Difficulty: Intermediate | Prerequisites: Part 1 of these study notes (framework and measurement).
Part 2 covers the evidence on the two main channels of international technology diffusion examined in Keller (2004): international trade (both imports and exports) and foreign direct investment (FDI). The central questions are whether trade and FDI cause technology to flow across borders, and how strong the evidence is for each channel.
Imports appear to be a significant channel of technology diffusion, particularly trade in differentiated capital goods, though the exact quantitative importance is still debated. Evidence for learning-by-exporting is weak: exporters are more productive, but this seems to reflect self-selection rather than learning. FDI spillovers exist but are concentrated in high-technology sectors, and recent micro-level studies suggest they are larger than earlier pessimistic surveys concluded.
Embodied technology diffusion
Technology transfer that occurs through the use of imported intermediate or capital goods. The technology is "embedded" in the physical product; the importing country uses it but does not necessarily gain full mastery of the underlying knowledge.
Think of it as buying a foreign machine and getting its productivity, without learning how to build it yourself.
Disembodied technology diffusion (R&D spillovers)
Technology transfer not tied to any physical product. Occurs when researchers in one country benefit from knowledge created elsewhere, through publications, conferences, reverse engineering, or personal interaction. A stronger form of diffusion than embodied transfer.
In simple terms, this is learning the idea itself, not just using the product it created.
Learning-by-exporting
The hypothesis that firms learn about foreign technology through the experience of selling to foreign customers, who impose higher quality standards and provide information on how to meet them.
Think of it as the idea that exporting to demanding foreign markets forces you to get better.
Self-selection into exporting
The alternative explanation for the observed productivity premium of exporters: firms that are already more productive choose to enter export markets, rather than exporting making them more productive.
Horizontal FDI spillovers
Productivity effects on domestic firms operating in the same industry as foreign-owned subsidiaries. These are the effects most commonly studied in the FDI spillover literature.
Vertical FDI spillovers
Productivity effects on domestic firms in upstream or downstream industries relative to foreign-owned subsidiaries. MNEs may transfer technology to local suppliers (backward linkages) or to local customers (forward linkages). Evidence from Kugler (2002) and Blalock and Gertler (2002) suggests these may be quantitatively important.
Technology sourcing versus technology transfer
MNE subsidiaries can either disseminate technology to domestic firms in the host country (inward FDI technology transfer) or pick up new technologies from host-country firms (outward FDI technology sourcing). Patent citation studies suggest the sourcing effect may be at least as strong as the transfer effect.
Eaton and Kortum (2002, 2001) combine their technology diffusion model with a Ricardian trade model. Trade gives countries access to foreign goods (and implicitly, the technology embedded in them). With transport costs increasing in distance, equipment goods prices are higher in remote countries, and productivity is correspondingly lower. Differences in the relative price of equipment account for about 25% of cross-country productivity differences in a sample of 34 countries.
A caveat: the model predicts that rich countries have lower equipment prices than poor countries, which is the opposite of what Summers and Heston's International Comparison Program data shows. This inconsistency is important enough to put a question mark behind the model's quantitative conclusions about trade as a diffusion channel.
Coe and Helpman (1995) estimate a positive and large effect of import-weighted foreign R&D on TFP for 22 OECD countries, and similar effects for 77 less developed countries (Coe, Helpman, and Hoffmaister 1997).
Keller (1998) challenged this by showing that randomly generated import shares produce similar regression results. This implies that import shares are not essential for the Coe-Helpman findings, so the results may reflect a common trend in R&D and TFP rather than a trade-specific mechanism.
Capital goods trade: Xu and Wang (1999) show that using differentiated capital goods imports (rather than overall trade) to weight foreign R&D accounts for about 10% more variation in productivity and outperforms Keller's counterfactual. This matters because the theory specifically predicts diffusion through differentiated intermediate and capital goods.
Indirect R&D spillovers: Lumenga-Neso, Olarreaga, and Schiff (2001) account for indirect spillovers (country A benefits from country C's R&D through country B, which imports from C and exports to A). Their specification outperforms both Coe-Helpman and Keller's models.
Patent citation evidence: Sjoholm (1996) finds a positive correlation between Swedish patent citations of foreign patents and bilateral imports, consistent with imports contributing to knowledge spillovers.
The evidence points to a significant role for imports in technology diffusion, but the various strands of literature leave questions open. We do not yet have a firm estimate of the quantitative importance of imports, nor do we know how much diffusion occurs through the embodied technology in traded goods versus other mechanisms associated with importing.
Do firms learn about foreign technology through exporting experience? Anecdotal and case-study evidence (particularly from East Asian export success stories in the 1960s onwards) supports the idea. The econometric evidence is considerably weaker.
Exporters are on average more productive than non-exporters. This is well established across many countries and datasets (Bernard and Jensen 1999, Clerides, Lach, and Tybout 1998). The question is whether this reflects learning from exporting (causation) or self-selection of already-productive firms into export markets.
Studied manufacturing plants in Colombia, Morocco, and Mexico during the 1980s. Estimated a dynamic discrete choice model for export participation alongside an autoregressive cost function that allows past exporting experience to reduce current costs. Results: generally no significant effects from past exporting on current performance. Where significant, the estimates go in the wrong direction (exporting raises costs). The authors' descriptive evidence (plots of average cost before and after export market entry) supports the null result.
Studied U.S. firms. Found that labour productivity growth is about 0.8 percentage points higher among exporters than non-exporters, but this estimate is small, shrinks over longer time horizons, and is likely downward-biased because it conditions on plant survival. Exporters are about 10 percentage points more likely to survive than non-exporters of the same size, which suggests the true productivity growth difference may be larger.
The conventional wisdom is that learning-by-exporting effects are non-existent or very weak. Keller notes this is consistent with the evidence but flags several unresolved issues:
The econometric evidence is at odds with the case-study literature, which is puzzling.
There may be heterogeneity across industries, especially in high-technology sectors.
The analysis could be improved with richer data on export destinations, the characteristics of foreign buyers, and whether exporting firms are engaged in R&D.
Firm-specific technology is transferred across borders when multinational parents share knowledge with their foreign subsidiaries (Markusen 2002). MNEs might generate technological externalities for domestic firms through labour training and turnover (Fosfuri, Motta, and Ronde 2001) or through provision of high-quality intermediate inputs (Rodriguez-Clare 1996). Governments worldwide spend heavily to attract MNE subsidiaries: Alabama spent $150,000 per job to attract a Mercedes-Benz plant in 1994.
Larrain, Lopez-Calva, and Rodriguez-Clare (2000) studied Intel's FDI into Costa Rica in the 1990s, interviewing top management and government officials. Two important consequences: (1) Intel funded schools teaching general (not Intel-specific) skills to local workers. (2) Intel's presence served as a signal to other potential foreign investors. This illustrates how FDI by a major high-technology company can trigger widespread changes in a small country, though whether these are spillovers in the strict sense is debatable.
Studies using patent citations to measure technology flows around MNEs (Globerman, Kokko, and Sjoholm 2000; Branstetter 2001a; Singh 2003) yield mixed results. MNE subsidiaries appear to source more technology from host-country firms than they transfer to them. However, MNE subsidiaries are larger and more technologically intensive than typical host-country firms, so the sourcing advantage may reflect firm heterogeneity rather than a causal pattern.
Earlier surveys (Hanson 2001; Gorg and Greenaway 2002) concluded there was no evidence for substantial FDI spillovers. Keller argues this view is overly pessimistic, based on more recent evidence:
Aitken and Harrison (1999): found a negative relationship between FDI and productivity for Venezuelan plants, likely reflecting increased competition rather than negative spillovers.
Haskel, Pereira, and Slaughter (2001) and Griffith, Redding, and Simpson (2003): found positive FDI spillovers in the United Kingdom, though the implied economic magnitudes were fairly small relative to the subsidies paid to attract FDI.
Keller and Yeaple (2003): found that FDI accounts for about 11% of U.S. manufacturing productivity growth, a much larger estimate than earlier studies. Two reasons for the difference:
FDI spillovers are much stronger in high-technology than in low-technology sectors. Pooling across all manufacturing masks the effect.
Better FDI data matters: using detailed industry-level employment data (rather than allocating all MNE employees to one "major industry") yields estimates seven times as large.
IV estimation: both Griffith et al. (2003) and Keller and Yeaple (2003) use instrumental variables and find no evidence of endogeneity bias. If anything, IV estimates suggest larger FDI spillovers than OLS.
MNEs may have stronger incentives to transfer technology to their local suppliers (vertical, backward linkages) than to local competitors (horizontal). Kugler (2002) and Blalock and Gertler (2002) find some evidence for vertical FDI spillovers. Javorcik (2004) provides further support.
Recent micro productivity studies tend to estimate positive, and in some cases economically large, spillovers associated with FDI. The effects are concentrated in high-technology sectors. Key open questions remain: how large are vertical compared to horizontal spillovers, which firms benefit most, whether results extend to poorer countries, and whether the effects are large enough to justify the subsidies governments pay to attract MNEs.
"Exporters are more productive because they learn from exporting." The evidence overwhelmingly suggests self-selection: productive firms choose to export, rather than exporting making firms productive. Do not confuse the observed cross-sectional productivity premium with a causal learning effect.
"There are no FDI spillovers." This was the conclusion of earlier survey literature, but more recent micro-level studies with better data and identification strategies find positive and sometimes large spillovers, particularly in high-technology sectors. The blanket "no spillovers" claim is outdated.
"All types of FDI generate the same spillovers." FDI motivated by access to cheap labour is unlikely to generate the same technology spillovers as FDI in high-technology sectors. Pooling across all manufacturing industries in spillover regressions can mask substantial heterogeneity.
"Trade diffuses technology because you can reverse-engineer imports." Embodied technology diffusion through trade is a relatively weak form of diffusion. The importing country uses the technology embedded in the product but does not necessarily gain mastery of the underlying knowledge. The manufactured outcome crosses borders; the blueprint may not.
⚠️ The distinction between self-selection and learning-by-exporting is a classic identification problem. Be prepared to explain both hypotheses and describe the Clerides-Lach-Tybout approach to distinguishing them.
⚠️ Know the Keller (1998) critique of Coe-Helpman. It is a clean example of how a simple robustness check (randomly generated weights) can undermine an influential empirical result.
⚠️ Understand why the FDI spillover literature has shifted from pessimism to cautious optimism. The key factors: better data (industry-level rather than major-industry FDI measures), sector heterogeneity (high-tech vs low-tech), and improved identification (IV estimation).
⚠️ Be able to distinguish embodied from disembodied technology diffusion and explain why the latter is a stronger form of diffusion.
⚠️ The distinction between horizontal and vertical FDI spillovers may appear as an exam question. Know why MNEs might have stronger incentives to transfer technology vertically (to suppliers) than horizontally (to competitors).
True or false: The Coe-Helpman foreign R&D variable is weighted by bilateral export shares. _____
Clerides, Lach, and Tybout (1998) found _____ (strong/weak/no) evidence of learning-by-exporting effects.
True or false: Keller and Yeaple (2003) found that FDI spillovers are roughly equal across high-tech and low-tech sectors. _____
Vertical FDI spillovers refer to productivity effects on firms in _____ (the same / upstream or downstream) industries relative to MNE subsidiaries.
True or false: IV estimation of FDI spillovers has generally produced smaller estimates than OLS. _____
Answers: 1. False (import shares). 2. No. 3. False (much stronger in high-tech). 4. Upstream or downstream. 5. False (IV estimates are at least as large as OLS).
Q: Describe the Clerides, Lach, and Tybout (1998) approach to testing for learning-by-exporting. What did they find?
A: They estimated a dynamic discrete choice model of export market participation alongside an autoregressive cost function for manufacturing plants in Colombia, Morocco, and Mexico. The cost function allowed past exporting experience to reduce current average variable costs. They found generally no significant effect of past exporting on current costs. Where significant, the coefficients went in the wrong direction (exporting raising costs). Descriptive evidence (cost plots before and after export entry) supported the null result.
Q: Why might the Coe-Helpman (1995) result overstate the importance of imports as a channel of technology diffusion?
A: Keller (1998) showed that replacing actual bilateral import shares with randomly generated shares produces similarly high regression coefficients and explained variation. This suggests the positive coefficient on the foreign R&D variable may capture a common trend in R&D and TFP across countries, rather than a trade-specific mechanism. The import shares may not contain the identifying variation.
Q: Explain why Keller and Yeaple (2003) find much larger FDI spillover estimates than earlier studies such as Haskel, Pereira, and Slaughter (2001).
A: Two main reasons. First, FDI spillovers are concentrated in high-technology sectors. Keller and Yeaple's Compustat sample has a higher share of high-tech firms than broader samples, so the average effect is larger. Second, they use detailed industry-level FDI employment data rather than allocating all MNE employees to one major industry, which avoids introducing noise that attenuates the estimated coefficient. Better measurement of FDI yields estimates seven times as large.
Q: What is the difference between horizontal and vertical FDI spillovers, and why might vertical spillovers be larger?
A: Horizontal spillovers accrue to domestic firms in the same industry as MNE subsidiaries. Vertical spillovers accrue to firms in upstream (supplier) or downstream (customer) industries. MNEs may have stronger incentives to transfer technology vertically because they want higher-quality inputs from local suppliers, whereas they have no incentive to help local competitors in the same industry.
The trade channel connects to trade theory (gravity models, Ricardian comparative advantage) and to the broader debate on whether openness promotes growth. The FDI material connects to industrial organisation (MNE theory, Markusen 2002), labour economics (worker mobility as a spillover channel), and development economics (whether FDI subsidies are welfare-improving). The self-selection versus learning debate connects to econometric methods for causal inference (IV, difference-in-differences, dynamic discrete choice).
imports technology diffusion, learning-by-exporting, self-selection exporters, FDI spillovers, horizontal spillovers, vertical spillovers, backward linkages, Coe Helpman, Keller critique, import-weighted R&D, embodied technology, disembodied spillovers, multinational enterprises MNE, technology sourcing, technology transfer, Intel Costa Rica case study, Clerides Lach Tybout, Bernard Jensen, Keller Yeaple, patent citations FDI, capital goods trade, Xu Wang, Javorcik, endogeneity FDI