We are unable to create an online viewer for this document. Please download the document instead.
EPRU Working Paper Series
Economic Policy Research Unit
Institute of Economics
University of Copenhagen
DK-1455 Copenhagen K
Tel: (+45) 3532 4411
Fax: (+45) 3532 4444
Technology Spillover through Trade and TFP Convergence:
120 Years of Evidence for the OECD Countries
Jakob Brøchner Madsen
The activities of EPRU are financed by a grant from
The National Research Foundation
Technology Spillover through Trade and TFP Convergence: 120 Years of
Evidence for the OECD Countries1
Jakob B. Madsen
Institute of Economics, EPRU and FRU
University of Copenhagen2
Abstract. Using a new dataset on imports of technology and total factor productivity (TFP) over more than a
century for the OECD countries, this paper tests for international technological transmission through trade. The
empirical estimates suggest that imports of knowledge have been responsible for an almost 200% increase in
TFP over the past century, but that the spillover effect has been highly unevenly distributed across countries,
but has contributed to TFP convergence among the OECD countries.
JEL classification: E13, E22, E23, O11, O3, O47.
Key words: Technology spillovers, Imports, TFP convergence.
Since the seminal paper of Coe and Helpman (1995), henceforth C&H, several studies have documented that
R&D cross-country spillovers, through the channel of trade flows, have been an important engine of TFP
growth in the industrialised countries (del Barrio-Castro et al., 2002, Coe et al., 1997, Crespo et al., 2004,
Engelbrecht, 1997, Frantzen, 2000, Guellec and Van Pottelsberghe de la Potterie, 2001, 2004, Lichtenberg and
Van Pottelsberghe de la Potterie, 1998, and Lumenga-Neso et al., 2001). Other studies have found the spillover
channel through imports to be less significant. Using the same model as C&H, Keller (1988) finds that
randomly generated bilateral trade shares in some instances gave better results than those of C&H and Kao et
al. (1999) find that the estimated coefficient of imports of knowledge is insignificant once the bias from their
OLS panel estimates is corrected for.
Common for all these empirical studies is that the stock of ideas is generated from R&D
expenditures. Although R&D stock is an excellent measure of knowledge, it has the limitation that R&D data
have only increasingly become available across OECD countries over the last couple of decades. R&D data for
the OECD countries are first scantly available from the late 1960s and, generally, were first consistently
collected on a fixed-term basis from the beginning of the 1980s, and first consistently collected on an annual
1 Support from an EPRU grant from the Danish government is gratefully acknowledged. Signe Skarequest and Philip Øksness
provided excellent research assistance.
2 Studiestraede 6, 1455 Copenhagen K, Denmark, Ph. +45 35 38 50 18, Fax + 45 35 32 30 00, e-mail: email@example.com.
basis from 1995.3 This data limitation has rendered it difficult to assess the effects of spillover on TFP growth
with confidence, particularly because the R&D stock and total factor productivity (TFP) have both been
increasing over the periods covered by the R&D based studies in almost all countries; thus lowering the
identifying variations in the data.
This paper seeks to overcome the data problem by using a new data set to assess the effects of
knowledge spillover through trade on TFP growth in 13 major OECD countries over more than a century, using
bilateral trade flow weights for 21 OECD countries.4 Patent data have been consistently collected on an annual
basis in almost all OECD countries since the Paris Convention was signed in 1883. While it is not claimed that
patent counts are superior to R&D data as a measure of the innovative activity, they are nevertheless a valuable
complement to the R&D based studies; particularly because imports of patent-based knowledge stock have
fluctuated markedly over the past 120 years. Knowledge imports to the OECD countries have, in fact,
fluctuated markedly over the past century; thus putting the C&H hypothesis to the real test. Similarly, TFP has
fluctuated over most of the past century and first stated to increase consistently for all the 13 OECD countries
considered since the beginning of the 1960, which again challenges the spillover framework to explain both ups
and downs. Furthermore, the industrialised countries are first now as open as they were in 1913.5 The Great
Depression, for instance, reduced openness, measured ad the import-GDP ratio, to almost a half in 1929 and,
therefore, lowered substantially the potential for technological spillovers through imports.
The TFP data set is constructed for the OECD countries extending more than one century back in
time using average factor shares for each individual country. Payments to the self employed have been imputed
into factor shares. Furthermore, the TFP has been based on hours worked as opposed to number of employed,
which is crucial since annual hours worked per worker have almost halved over the past 130 years. The import
weighting scheme follows the suggestions of Coe et al. (1997) and Xu and Wang (1999). Since technological
spillover through the channel of imports is most likely to take place through technologically sophisticated
product imports, the bilateral import weights are based on products that are classified by the OECD as high
technological such as machinery, equipment and chemical products.
The paper, furthermore, assesses whether technological spillovers through trade have been a
contributing factor to the TFP convergence of OECD countries as documented by Dowrick and Nguyen (1989)
and Wolff (1991) using a much more limited data sample than the one used here. Overall, this paper finds that
3 Only the US has consistent R&D data available on an annual basis from 1953.
4 The following 13 OECD countries are considered: Canada, the US, Japan, Australia, Denmark, Finland, France, Germany, Italy, the
Netherlands, Norway, Sweden and the UK.
5 The import-GDP ratio is on average the same today as in 1913 for the 13 countries in the sample. The following six countries have
experienced a decline in the import-GDP ratio since 1913: Japan, Denmark, Germany, Italy, the Netherlands, Norway, and the UK.
international technology spillovers through trade have been a significant contributor to the TFP growth in the
OECD countries over the past century and have contributed to TFP convergence, particularly before WWI.
2 Empirical framework
To estimate the effects on TFP of imports of technology the following error-correction and cointegration
models are estimated using pooled cross section and time-series analysis:
TFP = ? + ? ln( d
S / pop )
+? ln f
S + TD + CD + ?
TFP = ? + ? ? ln( d
S / pop )
+ ? ? ln f
S + ? ˆ
+ TD + ?
where Sd is the stock of domestic patents, f
S is the stock of imports of technology, CD is fixed effect country
dummies, TD is time-dummies, pop is size of population, ˆ
?2i,t 1? is the error-correction term, which is the
lagged residual from estimates of (1), ? is a disturbance term, and the subscripts t and i signify time and
country. One-period lags of the explanatory variables are included in the estimates of (2). Estimated
coefficients of further lags and lags of the dependent variable were insignificant.
Time-dummies are included in the estimates to allow for the influence on TFP of excluded
variables that are common across countries. As shown in the empirical section, the inclusion of time-dummies
is of major importance in the cointegration estimates when data for more than a century is considered. This is
because factors other than the stock of patents are likely to have influenced TFP, but have not been accounted
for in the model, such as human capital, organisational capital, and shifts from low to high productivity sectors.
Models (1) and (2) deviate from most other empirical estimates in the following four respects.
S is usually multiplied by the openness of the economy based on the premise that the economy’s
capacity to tap into the world knowledge depends on the openness of the economy. However, the measure of Sf
used here incorporates import volume and, therefore, implicitly incorporates the openness of the economy. The
empirical estimates below show that much better results are obtained when Sf is not multiplied by openness.
Second, the coefficient of Sd is not allowed to differ from the sample average for the G7 countries because their
estimated coefficients were not significantly different from zero, as shown in the empirical section.
Third, domestic stock of knowledge is normalised by population to obtain a logical
correspondence to TFP. Since TFP is normalised relative to factor inputs, it follows that imports of technology
also need to be normalised by population. In the studies referred to above this lack of normalisation is probably
not crucial to the results because the considered time-span is relatively short. However, when more than a
century of data is used, the normalisation becomes paramount. The populations of Australia and the USA, for
instance, have increased more than eight-fold and five-fold, respectively, over the period from 1883 to 2002.
Without the normalisation by population in the estimates, this would have meant that the stock of knowledge in
Australia would have increased eight times as much as it would otherwise have done, purely due to population
growth, provided that the propensity to patent is independent of the size of the population. Consequently, the
population adjustment becomes vital in estimates with long data. Fourth, Sf is also normalised with population
as shown in Section 2.1.
Caution has to be exercised in the interpretation of the panel cointegration estimates because their
statistical properties are derived under the assumption of cross-country independence. Clearly this is not a
tenable assumption since countries have been constantly hit by the same shocks such as the oil price shocks, the
monetary and agricultural shocks during the Great Depression, technological revolutions, exchange rate shocks,
world wars, and so on. It cannot, therefore, be assumed that TFP growth in the OECD countries has been driven
by forces that are entirely independent. The consequence of violation of the independence assumption is that
the test statistics are biased and that the coefficient estimates may not be super consistent as usually assumed in
panel cointegration estimates. The alternative of estimating the cointegration equations for each individual
country is not considered here as a viable alternative to the panel estimates because the time-dummies cannot
be included in single-country estimates and the findings below show that the time-dummies explain the
majority of the increase in TFP.
To shed supplementing light on the long-run relationship between the variables Equation (2) is
also estimated in non-overlapping five-year differences without the error-correction term to filter out the
potential influence of cyclical forces. Baltagi and Griffin (1984), for instance, advocate long differences as a
method to capture long-run relationships because they shave off short-run influences from the estimates.
As stated in the introduction, the data cover the following 13 OECD countries over the period from 1883 to
2002: Canada, the US, Japan, Australia, Denmark, Finland, France, Germany, Italy, the Netherlands, Norway,
Sweden and the UK. TFP data are backdated for Japan before 1885 using TFP growth for Australia and
backdated for France before 1894 using TFP growth for the UK in the graphical presentations. Details of data
availability and data sources are relegated to the data appendix. The 13 OECD countries are referred to as the
OECD countries for simplicity.
Imports of knowledge from country j to country i are computed from the following weighting
S = ?
i ? j
m is imports volume of high technological products from country j to country i,
S is the stock of
domestic knowledge in country j, and r
Y is real GDP at USD purchasing power parity for country j. The stock
of knowledge is computed using the inventory perpetual method on the domestic patent applications with a
20% geometric depreciation rate. This depreciation rate for patents and R&D follow the estimates by Parkes
and Schankerman (1984).
m is computed as nominal imports of high technological products deflated by the
economy-wide value added price-deflator because the available historical import price-deflators are severely
biased by being computed as import value divided by the metric weight of imports. This computation implies
that compositional changes and quality changes of imported products are not embodied into the import price
deflators. The bias is likely to be particularly serious in long-term analysis where imports have changed from
being predominantly low value added products to highly sophisticated merchandise over the course of the past
century. While the economy-wide deflator is far from being an ideal import deflator it, nevertheless, allows for
quality changes. Furthermore, the tradable sector has a disproportionally high weight in historical GDP data
(Maddison, 1982). The high technological products are SITC Section 5, chemicals and related products,
Section 7, machinery and transport equipment, and Section 8.7, professional and scientific instruments. The
OECD (1995) finds these sections to have a high or medium-high R&D intensity.
The weighting scheme given by (3) follows the suggestion of Lichtenberg and van Potterie (1998)
except that imports are weighted by population in their scheme. Import volume is divided by population of the
importing country in (3) to avoid the problem that imports of technology are proportional to the size of the
country and, therefore, to achieve a reasonable correspondence between Sf and TFP. As argued above, this
normalisation is crucial when the size of the population changes significantly in the estimation period. GDP or
import volume could, alternatively, have been used as normalising factors. However, this would have resulted
in an Sf that would converge to an almost constant level in the long run because the propensity to import
converge to a constant in the long run and the Sd/Yr-ratio converges to a constant in the steady state as predicted
by the Solow-Ramsey model of economic growth under the assumption of a constant discount factor.
Consequently, Sf normalised by GDP or total imports would not be able to account for some of the TFP
increase experienced in the OECD countries over the past century or more.
From (3) it can now be seen why ln
S is not multiplied by openness in (1) and (2). The term
(mr/pop) in (3) is a measure of openness, although not in the usual meaning of the ratio of imports to total
income. Multiplying ln
S with openness would, consequently, impact on TFP with the power of two.
Figure 1 displays imports of knowledge for the OECD countries following the weighting scheme
given by (3). The figure shows that imports of knowledge have fluctuated substantially over the past 120 years.
Three periods stand out: the stagnation and decline from 1913 to 1973 and the two periods of almost
uninterrupted increase from 1883 to 1913 and from 1973 to the present. In the two periods of increase trade
expanded and industrial revolutions took place (Jovanovic and Rousseau, 2003, Perez, 2002, Gordon, 2000).
The industrial revolution before 1913 was within the steel and chemical industries and the great inventions such
as the dynamo, light bulb, the automobile among were introduced at that time. The recent ICT revolution has
also been related to infrastructure and was initiated by the invention of the microprocessor by Intel in 1971
(Jovanovic and Rousseau, 2003).
Figure 1. Imports of Knowledge, OECD
1883 1893 1903 1913 1923 1933 1943 1953 1963 1973 1983 1993
Note. Unweighted average of the OECD countries used here.
Consider the period in-between the Great Depression and WWII, which shows a marked decline in the imports
of knowledge to a low in 1945. The fall during the Great Depression was not because of a slowdown in the
innovative activity on a worldwide scale but because imports imploded to almost a third during the first years
of the Depression. The decline during WWII, which was due to both lower innovative activity and declining
trade, was reversed only after a few years after the war. Despite being the golden period for the OECD
countries in terms of per capita growth, imports of knowledge remained almost constant in the 1950-1973
period due to stagnating world innovative activity. The literature on innovations and technological revolutions
generally agree that the postwar period up to the beginning of the 1970s was a period of low innovative activity
(Jovanovic and Rousseau, 2003, Perez, 2002, Gordon, 2000). That per capita output growth was so spectacular
in the OECD countries in the 1950-1973 period was predominantly because the great inventions in the past
were improved upon and were put to use (Gordon, 2000).
Figure 2. Imports of Knowledge, G7
1883 = 1
Over the whole period considered the imports of knowledge have increased 13-fold for the average country;
however, the increase has been quite different across countries (Figures 2 and 3). The countries experiencing
the largest growth in imports of knowledge are the US, Japan, the Scandinavian countries, and France. The
increase in France and the US, particularly, has been due to increasing propensity to import whereas Japan and
the Scandinavian countries have imported from countries that have experienced a large growth in knowledge.
Interestingly, during the period considered Japan has lowered its propensity to import, which has deterred some
of the knowledge spillover benefits from its trade partners. The countries that have experienced the least
increase in imports of knowledge are Germany, the UK, Australia, the Netherlands, Canada and Italy. The
significance of the cross-country variation over the whole period is that Japan, the US and Norway have
experienced a four times as strong increase in imports of knowledge as Canada, the Netherlands and Italy. As
shown in the figures below these trends are reflected in the TFP growth rates.
Figure 3. Imports of Knowledge, Non-G7
1883 = 1
The construction of the TFP data based on homogenous Cobb-Douglass technology where factors shares are
allowed to vary across countries
TFP = Y /
where L is labour inputs measured as annual hours worked per worker times economy-wide employment, K is
capital stock based on the inventory perpetual method for economy-wide real investment, and L
? is labour’s
average income share over the data period. Labour’s income share is calculated as the economy-wide
compensation to employees divided by nominal GDP, where compensation is corrected for imputed payments
to the self-employed. The problem associated with national account factor shares is that earnings from self-
employment are counted as profits although the labour of the self-employed should be counted as labour
income. To correct for this bias the average earnings per employee multiplied by the number of self-employed
is added to the compensation to employees. Capital stock and GDP are measured at purchasing power parities
in USD, which is important when the TFP convergence is examined.
Figure 4. TFP, G7
The TFP series are displayed in Figures 4 and 5. Again there is a marked cross-country variation in TFP growth
over the whole period considered. The UK and Australia stand out as the low performers, whereas Japan stands
out as the top performer. The Scandinavian countries, Germany, France and Italy have also performed well.
From looking at the figures it appears that the trade spillover can explain at least some of the cross-country
variations. The countries that have experienced the st