Difficulty: Intermediate | Prerequisites: Basic macroeconomics, familiarity with production functions and growth theory.
This is Part 1 of 3 covering Keller (2004), a survey article on international technology diffusion published in the Journal of Economic Literature. Part 1 covers the conceptual framework, how technology is measured, and empirical methods. Parts 2 and 3 cover the channels of diffusion (trade, FDI) and the determinants and policy implications. The paper matters because most countries depend on foreign technology for the bulk of their productivity growth, and understanding how that technology moves across borders is central to questions of global income convergence.
For most countries, foreign sources of technology account for 90 percent or more of domestic productivity growth, yet there is no global pool of technology. Keller (2004) surveys the evidence on how technology crosses borders, finding that imports and FDI are significant channels, that diffusion is geographically localised (it decays sharply with distance), and that domestic absorptive capacity (human capital, R&D) is necessary to benefit from foreign knowledge. The question of whether diffusion leads to convergence or divergence in incomes depends on how effectively countries engage with the international economy.
Non-rival good
A good whose use by one agent does not prevent use by another. The marginal cost of an additional user is negligible. Technology is non-rival; physical capital and labour are rival.
In simple terms, once an idea exists, letting one more person use it costs almost nothing.
Knowledge spillovers (technology spillovers)
Benefits from technological investments that accrue to parties other than the original inventor, without full compensation. They arise because technology is only partially excludable.
Think of it as the free productivity boost a firm gets when a competitor's R&D inadvertently teaches it something.
Total factor productivity (TFP)
The portion of output not explained by measured inputs (labour, capital). Computed as a residual: TFP = Output / (weighted inputs). Often used as a proxy for the level of technology.
In simple terms, TFP captures "how efficiently" an economy turns inputs into output, after you have accounted for how much labour and capital it uses.
R&D stock
The cumulative, depreciation-adjusted total of past R&D expenditures. Constructed using the perpetual inventory method: S(t) = R(t) + (1 - d) * S(t-1), where d is the depreciation rate (typically 0 to 10 percent).
Think of it as the running total of a country's investment in new knowledge, discounted for the fact that older research becomes obsolete.
Tacit knowledge
Knowledge that cannot be fully codified or written down. The person doing the work may not be able to articulate exactly what they know. Polanyi (1958) argued it can be passed on only "by example from master to apprentice."
In simple terms, it is know-how you can demonstrate but cannot put into a manual.
Absorptive capacity
The ability of a firm or country to identify, assimilate, and exploit external knowledge. Depends on human capital (Nelson and Phelps 1966) and on R&D spending (Cohen and Levinthal 1989).
Think of it as the skill base you need before foreign technology can do you any good.
Endogenous technical change
Models in which technological progress is the result of deliberate investment decisions (R&D), rather than falling from the sky. Key contributors: Romer (1990), Grossman and Helpman (1991), Aghion and Howitt (1992).
In simple terms, technology improves because firms and researchers choose to spend resources making it improve.
Diffusion lag (epsilon)
In the Eaton-Kortum framework, the parameter epsilon(ni) governs the speed at which technologies invented in country i become available in country n. A higher epsilon means faster diffusion.
Think of it as the rate at which a foreign invention shows up and becomes usable in your country.
Foreign R&D stock (import-weighted)
A country's foreign knowledge stock, constructed as the bilateral import-share-weighted sum of its trading partners' R&D stocks. Used in Coe and Helpman (1995) to test whether trade is a channel of technology diffusion.
In simple terms, it measures how much R&D your trade partners have done, weighted by how much you import from each of them.
Productivity differences explain most of the variation in incomes across countries, and technology is a key determinant of productivity.
For most countries, foreign sources of technology account for 90% or more of domestic productivity growth.
The G-7 countries account for roughly 84% of global R&D spending but only 64% of world GDP. The pattern of worldwide technical change is therefore driven largely by how technology diffuses from these few innovating countries to the rest.
Whether countries' incomes converge or diverge over time depends on whether technology diffusion is global or local.
Non-rivalry: the marginal cost for an additional agent to use a technology is negligible. This distinguishes technology from rival inputs like physical and human capital.
Partial excludability: returns to technological investment are partly private and partly public. Private returns keep innovation going; the public portion creates spillovers that benefit others.
Keller builds on Eaton and Kortum (1999, 1997) to lay out the core model. Key features:
N countries, each producing output via a Cobb-Douglas production function over a continuum of intermediate inputs.
Each technology has three dimensions: quality (random draw from a Pareto distribution), use (assigned to one intermediate sector), and a diffusion time lag.
Diffusion is stochastic: technologies invented in country i arrive in country n with a mean lag of epsilon(ni)^(-1). Higher epsilon = faster diffusion.
A technology may be superseded by a better one before it diffuses, so past R&D is effectively discounted.
In steady state, relative TFP levels across countries are determined by relative technology stocks, which in turn depend on diffusion rates.
Foreign R&D raises domestic TFP: more research abroad means a greater inflow of technologies.
Faster diffusion amplifies the effect: for a given level of foreign R&D, domestic TFP rises more when technologies cross borders quickly.
Global sources of growth: a country's importance for world growth depends on (i) its share of global research labour and technology stock, and (ii) how rapidly its innovations diffuse to others.
Technology is only partially codified. The remainder is tacit (Polanyi 1958), meaning it cannot be fully written down and is best transferred through face-to-face interaction. This has two consequences:
Technology diffusion will be incomplete, and technology stocks will vary across countries.
Diffusion will be geographically localised: the higher the tacit share, the more it matters that teacher and student are in the same location, which is costly to arrange over long distances.
Technology is intangible and difficult to measure directly. Three widely used proxies exist, each with trade-offs.
1. R&D expenditures (input measure)
Internationally comparable data published by the OECD since roughly 1965.
Only about two dozen rich countries report substantial R&D, because the OECD definition captures resources spent on innovation, not imitation or adoption.
Drawback: R&D spending is a noisy measure of technology improvement in any given period, because innovation is stochastic. Many authors therefore construct R&D stocks from flows using the perpetual inventory method.
Returns to R&D vary across agents and over time, and publicly funded R&D tends to have lower measured returns than privately funded R&D.
2. Patents (output measure)
A patent gives its holder a temporary legal monopoly in exchange for public disclosure of technical information.
Patent data has been collected for over 150 years and covers poorer countries as well.
Drawbacks: a small number of patents account for most of the total value; simple patent counts may not capture technology output well. Citation-weighted patent data (Jaffe and Trajtenberg 2002) partly addresses this. Many innovations are never patented, and non-codifiable technology is missed entirely.
3. Total factor productivity (TFP) (effect measure)
Computed as output minus the contribution of measured inputs. The residual is attributed to technology.
TFP is a derived measure, so it inherits measurement error from the underlying data on inputs and output. The appropriate physical-quantity data is rarely available; researchers typically use nominal revenues and depreciated capital spending instead.
Higher measured TFP may reflect higher mark-ups rather than higher efficiency (Katayama, Lu, and Tybout 2003).
Strategies to reduce contamination: use TFP changes rather than levels (helps if spurious factors are stable over time), or pair TFP with R&D data (Griliches 1984) to establish a relationship with its presumed cause.
Market transactions (royalties, licence fees, patent payments) are recorded in the international services balance. Data is available for major countries.
Most economists believe the bulk of international technology diffusion occurs through externalities (spillovers) rather than market transactions, because technology is not fully codifiable and complete contracts cannot be written.
Direct data on spillovers does not exist. Indirect measures include patent citations (if patent B cites patent A, B's inventor likely benefited from A's technology, though the net benefit is unknown) and international R&D spillover regressions (correlating foreign R&D with domestic TFP).
Offer rich description of the setting and the factors behind technology diffusion in a particular instance (e.g. Larrain, Lopez-Calva, and Rodriguez-Clare (2000) on Intel's FDI into Costa Rica).
Main limitation: it is impossible to know how general a single case is. Most useful when backed by quantitative evidence.
Test whether a specific foreign activity (FA), such as the industry share of employment in foreign-owned firms, is correlated with a domestic technology outcome (DTO), such as domestic firm productivity growth.
General and easy to apply across different models, but the use of a foreign activity variable rather than a foreign technology variable makes precise interpretation difficult and opens the door to spurious correlation.
Key threats to identification: business cycle effects (addressed with time fixed effects), unobserved heterogeneity (addressed with entity fixed effects, though over-differencing exacerbates measurement error), and endogeneity (foreign investors may target high-productivity-growth industries). Instrumental variable (IV) approaches are beginning to be used but good instruments for R&D stocks are hard to find.
Replace the generic foreign activity variable with a foreign technology variable (e.g. foreign R&D stock) and add a specific channel of diffusion (e.g. bilateral imports).
The Coe and Helpman (1995) specification is the leading example: TFP is regressed on domestic R&D and an import-share-weighted foreign R&D stock. More structure aids interpretation, but what determines R&D is typically not modelled, so endogeneity remains a concern.
A second variant puts more emphasis on econometric specification (e.g. Clerides, Lach, and Tybout 1998 on learning-by-exporting): simultaneous estimation of an export participation equation and a cost function, controlling for self-selection of productive firms into exporting.
Eaton and Kortum (1996, 1997, 1999) use fully specified general equilibrium models to study technology diffusion. The full model pins down all endogenous variables, so single-equation endogeneity is no longer an issue.
Costs: strong distributional assumptions are needed (Pareto for quality, exponential for diffusion lags) and are hard to test within the model. Some parameters are calibrated rather than estimated, so results are partly simulated. Judging empirical success is difficult: if a model fits the data within a 15% error margin, how much of that fit comes from free parameters versus structural assumptions?
\ln(Y_{nt} / J) = J^{-1} \int_0^J \ln[Z_{nt}(j) X_{nt}(j)] \, djProduction function: output in country n at time t is a Cobb-Douglas aggregate over a continuum of intermediate inputs, where Z is quality and X is quantity.
\frac{A_{nt}}{A_{Nt}} = \left( \frac{\mu_{nt}}{\mu_{Nt}} \right)^{1/\theta}Steady-state relative TFP: a country's TFP relative to the frontier country N is determined by its relative technology stock (mu), which depends on diffusion rates and research efforts.
\ln f_c = \alpha_c + \beta^d \ln S_c + \beta^f \ln S_c^f + \varepsilon_cCoe-Helpman regression (equation 12): TFP of country c is a function of domestic R&D stock (S_c) and import-weighted foreign R&D stock (S_c^f). If beta^f > 0, foreign R&D raises domestic productivity, consistent with technology diffusion through trade.
\ln TFP_{cit} = \beta \ln \left( S_{cit} + \gamma \sum_{c' \in G5} S_{c'it} \times e^{-\delta D_{c'c}} \right) + \alpha' X + \varepsilon_{cit}Keller (2002a) distance-decay specification (equation 16): the weight given to foreign R&D declines exponentially with geographic distance D. If delta > 0, technology diffusion is geographically localised.
The framework matters because it connects directly to policy debates about openness and development. Countries that engage more with the international economy (through trade, FDI, and person-to-person contact) tend to gain better access to foreign technology, which is the dominant source of productivity growth for all but the largest innovating nations. This is why a closed-off international economic regime has detrimental welfare consequences: it cuts a country off from the technology that drives its growth.
The geographic localisation result has practical implications too. Technology transfer costs (roughly 20% of total project costs even for within-firm transfers, per Teece 1977) help explain why multinational firms exist and why governments compete to attract them. The distance-decay finding also explains why countries near major R&D centres (Western Europe near the US and each other, East Asia near Japan) tend to benefit more from spillovers than geographically remote countries like Australia or New Zealand.
"Technology is global because the internet exists." Students often assume that because computer programs and designs can be transmitted digitally at near-zero cost, there is effectively a global pool of technology. Keller's evidence shows this is wrong: technology remains highly geographically localised, largely because a significant component is tacit and requires face-to-face interaction to transfer.
"More R&D always means more technology." R&D expenditure is a noisy input measure. Innovation is stochastic, returns vary enormously across projects and over time, and publicly funded R&D tends to have lower measured returns than private R&D. A high R&D spend does not guarantee proportional technology output.
"A positive correlation between FDI and productivity proves spillovers." Correlation is not causation. Foreign investors may choose to enter industries that are already growing fast (endogeneity), and omitted variables (such as trade costs that vary across industries) can generate spurious positive coefficients.
"International technology diffusion is automatic." Keller repeatedly stresses that diffusion is neither inevitable nor automatic. Countries need domestic absorptive capacity (human capital, R&D) to benefit from foreign technology. Simply being open to trade or FDI is necessary but not sufficient.
⚠️ Be prepared to explain the two key properties of technology (non-rivalry, partial excludability) and why they matter for diffusion. This distinction between technology and rival factor inputs is foundational.
⚠️ Know the three measures of technology (R&D, patents, TFP) and be able to discuss at least one limitation of each. A common exam question asks you to evaluate which measure is most appropriate for a given research question.
⚠️ The Coe-Helpman (1995) specification is one of the most cited empirical frameworks in this literature. Understand how the import-weighted foreign R&D stock is constructed and what a positive coefficient on it implies. Also know Keller's (1998) critique: randomly generated import shares produce similar results, casting doubt on whether imports per se are the channel.
⚠️ Understand the difference between technology embodied in traded goods (a relatively weak form of diffusion, since only the manufactured outcome crosses borders) and disembodied R&D spillovers (a stronger form, implying full mastery of the underlying technology). This distinction is central to interpreting the evidence on trade as a diffusion channel.
⚠️ The distinction between association studies and structure studies is likely to appear on methods-oriented exams. Know why endogeneity is a particular concern and how IV estimation, fixed effects, and difference-in-differences approaches attempt to address it.
True or false: Technology is a rival good, like physical capital. _____
The G-7 countries account for approximately _____% of global R&D spending.
True or false: The perpetual inventory method is used to construct R&D stocks from annual R&D expenditure flows. _____
In the Coe-Helpman specification, the foreign R&D stock is weighted by bilateral _____ shares.
True or false: Keller (1998) showed that randomly generated trade weights produce results similar to actual import shares in the Coe-Helpman regression. _____
Answers: 1. False (technology is non-rival). 2. 84%. 3. True. 4. Import. 5. True.
Q: Explain the two key properties of technology emphasised in endogenous growth theory and why each matters for international technology diffusion.
A: Technology is non-rival (the marginal cost of an additional user is negligible) and partially excludable (returns are partly private, partly public). Non-rivalry means technology can in principle spread to every country without being "used up." Partial excludability means that some benefits leak to others as spillovers, creating the potential for international diffusion, but also that inventors have incentives to restrict access.
Q: What are the main advantages and disadvantages of using TFP as a measure of technology?
A: TFP captures the effect of technology on output, making it a more direct measure of what matters (productivity) than R&D (an input) or patents (an output that misses non-codifiable knowledge). Its main disadvantages are that it is a residual and may capture spurious influences (mark-up variation, measurement error in inputs), and that the ideal physical-quantity data on inputs and output is rarely available.
Q: Describe the Coe and Helpman (1995) approach to testing whether trade is a channel of technology diffusion. What is Keller's (1998) critique?
A: Coe and Helpman regress country-level TFP on domestic R&D and a foreign R&D variable constructed as the bilateral import-share-weighted sum of trade partners' R&D stocks. A positive coefficient on the foreign variable is interpreted as evidence that imports channel foreign technology. Keller (1998) showed that replacing actual import shares with randomly generated shares produces similarly high coefficients and explained variation, suggesting that import shares may not be essential and that the result could reflect a common trend in R&D and TFP rather than a trade-specific mechanism.
Q: Why might association studies produce biased estimates of FDI spillovers? Describe at least two sources of bias.
A: First, endogeneity: foreign investors may target industries with high productivity growth (because they are profitable), generating a spurious positive correlation between FDI presence and domestic productivity. Second, omitted variable bias: if FDI is concentrated in high-trade-cost industries that also happen to have high average productivity, the FDI coefficient captures the trade-cost effect. Third, unobserved heterogeneity across firms or industries that is correlated with both FDI presence and productivity.
This material connects to growth theory (Solow residual, endogenous growth models of Romer and Aghion-Howitt) because technology diffusion is the mechanism through which the frontier expands for non-innovating countries. It also connects to international trade theory (Ricardian comparative advantage, gravity models) because trade is one of the channels through which technology crosses borders. The concept of absorptive capacity links to human capital theory (Nelson and Phelps 1966) and the economics of education.
technology diffusion, knowledge spillovers, R&D spillovers, total factor productivity, TFP, non-rival goods, tacit knowledge, codified knowledge, absorptive capacity, Coe and Helpman, Eaton and Kortum, endogenous growth, international R&D, foreign knowledge stock, import-weighted R&D, patent citations, technology transfer, convergence, divergence, Polanyi, perpetual inventory method, Griliches, association studies, structure studies, general equilibrium, FDI spillovers, Keller 2004