R&D industries.
4.3 Comparision with related literature
Our paper relates most closely to Keller (2002a) who also considers the transmission of technology
via intermediate input trade. Keller�s (2002a) sample only includes eight OECD countries and all
input-output relations are derived from the US. In contrast, we use a sample of 32 countries and
country-speci�c input-output tables. Moreover, Keller derives the input-output table for imports
from US data, while the split between domestic and foreign inputs we use is provided by the input-
output tables of each country. Finally, we propose the concept of R&D content of intermediates as
the one that most accurately re�ects the R&D knowledge embodied in inputs, while Keller follows
Coe and Helpman (1995) in constructing foreign R&D stock measures.
Keller (2002a) �nds that technology transmission through intermediate purchases is important.
Speci�cally, the contribution of own-industry domestic R&D is about 50% of the total e¤ect on
TFP, followed by other-industry domestic R&D (30%) and foreign R&D (20%). Keller estimates the
e¤ect of own-industry domestic R&D on TFP using OLS. The other components (other-industry
domestic and foreign) are added under a non-linear speci�cation. He �nds that relative to the
OLS speci�cation with own domestic R&D only, the non-linear speci�cation with the additional
R&D measures does not add much to the empirical �t. This explains his �nding that own-industry
domestic R&D stock is the most important component of the total e¤ect on TFP.
Our results are quite di¤erent from Keller�s. A distinct feature of our �ndings is that we dis-
19
tinguish high from low-R&D industries. As Keller, we do �nd that our own-industry R&D content
measure is signi�cantly correlated with TFP but only among high-R&D industries. Moreover,
di¤erent from Keller we �nd TFP to have an even larger elasticity with respect to own-industry
foreign R&D content, particularly for inputs purchased from G5 countries. These e¤ects are not
present among low-R&D industries. The types of e¤ects we identify may be hard to detect in
Keller�s sample, which only has three non-G5 in addition to the G5 countries. Finally, notice that
what allows us to identify statistically signi�cant roles for own-industry foreign and domestic R&D
content is the separate analysis of high and low-R&D industries. These coe¢ cients (�5 and �6
respectively) are not signi�cant in the full sample of 13 manufacturing industries.16
Our paper also relates to an earlier literature, especially to Scherer (1982) and Griliches and
Lichtenberg (1984). Although these papers only involved US data, they relate to ours in that they
also focus on intermediate input trade across industries. Griliches and Lichtenberg (1984) use US
data for manufacturing industries between 1959 and 1978 to show that own-product R&D is a
signi�cant determinant of measured TFP. Di¤erent from Scherer (1982), they �nd that the e¤ect of
own-product R&D on TFP is more important than the e¤ect of embodied-R&D in inputs purchased
from other industries. They conclude that R&D spillovers are tenuous and attribute this result to
the fact that relatively speaking, while manufacturing industries do not "import" much R&D, they
"export" a lot to non-manufacturing industries. Our results have actually a very similar taste to
Griliches and Lichtenberg�s, even if we focus only on manufacturing industries. Our main message
is that R&D spillovers are quite limited because they are fairly industry speci�c. Own-industry
R&D content is signi�cantly correlated with TFP, while other-industry R&D content is not. This
is quite similar to Griliches and Lichtenberg. They do not use international data, so they cannot
say anything about embodied R&D in foreign inputs. But even then, we �nd that the foreign R&D
that matters is the one embodied in own-industry inputs coming from G5 countries. Thus, it is
not only that the most important sources of R&D embodied in inputs for high-R&D industries
are the very same high-R&D industries, but that there is just a handful of countries the ones that
represent the sources of knowledge.17
16We estimated non-linear least squares regressions similar to those in Keller (2002a) but using our measure ofR&D content. Our results are robust to this non-linear speci�cation.17 It is important to acknowledge that there are a number of issues involved in measuring R&D spillovers. Griliches
wrote extensively about this. For instance, Griliches and Lichtenberg (1984) argue that R&D spillovers emerge from"errors" in output de�ators of supplying industries. These "errors" correspond "to the failure of these de�ators tore�ect accurately changes in the user value, or marginal productivity, or quality, of the commodities whose prices they
20
5 Robustness and discussion
In this section we present a number of robustness checks and perform additional empirical exercises
to further discuss the �ndings presented above. Speci�cally, we are interested in further exploring
the di¤erences among high and low-R&D industries.
5.1 Robustness
In this section we report robustness checks for our estimates on Table 3, speci�cally those reported
for equations (A) and (D). Robustness exercises are presented on Tables 4 and 5 respectively. Panel
1 on each of these tables reports results for all 13 manufacturing industries, panel 2 does for high-
R&D industries, and panel 3 for low-R&D sectors. Column (1) across panels on Table 4 replicates
the estimates of our benchmark equation (A) on Table 3. Results in column (2) correspond to an
estimation of equation (A) in which rather than using the total (direct and indirect) R&D content
measure, we use only the direct R&D content F diht as given in equation (5). The results indicate
that indirect input purchases as captured by the Leontief input-output framework turn out not be
quantitatively relevant for the 13 manufacturing industries as a whole, but they are for high-R&D
industries. When only direct R&D content is included we estimate � = 0:091 for these industries,
indicating that indirect R&D content is perhaps lower among these industries.
In column (3) we exclude from the sample non-OECD countries (China, Taiwan and South
Africa). These are countries for which there was relatively more missing data in real R&D ex-
penditures, so we were forced to interpolate some data. Our benchmark results in column (1)
are con�rmed in column (3). Next, column (4) estimates equation (A) by constructing our R&D
content measure F under the assumption that the relevant R&D stock is lagged �ve years. It may
be the case that the stock of knowledge in any give industry at a speci�c point in time takes some
years in showing up as embodied quality improvements of the intermediate inputs produced by
the industry. Given data constraints we can only consider �ve-year lags. Relative to column (1),
are supposed to index" (p. 325). In addition, they hypothezise these errors are correlated with supplier�s performance(or R&D). Along the same lines, Scherer (1982) states that "three-fourths of all U.S. industrial R&D is concernedwith creating new or improved externally-sold products.....Under imperfectly monopolistic pricing and, a fortiori,vigorous competition, much of the bene�t from industry i�s superior new products will be passed on to buyers, e.g.,in industry j. This tendency is likely to be mirrored in the distribution of measured productivity growth as pricede�ators systematically underestimate the hedonic value of new products and hence undervalue the inputs used byan innovating industry�s customers" (p. 627). In sum, both these papers suggest that there are bene�ts from R&Dthat the buyer of a new product enjoys but are not captured in the price either because de�ators are not hedonic, orbecause the market structure may limit monopoly pricing. This discussion applies to our analysis as well.
21
results in column (4) con�rm our main results once more.
Column (5) addresses potential endogeneity problems by using two-stage least squares (2SLS).
We use �ve-year lag values of the variables as instruments.18 The results from 2SLS are reported
in column (5). While we obtain a relatively higher value of � = 0:088 for high-R&D intensive
industries, overall results are similar to the benchmark estimation in column (1). In conclusion,
Table 4 indicates that our results are robust to a number of alternative speci�cations. Last, column
(6) checks whether the addition of one more "country" representing the rest of the world (ROW)
a¤ects the benchmark results. Our benchmark estimation ignores transactions of intermediate
inputs with countries outside of our sample of 32 countries. The ROW "country" is constructed
as an aggregate of a few countries for which input-output data are available: Brazil, India and
Indonesia.19 As column (6) indicates, our results are robust to this modi�cation.
Table 5 is analogous to Table 4, except that the robustness analysis is performed on the estimates
of equation (D) in Table 3. An inspection of Table 5 reveals that the estimates of this equation are
overall quite robust. Notice that among high-R&D intensive industries (panel 5.2), the coe¢ cient
on G5-foreign own-industry R&D content in column (5) under 2SLS is larger than in column (1): �6
goes from 0:075 in column (1) to 0:106 in column (5).20 The same is true for all 13 manufacturing
industries (panel 5.1): while �6 is not statistically signi�cant in column (1), it becomes signi�cant
at 5% con�dence with a value of 0:042.
5.2 Source countries
Although estimates of equation (D) on Table 3 clearly identify G5 countries as the key sources of
embodied R&D for high-R&D intensive industries, no clear potential source countries are identi�ed
for low-R&D sectors. It is still possible that countries other than the G5 are the relevant R&D
sources for low-R&D industries. We explore this issue by examining all possible sets of source and
18We examine the likelihood of endogeneity of each variable of concern by using the Durbin-Wu-Hausman test. Inthe case of Table 4, lnFiht was found to be endogenous under the Durbin-Wu-Hausman test for all manufacturingindustries and among low-R&D industries. Detailed statistics of these tests are available upon request. We uselnFih;t�5, lnLih;t�5, and lnKih;t�5 as instruments.19Brazil, India and Indonesia do not have R&D investment data at the industry level and as a result, are not
included in our sample of 32 countries. When including these countries in the ROW aggregate we are only able tocapture the indirect R&D content of intermediates. For instance, if India purchases inputs from the US, which arethen used in the production of other inputs India exports to Mexico, these inputs then "indirectly" embody US R&Dknowledge that will be used in Mexico.20 In this case we cannot reject the null hypothesis of endogeneity only for own-industry domestic R&D content.
We use ln[Di;t�5(h)Aii;t�5(h; h)Qi;t�5(h)], lnLih;t�5, and lnKih;t�5 as instruments.
22
recipient countries in our sample of 32 countries. Speci�cally, we create all possible combinations
of �ve-country sources of R&D knowledge, with the remaining 27 countries being recipients. There
are 201,376 combinations of this sort. We then estimate equations (D) and (E) separately for high
and low-R&D industries for each of these combinations. In estimating equations (D) and (E),
the former role of G5 countries (j 6= i, j 2 G5) will now be taken by just the random �ve-country
sources created in the respective combination. Since there is a large number of equations estimated,
here we only report the statistical signi�cance of �6 which is the TFP elasticity to the R&D content
from own industry intermediates produced in the �ve-country source group, and �8 which is the
elasticity with respect to intermediates produced by other industries in the �ve-country group.
Figure 3.1 reports the probability density functions of the t-statistics of coe¢ cients �6 and �8
for the high-R&D industries, while Figure 3.2 displays the corresponding densities for low-R&D
industries. Figure 3.1 suggests that among high-R&D industries, there is a clear di¤erence between
the importance of R&D content coming from own industry, and that from other industries. In fact,
even from a random selection of �ve-country sources, the �gure suggests that foreign inputs from
other industries are not associated with TFP, while there is a larger number of cases in which foreign
inputs from own industry are positively associated with TFP (at 5% signi�cance level). Speci�cally,
�6 is signi�cant in 80% of the cases, while �8 is in only 0:3% of the cases. Things look quite di¤erent
on Figure 3.2 for low-R&D industries. Interestingly, the densities of the t-statistics for �6 and �8
do not di¤er signi�cantly. In this case, 27% of all the combinations provide statistically signi�cant
TFP elasticities to own-industry foreign R&D content, while it is 42% for other-industry foreign
R&D content.
To further explore the di¤erences between high and low-R&D industries found on Figures 3.1
and 3.2, we now report the fraction of cases in which each speci�c country was rejected as one of
the �ve-source countries whose inputs had a statistically signi�cant association with the TFP of
recipient countries. Figures 4.1 and 4.2 report the cases of �6 for high and low-R&D industries
respectively. If G5 countries are dominant sources of R&D embodied in intermediate inputs, then
coe¢ cient �6 would tend to be statistically signi�cant when the randomly selected �ve countries
include several G5 countries. Turning �rst to Figure 4.1, it turns out that the lowest probabilities
of being rejected as an R&D source country for high-R&D industries are precisely assigned to G5
countries. Again, things are quite di¤erent for the case of low-R&D industries in Figure 4.2. In
23
this case, all countries appear to have the same probability of being rejected as sources of R&D
knowledge. We interpret this di¢ culty in identifying speci�c sources of knowledge for low-R&D
industries as an indication that intermediate input trade is not a relevant channel of di¤usion among
low-R&D industries.
6 Concluding comments
The idea that technology transfers across borders are a powerful engine to increase productivity
levels around the world continues to resonate among academics and policy makers. Technological
knowledge embodied in intermediate inputs traded across industries and countries constitutes one of
the channels of technology di¤usion. This channel has been the subject of study in both theoretical
and empirical models. To the extent of our knowledge, this is the �rst paper to employ recently
available international input-output data in order to examine the correlation between industry-level
productivity and the R&D knowledge contained in intermediate inputs. Our analysis is distinct in
at least two respects: we propose the concept of R&D content of intermediates in order to capture
embodied R&D knowledge, and we separately analyze high and low-R&D industries.
Our results suggests that R&D knowledge spillovers through intermediate input trade are quite
limited. In fact, we �nd that this channel is relevant only among high-R&D industries. Gains in
productivity among high-R&D industries are associated with the use of own-industry intermediate
inputs, both domestic and imported from G5 countries. We do not �nd this channel to be rele-
vant among low-R&D industries. We conclude that knowledge embodied in intermediates is quite
industry speci�c.
Although we use the best available data to perform our analysis, there remain data limitations
that could be solved as more data become available. The elasticities we estimate here may be
interpreted as short term only, as our data cover a ten-year span. Future work could bene�t from
more re�ned data collection protocols across countries, and could also revisit the di¤erences between
high and low-R&D industries in the long run.
24
References
[1] Aghion, P., and P. Howitt. 1992. �A model of growth through creative destruction.�Econo-
metrica 60: 323�351.
[2] Basu, S. and J. Fernald. 1995. �Are apparent productivity spillovers a �gment of speci�cation
error?�Journal of Monetary Economics, 36: 165-188.
[3] Bruno, M. 1978. �Duality, intermediate inputs, and value-added.� In: M. Fuss and D. Mc-
Fadden, eds., Production economics: A dual approach to theory and practice, Vol. 2, North-
Holland, Amsterdam.
[4] Caves, D. W., L. Christensen and E. Diewert. 1982. �Multilateral comparisons of output, input,
and productivity using superlative index numbers.�Economic Journal, 92: 73-86.
[5] Coe, D. T., and E. Helpman. 1995. �International R&D spillovers.�European Economic Re-
view, 39: 859�887.
[6] Coe, D. T, E. Helpman and A. Ho¤maister. 2009. �International R&D spillovers and institu-
tions.�European Economic Review, 53: 723-741.
[7] Goto, A. and K. Suzuki. 1989. �R&D capital, rate of return on R&D investment, and spillover
of R&D in Japanese manufacturing industries.�Review of Economics and Statistics, 71 (4):
555-564.
[8] Griliches, Z. and F. Lichtenberg. 1984. �Interindustry Technology Flows and Productivity
Growth: A Reexamination.�Review of Economics and Statistics, 66 (2): 324-329.
[9] Grossman, G. M., and E. Helpman. 1991. Innovation and Growth in the Global Economy.
Cambridge, MA: The MIT Press.
[10] Halpern, L., M. Koren and A. Szeidl. 2011. �Imported Inputs and Productivity.�Manuscript.
[11] Harrigan, J. 1997. �Technology, Factor Supplies, and International Specialization: Estimating
the Neoclassical Model.�American Economic Review, 87: 475-494.
[12] Heston, A., R. Summers and B. Aten. 2006. Penn World Table Version 6.2. Center for Inter-
national Comparisons of Production, Income and Prices. University of Pennsylvania.
25
[13] Kasahara, H. and J. Rodrigue. 2008. �Does the use of imported intermediates increase pro-
ductivity? Plant level evidence.�Journal of Development Economics, 87: 106-118.
[14] Keller, W. 2002a. �Trade and the Transmission of Technology.�Journal of Economic Growth,
7: 5-24.
[15] Keller, W. 2002b. �Geographic localization of international technology di¤usion.�American
Economic Review, 92: 120-142.
[16] Litchtenberg, F. and B. van Pottelsberghe de la Potterie. 1998. �International R&D spillovers:
A comment.�European Economic Review, 42: 1483-1491.
[17] Madsen, J. 2007. �Technology spillover through trade and TFP convergence: 135 years of
evidence for the OECD countries.�Journal of International Economics, 72: 464-480.
[18] Puzzello, L. 2012. �A Propotionality Assumption and Measurement Biases in the Factor Con-
tent of Trade.�Journal of International Economics, 87 (1): 105-111.
[19] Romer, P. 1990. �Endogenous Technological Change.� Journal of Political Economy, 98 (5):
S71-S102.
[20] Scherer, F.M. 1982. �Inter-industry technology �ows and productivity growth.� Review of
Economics and Statistics, 64 (4): 627-634.
[21] Schmookler, J. 1966. Invention and Economic Growth. Cambridge: Harvard University Press.
[22] Terleckyj, N. 1974. E¤ects of R&D on the Productivity Growth of Industries: An Exploratory
Study. National Planning Association, Washington, D.C.
[23] Tre�er, D. and S. Zhu. 2010. �The structure of factor content predictions.�Journal of Inter-
national Economics, 82 (2): 195-207.
[24] van Ark, B. and M. Timmer. 2006. �Reader on Productivity Levels.�Manuscript.
[25] van Pottelsberghe de la Potterie, B. and F. Litchtenberg. 2001. �Does foreign direct investment
transfer technology across borders?�Review of Economics and Statistics, 83 (3): 490�497.
[26] Xu, B. and J. Wang. 1999. �Capital goods trade and R&D spillovers in the OECD.�Canadian
Journal of Economics, 32 (5): 1258-1274.
26
A Data appendix
Our data set includes 32 countries and 27 industries in years 1995, 2000, and 2005.21 The numberof available countries, industries, and years in our study depends on the data availability of R&Dexpenditures and input-output tables.
A.1. Knowledge capital Real business R&D expenditures (2000 prices and 2000 PPP dollars)are obtained from the OECD STAN R&D Expenditure of Industry (2010) and the OECD Scienceand Technology Statistics (2010). The OECD uses the implicit GDP de�ator and PPP conversionsto compute real R&D expenditures. These de�ators provide an approximate measure of the averagereal opportunity cost of carrying out R&D (see Annex 9, 2002 OECD Frascati Manual: ProposedStandard Practice for Surveys on Research and Experimental Development). Real business R&Dexpenditures for 16 countries (Australia, Belgium, Canada, Denmark, Finland, France, Germany,Ireland, Italy, Japan, the Netherlands, Norway, Spain, Sweden, the United Kingdom, and theUnited States) are available from 1987 to 2006. For these countries, the 1987 value is used asthe initial value of the knowledge capital stock. For other countries in the data set, initial valuesdi¤er according to the availability of the data (1989 for Austria, 1991 for China, 1992 for the CzechRepublic, 1988 for Greece, 1993 for Hungary, 1991 for Israel, 1995 for South Korea, 1992 for Mexico,1989 for New Zealand, 1994 for Poland, 1998 for Portugal, 1992 for the Slovak Republic, and 1993for Slovenia, 1995 for South Africa, 1995 for Chinese Taipei, and 1990 for Turkey). We use zeroknowledge capital for the initial year when the initial value of the real business R&D expenditureis zero or missing. All the unreported values between two years are interpolated from the averageannual growth rates of these two years. We estimate all the remaining unreported values from thegrowth rates of total manufacturing, total services, or total business R&D expenditures. Data forsome countries such as China and South Africa are subject to signi�cant interpolation. The realR&D stock for industry h in country c is computed from the values of real R&D expenditures, itsaverage growth rate, and the depreciation rate (15%) using the perpetual inventory method. Forthe empirical exercises we use knowledge capital data for years 1995, 2000, and 2005 only.
A.2. Input-Output tables Input-output tables (nominal $US basis) are obtained from theOECD Input-Output Tables (updated in November 2010) for years around 1995, 2000, and 2005. Tobe consistent with the 13 manufacturing industries from the OECD STAN Database (2011), we ag-gregate each of input-output table into 27 industries: 13 manufacturing and 14 non-manufacturing.While the aggregation might eliminate the within-industry heterogeneity, it is essential to havethe same categories across the several databases (e.g., the OECD STAN, the OECD Science andTechnology Statistics, and the OECD Input-Output Tables). Since the input-output tables are notavailable for certain years (Israel, 2000; South Korea, 1995; Mexico, 1995 and 2000; New Zealand,2005; South Africa, 2005), we use the nearest year�s input-output tables as their proxies to havea complete set of 32 countries for each year. Industry-level output values are adjusted to years1995, 2000, or 2005 using industry-level (from the OECD STAN Database) or country-level (fromthe OECD National Accounts Statistics (2011)) growth rates of nominal value added ($US basis).For example, the input-output table of Mexico for year around 2005 is actually that of year 2003.We multiply it by the industry-level growth rate of value added from 2003 to 2005. We apply theadjustment for Australia (1995, 2000, and 2005), South Korea (1995), Mexico (1995, 2000, and2005), New Zealand (1995, 2000, and 2005), Turkey (1995, 2000, and 2005), Chinese Taipei (1995,2000, and 2005), and South Africa (1995 and 2005). Finally, it is important to notice that in the
21We have the following missing observations: IRL (food products, paper products and chemicals), NZL (electricalmachinery and other transport), MEX (other transport), and POL (other transport).
27
OECD input-output tables, R&D is treated as an independent input activity and is categorizedas an item in real estate and business services for 26 countries (all 32 countries in our sampleexcept for Australia, Canada, China, Israel, Mexico, Poland, and Portugal). To avoid the possibledouble counting of R&D activities, we remove the sub-category of R&D from the original OECDinput-output tables.
A.3. TFP indexes We construct industry-level TFP indexes using real value added (Yiht), laborcompensation shares (�iht), physical capital (Kiht), and labor (Liht) for each country i, industry hand period t. We obtain these variables from the OECD STAN Database (2011), the OECD Input-Output Tables (2010), ILO Laborsta Internet, the OECD National Accounts Statistics (2011), andNational Statistics (Chinese Taipei and Japan). We now report some details for each of thesevariables.
Value added Nominal value added are taken from the input-output tables described above.Value added de�ators are from the OECD STAN Database and the OECD National AccountsStatistics. We can obtain value added de�ators based on double-de�ation methods for all 32countries in the sample except for the non-OECD countries (China, Chinese Taipei and SouthAfrica), Australia, Ireland, Turkey, and the UK. The number of unreported value added de�atorsis much smaller than that of previous OECD databases. Since most of the unreported data for realvalue added are due to the lack of value added de�ators, we employ the de�ators of the groupedindustries to obtain real value added. We convert real value added (local currency, 2000 price) into2000 international dollars by purchasing power parity rates (PPP) from either the OECD NationalAccounts Statistics or Penn World Table 6.2.
Labor compensation shares Labor compensation values (nominal) are from the OECDInput-Output Tables. We calculate labor compensation shares from labor compensation (nominal)divided by value added (nominal). We smooth labor compensation shares using an equation similarto Harrigan (1997): �iht = �i + �h + �t + �h ln(Kiht=Liht).
Physical capital stocks To construct industry-level capital stocks we obtain real gross �xedcapital formation (GFCF) values from 1980 to the latest available year for Australia, Austria,Belgium, Canada, Finland, France, Germany, Italy, South Korea, the Netherlands, New Zealand,Norway, Spain, the United Kingdom, and the United States. For other countries in the data set,the initial values di¤er according to the availability of the data and missing data are supplementedfrom various data sets. We take as many real GFCF data as possible from the OECD STANDatabase, the OECD National Accounts Statistics, and national statistics, but there are still someunavailable data. To interpolate these unreported values, we implement the following procedure.We �rst interpolate the unreported values of nominal GFCF. If industries such as �motor vehicles�and �other transport equipment�are unavailable but their sub-totals �transportation equipment�are available for certain years, we use the share of the nearest year to distribute the sub-totalsto each detailed sector. Second, we interpolate the unreported values of GFCF de�ators. If theindustry-level de�ator is unavailable for certain years but their sub-total sector�s de�ator is avail-able for all years, we use the available growth rates of subtotal sectors to approximate the pricelevels of each industry for missing years. If the industry-level de�ator is unavailable for all years,we use the GFCF de�ators of subtotal industries. Real GFCF is nominal GFCF divided by GFCFde�ator. Finally, for eight countries (Australia, China, Israel, South Korea, Mexico, Turkey, Chi-nese Taipei, and South Africa), we �rst compute the capital stock in an aggregate manufacturingsector (if not available, we use the country-level total) from real GFCF data and allocate them to
28
each industry according to its share of gross operation surplus (value added minuses labor com-pensation) in each country. This method assumes that the return to capital stock is identicalacross industries within a country. Manufacturing GFCF values are taken from the OECD STANDatabase (Australia, Israel, and South Korea), the OECD National Accounts Statistics (Mexico,Turkey, and South Africa), the World Development Indicators (China), or national statistics (Chi-nese Taipei). Japanese industry-level data on real GFCF expenditures are from the Histat (Instituteof Economic Research, Hitotsubashi University). We use the perpetual inventory method (with adepreciation rate of 13.33%) to compute the capital stock. The real capital stock is converted into2000 international dollars using 2000 PPP (investment) rates.
Employment Labor input (total employment) is from the OECD STAN Database and ILOLaborsta Internet. For China and Turkey, employment in an aggregate manufacturing sector isallocated to each industry according to its share of labor compensation in each country. We adjustthe international di¤erence in average annual working hours for all 32 countries but China, ChineseTaipei, and South Africa. The data is obtained from the OECD Labor Force Statistics (2010). Wenormalize U.S. working hours for each year to one.
29
Table 1. Input-output structure of 32 countries - 2000
1.1. High R&D sectors
G5 countries Non-G5 countries
average st. dev max value min value average st. dev max value min value
Intermediate inputs (input spending/output) 0.658 0.065 0.718 (FRA) 0.610 (USA) 0.684 0.088 0.765 (SVK) 0.505 (ISR)
Manufacturing inputs (input spending/output) 0.447 0.094 0.514 (FRA) 0.415 (GBR) 0.476 0.126 0.586 (CHN) 0.304 (GRC)
Own industry (input spending/output) 0.260 0.070 0.307 (FRA) 0.229 (USA) 0.270 0.109 0.382 (CAN) 0.152 (GRC)
Domestic share 0.571 0.192 0.785 (JPN) 0.348 (GBR) 0.308 0.221 0.842 (CHN) 0.041 (SVK)
Foreign share 0.429 0.192 0.652 (GBR) 0.215 (JPN) 0.692 0.221 0.959 (SVK) 0.158 (CHN)
Other mfg industries (input spending/output) 0.187 0.086 0.207 (FRA) 0.178 (GER) 0.206 0.110 0.304 (DNK) 0.114 (CAN)
Domestic share 0.479 0.170 0.578 (GBR) 0.447 (GER) 0.596 0.213 0.851 (SVN) 0.368 (TWN)
Foreign share 0.521 0.170 0.553 (GER) 0.422 (GBR) 0.404 0.213 0.632 (TWN) 0.149 (SVN)
1.2. Low R&D sectors
G5 countries Non-G5 countries
average st. dev max value min value average st. dev max value min value
Intermediate inputs (input spending/output) 0.614 0.062 0.655 (FRA) 0.590 (JPN) 0.662 0.071 0.728 (HUN) 0.580 (CAN)
Manufacturing inputs (input spending/output) 0.333 0.079 0.365 (FRA) 0.304 (GBR) 0.379 0.101 0.463 (HUN) 0.305 (NZL)
Own industry (input spending/output) 0.180 0.085 0.210 (GER) 0.152 (GBR) 0.196 0.102 0.258 (KOR) 0.136 (IRL)
Domestic share 0.755 0.151 0.877 (JPN) 0.674 (GBR) 0.566 0.251 0.857 (CHN) 0.268 (BEL)
Foreign share 0.245 0.151 0.326 (GBR) 0.123 (JPN) 0.434 0.251 0.732 (BEL) 0.143 (CHN)
Other mfg industries (input spending/output) 0.153 0.083 0.167 (USA) 0.141 (JPN) 0.183 0.114 0.247 (CHN) 0.138 (ISR)
Domestic share 0.446 0.190 0.491 (USA) 0.385 (GER) 0.483 0.242 0.606 (BEL) 0.367 (PRT)
Foreign share 0.554 0.190 0.615 (GER) 0.509 (USA) 0.517 0.242 0.633 (PRT) 0.394 (BEL)
Notes : Maximum and minimum values are based on the averages across industries within each country.
Table 2. Summary statistics
2.1. R&D content of domestic intermeidates (stock)
Year 2000 Growth rate (%) between 1995 and 2005
min value max value average st.dev min value max value average st.dev
Food products 14.3 (SVN) 15763 (USA) 1164 2935 -15.9 (SVK) 31.5 (CHN) 2.98 9.42
Textiles 3.7 (SVK) 8549 (USA) 664 1635 -23.1 (CZE) 31.7 (CHN) -0.95 10.89
Wood products 3.2 (MEX) 3169 (USA) 211 575 -10.4 (TWN) 51.1 (CHN) 2.80 10.85
Paper products 5.4 (IRL) 16562 (USA) 963 3007 -14.7 (SVK) 39.8 (CHN) 2.24 9.74
Chemicals* 6.0 (SVN) 36569 (USA) 2634 7000 -17.9 (SVK) 40.0 (CHN) 4.65 10.94
Rubber and plastics 4.4 (IRL) 16104 (USA) 1208 3392 -12.8 (SVN) 29.8 (CHN) 3.60 9.69
Mineral products 4.0 (SVN) 2945 (USA) 334 692 -9.1 (SVK) 23.6 (MEX) 2.85 7.29
Basic metals 0.6 (IRL) 12191 (JPN) 905 2312 -20.6 (SVK) 44.5 (CHN) 1.76 11.11
Fabricated metal 4.3 (IRL) 9514 (USA) 804 1887 -13.6 (SVK) 38.4 (CHN) 3.47 8.80
Machinery 8.0 (GRC) 23505 (JPN) 1989 5081 -14.2 (SVK) 34.7 (CHN) 3.32 8.81
Electrical machinery* 10.7 (SVK) 65152 (USA) 5634 15216 -10.3 (POL) 43.4 (CHN) 4.05 10.53
Motor vehicles* 1.2 (GRC) 60297 (JPN) 5088 14003 -8.8 (ISR) 38.1 (CHN) 5.51 9.44
Other transport* 0.9 (HUN) 28689 (USA) 1674 5310 -14.2 (CZE) 40.3 (CHN) 5.29 9.30
2.2. R&D content of foreign intermediates (stock)
Year 2000 Growth rate (%) between 1995 and 2005
min value max value average st.dev min value max value average st.dev
Food products 43.2 (SVN) 3443 (USA) 798 762 -3.4 (NLD) 13.4 (CHN) 3.71 3.79
Textiles 19.2 (NZL) 3832 (CHN) 608 812 -14.7 (NLD) 16.7 (TUR) 0.73 6.14
Wood products 7.3 (ISR) 811 (USA) 143 177 -4.5 (SVN) 30.9 (CHN) 5.27 6.54
Paper products 39.4 (SVN) 4050 (USA) 567 743 -4.5 (NLD) 23.2 (CHN) 3.59 5.71
Chemicals* 27.9 (NZL) 8267 (USA) 2155 2285 -7.7 (NZL) 23.5 (CHN) 5.51 5.08
Rubber and plastics 54.1 (NZL) 3696 (USA) 827 928 -4.5 (ISR) 23.4 (CZE) 5.30 5.78
Mineral products 9.2 (NZL) 1367 (CHN) 238 278 -6.1 (IRL) 12.3 (TUR) 4.87 4.37
Basic metals 16.2 (NZL) 2695 (CHN) 581 674 -10.8 (NLD) 19.8 (CHN) 4.44 6.20
Fabricated metal 31.7 (NZL) 2794 (USA) 517 616 -5.2 (NLD) 16.4 (POL) 6.16 4.93
Machinery 60.0 (GRC) 5495 (GER) 1195 1431 -6.9 (IRL) 16.0 (HUN) 6.17 4.91
Electrical machinery* 126.2 (GRC) 23057 (MEX) 5686 6701 -2.8 (NLD) 36.6 (CHN) 6.71 8.79
Motor vehicles* 22.9 (ISR) 17358 (USA) 2988 4203 -0.6 (DNK) 27.2 (CZE) 8.18 7.12
Other transport* 25.9 (GRC) 7352 (FRA) 1316 1969 -8.1 (NLD) 20.9 (CHN) 2.03 5.71
Note : The units of R&D stock is millions of 2000 international dollars. (*) indicates high-R&D industries.
Table 2. Summary statistics - Cont´d
2.3. TFP index
Year 2000 Growth rate (%) between 1995 and 2005
min value max value average st.dev min value max value average st.dev
Food products 0.461 (CHN) 1.896 (USA) 1.064 0.335 -4.4 (USA) 5.5 (KOR) 0.02 2.81
Textiles 0.482 (CHN) 2.134 (NLD) 1.091 0.442 -5.0 (JPN) 5.5 (KOR) 0.08 2.72
Wood products 0.461 (CHN) 1.811 (CAN) 1.062 0.339 -8.5 (USA) 8.8 (KOR) 0.14 3.85
Paper products 0.398 (CHN) 1.767 (USA) 1.063 0.327 -4.2 (PRT) 7.7 (KOR) 0.14 2.76
Chemicals* 0.398 (CHN) 1.712 (FRA) 1.062 0.314 -7.6 (HUN) 4.0 (CHN) 0.07 2.54
Rubber and plastics 0.427 (CHN) 1.553 (AUS) 1.068 0.344 -4.6 (TWN) 8.7 (CZE) 0.10 3.01
Mineral products 0.410 (CHN) 1.673 (NLD) 1.062 0.321 -4.4 (JPN) 7.0 (SVK) 0.10 2.86
Basic metals 0.397 (CHN) 1.783 (NOR) 1.056 0.326 -6.7 (DNK) 9.4 (IRL) 0.18 3.31
Fabricated metal 0.386 (CHN) 1.840 (CAN) 1.079 0.374 -4.4 (SWE) 8.8 (PRT) 0.09 2.88
Machinery 0.330 (SVK) 1.966 (CAN) 1.095 0.417 -6.3 (USA) 14.5 (SVK) 0.14 3.87
Electrical machinery* 0.373 (CHN) 2.376 (FIN) 1.110 0.477 -6.1 (NLD) 19.0 (SWE) 0.20 5.16
Motor vehicles* 0.373 (CHN) 1.781 (CAN) 1.049 0.308 -6.3 (HUN) 6.8 (CZE) 0.21 3.25
Other transport* 0.301 (CHN) 2.243 (CAN) 1.114 0.464 -7.7 (DNK) 8.0 (GER) 0.45 3.55
Note : (*) indicates high-R&D industries.
Table 3. TFP elasticity to R&D content of intermediates
3.1. Recipient industries: All manufacturing
Equation (A) Equation (B) Equation (C) Equation (D) Equation (E)
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
All countries, all industries ρ 0.041 (3.214)**
All countries, industry h ρ1 0.025 (1.953)* 0.019 (1.374)
All countries, other industries ( ≠ h ) ρ2 0.011 (0.634) 0.006 (0.329)
Country i , all industries ρ3 -0.018 (-1.141)
Countries ( ≠ i ), all industries ρ4 0.062 (3.316)**
Country i , industry h ρ5 0.008 (1.077)
G5, industry h ρ6 0.017 (1.293)
Country i , other industries ( ≠ h ) ρ7 -0.041 (-1.616)
G5, other industries ( ≠ h ) ρ8 0.051 (1.796)*
Obs 1227 1227 1227 997 1032
Adjusted R-2 0.698 0.697 0.699 0.691 0.689
SIC 0.352 0.358 0.353 0.431 0.402
3.2. Recipient industries: High R&D
Equation (A) Equation (B) Equation (C) Equation (D) Equation (E)
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
All countries, all industries ρ 0.082 (3.906)**
All countries, industry h ρ1 0.078 (2.344)** 0.078 (2.174)**
All countries, other industries ( ≠ h ) ρ2 -0.004 (-0.102) -0.042 (-1.189)
Country i , all industries ρ3 0.017 (0.698)
Countries ( ≠ i ), all industries ρ4 0.075 (2.921)**
Country i , industry h ρ5 0.038 (2.625)**
G5, industry h ρ6 0.075 (2.719)**
Country i , other industries ( ≠ h ) ρ7 0.007 (0.130)
G5, other industries ( ≠ h ) ρ8 -0.006 (-0.108)
Obs 369 369 369 303 309
Adjusted R-2 0.648 0.650 0.651 0.671 0.649
SIC 1.315 1.321 1.319 1.300 1.331
3.3. Recipient industries: Low R&D
Equation (A) Equation (B) Equation (C) Equation (D) Equation (E)
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
All countries, all industries ρ 0.004 (0.291)
All countries, industry h ρ1 -0.003 (-0.246) -0.009 (-0.596)
All countries, other industries ( ≠ h ) ρ2 0.009 (0.445) 0.003 (0.125)
Country i , all industries ρ3 -0.003 (-0.115)
Countries ( ≠ i ), all industries ρ4 0.013 (0.447)
Country i , industry h ρ5 0.008 (0.754)
G5, industry h ρ6 -0.016 (-1.189)
Country i , other industries ( ≠ h ) ρ7 -0.040 (-1.339)
G5, other industries ( ≠ h ) ρ8 0.051 (1.543)
Obs 858 858 858 694 694
Adjusted R-2 0.734 0.734 0.734 0.726 0.726
SIC 0.354 0.362 0.362 0.443 0.442
Notes : t -statistics are based on robust standard errors. ** (*) indicates statistically significant at the 5% (10%) level.
Table 4. Robustness checks: Estimates of TFP elasticities from equation (A)
4.1. Recipient industries: All manufacturing
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients All All w/o non-OECD All All All
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ 0.041 (3.214)** 0.041 (3.314)** 0.043 (3.141)** 0.040 (3.143)** 0.028 (1.803)* 0.041 (3.214)**
Obs 1227 1227 1110 1227 818 1227
Adjusted R-2 0.698 0.698 0.637 0.698 0.698 0.698
SIC 0.352 0.352 0.392 0.353 - 0.352
4.2. Recipient industries: High R&D
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients All All w/o non-OECD All All All
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ 0.082 (3.906)** 0.091 (4.756)** 0.083 (3.755)** 0.083 (3.943)** 0.084 (3.128)** 0.082 (3.903)**
Obs 369 369 333 369 246 369
Adjusted R-2 0.648 0.643 0.562 0.648 0.673 0.648
SIC 1.315 0.541 1.341 1.315 - 1.315
4.3. Recipient industries: Low R&D
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients All All w/o non-OECD All All All
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ 0.004 (0.291) 0.003 (0.202) 0.005 (0.310) 0.005 (0.330) -0.031 (-1.686)* 0.005 (0.301)
Obs 858 858 777 858 572 858
Adjusted R-2 0.734 0.734 0.686 0.734 0.721 0.734
SIC 0.354 0.354 0.388 0.354 - 0.354
Notes : t -statistics are based on robust standard errors. ** (*) indicates statistically significant at the 5% (10%) level.
Table 5. Robustness checks: Estimates of TFP elasticities from equation (D)
5.1. Recipient industries: All manufacturing
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients non G5 non G5 w/o non-OECD non G5 non G5 non G5
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ2 0.006 (0.329) -0.010 (-0.574) 0.015 (0.733) 0.012 (0.638) 0.025 (1.040) 0.006 (0.343)
ρ5 0.008 (1.077) 0.011 (1.521) 0.007 (0.850) 0.002 (0.205) -0.012 (-1.001) 0.008 (1.070)
ρ6 0.017 (1.294) 0.021 (1.802)* 0.014 (0.955) 0.020 (1.549) 0.028 (1.739)* 0.017 (1.280)
Obs 997 997 898 985 663 997
Adjusted R-2 0.691 0.691 0.622 0.691 0.684 0.691
SIC 0.431 0.431 0.469 0.441 - 0.431
5.2. Recipient industries: High R&D
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients non G5 non G5 w/o non-OECD non G5 non G5 non G5
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ2 -0.042 (-1.189) -0.032 (-1.151) -0.033 (-0.830) -0.037 (-1.090) -0.079 (-1.864)* -0.042 (-1.192)
ρ5 0.038 (2.625)** 0.037 (2.690)** 0.033 (2.230)** 0.036 (2.375)** 0.031 (1.421) 0.038 (2.628)**
ρ6 0.075 (2.719)** 0.067 (2.999)** 0.072 (2.228)** 0.073 (2.677)** 0.111 (3.845)** 0.075 (2.714)**
Obs 303 303 270 301 202 303
Adjusted R-2 0.671 0.672 0.578 0.672 0.675 0.671
SIC 1.300 1.297 1.341 1.297 - 1.300
5.3. Recipient industries: Low R&D
(1) (2) (3) (4) (5) (6)
Notes Benchmark DBQ - 5 year lag 2SLS ROW
Recipients non G5 non G5 w/o non-OECD non G5 non G5 non G5
Types of R&D Total Direct Total Total Total Total
coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat coef t -stat
ρ2 0.003 (0.125) -0.019 (-0.883) 0.006 (0.219) 0.014 (0.645) 0.044 (1.438) 0.003 (0.147)
ρ5 0.008 (0.754) 0.011 (1.093) 0.008 (0.666) -0.001 (-0.128) -0.019 (-1.083) 0.008 (0.748)
ρ6 -0.016 (-1.189) -0.012 (-0.921) -0.019 (-1.231) -0.009 (-0.618) -0.013 (0.690) -0.017 (-1.199)
Obs 694 694 628 684 461 694
Adjusted R-2 0.726 0.726 0.671 0.725 0.712 0.726
SIC 0.443 0.442 0.467 0.463 - 0.443
Notes : t -statistics are based on robust standard errors. ** (*) indicates statistically significant at the 5% (10%) level.
Figure 1.1. Industry shares on total R&D stock (32 countries) - 2000
Figure 1.2. Shares of G5 countries in total R&D stock by industry - 2000
0
5
10
15
20
25
30
35
40
45F
oo
d p
rod
ucts
Tex
tile
s
Wo
od
pro
duct
s
Pap
er p
rod
ucts
Ch
emic
als
Ru
bb
er a
nd
pla
stic
s
Min
eral
p
rod
uct
s
Bas
ic m
etal
s
Fab
rica
ted
m
etal
Mac
hin
ery
Ele
ctri
cal
mac
hin
ery
Mo
tor
veh
icle
s
Oth
er tr
ansp
ort
0
10
20
30
40
50
60
70
80
90
100
Fo
od
pro
duc
ts
Tex
tile
s
Wo
od
pro
duct
s
Pap
er p
rod
ucts
Ch
emic
als
Ru
bb
er a
nd
pla
stic
s
Min
eral
p
rod
uct
s
Bas
ic m
etal
s
Fab
rica
ted
m
etal
Mac
hin
ery
Ele
ctri
cal
mac
hin
ery
Mo
tor
veh
icle
s
Oth
er tr
ansp
ort
Figure 2.1. R&D content of intermediates in China - 2000
0
2000
4000
6000
8000
10000
12000
Fo
od
pro
du
cts
Tex
tile
s
Wo
od
pro
du
cts
Pa
per p
rod
ucts
Ch
em
ica
ls
Ru
bb
er a
nd
pla
stic
s
Min
era
l
pro
du
cts
Ba
sic m
eta
ls
Fa
bri
ca
ted
meta
l
Ma
ch
inery
Ele
ctr
ica
l
ma
ch
inery
Mo
tor v
eh
icle
s
Oth
er
tra
nsp
ort
(millio
ns,
R&
D S
tock
)
Domestic, industry j
G5, industry j
Domestic, other industries
G5, other industries
Figure 2.2. R&D content of intermediates in the Czech Republic - 2000
0
100
200
300
400
500
600
700
800
900
1000
Fo
od
pro
du
cts
Tex
tile
s
Wo
od
pro
du
cts
Pa
per p
rod
ucts
Ch
em
ica
ls
Ru
bb
er a
nd
pla
stic
s
Min
era
l
pro
du
cts
Ba
sic m
eta
ls
Fa
bri
ca
ted
meta
l
Ma
ch
inery
Ele
ctr
ica
l
ma
ch
inery
Mo
tor v
eh
icle
s
Oth
er
tra
nsp
ort
(millio
ns,
R&
D S
tock
)
Domestic, industry j
G5, industry j
Domestic, other industries
G5, other industries
Figure 2.3. R&D content of intermediates in the UK - 2000
0
1000
2000
3000
4000
5000
6000
7000
Fo
od
pro
du
cts
Tex
tile
s
Wo
od
pro
du
cts
Pa
per p
rod
ucts
Ch
em
ica
ls
Ru
bb
er a
nd
pla
stic
s
Min
era
l
pro
du
cts
Ba
sic m
eta
ls
Fa
bri
ca
ted
meta
l
Ma
ch
inery
Ele
ctr
ica
l
ma
ch
inery
Mo
tor v
eh
icle
s
Oth
er
tra
nsp
ort
(millio
ns,
R&
D S
tock
)
Domestic, industry j
G5 (w/o GBR), industry j
Domestic, other industries
G5 (w/o GBR), other industries
Figure 2.4. R&D content of intermediates in the US - 2000
0
10000
20000
30000
40000
50000
60000
Fo
od
pro
du
cts
Tex
tile
s
Wo
od
pro
du
cts
Pa
per p
rod
ucts
Ch
em
ica
ls
Ru
bb
er a
nd
pla
stic
s
Min
era
l
pro
du
cts
Ba
sic m
eta
ls
Fa
bri
ca
ted
meta
l
Ma
ch
inery
Ele
ctr
ica
l
ma
ch
inery
Mo
tor v
eh
icle
s
Oth
er
tra
nsp
ort
(millio
ns,
R&
D S
tock
)
Domestic, industry j
G5 (w/o USA), industry j
Domestic, other industries
G5 (w/o USA), other industries
Figure 3.1. Densities of t-statistics – Random choices of 5-source countries – High-R&D industries
Figure 3.2. Densities of t-statistics – Random choices of 5-source countries – Low-R&D industries
0
5
10
15
20
25
-1.9
-1.7
-1.5
-1.2
-1.0
-0.8
-0.5
-0.2
0.1
0.3
0.6
1.0
1.3
1.6
2.0
2.3
2.7
3.0
3.4
3.8
4.2
4.6
5.0
PD
F (
%)
t-statistics on R&D stock embodied from 5 countries
own
other
0
2
4
6
8
10
12
14
16
-1.9
-1.7
-1.5
-1.2
-1.0
-0.8
-0.5
-0.2
0.1
0.3
0.6
1.0
1.3
1.6
2.0
2.3
2.7
3.0
3.4
3.8
4.2
4.6
5.0
PD
F (
%)
t-statistics on R&D stock embodied from 5 countries
own
other
Figure 4.1. Probability of rejection as R&D-source country (High-R&D industries, Own)
Figure 4.2. Probability of rejection as R&D-source country (Low-R&D industries, Own)
0%
1%
2%
3%
4%
5%
6%
NZ
L
GR
C
PR
T
SV
N
HU
N
SV
K
ME
X
ZA
F
TU
R
IRL
NO
R
PO
L
CZ
E
DN
K
ISR
AU
S
AU
T
FIN
ES
P
BE
L
TW
N
NL
D
SW
E
CA
N
CH
N
ITA
KO
R
GB
R
FR
A
DE
U
JPN
US
A
(%)
0.0%
0.5%
1.0%
1.5%
2.0%
2.5%
3.0%
3.5%
4.0%
NZ
L
GR
C
PR
T
SV
N
HU
N
SV
K
ME
X
ZA
F
TU
R
IRL
NO
R
PO
L
CZ
E
DN
K
ISR
AU
S
AU
T
FIN
ES
P
BE
L
TW
N
NL
D
SW
E
CA
N
CH
N
ITA
KO
R
GB
R
FR
A
DE
U
JPN
US
A
(%)