Revisiting international knowledge spillovers: the role of GVCs

Abstract

The diffusion of knowledge is an important determinant of economic development. International trade has been established as a key mechanism in facilitating diffusion. The rise of global value chains (GVCs) has transformed trade in recent years. Yet the role of GVCs in giving rise to knowledge spillovers remains under-explored. In this paper, we study the elasticity of industry-level total factor productivity (TFP) to technology that is imported through intermediate trade in GVCs. To do so, we combine novel input–output decomposition methods with recent insights from the literature on the factor content of trade. We focus on a panel of 32 countries and 39 sectors over the 2000–2014 period using WIOD and OECD data. We find that domestic TFP is elastic to knowledge flows arising from GVCs and that the magnitude of this effect is larger relative to all other knowledge flows. We also find that GVC participation is particularly conducive to technology upgrading in countries that are far away from the technology frontier, and that GVC-related spillovers persist over large geographical distances.

1. Introduction

The international diffusion of knowledge is an important determinant of productivity growth and, consequently, economic development. Cross-country differences in the uptake of new technology have been estimated to account for approximately 25% of per capita income differences (Comin and Hobijn, 2010). Trade is one of the channels through which knowledge can “spill over” across firms and industries in different countries (Grossman and Helpman, 1991; Keller, 2004; Cai et al., 2022). In recent years, the organization of international trade has been dramatically altered by global value chains (GVCs)—a form of international industrial organization in which production is a geographically fragmented, yet closely coordinated process (Gereffi et al., 2005; Baldwin, 2011; Antràs and Chor, 2013).

While a large empirical literature has studied spillovers from international trade more generally—finding that exposure to knowledge produced by one’s trading partners does enhance domestic productivity in recipient economies (Coe and Helpman, 1995; Lichtenberg and van Pottelsberghe de la Potterie, 1998; Xu and Wang, 1999; Keller, 2002b; Nishioka and Ripoll, 2012; Fracasso and Vitucci Marzetti, 2015)—there is little cross-country evidence on GVC-related spillovers specifically. Yet the emergence of GVCs is often described as having substantial implications for the international diffusion of knowledge.¹ This paper contributes to filling this gap. We ask whether trade in GVCs gives rise to knowledge spillovers, and what is their magnitude relative to those arising from other channels.

Typically organized by multinational enterprises (MNEs) through networks of suppliers and affiliates in host countries (Antras and Yeaple, 2013; Cadestin et al., 2018), GVCs are underpinned by particularly close relationships between customers and suppliers. This stands in contrast with the type of bilateral, arms-length trading relationships which literature on international research and development (R&D) spillovers has focused so far. These “sticky” relationships are thought to be one of the key mechanisms enabling the cross-border spill over of knowledge—in the form of new technology, input varieties, and standards (Gereffi et al., 2005; Lema et al., 2019).²

Firm-level studies provide support to this idea. There is evidence that GVC participation enhances the productivity of domestic suppliers in North Africa (Del Prete et al., 2017) and Latin America (Montalbano et al., 2018), and that it facilitates the transfer of production standards and know-how in specific industries (Saliola and Zanfei, 2009; Alcacer and Oxley, 2014).³ Yet a large literature of case studies is more skeptical, arguing that a skewed distribution of power within GVCs can limit the extent to which spillovers occur (Humphrey and Schmitz, 2002), and that, moreover, knowledge transfer in GVCs is an affair largely conditional on local absorptive capacities (Morrison et al., 2008) and institutional settings (Pietrobelli and Rabellotti, 2011).

In this paper, we contribute to the literature on GVCs and technology diffusion by providing novel evidence on the aggregate-level relationship between domestic productivity and knowledge sourcing through the channel of GVCs. To do so, we combine insights from two distinct strands of literature. First is literature on inter-industry and international R&D spillovers, which provides the estimation framework for this study. The second is literature on GVCs, and particularly work which uses input–output decompositions to trace global flows of capital, labor, and knowledge (Wang et al., 2017).⁴ This literature provides the theoretical foundation of our paper.

More specifically, to provide evidence on the role of knowledge flows in GVCs in a cross-country, internationally comparable setting, we study GVC-related spillovers in a Coe and Helpman Coe and Helpman (1995)-type framework, combining data on business enterprise R&D with global input–output tables for a panel of 32 countries and 39 industries over the 2000-2014 period.⁵ We estimate the elasticity of domestic industry-level total factor productivity (TFP) to measures of the R&D that countries source through participation in GVCs. We then compare these elasticities to those arising from importing and domestic sourcing.

Our measures have two advantages. First, they enable us to directly capture, for any given industry in any country, the amount of trade-embodied R&D which is sourced and supplied through GVCs, without making use of weighting schemes which can introduce bias in the estimation (Keller, 1998; Coe and Hoffmaister, 1999).⁶ Another distinguishing feature of our measures is that they capture direct (originating from one’s own trading partner) as well as indirect (originating from all the trading partners of one’s own partner) knowledge flows—an important requirement in literature on spillovers since at least the work of Lumengo-Neso et al. Lumengo-Neso et al. (2005).

We construct our measures for each of the country-sector pairs in our sample and for each year between 2000 and 2014. We observe that for country-sectors in our panel, the import of embodied knowledge via GVCs is at least an order of magnitude smaller than domestically sourced R&D. Yet when estimating elasticities, we find that the sourcing of R&D via GVCs has substantial effects on domestic productivity. In our benchmark model, we estimate an elasticity of 0.16—against one of 0.07 for domestic R&D flows and one of 0.01 for non-GVC international R&D flows. Our estimate suggests that a 10% increase in R&D sourcing through GVCs boosts domestic productivity by 1.6%, and that this effect is larger than that associated with any other R&D flow.

We also find that these effects are stronger for countries and industries which sit further away from the technology frontier. The elasticity of domestic TFP to knowledge imported through GVCs tends to be approximately 0.5 percentage points higher for industries toward the bottom of the distribution of R&D capabilities in our panel relative to industries toward the top of the distribution. Finally, we find that while geographical distance does dampen international R&D spillovers, GVC-related spillovers are surprisingly robust over large distances. When taking into account the distance between trading partners, GVC-related knowledge flows remain positively and significantly associated with domestic productivity, although their elasticity tends to be, on average, about a third than that estimated without taking distance into account.

Moreover, we rule out that our results are driven by reverse causality in the relationship between domestic productivity and knowledge sourced through international trade in GVCs or that they are driven by the inclusion of large R&D spenders, such as China, Japan, and the United States in our sample. We also provide evidence that R&D spillovers from GVC participation are a feature not only of manufacturing industries, but also of knowledge-intensive business services (KIBS). This finding supports recent work on the role of KIBS and services more generally in international trade and globalization (Baldwin, 2019; Savona, 2021).

Overall, our findings provide evidence that participation in GVCs has an important effect on industry-level productivity. While the estimates we report appear to be small, they fall toward the middle of the distribution of elasticities reported in directly comparable studies, which range from almost 0.3 (Keller, 2002b) to 0.04 (Nishioka and Ripoll, 2012) and 0.07 (Foster-McGregor et al., 2017). These differences are chiefly related to the period under consideration, and to differences in sample size. The largest estimates are reported by Keller Keller (2002b), whose work focuses on the period from the mid-1970s to the mid-1990s—arguably the heyday of economic globalization. Moreover, all these studies focus on manufacturing only, whereas we study trade across all sectors.

Our results contribute to corroborate theoretical work and micro-level empirical studies, which point to the importance of GVCs as a channel of knowledge diffusion across countries and industries. Moreover, we provide evidence that GVC participation is particularly beneficial for countries relatively far away from the technology frontier.

The remainder of the paper is organized as follows. Section 2 provides a brief overview of the literature on international R&D spillovers and GVCs. Section 3 outlines our methodology. Section 4 describes our data. Section 5 provides descriptive statistics and presents our results, placing them within the framework of related literature. Section 6 concludes.

2. Related literature

2.1 Inter-industry and international R&D spillovers

Empirical studies of inter-industry R&D spillovers are the first strand of literature which our paper relates to.⁷ The first wave of work in this area focused on single economies, and particularly the United States (Scherer, 1982; Griliches and Lichtenberg, 1984; Griliches, 1998a). This line of work focused on estimating, in a regression framework, the elasticity of domestic productivity—typically defined in terms of TFP—to a given stock of cumulative R&D expenditure that is external to the firm, sector, or country under consideration (for a comprehensive review of this literature, see Hall et al. (2009)). Literature on international inter-industry spillovers followed, premised on the idea that in open economies, the stock of external knowledge upon which firms can build lies as much at home as it does abroad.⁸

Yet not all external knowledge will be useful to everyone. Firms are likely to rely upon different amounts of knowledge from different sources, according to how distant—in economic, technological, and geographical space—these are relative to them (Griliches, 1998b; Keller, 2002a). Devising a weighting function to capture the distance between firms endowed with different “pieces” of knowledge is thus a central theme in this literature.⁹ Assuming that the volume of trade is an important determinant, one approach uses bilateral imports as weights. These are then normalized using either the recipient’s total imports (Coe and Helpman, 1995) or its trading partners’ economic size (Lichtenberg and van Pottelsberghe de la Potterie, 1998; Fracasso and Vitucci Marzetti, 2015).

Subsequent studies zoom into different types of imports.¹⁰ The mechanisms for knowledge transfer might differ when considering final as opposed to intermediate or capital goods.¹¹ Focusing on capital goods imports, for instance, Xu and Wang Xu and Wang (1999) estimate a larger elasticity of domestic productivity to the import of foreign R&D relative to overall imports. In a further extension of this literature, Lumengo-Neso et al. Lumengo-Neso et al. (2005) argue that open and interconnected economies should benefit not only from their own trading partners’ knowledge, but also from knowledge produced in all the other countries with which their partners trade. They suggest using Leontief’s inverse to capture these “indirect” spillovers (Lumengo-Neso et al., 2005).

That countries may benefit from indirect as well as direct spillovers is an important concern for studies of international R&D spillovers which rely on input–output data (Keller, 2002b; Nishioka and Ripoll, 2012; Foster-McGregor et al., 2017).¹² Whereas Lumengo-Neso et al. Lumengo-Neso et al. (2005) weight their knowledge stocks using matrices constructed from gross import data, this strand of literature uses cross-country input–output tables. These are used to construct matrices of intermediate inputs purchases, which serve to measure the distance separating countries and industries from each other. The assumption here is that closeness between industries is proportional to their purchases from each other—an idea first introduced in work using input–output data to estimate inter-industry spillovers in the US economy (Schmookler, 1966; Terleckyj, 1980; Scherer, 1982; Griliches and Lichtenberg, 1984).

Much of this early work is concerned with understanding the direction of inter-industry R&D spillovers. Scherer’s Scherer (1982); Scherer (1984) technology flow matrix, for instance, relies on the sorting of all available R&D data into industries’ “own” R&D and “used” R&D sourced from other industries. To construct the latter component, other-industry R&D data is redistributed using the proportion of source industry’s patents that are classified, ex ante, as being intended for use in recipient industries. Griliches and Lichtenberg Griliches and Lichtenberg (1984) employ a similar methodology, though their estimation relies on more detailed TFP data. By contrast, Nishioka and Ripoll’s Nishioka and Ripoll (2012) approach is similar to Lumengo-Neso et al. Lumengo-Neso et al. (2005) in that it relies on the Leontief model to trace intermediate sourcing patterns within an international input–output framework.

These strands of work have contributed to establish that international trade is a channel for the international diffusion of knowledge. Yet the role of GVCs—which have revolutionized the way trade and production are organized internationally—remains unexplored in this literature. Our paper addresses this knowledge gap. To do so, we leverage internationally comparable input–output data (Timmer et al., 2015). We also build on a key intuition of older work on inter-industry R&D spillovers—the idea that matrices of input–output relationships provide an ideal weighting scheme to study the nature and evolution of knowledge flows between industries.¹³

2.2 GVCs and knowledge diffusion

Our paper also relates to the literature on GVCs and knowledge diffusion. Literature on international R&D spillovers has primarily focused on bilateral trade. Yet the organization of international trade has changed substantially in recent years. Unleashed by increasingly complex and comprehensive trade and investment agreements, the falling cost of transport, the rise of modularity, and the pervasiveness of information and communication technology (ICT), production activities have increasingly become fragmented across borders (Antras and Yeaple, 2013). As production moved, knowledge also increasingly did so. Management, production standards, and physical technology all started traveling at a faster pace than in previous waves of globalization (Baldwin, 2011).

According to scholars of GVCs, the implications for the diffusion of knowledge are profound. A consensus exists around the idea that exposure to international trade is an important determinant of international knowledge diffusion (Keller, 2004; Cai et al., 2022). Yet literature on GVCs takes a step further, pointing to the highly coordinated nature of trade in GVCs. Trade in GVCs typically involves the exchange of highly customized components, which commit buyers and suppliers to each other (Antràs and Chor, 2022). The result is the development of close relationships between participants to a value chain. The need for coordination and commitment, goes the argument, gives rise to “sticky” relationships (Martin et al., 2021), which, in turn, may be particularly conducive to the diffusion of knowledge.

The role of stickiness has been investigated, in particular, in studies of specific industries including the mobile telecommunications (Alcacer and Oxley, 2014) and the business services (Keijser et al., 2021) industries. Studies using panel data have also found that GVC participation is linked to the productivity and innovation propensity of domestic suppliers, in both developing (Del Prete et al., 2017; Montalbano et al., 2018; Torres Mazzi and Foster-McGregor, 2021) and industrialized economies (Brancati et al., 2017).¹⁴ GVCs have also been found to facilitate knowledge transfer, understood in terms of managerial practices, production standards, and other forms of know-how (Saliola and Zanfei, 2009).¹⁵ These effects may be particularly important for developing economies, where local labor markets and universities provide less of a fertile ground to develop a robust knowledge base (Fu et al., 2011).

With this paper, we complement firm-level evidence.¹⁶ We do so by focusing on an internationally comparative setting which covers over 70% of world output (Timmer et al., 2012). By contrast, many of the findings emerging from the micro-level literature tend to be context-dependent. Moreover, while firm-level literature uses rather indirect proxies of GVC participation—typically defined using a dummy variable reflecting a firm’s status as a two-way trader (Montalbano et al., 2018) or as an exporter of customized products (Torres Mazzi and Foster-McGregor, 2021)—we are able to directly trace countries’ and sector’s involvement in GVCs. Our measures allow us to track countries’ intermediate sourcing patterns, as well as their exports, with all other countries in the world.

3. Methodology

3.1 International R&D spillovers

We are interested in estimating the elasticity of industry-level productivity to the import of embodied knowledge through GVC linkages, controlling for domestic absorptive capacities. Our starting point is the following equation from Coe and Helpman Coe and Helpman (1995), with subscripts i, h, and t indicating indices for countries, industries and time:

where TFP is productivity in country i and industry h, and the α terms indicate fixed effects. The two S terms reflect the amount of industrial R&D produced domestically and abroad that recipient industries are able to access by trading with all other industries. Literature on international R&D spillovers typically measures the S terms by constructing industry-level knowledge stocks, which are then weighted using the volume of bilateral trade. This approach has two drawbacks, however: it lumps together final and intermediate purchases, and it does not take into account the input–output structure of international economic transactions.

To overcome these difficulties, we leverage recent work on the factor content of trade in an input–output context (Trefler and Zhu, 2010; Nishioka and Ripoll, 2012) to construct a matrix of R&D embodied in intermediate and final goods trade at the world level. In this matrix, row sums capture the R&D embodied in intermediates by source country-sector and column sums the amount of embodied R&D that recipient country-sector source from further upstream.¹⁷ We then build on Wang et al.’s Wang et al. (2017) input–output decomposition framework to decompose our global matrix of trade in embodied R&D along several dimensions.

This approach, described in detail in Section 3.2, enables us to measure knowledge flows arising from three distinct sources: domestic sourcing; international sourcing; and finally, integration within GVCs. The novel methodological contribution of our paper stems from our consideration of these different channels for knowledge diffusion within a single analytical framework. In Section 3.3, we discuss how we integrate the three channels within Coe and Helpman’s framework in our estimation strategy.

3.2 Measuring the R&D embodied in international trade

Consider a world economy whose structure can be represented within an input–output framework. Let i and |$j = 1,{\ldots}, N$| indices for countries and g and |$h = 1, {\ldots}, G$| for industries. Assuming that the output of every industry is consumed either in final or intermediate form, total gross output at the world level can be disaggregated into intermediate and final products,

where |$A = ZX^{-1}$|⁠, Z is a matrix of intermediate input flows, X is a NG × NG diagonal matrix with the output vector X in its diagonal, and Y is a is a NG × 1 vector of final goods and service consumption. Based on these definitions, we can define our global matrices as follows:

Rearranging terms in (2) yields the Leontief equation, X = BY, where |$B = (I - A)^{-1}$| is the Leontief inverse matrix for the global economy. Since our world economy is characterized by open economies, our gross output production and use balance (2) can be disaggregated into the domestic and foreign inputs—represented, here, by subscripts i and j—that are employed to cater to both domestic and foreign consumers:

where A_i is a NG × NG diagonal block matrix of domestic input coefficients, |$A_{j} = A - A_{i}$| is a NG × NG off-diagonal block matrix of imported input coefficients, Y is a NG × 1 vector of final goods and services production, Y_i is a NG × 1 vector of final goods and services production for domestic consumption, |$Y_{j} = Y - Y_{i}$| is a NG × 1 vector of final goods and services exports, and E is a NG × 1 vector of gross exports. Rearranging this equation yields:

where L, a NG × NG diagonal block matrix, indicates the local Leontief inverse |$L = (I - A_{i})^{-1}$|⁠, implying the involvement of domestic inputs only in production activities. Before we proceed, we follow Nishioka and Ripoll Nishioka and Ripoll (2012) and construct K, a |$1 \times G$| row vector whose gth element is the industrial R&D stock used directly by industry g in country j. Since we assume that all factors are fully employed, we can construct D, a |$1 \times G$| row vector whose gth element is the R&D stock per unit of good produced by industry g. Then, D is such that it satisfies:

We can now define D to be a |$1 \times NG$| global vector of direct R&D requirements, |$D = [D_{1}, D_{2}, {\ldots}, D_{N}]$|⁠, and K to be a |$1 \times NG$| global vector of domestic R&D stocks, |$K = [K_{1}, K_{2}, {\ldots}, K_{N}]$|⁠, where K = DX. Similar to work on the factor content of trade, we can now define the R&D embodied in intermediate inputs as the total R&D stock that is embodied in AX. The total—direct and indirect—intermediate requirements needed to deliver AX are given, according to Leontief’s insight, by |$(I - A)^{-1}AX$|⁠, or BAX.¹⁸

Transforming row vector D of direct R&D requirements into a NG × NG diagonal matrix, |$\hat{D}$|⁠, replacing X with BY and further converting the three final goods and service production vectors Y, Y_i , and Y_j into NG × NG diagonal matrices |$\hat{Y}$|⁠, |$\hat{Y}_{i}$|⁠, and |$\hat{Y}_{j}$|⁠, we can derive a simultaneous decomposition of R&D production and final goods consumption:¹⁹

Each element in the |$\hat{D}B \hat{Y}$| matrix can be understood as capturing the R&D produced in a country j, and embodied in the intermediate exports from industry g, that is directly or indirectly employed in country i for the production of industry h.

Following Wang et al. Wang et al. (2017), the |$\hat{D}B\hat{Y}$| matrix can be further decomposed into four NG × NG matrices. The first two terms in the final row of (6) capture domestic flows of R&D embodied in intermediates.²⁰ From these two matrices, we construct the following term for the estimation:

Equation (7) represents the sum of embodied R&D that industries in recipient countries source domestically, either from themselves or from other upstream industries. Relating |$S_{i}^{d}$| to industry-level TFP yields the elasticity of domestic industries to the embodied R&D that they source domestically.

Of the latter two terms in equation (6), the |$\hat{D}L\hat{A}_{J}L\hat{Y}_{i}$| matrix involves the R&D embodied in intermediate imports that recipient country-sectors use to produce final goods and services that are eventually consumed within the same country.²¹ The resulting term we employ in our estimation is the following:

Equation (8) captures R&D produced abroad, and embodied in the intermediate inputs which countries use for domestic production and consumption. An example would be R&D spent in coming up with a new computer chip design in the United States, which is then imported and used to manufacture laptops that are produced and consumed within Taiwan. Effectively, to the extent that a positive elasticity is found between this term and domestic TFP, this would reflect a process of learning by importing.

Finally, the |$\hat{D}L\hat{A}_{j}(B\hat{Y} - L\hat{Y}_{i})$| matrix reflects GVC activities, or activities involving production sharing which cross borders more than once. It captures the use of (domestic and foreign) R&D embodied in intermediates that are imported from abroad, processed domestically, and then exported. Matrix |$\hat{D}L\hat{Y}_{j}(B\hat{Y} - L\hat{Y}_{i})$| yields the following term for the estimation:

Equation (9) captures the sum of embodied R&D that industries in domestic countries source upstream and then use to produce for export within GVCs. An example of this type of linkage would be the industrial R&D which is embodied in the car parts that German manufacturers export to Poland to be assembled by the local automotive industry, whence it is subsequently exported for final consumption—either back in Germany, or in a third country.

3.3 Estimation strategy

We start from Equation (1), and adapt it into the following estimating equation:

where the two S terms reflect, respectively, the total sum of embodied R&D that country-sectors source domestically and the total sum of R&D that they source abroad.²² While we expect the elasticity of domestic TFP to be generally positive with relation to both the R&D that is sourced domestically (Griliches, 1998a) and to that which is sourced abroad (Keller, 2002b), we are interested in the relative magnitude of the two elasticities.

As is customary in literature on international R&D spillovers, we also include A, a vector of control variables reflecting absorptive capacities. At the country level, we proxy these with the level of secondary school attainment from the Barro and Lee database. To capture absorptive capacities at the industry level, we include sectors’ R&D stocks. We also include fixed effects to control for time-invariant country- and industry-specific characteristics, including the institutional environment, the size of the market, and the level of technological sophistication, as well as for time-variant shocks at the global level or at the level of specific industries.

We then decompose S^f into its two main components, as discussed in Section 3.2. The first-term (8) in the decomposition presented above—reflects the sum of embodied R&D that country-sectors source abroad and subsequently used to cater to domestic consumers. The second-term (9) in Section 3.2—captures the sum of embodied R&D which is sourced abroad to be used to produce for export. This is embodied R&D which crosses borders at least twice, and which is therefore inherently part of a coordinated global production chain.

We plug these two terms in lieu of S^f, resulting in the following estimating equation:

That imports are associated with a learning premium is another well-established finding (Coe and Helpman, 1995; Keller, 2002a; Nishioka and Ripoll, 2012). This is why we are mainly interested in coefficient γ₂, which represents the elasticity of domestic TFP to the import of embodied R&D through GVC linkages. We are particularly interested in studying whether intermediate purchases which occur in a context of GVCs offer a greater productivity premium relative to those purchases which are used, say, to produce goods for the domestic market. To the best of our knowledge, our paper is the first attempt at estimating and comparing these different learning channels within a unified framework.

4. Construction of the database

Our data is constructed from the OECD ANBERD database and from WIOD input–output tables and socio-economic accounts. In what follows, we start by describing the construction of our measure of TFP and of industry-level knowledge stocks K_j. We then describe the construction of the S terms described in Section 3.2.

To measure TFP at the industry level, we follows Levinsohn and Petrin’s Levinsohn and Petrin (2003) approach. We depart from the assumption that technology is Cobb-Douglas, with an equation of the following form:

where y_iht is the log of output in country i and industry h at time t, l_iht and k_iht are the labor and capital inputs, and m_iht reflects intermediate inputs. Levinsohn and Petrin Levinsohn and Petrin (2003) suggest using data on intermediate inputs to account for sector specific productivity shocks, which would push firms to endogenously scale up (or cut down) inputs use. We use data on labor compensation, physical capital, and intermediate input purchases from the WIOD SEA (socio-economic accounts) database; data on output comes from WIOD’s input–output tables. We use industry-level deflators from WIOD.

WIOD is also the key source we use to construct our measures capturing the R&D that industries in different countries source through domestic and international linkages. To construct these measures, we employ the 2016 release of WIOD, which covers 43 countries and 56 industrial sectors (Timmer et al., 2015). We remove eleven countries because we do not have reliable R&D data for them, and aggregate a number of industries to ensure concordance with ANBERD data (see Table A1 in the Annex for details on the aggregations).

WIOD data is based on the collection, harmonization, and standardization of national supply and use tables (SUTs). These tables report information on the supply and use of different products in 56 industries, together with information on final use by product, and value-added and gross output by industry.²³ The construction of WIOD also relies on trade data to differentiate between domestic and imported inputs in the use tables, with the latter being further split up by source country. Imports are further differentiated by use category—that is, intermediates, consumption, and investment goods. The SUTs on which WIOD is built also include data on services trade. The resulting set of SUTs are then transformed in international input–output table (Timmer et al., 2015).

To construct our measures of the R&D embodied in intermediates which are sourced through different domestic and international channels, we start by extracting global matrices X, A, as well as the final demand vectors Y, Yⁱ, and Y^j described in Section 3.2, for each year in the 2000-2014 period. Using the R&D stock data calculated as described below, we then construct vector D in line with equation (5). Based on these matrices, we subsequently construct our four S variables according to equations (7)–(9).

Finally, to construct industry-level knowledge stocks for each country, we employ business enterprise R&D (BERD) data from the OECD ANBERD database. The domestic R&D stock K_jgt for country j, sector g at time t is computed using the perpetual inventory method,

where δ is the depreciation rate to account for the obsolescence of knowledge, and R_jgt is real business R&D investment. We use δ = 0.10. The initial value of the R&D stock is calculated according to the following equation,

where π_jg is the average growth rate of real business R&D investment for industry g in country j over the whole period for which data are available (2000–2014). After imputing a small number of values, and aggregating two industries to ensure concordance with WIOD data (see Table A3 for details on these aggregations), we have data for 32 economies and 39 industries—including manufacturing, services, and mining industries—for the period 2000 to 2014.²⁴

5. Results and discussion

5.1 Descriptive evidence

Table 1 shows the distribution of the total R&D stock across industrial sectors in 2000 and 2014, alongside data on TFP growth. Knowledge capital is quite concentrated in our sample. In manufacturing, the computer, transport, pharmaceutical, machinery, and chemical industries account, together, for over 52% of the total R&D stock, with limited change in their share over time. Among services, the group of KIBS (corresponding to ISIC Rev. 4, Division 69-75), which includes legal, management consulting, and R&D activities, also accounts for a large share of the R&D stock in our sample—over 10% of the total R&D stock. Other knowledge-intensive service industries include programming and other ICT services, and the publishing sector.

Table 2 shows the distribution of total business R&D stock across countries in our data, again for 2000 and 2014, alongside data on TFP growth over the course of the 2000–2014 period. The distribution of knowledge capital across countries is even more concentrated than that across industrial sectors. Unsurprisingly, G7 countries account for the lion’s share in both 2000 and 2014. The United States emerges as the clear leader, accounting for over 30% of the total knowledge stock in 2014—almost the same share it had at the start of the period. Followers include Japan, with about 18% (down from 25% in 2000), and China, which jumped from accounting for under 4% of the total R&D stock to over 12%. European countries, such as Germany, Great Britain, and France, follow at below 10% of the total, and so does Canada.

Figures 1 and 2 show the distribution of our S measures across a selection of industries, in 2000 and 2014. We focus on manufacturing industries and KIBS, because these sectors account for a large share of the stock of R&D in our panel—approximately 78% of the total (Table 1). The figures show each sourcing channel’s share of total embodied R&D that industries sourced domestically and internationally, distributed across a selection of industries. They provide an indication of the relative magnitude of these three inter-industry knowledge flows, but also of the relative reliance of different industries on R&D sourced through domestic, international, and GVC linkages.

The figures suggest a number of observations. First, while industries which are large R&D spenders tend to also source embodied knowledge from other industries, this is not always the case. Large R&D spenders, such as the computer and transport industries are, for instance, on the receiving end of the largest shares of embodied R&D in our panel. By contrast, the pharmaceutical and chemical sectors—which together account for over 15% of the R&D stock in our sample in 2014—appear to rely relatively less on externally sourced R&D. KIBS, despite accounting for a large share of R&D, also do not seem to rely on embodied knowledge. Presumably, these differences are due to differences in tradability. The computer and transport industries are highly traded, more so than, say, business services. These differences may also be due to the nature of knowledge production in different industries. Pharmaceutical firms may not be as strongly reliant on intermediate inputs, as innovation in pharmaceutical and chemical products tends to rely more on disembodied, scientific advances.

Another observation concerns the relative magnitudes of the three sources of embodied R&D we consider. Industries in our panel predominantly rely on domestically sourced R&D, which is used to produce for domestic and foreign consumers. The main difference between 2014 and 2000 is that the international sourcing of knowledge, particularly through GVCs, increases its share in total incoming R&D across all industries, suggesting that GVCs have grown substantially between the two points in time. Yet it is interesting to note that it remains noticeably lower than domestic sources across the board. Some industries, however, are more exposed to GVC linkages than others. These are again, the Computer and Transport industries, followed by the Machinery and equipment sector. These industries account for a larger share of embodied R&D sourced through international channels relative to other sectors, and are also more integrated within GVCs than other industries.

Figures 3 and 4 show the distribution of incoming R&D embodied in trade sourced through different channels, again in 2000 and 2014, but this time across countries. The preponderance of domestic knowledge sourcing in our panel is partly explained by the role of large countries, such as the United States, Japan and—particularly in recent years—China. This reflects the fact that international linkages are less important in larger countries, where industries tend to source relatively more inputs domestically than they do abroad. Differences between 2000 and 2014 are less marked when one compares the distribution across countries. Indeed, the relative importance of domestic sources of knowledge in our panel does not appear to decrease over time, although the shares of embodied R&D that economies purchase abroad—either via arms’ length or GVC relationships—does show an increase between 2000 and 2014.

There are also differences among the medium-sized economies in our sample. Generally, larger R&D spenders in this category tend to be more similar to their large economy counterparts. Countries, such as Germany, France, and the UK do source more knowledge inputs domestically than they do internationally, and so do smaller, open countries, such as the Netherlands. By contrast, for smaller R&D spenders among medium- and small-sized countries, such as the Czech Republic, Romania, and Turkey the international sourcing of knowledge inputs is as, if not more, important than domestic sources. Plausibly, this reflects the under-development of the domestic scientific and technological bases of emerging economies, which are therefore more likely to rely on embodied knowledge from abroad. In the analysis below, we explore some of these differences in greater detail.

5.2 Econometric results

Table 3 reports results from estimating equations (A) and (B) on our data. Columns (1) and (3) report results from estimating these equations without sector-year fixed effects. Columns (2) and (4), our preferred specifications, include sector-year fixed effects.

We find that domestic productivity is generally elastic to the sourcing of embodied R&D. Domestic TFP in recipient industries is elastic to both domestic and foreign sourcing, though the coefficient we estimate for the foreign sourcing component is over three times larger than that for domestic sourcing (see Column (1)). The coefficient we estimate for S^d falls within the range of values that is common for these regressions (Griliches, 1998a; Keller, 2002b). When including industry-year fixed effects, the estimated elasticity of domestic TFP to the sourcing of embodied knowledge from abroad is 0.17. Our estimated elasticity is approximately half that reported in Keller Keller (2002b) but approximately four times larger than that reported by Nishioka and Ripoll Nishioka and Ripoll (2012)—two of the more readily comparable studies in the literature.

In Columns (3) and (4), we disaggregate S^f into its two components—one which captures imports directed at domestic consumption, and the other which belongs to GVCs. We find that domestic TFP is elastic to both these components, with coefficients on both terms being statistically significant at the 1% level. Yet the estimated elasticity of domestic productivity to the import of embodied knowledge through GVC linkages is larger—by almost an order of magnitude—than the elasticity to the import of embodied R&D through non-GVC trade.²⁵ In our preferred specification (Column (4)), we estimate an elasticity of 0.16 to the import of embodied knowledge via GVCs. Thus, a 10% increase in knowledge sourcing via GVC relationship is associated with a 1.6% boost to domestic productivity.

While these effects appear to be relatively small in magnitude, they fall approximately in the middle of the distribution of elasticities reported in the literature (Keller, 2002b; Nishioka and Ripoll, 2012; Foster-McGregor et al., 2017).²⁶ Piermartini and Rubínová Piermartini and Rubínová (2021)—the only other study to focus on GVC spillovers specifically—study the elasticity of domestic patenting to countries’ exposure to a pool of foreign R&D weighted by the value of GVC trade. They estimate an average elasticity of 0.01. However, their focus on patent applications and use of a weighting scheme—in contrast to our reliance on measures based on the “content of trade” literature—makes it difficult to directly compare effect sizes.

Overall, our findings indicate that domestic productivity is positively and significantly elastic to the import of embodied R&D through international trade, and that this effect is primarily driven by the knowledge stock embodied in intermediates that countries import as part of cross-country GVC activities. These results complement the findings of Piermartini and Rubínová Piermartini and Rubínová (2021) and Tajoli and Felice Tajoli and Felice (2018) on the role of GVCs in domestic innovation, and provide additional support to micro-level literature on the relationship between GVC participation, cross-country knowledge diffusion, and productivity growth in recipient countries (Saliola and Zanfei, 2009; Alcacer and Oxley, 2014; Montalbano et al., 2018).

Indeed, our results suggest that among the various sourcing channels that are available to industries in different countries, participation in GVCs—entailing the import of intermediate inputs embodying foreign R&D for the purpose of domestic processing and further exporting—is the single most important driver of domestic productivity. This is particularly interesting in light of the descriptive data we report in Section 5.1. As a percentage of total R&D that industries source domestically and abroad, the R&D that is imported, in embodied form, through GVC linkages is at least an order of magnitude smaller relative to the other channels—and particularly domestically sourced R&D. Yet its effects on domestic productivity tend to trump all other sourcing channels.

5.3 Heterogeneity in R&D spillovers from GVC participation

5.3.1 The role of absorptive capacities: do GVCs help bridge the technology gap?

We now ask whether the elasticities we estimate are heterogeneous across industries and countries in our panel. We are particularly interested in testing whether there are differences between country-sectors endowed with higher absorptive capacities relative to their peers with lower capabilities. An oft-repeated claim is that exposure to GVC trade can help firms in developing and emerging economies bridge the technology gap vis-‘a-vis industrialized countries, by providing access to critically needed inputs and technologies (Lema et al., 2019; World Bank, 2020). To test this idea, we subsequently interact, within equation (B), terms S^import and S^gvc with R&D stocks at the country-sector level.

By introducing within our estimating equation an interaction term between the variables capturing the amount of embodied R&D sourced through different international trading channels and R&D stocks, we assess whether the relationship between GVC trade and domestic productivity changes at different levels of technological prowess. Proxying for technological capabilities using the accumulated stock of R&D is relatively common in the literature, with work in innovation studies (Criscuolo and Narula, 2008; Castellani et al., 2019) as well as in literature on international trade and knowledge spillovers more specifically (Foster-McGregor et al., 2017; Lee, 2020).

The effect of having increased access to foreign sources of knowledge for countries that sit far away from the technology frontier is theoretically ambiguous. On the one hand, in a world where technological “backwardness” does yield some advantages, with imitation of the frontier’s products and technology helping to reduce the costs of domestic innovation, one would expect less technologically advanced countries to experience a premium from their engagement in trade. Should participation in GVCs be indeed more important at lower levels of technological development, we would expect the interaction between absorptive capacities and the import of foreign knowledge to be negative. Within our framework, this would mean that the elasticity of domestic TFP to the import of embodied R&D from abroad is higher at lower levels of technological capability.

Yet on the other hand, endogenous growth theory suggests that technologically prowess comes with its own, more permanent advantages—and specifically the idea that the more one knows, the more one can learn. In this world, exposure to GVCs should benefit countries that are further up the capability ladder rather than emerging economies. We would then expect a positive elasticity of domestic productivity to GVC exposure, suggesting that the effect is higher at higher levels of development—proxied, here, by country-sectors’ R&D stocks. Which of these two forces is likely to dominate at any given point in time is an empirical question.²⁷ Moreover, we would also generally expect the magnitude of the coefficient of the interaction between our GVC term and R&D stocks to be larger than that between R&D stocks and non-GVC imports, as we would expect the “GVC premium” we reported in Table 3 to remain. Table 4 below reports results from introducing the two interaction terms.

Signs on our interaction terms are negative, although only the interaction between R&D stocks and GVC participation is statistically significant (albeit only at the 5% level). Our findings suggest that there might indeed be a small learning premium connected with international trade and, more specifically, GVC participation for relatively less technologically advanced economies. This premium is small, however. On average, a 10% increase in a country-sector’s R&D stock is associated with a decrease in the elasticity of its domestic productivity to the import of R&D through GVC linkages of approximately 0.04%.

To put these figures in context, the elasticity we estimate in Column (4) of Table 4 for the Czech wood and wood products industry—which sits toward the bottom of the distribution of R&D stocks in our panel—is approximately 0.17, whereas that of the US motor vehicles sector—which is among the high R&D-intensity industries—is approximately 0.13. While this is an economically small difference, making it extremely unlikely that countries’ participation in GVCs may, in and by itself, be sufficient to achieve technological catch-up, these results do suggest a small role for GVCs as conduits of knowledge diffusion from more toward less technologically advanced economies.

5.3.2 GVC participation and geographical proximity

We now ask whether the magnitude of knowledge spillovers arising from GVC participation is sensitive to geographical proximity. An established finding in the literature on knowledge spillovers is that these tend to be local, rather than global; and that they decay with distance (Jaffe et al., 1993; Keller, 2002a).²⁸ Yet an argument could be made that, by virtue of its “global” nature, trade in GVCs has reduced the importance of geographical proximity for the transfer of knowledge.²⁹

There are at least two versions of this argument. The first revolves around the content of intermediate inputs. In a world characterized by GVCs, firms in any given location can gain access to different “bits” of foreign knowledge embodied in inputs. Electronics producers in China may be purchasing French or Dutch knowledge through their purchase of Korean intermediates, for instance. A second version of the argument concerns multinational investment. By transferring knowledge embodied in machinery or people abroad, MNCs are likely to be making the importance of distance disappear. A Volkswagen factory in Brazil will be sourcing the same inputs as a factory based in Slovakia, following the same production routines, with the same managerial know-how: learning and productivity increases might take place in equal measures—regardless of geographical distance.

To test this idea, we run our estimations using modified versions of the S terms in equation (B). To do so, we build a global matrix of bilateral distances, P, which we construct starting from the CEPII Gravity database. In line with the literature, we calculate bilateral distances as |$exp(-Dist_{ij})$| where Dist reflects the distance—in kilometers—between two countries’ capitals.³⁰ We then insert our matrix of bilateral distances P within equation (6)—which yields a distance-weighted global matrix, |$\hat{D}PB\hat{Y}$|—and proceed with the decomposition described in Section 3.2. Table 5 reports results from estimating our equation by taking geographical distance into account. Naturally, S^d is not affected by this transformation; only the terms capturing foreign knowledge are.

Our findings suggest that, as argued by literature on knowledge spillovers more generally, geographical distance between trading partners does dampen international R&D spillovers—including those arising from GVCs. The estimated elasticity of domestic TFP to the import of embodied R&D via GVCs falls from 0.16 to 0.06 when taking proximity into account. At the same time however, the elasticity remains positive and statistically significant—contrary to that we estimate for the sourcing of R&D via arms-length trade. These results suggest that GVCs continue playing an important role in the cross-border diffusion of knowledge, even when geographical distance is explicitly taken into account.

5.3.3 Do R&D spillovers differ across sectors?

In this section, we ask whether our results are driven by recipient industries in a single sector—manufacturing. There is consensus that GVCs have transformed the nature of manufacturing production to a far larger extent than in other economic activities (Gereffi et al., 2005; Antràs and Chor, 2013). Historically, the revolutions in modularity and transport—which, together with the ICT revolution, unleashed the international fragmentation of production—primarily affected trade in manufacturing goods (Baldwin, 2011). Moreover, as we show in Table 1, manufacturing industries tend to be more knowledge-intensive—a point long recognized in studies of structural change (Foellmi and Zweimüller, 2008; Cantore et al., 2017).³¹ Inputs in manufacturing are also more likely to be traded than in agriculture, mining, or services (see Figures 1 and 2). Manufacturing industries may thus be more likely to source knowledge-intensive inputs, and to increase their productivity as a result.

Given the preeminent role of manufacturing in the organization of GVCs and in knowledge production globally, it is possible that our results are primarily driven by manufacturing. At the same time, however, there are arguments concerning the growing tradeability of services—particularly KIBS (Savona, 2021). Trade in services has been described as the next “unbundling” of globalization (Baldwin, 2019).³² The case for agriculture is complicated by the high level of protection afforded to agricultural activities, especially by industrialized economies (Boestel et al., 2002). Recent evidence, however, points to an increase of GVC participation of the agri-food sectors with rising trade in both intermediate and final goods (Nenci et al., 2022).

Against this backdrop, in Table 6, we estimate equation (B) on three sub-samples—agriculture and mining; manufacturing; and services. We focus on the elasticity of domestic productivity in these three broad recipient sectors to the inflow of knowledge sourced through GVCs and other channels. Results for manufacturing are hardly surprising. Domestic productivity in manufacturing is highly elastic to knowledge flows arising from GVCs, the magnitude of which is higher than for R&D sourced domestically or via non-GVC imports. Results for services are particularly interesting: trading in complex value chains is an important driver of productivity in services, with an estimated elasticity of close to 0.2—a result which is statistically significant at the 1% level. Productivity in agriculture is also elastic to knowledge sourced through trade in trade in GVCs, but the coefficient is far less precisely estimated.

Overall, we find evidence that knowledge flows arising from trade in complex and coordinated networks of exchange are linked to a productivity premium across broad economic sectors. Productivity in manufacturing industries is not alone in its elasticity to trade in GVCs. Services and, to a lesser extent, agriculture appear to gain a productivity premium from their participation in GVCs. In light of the descriptive evidence we report in Figures 1 and 2, the results we observe across services industries are likely driven by the KIBS sub-sector—the focus of recent work on the role of services in globalization (Baldwin, 2019; Baldwin, 2022).

5.3.4 Do R&D spillovers differ across time

Next, we look at differences in spillover size across time. GVCs are a dynamic phenomenon, and one can expect the nature of international trade spillovers to change not just across sectors (Foellmi and Zweimüller, 2008) and across space (Jaffe et al., 1993), but also across time. Table 7 reports results from estimating equation (B) on two sub-samples—from 2000 to 2007 and from 2007 to 2014. Our results suggest that overall, the magnitude of domestic and trade-related R&D spillovers decreases over time. This is particularly the case for the elasticity of domestic productivity to the import of embodied knowledge through GVCs, which is almost a third lower in the first than in the second period. This finding is likely to reflect the long-lasting impact of the Great Recession on incomes, aggregate demand, and international trade flows.

5.4 Robustness checks

5.4.1 Addressing reverse causality

Our analysis so far suggests that knowledge flows arising from GVC participation play an important role in raising the productivity of domestic industries; that these results are driven primarily by GVC integration in the manufacturing sector and by KIBS; and that spillovers are surprisingly robust across space. An important concern, however, is reverse causality. The concern is that the results may be not be driven by industries becoming more productive as a result of being exposed to knowledge flows in GVCs, but rather by a productivity effect—such as a positive shock to productivity in a given industry. This might increase international trade and, consequently, the level of R&D traded internationally, from which industries benefit.

There may therefore be a two-way relationship between TFP and our S variables. To rule out this possibility, we estimate equations (A) and (B) with lagged versions of our measures of the R&D embodied in trade. Should our results be robust to the introduction of lags, we could be confident in our interpretation that global knowledge flows drive productivity improvements in recipient economies—rather than the other way around. Table 8 reports results with our lagged variables, at t − 1. Results are also robust to the introduction of longer lags, which we do not report here.³³ The magnitudes of the effects we find remain practically unchanged—and so does their direction.

The estimates presented in Table 8 suggest that we can rule out the possibility of our results being driven by reverse causality in the relationship between domestic industry-level TFP and knowledge flows arising from international trade in GVCs.

5.4.2 Are spillovers trade-related?

An important concern is that the elasticities we estimate might be statistical artifacts. Keller Keller (1998) observed that the use of random weights—later shown to capture, in fact, the average of a country’s trading partners’ domestic R&D capital stocks (Coe and Hoffmaister, 1999)—to construct spillover indicators performed just as well in predicting countries’ TFP as did bilateral import weights taken from actual trade data, as used initially by Coe and Helpman Coe and Helpman (1995). Coe and Hoffmaister Coe and Hoffmaister (1999) later showed that when alternative sets of random weights are used to measure R&D spillovers, the estimated elasticities are, as one would expect from the use of truly random weights, extremely small.

We revisit this debate by randomizing the elements of global matrices A and Y in equation 3 before proceeding with the decomposition described in Section 3.2. These two matrices reflect, respectively, global input coefficients and global demand for products and services. We set each element as a random number falling within the (0, 1) interval. By randomizing the elements of these two global matrices for each year in the 2000–2014 period and then taking, for each country-sector in our sample, the column sums of the resulting global and domestic matrices, we are weighing country-sector’s exposure to foreign R&D in a way which is fully random, in the spirit of Coe and Hoffmaister Coe and Hoffmaister (1999).

Table 9 reports results from estimating equations (A) and (B) using randomized S terms. As expected, the elasticities we estimate are substantially smaller in size, and generally negative, as compared to those estimated using terms extracted from the “true” A and Y matrices. Coefficients for our variables proxying for absorptive capacities remain similar to those estimated above. These results strongly suggest that the elasticities we estimate are not random, as elasticities estimated using randomly generated trade patterns do not come close—in terms of sizes and signs—to resembling elasticities estimated using actual trade patterns.

5.4.3 Are results driven by big players?

A final concern that is important to address relates to the role of major countries in our sample. As we show in Table 2, R&D spending in our sample tends to be concentrated among three large countries—the United States, Japan, and, increasingly, China. To rule out this concern, in Table 10 below, we re-estimate equation (B) on a sub-sample of countries which excludes the United States, Japan, and China. We find that the elasticity of domestic TFP to knowledge flows arising from GVCs remains unchanged. This finding provides evidence that our results are not driven by the main world investors in R&D. Our estimates of the role of domestic and non-GVC knowledge flows, however, are a bit smaller in magnitude—reflecting the large role that the domestic economies of large countries play in fostering their productivity performance.

6. Concluding remarks

There is growing discussion into the implications of the fragmentation of production for development outcomes, including productivity growth (Pahl and Timmer, 2020), employment (Rodrik, 2018a; Bontadini et al., 2022), and industrial catch up (Fagerberg et al., 2018; Gereffi, 2019). An important aspect of this debate concerns the potential role of GVCs as conduits for the international diffusion of knowledge (Taglioni and Winkler, 2014; Lema et al., 2019; World Bank, 2020). While firm-level evidence points to productivity and, to a lesser extent, innovation gains from engaging in GVCs (Saliola and Zanfei, 2009; Montalbano et al., 2018), something of a consensus exists that these benefits tend to be contextual, and largely conditional on pre-existing capabilities and institutional settings (Pietrobelli and Rabellotti, 2011).

With this chapter, we contribute to this debate by estimating, for a panel of advanced and emerging economies over the 2000–2014 period, knowledge spillovers arising from GVCs. To do this, we build on a large strand of literature concerned with the estimation of inter-industry spillovers in an input–output setting, both within single economies (Schmookler, 1966; Terleckyj, 1980; Scherer, 1982; Griliches and Lichtenberg, 1984) and across countries (Keller, 2002b; Nishioka and Ripoll, 2012; Foster-McGregor et al., 2017). We revisit this literature, however, by focusing on the role of GVCs specifically. To do so, we combine insights from literature on the factor content of trade (Trefler and Zhu, 2010; Nishioka and Ripoll, 2012) with recent advances in input–output decomposition techniques (Wang et al., 2017).

Our findings suggest that while domestic industry-level TFP is generally elastic to the R&D embodied in inputs sourced both domestically and abroad, there does seem to be a premium associated with GVC participation. While the descriptive evidence we provide shows that GVCs still account for a relatively small share of the R&D embodied in trade, we find that international spillovers from GVC participation are substantially larger than spillovers arising from other channels. In our benchmark estimation, we estimate an elasticity of domestic TFP to the embodied R&D imported via GVCs of 0.16. This value is approximately double the estimated elasticity to R&D sourced domestically, and an order of magnitude larger than the elasticity we estimate for embodied knowledge sourced via non-GVC international trade. This suggests that while sourcing knowledge through GVCs remains a quantitatively small phenomenon relative to knowledge sourcing through domestic channels, it is qualitatively very important.

The spillovers we estimate appear to exist across all sectors of an economy, and to be robust to the inclusion of a weighting scheme based on geographical proximity. We find that while geography dampens the magnitude of GVC-related R&D spillovers, it does not cancel them out altogether. This stands in contrast with non-GVC, international R&D spillovers, which do decay—completely—with distance. These findings suggest that the coordinated nature of exchange which underpins GVCs may have reduced the importance of geography for the international transfer of knowledge. This is not necessarily surprising, as MNCs tend to offshore technology, standards, and know-how alongside with production (Baldwin, 2011).

Our paper also has a number of limitations. We do not, for instance, explore the role of an industry’s positioning—defined in terms of its distance from final demand—within GVCs (Antràs and Chor, 2022). Recent research points to the complex effects of positioning on wages (Mahy et al., 2022) and value added (Montalbano and Nenci, 2022). In the case of knowledge flows, industries which are relatively downstream, that is, relatively close to final consumers, may be on the receiving end of a larger portion of knowledge relative to their more upstream counterparts. By the same token, industries located upstream may be more likely to be a source of knowledge flows: the extent to which an industry sells intermediate inputs to other sectors in the world may then be indication of its role in global knowledge flows. We believe that this would be a fruitful avenue for further research on the role of GVCs in global knowledge flows.

Moreover, the policy implications of our results are not clear-cut. On the one hand, the evidence we present here suggests that countries—particularly emerging economies—may benefit from entering trade agreements (Mattoo et al., 2020). Regional and preferential trade agreements in particular have been found to facilitate GVC integration, especially through those provisions which apply behind the border and pave the way for a greater degree of regulatory harmonization between countries (Delera and Foster-McGregor, 2020; Baccini et al., 2021). Insofar as they grant access to new varieties of inputs embodying foreign technology, our paper suggests that trade agreements are an important policy tool to stimulate knowledge diffusion across borders and raise productivity levels.³⁴

Trade agreements, however, can only go so far. Whether the “vertical” spillovers we identify give way to “horizontal” ones—benefiting local firms—remains something of an open question. The existing evidence remains less than sanguine (Rojec and Knell, 2018). These questions are particularly relevant for our finding that GVC participation appears to be particularly beneficial for industries and countries which sit somewhat far away from the technology frontier. To some extent, this notion is borne out by the staggering economic success of countries, such as China or Poland, which, despite being very far apart geographically and institutionally, share a history of growing integration within manufacturing GVCs over the past three decades (Piatkowski, 2018). Yet at the same time, these very success stories point to the role of policy not only in attracting foreign investment in strategic industries, but also in strengthening national innovation systems and thereby facilitating the technological upgrading of domestic suppliers (Pietrobelli et al., 2021).

Indeed, while our paper points to the productivity-enhancing role of intermediate sourcing within GVCs, important questions remain. Whether the knowledge flows arising from the cross-border fragmentation of production will contribute to convergence or, rather, reinforce existing asymmetries in the global economy (Bontadini et al., 2022), remains to be seen. Existing evidence, however, invites a degree of skepticism on the role of global integration in contexts characterized by weak domestic science and technology capabilities (Castellacci, 2011).

Filling these gaps would require strengthening research efforts aimed at tracing the mediating role of industrial policy and science and technology institutions in the context of GVC trade; understanding the use of specific provisions in trade agreements, such as provisions aimed at the promotion of foreign investment; and studying the experiences of countries using policies aimed at retaining, and not simply attracting, investment. Helping domestic firms grow and capture an increasingly large share of value-added activities within GVCs will likely remain one of the key policy challenges of the 21st century.

Footnotes

References

Appendix 1 Annexes

1.1. Annex I: Data construction

1.2. Annex II: Additional descriptive statistics

Author notes